{"id":5287,"date":"2026-02-24T04:14:15","date_gmt":"2026-02-24T04:14:15","guid":{"rendered":"https:\/\/research.undetectable.ai\/?p=5287"},"modified":"2026-02-24T04:14:50","modified_gmt":"2026-02-24T04:14:50","slug":"i-tested-5-ai-image-detectors-these-are-the-best-in-2026","status":"publish","type":"post","link":"https:\/\/research.undetectable.ai\/pt\/testei-5-detectores-de-imagens-de-ia-estes-sao-os-melhores-em-2026\/","title":{"rendered":"Testei 5 detectores de imagem de IA. Estes s\u00e3o os melhores em 2026:"},"content":{"rendered":"

\u00c9 prov\u00e1vel que j\u00e1 tenha visto mais fotografias geradas por IA na Internet do que imagina. <\/strong><\/p>\n\n\n\n

Por vezes, \u00e9 \u00f3bvio que uma imagem foi gerada por IA, mas \u00e9 cada vez mais dif\u00edcil de perceber \u00e0 medida que as ferramentas generativas de imagem e v\u00eddeo melhoram. Novas ferramentas como a Nano Banana Pro<\/a>e as actualiza\u00e7\u00f5es da Modelo de imagem ChatGPT<\/a> permitem aos utilizadores gerar rapidamente imagens sint\u00e9ticas que espelham imagens reais. Estudos anteriores revelaram que 85% dos americanos dizem Os deepfakes est\u00e3o a minar a confian\u00e7a online<\/a>.<\/p>\n\n\n\n

V\u00e1rias ferramentas afirmam detetar imagens geradas por IA, mas ser\u00e1 que funcionam? <\/strong><\/h2>\n\n\n\n

Efectuei 50 verifica\u00e7\u00f5es de dete\u00e7\u00e3o em cinco dos detectores de imagens com IA mais populares e documentei os resultados. N\u00e3o s\u00f3 apresentarei todos os dados neste artigo e explic\u00e1-los-ei, como tamb\u00e9m farei uma liga\u00e7\u00e3o \u00e0 documenta\u00e7\u00e3o no final.<\/p>\n\n\n\n

Os cinco detectores que utilizei para estes testes foram: TruthScan<\/a>, IA ou n\u00e3o<\/a>, Motor de vis\u00e3o<\/a>, EraIssoAI<\/a>e Winston AI<\/a>. Testei essas ferramentas fazendo com que cada uma analisasse duas imagens geradas pelo ChatGPT, seis imagens geradas pelo Nano Banana e duas imagens geradas pelo Midjourney. Utilizei v\u00e1rios estilos e t\u00e9cnicas de est\u00edmulo (sobre os quais falarei em pormenor mais tarde, quando analisar cada imagem analisada).<\/p>\n\n\n\n

Os cinco detectores foram incumbidos de detetar imagens criadas por IA em categorias como fraude, desinforma\u00e7\u00e3o, fotografia geral e deepfakes. N\u00e3o \u00e9 de surpreender que nem todos os detectores tenham tido um bom desempenho. O TruthScan foi o \u00fanico detetor a classificar de forma consistente todo o conte\u00fado que enviei com 97% ou superior.<\/strong><\/p>\n\n\n\n\n\n\n\n\nResultados do teste do Detetor de Imagens AI<\/title>\n<style>\n @import url('https:\/\/fonts.googleapis.com\/css2?family=DM+Sans:wght@400;500;600;700&family=Outfit:wght@600;700;800&display=swap');\n\n .ai-detector-wrap {\n max-width: 780px;\n margin: 2rem auto;\n font-family: 'DM Sans', sans-serif;\n -webkit-font-smoothing: antialiased;\n }\n\n .ai-detector-wrap h2 {\n font-family: 'Outfit', sans-serif;\n font-size: 1.75rem;\n font-weight: 800;\n color: #1a1a2e;\n margin-bottom: 1.25rem;\n letter-spacing: -0.02em;\n }\n\n .ai-detector-table {\n width: 100%;\n border-collapse: separate;\n border-spacing: 0;\n border-radius: 14px;\n overflow: hidden;\n box-shadow:\n 0 1px 3px rgba(0,0,0,0.06),\n 0 8px 24px rgba(0,0,0,0.08),\n 0 0 0 1px rgba(0,0,0,0.04);\n }\n\n .ai-detector-table thead th {\n background: #1a1a2e;\n color: #ffffff;\n font-family: 'Outfit', sans-serif;\n font-weight: 700;\n font-size: 0.85rem;\n text-transform: uppercase;\n letter-spacing: 0.06em;\n padding: 16px 20px;\n text-align: center;\n border: none;\n }\n\n .ai-detector-table thead th:first-child {\n text-align: left;\n padding-left: 24px;\n }\n\n .ai-detector-table tbody tr {\n transition: background 0.2s ease;\n }\n\n .ai-detector-table tbody tr:hover {\n background: rgba(0,0,0,0.02);\n }\n\n .ai-detector-table tbody td {\n padding: 14px 20px;\n text-align: center;\n font-size: 0.95rem;\n font-weight: 500;\n color: #2d2d3a;\n border-bottom: 1px solid #eee;\n }\n\n .ai-detector-table tbody tr:last-child td {\n border-bottom: none;\n }\n\n .ai-detector-table tbody td:first-child {\n text-align: left;\n padding-left: 24px;\n font-weight: 600;\n color: #1a1a2e;\n }\n\n \/* Colored cells *\/\n .cell-green {\n background: #22c55e;\n color: #fff !important;\n font-weight: 700 !important;\n border-radius: 6px;\n padding: 6px 12px;\n display: inline-block;\n min-width: 32px;\n text-align: center;\n font-size: 0.9rem;\n }\n\n .cell-red {\n background: #ef4444;\n color: #fff !important;\n font-weight: 700 !important;\n border-radius: 6px;\n padding: 6px 12px;\n display: inline-block;\n min-width: 32px;\n text-align: center;\n font-size: 0.9rem;\n }\n\n \/* Accuracy badges *\/\n .badge-100 { background: #16a34a; color: #fff; }\n .badge-80 { background: #a3e635; color: #1a1a2e; }\n .badge-70 { background: #facc15; color: #1a1a2e; }\n .badge-60 { background: #fb923c; color: #fff; }\n .badge-30 { background: #ef4444; color: #fff; }\n\n .accuracy-badge {\n font-weight: 800;\n font-family: 'Outfit', sans-serif;\n font-size: 0.95rem;\n border-radius: 8px;\n padding: 6px 16px;\n display: inline-block;\n min-width: 60px;\n text-align: center;\n letter-spacing: 0.02em;\n }\n\n \/* Row rank indicator *\/\n .rank-icon {\n margin-right: 6px;\n font-size: 1rem;\n }\n\n \/* Asterisk note *\/\n .table-note {\n margin-top: 0.75rem;\n font-size: 0.8rem;\n color: #888;\n padding-left: 4px;\n }\n\n \/* Responsive *\/\n @media (max-width: 600px) {\n .ai-detector-table thead th,\n .ai-detector-table tbody td {\n padding: 10px 10px;\n font-size: 0.8rem;\n }\n .ai-detector-wrap h2 {\n font-size: 1.3rem;\n }\n }\n<\/style>\n<\/head>\n<body style=\"background:#f8f8fa; padding: 2rem;\">\n\n<div class=\"ai-detector-wrap\">\n <h2>Resultados do teste do detetor de imagens AI<\/h2>\n\n <table class=\"ai-detector-table\">\n <thead>\n <tr>\n <th>Ferramenta de dete\u00e7\u00e3o<\/th>\n <th>Total de ensaios<\/th>\n <th>Detectado corretamente<\/th>\n <th>Falhou (Fails)<\/th>\n <th>Exatid\u00e3o<\/th>\n <\/tr>\n <\/thead>\n <tbody>\n <tr>\n <td><span class=\"rank-icon\">\ud83e\udd47<\/span>TruthScan<\/td>\n <td>10<\/td>\n <td><span class=\"cell-green\">10<\/span><\/td>\n <td><span class=\"cell-green\">0<\/span><\/td>\n <td><span class=\"accuracy-badge badge-100\">100%<\/span><\/td>\n <\/tr>\n <tr>\n <td><span class=\"rank-icon\">\ud83e\udd48<\/span>IA ou n\u00e3o<\/td>\n <td>10<\/td>\n <td><span class=\"cell-green\">8<\/span><\/td>\n <td><span class=\"cell-red\">2*<\/span><\/td>\n <td><span class=\"accuracy-badge badge-80\">80%<\/span><\/td>\n <\/tr>\n <tr>\n <td><span class=\"rank-icon\">\ud83e\udd49<\/span>Motor de vis\u00e3o<\/td>\n <td>10<\/td>\n <td><span class=\"cell-green\">7<\/span><\/td>\n <td><span class=\"cell-red\">3<\/span><\/td>\n <td><span class=\"accuracy-badge badge-70\">70%<\/span><\/td>\n <\/tr>\n <tr>\n <td><span class=\"rank-icon\">   <\/span>EraIssoAI<\/td>\n <td>10<\/td>\n <td><span class=\"cell-green\">6<\/span><\/td>\n <td><span class=\"cell-red\">4<\/span><\/td>\n <td><span class=\"accuracy-badge badge-60\">60%<\/span><\/td>\n <\/tr>\n <tr>\n <td><span class=\"rank-icon\">   <\/span>Winston AI<\/td>\n <td>10<\/td>\n <td><span class=\"cell-green\">3<\/span><\/td>\n <td><span class=\"cell-red\">7<\/span><\/td>\n <td><span class=\"accuracy-badge badge-30\">30%<\/span><\/td>\n <\/tr>\n <\/tbody>\n <\/table>\n\n <p class=\"table-note\">* AI or Not rotulou uma gera\u00e7\u00e3o AI como 78%, uma como 85% e uma como 89%. Uma vez que o teste exige pelo menos 90%, contamos as duas primeiras como falhas, mas tratamos a 89% como uma quase falha, da\u00ed o asterisco...<\/p>\n<\/div>\n\n<\/body>\n<\/html>\n\n\n\n<p style=\"font-size:17px\">O AI or Not tamb\u00e9m teve um desempenho muito bom. De facto, embora o AI or Not tenha classificado v\u00e1rios itens gerados por IA abaixo da certeza 90%, foi o segundo detetor mais preciso e consistente durante os meus testes, atr\u00e1s do TruthScan. Mesmo nos casos em que o AI or Not falhou, essas falhas n\u00e3o foram catastr\u00f3ficas, mas ainda h\u00e1 espa\u00e7o para melhorias. <\/p>\n\n\n\n<p style=\"font-size:17px\">O resto dos detectores de imagens de IA que experimentei tiveram um desempenho muito pior. Por exemplo, o Sight Engine classificou incorretamente 3 imagens de IA relacionadas com fraude como aut\u00eanticas.<\/p>\n\n\n\n<p style=\"font-size:17px\">Agora vou mostrar-vos cada imagem (das 10 que gerei), explicar como a criei e mostrar como cada modelo a pontuou.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">#1. \"Man On a Ledge\" (gerado por ChatGPT)<\/h2>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"683\" height=\"1024\" src=\"https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/CGPT-Generation-1-Man-on-ledge-683x1024.jpg\" alt=\"\" class=\"wp-image-5289\" title=\"\" srcset=\"https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/CGPT-Generation-1-Man-on-ledge-683x1024.jpg 683w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/CGPT-Generation-1-Man-on-ledge-200x300.jpg 200w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/CGPT-Generation-1-Man-on-ledge-768x1152.jpg 768w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/CGPT-Generation-1-Man-on-ledge-8x12.jpg 8w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/CGPT-Generation-1-Man-on-ledge.jpg 1024w\" sizes=\"auto, (max-width: 683px) 100vw, 683px\" \/><\/figure>\n\n\n\n<style>\n @import url('https:\/\/fonts.googleapis.com\/css2?family=DM+Sans:wght@400;500;600;700&family=Outfit:wght@600;700;800&display=swap');\n\n .detector-scores-wrap {\n max-width: 780px;\n margin: 2rem auto;\n font-family: 'DM Sans', sans-serif;\n -webkit-font-smoothing: antialiased;\n }\n\n .detector-scores-wrap h3 {\n font-family: 'Outfit', sans-serif;\n font-size: 1.4rem;\n font-weight: 800;\n color: #1a1a2e;\n margin-bottom: 1.25rem;\n letter-spacing: -0.02em;\n }\n\n .detector-scores-table {\n width: 100%;\n border-collapse: separate;\n border-spacing: 0;\n border-radius: 14px;\n overflow: hidden;\n table-layout: fixed;\n box-shadow:\n 0 1px 3px rgba(0,0,0,0.06),\n 0 8px 24px rgba(0,0,0,0.08),\n 0 0 0 1px rgba(0,0,0,0.04);\n }\n\n .detector-scores-table thead th {\n background: #1a1a2e;\n color: #ffffff;\n font-family: 'Outfit', sans-serif;\n font-weight: 700;\n font-size: 0.7rem;\n text-transform: uppercase;\n letter-spacing: 0.04em;\n padding: 14px 6px;\n text-align: center;\n border: none;\n }\n\n .detector-scores-table thead th:first-child {\n text-align: left;\n padding-left: 14px;\n width: 18%;\n }\n\n .detector-scores-table thead th:nth-child(2) {\n width: 12%;\n }\n\n .detector-scores-table tbody tr {\n transition: background 0.2s ease;\n }\n\n .detector-scores-table tbody tr:hover {\n background: rgba(0,0,0,0.02);\n }\n\n .detector-scores-table tbody td {\n padding: 14px 6px;\n text-align: center;\n font-size: 0.9rem;\n font-weight: 500;\n color: #2d2d3a;\n border-bottom: 1px solid #eee;\n }\n\n .detector-scores-table tbody tr:last-child td {\n border-bottom: none;\n }\n\n .detector-scores-table tbody td:first-child {\n text-align: left;\n padding-left: 14px;\n font-weight: 600;\n color: #1a1a2e;\n }\n\n .score-pill {\n font-weight: 700;\n font-family: 'Outfit', sans-serif;\n font-size: 0.78rem;\n border-radius: 6px;\n padding: 5px 8px;\n display: inline-block;\n min-width: 0;\n text-align: center;\n white-space: nowrap;\n }\n\n .score-pass {\n background: #22c55e;\n color: #fff;\n }\n\n .score-warn {\n background: #fb923c;\n color: #fff;\n }\n\n .score-fail {\n background: #ef4444;\n color: #fff;\n }\n\n @media (max-width: 600px) {\n .detector-scores-table thead th,\n .detector-scores-table tbody td {\n padding: 10px 4px;\n font-size: 0.7rem;\n }\n .score-pill {\n padding: 4px 5px;\n font-size: 0.68rem;\n }\n .detector-scores-wrap h3 {\n font-size: 1.15rem;\n }\n }\n<\/style>\n\n<div class=\"detector-scores-wrap\">\n <h3>Como os detectores de imagem pontuaram<\/h3>\n\n <table class=\"detector-scores-table\">\n <thead>\n <tr>\n <th>Modelo de imagem<\/th>\n <th>Categoria<\/th>\n <th>TruthScan<\/th>\n <th>Motor de vis\u00e3o<\/th>\n <th>IA ou n\u00e3o<\/th>\n <th>Winston AI<\/th>\n <th>EraIssoAI<\/th>\n <\/tr>\n <\/thead>\n <tbody>\n <tr>\n <td>ChatGPT <span style=\"font-size:0.75rem; color:#888; display:block;\">Gera\u00e7\u00e3o completa<\/span><\/td>\n <td>Geral<\/td>\n <td><span class=\"score-pill score-pass\">99.00% AI<\/span><\/td>\n <td><span class=\"score-pill score-pass\">99.00% AI<\/span><\/td>\n <td><span class=\"score-pill score-warn\">77.99% AI<\/span><\/td>\n <td><span class=\"score-pill score-fail\">0,98% AI<\/span><\/td>\n <td><span class=\"score-pill score-fail\">1.00% AI<\/span><\/td>\n <\/tr>\n <\/tbody>\n <\/table>\n<\/div>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p style=\"font-size:17px\">Para mim (e vejo diariamente muito conte\u00fado gerado por IA), a imagem parece bastante convincente \u00e0 primeira vista. Se clic\u00e1ssemos no perfil de uma pessoa nas redes sociais, percorr\u00eassemos o seu feed e v\u00edssemos esta imagem, ser\u00e1 que se destacaria imediatamente como uma falsifica\u00e7\u00e3o gerada por IA? Para ser justo, tentei ser criativo com a pergunta. Aqui est\u00e1 a que utilizei para gerar esta imagem:<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"has-text-align-center has-vivid-cyan-blue-color has-text-color has-link-color has-medium-font-size wp-elements-99b76fc78211de13a760aff177dd504c\">“<em><strong>Crie uma fotografia ao estilo de uma c\u00e2mara instant\u00e2nea de um homem em cima de um telhado a olhar para a rua, \u00e0 noite, iluminado pelo flash da c\u00e2mara, com vest\u00edgios de luz dos carros l\u00e1 em baixo, uma mulher atr\u00e1s dele com a m\u00e3o nos l\u00e1bios, a sorrir, a est\u00e9tica da fotografia \u00e9 a de uma fotografia tirada numa c\u00e2mara instant\u00e2nea em 2009 por amigos da faculdade a brincar.<\/strong><\/em>“<\/p>\n<\/blockquote>\n<\/blockquote>\n<\/blockquote>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<p style=\"font-size:17px\">Independentemente da minha criatividade ou do n\u00famero de detalhes que inclu\u00ed na minha pergunta, o TruthScan e o Sight Engine continuaram a classificar corretamente o resultado como gerado por IA. O detetor AI or Not chegou perto, mas ainda assim falhou o alvo. <\/p>\n\n\n\n<p style=\"font-size:17px\">Winston e WasItAI estavam completamente errados - classificando a imagem do ChatGPT como real.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">#2. \"O recibo falso\" (gerado por ChatGPT)<\/h2>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"683\" height=\"1024\" src=\"https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/CGPT2Bestbuyreceipttest-683x1024.jpg\" alt=\"Uma imagem que mostra o que parece ser um recibo manchado\" class=\"wp-image-5291\" title=\"\" srcset=\"https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/CGPT2Bestbuyreceipttest-683x1024.jpg 683w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/CGPT2Bestbuyreceipttest-200x300.jpg 200w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/CGPT2Bestbuyreceipttest-768x1152.jpg 768w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/CGPT2Bestbuyreceipttest-8x12.jpg 8w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/CGPT2Bestbuyreceipttest.jpg 1024w\" sizes=\"auto, (max-width: 683px) 100vw, 683px\" \/><\/figure>\n\n\n\n<style>\n @import url('https:\/\/fonts.googleapis.com\/css2?family=DM+Sans:wght@400;500;600;700&family=Outfit:wght@600;700;800&display=swap');\n\n .detector-scores-wrap {\n max-width: 780px;\n margin: 2rem auto;\n font-family: 'DM Sans', sans-serif;\n -webkit-font-smoothing: antialiased;\n }\n\n .detector-scores-wrap h3 {\n font-family: 'Outfit', sans-serif;\n font-size: 1.4rem;\n font-weight: 800;\n color: #1a1a2e;\n margin-bottom: 1.25rem;\n letter-spacing: -0.02em;\n }\n\n .detector-scores-table {\n width: 100%;\n border-collapse: separate;\n border-spacing: 0;\n border-radius: 14px;\n overflow: hidden;\n table-layout: fixed;\n box-shadow:\n 0 1px 3px rgba(0,0,0,0.06),\n 0 8px 24px rgba(0,0,0,0.08),\n 0 0 0 1px rgba(0,0,0,0.04);\n }\n\n .detector-scores-table thead th {\n background: #1a1a2e;\n color: #ffffff;\n font-family: 'Outfit', sans-serif;\n font-weight: 700;\n font-size: 0.7rem;\n text-transform: uppercase;\n letter-spacing: 0.04em;\n padding: 14px 6px;\n text-align: center;\n border: none;\n }\n\n .detector-scores-table thead th:first-child {\n text-align: left;\n padding-left: 14px;\n width: 18%;\n }\n\n .detector-scores-table thead th:nth-child(2) {\n width: 12%;\n }\n\n .detector-scores-table tbody tr {\n transition: background 0.2s ease;\n }\n\n .detector-scores-table tbody tr:hover {\n background: rgba(0,0,0,0.02);\n }\n\n .detector-scores-table tbody td {\n padding: 14px 6px;\n text-align: center;\n font-size: 0.9rem;\n font-weight: 500;\n color: #2d2d3a;\n border-bottom: 1px solid #eee;\n }\n\n .detector-scores-table tbody tr:last-child td {\n border-bottom: none;\n }\n\n .detector-scores-table tbody td:first-child {\n text-align: left;\n padding-left: 14px;\n font-weight: 600;\n color: #1a1a2e;\n }\n\n .score-pill {\n font-weight: 700;\n font-family: 'Outfit', sans-serif;\n font-size: 0.78rem;\n border-radius: 6px;\n padding: 5px 8px;\n display: inline-block;\n min-width: 0;\n text-align: center;\n white-space: nowrap;\n }\n\n .score-pass {\n background: #22c55e;\n color: #fff;\n }\n\n .score-warn {\n background: #fb923c;\n color: #fff;\n }\n\n .score-fail {\n background: #ef4444;\n color: #fff;\n }\n\n @media (max-width: 600px) {\n .detector-scores-table thead th,\n .detector-scores-table tbody td {\n padding: 10px 4px;\n font-size: 0.7rem;\n }\n .score-pill {\n padding: 4px 5px;\n font-size: 0.68rem;\n }\n .detector-scores-wrap h3 {\n font-size: 1.15rem;\n }\n }\n<\/style>\n\n<div class=\"detector-scores-wrap\">\n <h3>Como os detectores de imagem pontuaram<\/h3>\n\n <table class=\"detector-scores-table\">\n <thead>\n <tr>\n <th>Modelo de imagem<\/th>\n <th>Categoria<\/th>\n <th>TruthScan<\/th>\n <th>Motor de vis\u00e3o<\/th>\n <th>IA ou n\u00e3o<\/th>\n <th>Winston AI<\/th>\n <th>EraIssoAI<\/th>\n <\/tr>\n <\/thead>\n <tbody>\n <tr>\n <td>ChatGPT <span style=\"font-size:0.75rem; color:#888; display:block;\">Gera\u00e7\u00e3o completa<\/span><\/td>\n <td>Fraude<\/td>\n <td><span class=\"score-pill score-pass\">99.00% AI<\/span><\/td>\n <td><span class=\"score-pill score-fail\">19.00% AI<\/span><\/td>\n <td><span class=\"score-pill score-pass\">94.48% AI<\/span><\/td>\n <td><span class=\"score-pill score-fail\">0,04% AI<\/span><\/td>\n <td><span class=\"score-pill score-fail\">1.00% AI<\/span><\/td>\n <\/tr>\n <\/tbody>\n <\/table>\n<\/div>\n\n\n\n<p style=\"font-size:17px\">Quis gerar um recibo falso com o ChatGPT para ver como ficaria realista. Pode notar que a imagem n\u00e3o tem um endere\u00e7o leg\u00edtimo, mas a textura ainda parece um pouco realista. Diria que esta imagem \u00e9 menos convincente do que a primeira que pedi ao ChatGPT para gerar, mas a maioria dos detectores ainda assim falhou ao analis\u00e1-la.<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li style=\"font-size:17px\">De todos os controlos que fiz a esta imagem, <strong>TruthScan<\/strong> foi o mais exato. <\/li>\n\n\n\n<li style=\"font-size:17px\"><strong>Motor de vis\u00e3o<\/strong> falhou realmente aqui, mostrando um mau julgamento numa imagem de IA relacionada com documentos.<\/li>\n\n\n\n<li style=\"font-size:17px\"><strong>IA ou n\u00e3o<\/strong> teve um desempenho visivelmente melhor nesta segunda an\u00e1lise.<\/li>\n\n\n\n<li style=\"font-size:17px\"><strong>Winston AI <\/strong>teve o pior desempenho, e o WasItAI tamb\u00e9m falhou completamente.<\/li>\n<\/ul>\n\n\n\n<p style=\"font-size:17px\">Dado que esta imagem est\u00e1 associada \u00e0 categoria de fraude, \u00e9 preocupante que a maioria dos detectores a tenham classificado incorretamente. <\/p>\n\n\n\n<h3 class=\"wp-block-heading\">O prompt que utilizei para gerar a imagem<\/h3>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<p class=\"has-vivid-cyan-blue-color has-text-color has-link-color wp-elements-b58fc16e4a944c203389dba029f958d9\"><strong>“<em>Gerar uma imagem de um recibo da Best Buy, que diz que o montante total gasto \u00e9 de mil milh\u00f5es de d\u00f3lares, e que tem uma mancha de caf\u00e9<\/em>“<\/strong><\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<p style=\"font-size:17px\">Claramente, eu escrevi um prompt r\u00e1pido e simples. Ainda assim, a qualidade da sa\u00edda do ChatGPT \u00e9 visualmente atraente a olho nu. N\u00e3o t\u00e3o atraente para dois dos cinco detectores que testei, ao que parece.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">#3: \"Hazard Package\" (gerado com Nano Banana)<\/h2>\n\n\n\n<p><strong><em>Em vez de deixar o Nano Banana gerar uma imagem original, dei-lhe uma para alterar. <\/em><\/strong><\/p>\n\n\n\n<p style=\"font-size:17px\">A ideia era demonstrar como algu\u00e9m poderia facilmente utilizar a IA para criar provas falsas e alegar que recebeu uma encomenda danificada\/perigosa. Primeiro, tirei uma fotografia com um iPhone 15 de uma embalagem vazia da Amazon que encontrei.<\/p>\n\n\n\n<p><strong>Aqui est\u00e1 a fotografia que tirei:<\/strong><\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"759\" height=\"1024\" src=\"https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/Amazon-package-759x1024.jpg\" alt=\"Uma fotografia de uma encomenda da Amazon\" class=\"wp-image-5293\" title=\"\" srcset=\"https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/Amazon-package-759x1024.jpg 759w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/Amazon-package-222x300.jpg 222w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/Amazon-package-768x1036.jpg 768w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/Amazon-package.jpg 1024w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/Amazon-package-9x12.jpg 9w\" sizes=\"auto, (max-width: 759px) 100vw, 759px\" \/><\/figure>\n\n\n\n<p style=\"font-size:17px\">Em seguida, carreguei a fotografia que tirei para a Nano Banana e dei-lhe a seguinte indica\u00e7\u00e3o:<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<p class=\"has-vivid-cyan-blue-color has-text-color has-link-color wp-elements-41ee9a66efaa63aab6aedf0d2e80cce1\"><strong>“<em>Quero que edites esta fotografia, n\u00e3o alteres nada, exceto o local onde se encontra a etiqueta, acrescenta-lhe danos e manchas pretas de lama na caixa que parecem mesmo manchas de gordura<\/em>“<\/strong><\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h3 class=\"wp-block-heading\">Sa\u00edda de Nano Bana:<\/h3>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"747\" height=\"1024\" src=\"https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/GV3-Amazon-Package-Damaged-747x1024.jpg\" alt=\"Uma fotografia de uma encomenda\" class=\"wp-image-5292\" title=\"\" srcset=\"https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/GV3-Amazon-Package-Damaged-747x1024.jpg 747w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/GV3-Amazon-Package-Damaged-219x300.jpg 219w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/GV3-Amazon-Package-Damaged-768x1052.jpg 768w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/GV3-Amazon-Package-Damaged-9x12.jpg 9w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/GV3-Amazon-Package-Damaged.jpg 864w\" sizes=\"auto, (max-width: 747px) 100vw, 747px\" \/><\/figure>\n\n\n\n<style>\n @import url('https:\/\/fonts.googleapis.com\/css2?family=DM+Sans:wght@400;500;600;700&family=Outfit:wght@600;700;800&display=swap');\n\n .detector-scores-wrap {\n max-width: 780px;\n margin: 2rem auto;\n font-family: 'DM Sans', sans-serif;\n -webkit-font-smoothing: antialiased;\n }\n\n .detector-scores-wrap h3 {\n font-family: 'Outfit', sans-serif;\n font-size: 1.4rem;\n font-weight: 800;\n color: #1a1a2e;\n margin-bottom: 1.25rem;\n letter-spacing: -0.02em;\n }\n\n .detector-scores-table {\n width: 100%;\n border-collapse: separate;\n border-spacing: 0;\n border-radius: 14px;\n overflow: hidden;\n table-layout: fixed;\n box-shadow:\n 0 1px 3px rgba(0,0,0,0.06),\n 0 8px 24px rgba(0,0,0,0.08),\n 0 0 0 1px rgba(0,0,0,0.04);\n }\n\n .detector-scores-table thead th {\n background: #1a1a2e;\n color: #ffffff;\n font-family: 'Outfit', sans-serif;\n font-weight: 700;\n font-size: 0.7rem;\n text-transform: uppercase;\n letter-spacing: 0.04em;\n padding: 14px 6px;\n text-align: center;\n border: none;\n }\n\n .detector-scores-table thead th:first-child {\n text-align: left;\n padding-left: 14px;\n width: 18%;\n }\n\n .detector-scores-table thead th:nth-child(2) {\n width: 12%;\n }\n\n .detector-scores-table tbody tr {\n transition: background 0.2s ease;\n }\n\n .detector-scores-table tbody tr:hover {\n background: rgba(0,0,0,0.02);\n }\n\n .detector-scores-table tbody td {\n padding: 14px 6px;\n text-align: center;\n font-size: 0.9rem;\n font-weight: 500;\n color: #2d2d3a;\n border-bottom: 1px solid #eee;\n }\n\n .detector-scores-table tbody tr:last-child td {\n border-bottom: none;\n }\n\n .detector-scores-table tbody td:first-child {\n text-align: left;\n padding-left: 14px;\n font-weight: 600;\n color: #1a1a2e;\n }\n\n .score-pill {\n font-weight: 700;\n font-family: 'Outfit', sans-serif;\n font-size: 0.78rem;\n border-radius: 6px;\n padding: 5px 8px;\n display: inline-block;\n min-width: 0;\n text-align: center;\n white-space: nowrap;\n }\n\n .score-pass {\n background: #22c55e;\n color: #fff;\n }\n\n .score-warn {\n background: #fb923c;\n color: #fff;\n }\n\n .score-fail {\n background: #ef4444;\n color: #fff;\n }\n\n @media (max-width: 600px) {\n .detector-scores-table thead th,\n .detector-scores-table tbody td {\n padding: 10px 4px;\n font-size: 0.7rem;\n }\n .score-pill {\n padding: 4px 5px;\n font-size: 0.68rem;\n }\n .detector-scores-wrap h3 {\n font-size: 1.15rem;\n }\n }\n<\/style>\n\n<div class=\"detector-scores-wrap\">\n <h3>Como os detectores de imagem pontuaram<\/h3>\n\n <table class=\"detector-scores-table\">\n <thead>\n <tr>\n <th>Modelo de imagem<\/th>\n <th>Categoria<\/th>\n <th>TruthScan<\/th>\n <th>Motor de vis\u00e3o<\/th>\n <th>IA ou n\u00e3o<\/th>\n <th>Winston AI<\/th>\n <th>EraIssoAI<\/th>\n <\/tr>\n <\/thead>\n <tbody>\n <tr>\n <td>Nano Banana <span style=\"font-size:0.75rem; color:#888; display:block;\">Edi\u00e7\u00e3o Deepfake<\/span><\/td>\n <td>Fraude<\/td>\n <td><span class=\"score-pill score-pass\">99.00% AI<\/span><\/td>\n <td><span class=\"score-pill score-pass\">98.00% AI<\/span><\/td>\n <td><span class=\"score-pill score-pass\">99.10% AI<\/span><\/td>\n <td><span class=\"score-pill score-fail\">31.98% AI<\/span><\/td>\n <td><span class=\"score-pill score-pass\">99.00% AI<\/span><\/td>\n <\/tr>\n <\/tbody>\n <\/table>\n<\/div>\n\n\n\n<p style=\"font-size:17px\">O lodo parecido com \u00f3leo que o Nano Banana gerou na sa\u00edda tinha um brilho sujo convincente, completo com um efeito de encharcamento - nada cartoon. Felizmente, a maioria dos detectores que utilizei rotularam corretamente a imagem como falsa. A \u00fanica exce\u00e7\u00e3o foi o Winston, que mais uma vez teve um desempenho fraco, classificando erradamente a imagem como aut\u00eantica. AI or Not, TruthScan e WasItAI foram os detectores mais precisos na dete\u00e7\u00e3o desta imagem.<\/p>\n\n\n\n<p style=\"font-size:17px\">\u00c9 preocupante como parece simples carregar uma imagem para um chatbot e alter\u00e1-la completamente. Consigo imaginar fraudadores e burl\u00f5es a utilizarem ferramentas de gera\u00e7\u00e3o de imagens para tentarem fazer pedidos de devolu\u00e7\u00e3o fraudulentos ou participarem em fraudes no mercado.<\/p>\n\n\n\n<p style=\"font-size:17px\">Se as plataformas de com\u00e9rcio eletr\u00f3nico apenas exigirem uma imagem como prova para iniciar uma reclama\u00e7\u00e3o de um pacote danificado, ent\u00e3o, sem qualquer forma fi\u00e1vel de detetar a adultera\u00e7\u00e3o por IA, ficam expostas a um importante vetor de ataque. Por exemplo, o s\u00edtio oficial da Amazon <a href=\"https:\/\/sellercentral.amazon.com\/help\/hub\/reference\/external\/GYW2EV9FEMJ3JJGR?locale=en-US\" target=\"_blank\" rel=\"noopener\">pol\u00edtica de devolu\u00e7\u00e3o de artigos<\/a> afirma literalmente<strong>, “<em>Os materiais perigosos, incluindo os l\u00edquidos ou gases inflam\u00e1veis, n\u00e3o s\u00e3o recuper\u00e1veis<\/em>.”<\/strong><\/p>\n\n\n\n<p style=\"font-size:17px\">Assim, se recebesse uma encomenda coberta de \u00f3leo ou de lama negra gordurosa, n\u00e3o teria de a devolver, mas poderia obter um reembolso. Se, por alguma raz\u00e3o, isso lhe acontecesse, a Amazon pediria provavelmente uma prova fotogr\u00e1fica. Est\u00e1 a ver o problema que as imagens geradas por IA podem colocar quando s\u00e3o indistingu\u00edveis das imagens reais?<\/p>\n\n\n\n<h2 class=\"wp-block-heading has-text-align-left\">#4. \"Refei\u00e7\u00e3o de barata\" (gerada com Nano Banana)<\/h2>\n\n\n\n<p style=\"font-size:17px\">Esta pr\u00f3xima imagem de IA pode fazer-te perder o apetite. Acontece que h\u00e1 casos documentados de pessoas que utilizam a IA para receber <a href=\"https:\/\/dailydot.com\/ai-generated-undercooked-food-doordash\" data-type=\"link\" data-id=\"https:\/\/dailydot.com\/ai-generated-undercooked-food-doordash\" target=\"_blank\" rel=\"noopener\">reembolsos fraudulentos<\/a> em aplica\u00e7\u00f5es de entrega de comida. Para este quarto teste, segui os mesmos passos que no teste tr\u00eas. Desta vez, tirei uma fotografia da minha caixa de comida vazia depois do almo\u00e7o.<\/p>\n\n\n\n<p><strong>Aqui est\u00e1 a imagem real:<\/strong><\/p>\n\n\n\n<figure class=\"wp-block-image size-large is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"768\" height=\"1024\" src=\"https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/IMG_6145-768x1024.jpg\" alt=\"Uma fotografia de uma caixa de comida para levar vazia.\" class=\"wp-image-5296\" style=\"width:462px;height:auto\" title=\"\" srcset=\"https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/IMG_6145-768x1024.jpg 768w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/IMG_6145-225x300.jpg 225w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/IMG_6145-scaled.jpg 1024w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/IMG_6145-9x12.jpg 9w\" sizes=\"auto, (max-width: 768px) 100vw, 768px\" \/><\/figure>\n\n\n\n<p style=\"font-size:17px\">Em seguida, coloquei a fotografia real que tirei no Nano Banana e dei-lhe o seguinte comando:<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<p class=\"has-vivid-cyan-blue-color has-text-color has-link-color wp-elements-e9f00ca4e893e7742bce463148a639f5\"><strong><em>\"Editar esta imagem. N\u00e3o altere nada, exceto acrescentar massa no local onde est\u00e1 a folha de alum\u00ednio e acrescentar um monte de pequenas baratas beb\u00e9s na comida. Faz com que tudo pare\u00e7a realista e n\u00e3o caricatural\". <\/em><\/strong><\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h3 class=\"wp-block-heading\">A sa\u00edda da Nano Banana:<\/h3>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"747\" height=\"1024\" src=\"https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/GV3-Cockroach-meal.jpg\" alt=\"Uma fotografia alterada pela IA da caixa de comida para levar com massa e insectos dentro da caixa\" class=\"wp-image-5295\" title=\"\" srcset=\"https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/GV3-Cockroach-meal.jpg 747w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/GV3-Cockroach-meal-219x300.jpg 219w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/GV3-Cockroach-meal-9x12.jpg 9w\" sizes=\"auto, (max-width: 747px) 100vw, 747px\" \/><\/figure>\n\n\n\n<style>\n @import url('https:\/\/fonts.googleapis.com\/css2?family=DM+Sans:wght@400;500;600;700&family=Outfit:wght@600;700;800&display=swap');\n\n .detector-scores-wrap {\n max-width: 780px;\n margin: 2rem auto;\n font-family: 'DM Sans', sans-serif;\n -webkit-font-smoothing: antialiased;\n }\n\n .detector-scores-wrap h3 {\n font-family: 'Outfit', sans-serif;\n font-size: 1.4rem;\n font-weight: 800;\n color: #1a1a2e;\n margin-bottom: 1.25rem;\n letter-spacing: -0.02em;\n }\n\n .detector-scores-table {\n width: 100%;\n border-collapse: separate;\n border-spacing: 0;\n border-radius: 14px;\n overflow: hidden;\n table-layout: fixed;\n box-shadow:\n 0 1px 3px rgba(0,0,0,0.06),\n 0 8px 24px rgba(0,0,0,0.08),\n 0 0 0 1px rgba(0,0,0,0.04);\n }\n\n .detector-scores-table thead th {\n background: #1a1a2e;\n color: #ffffff;\n font-family: 'Outfit', sans-serif;\n font-weight: 700;\n font-size: 0.7rem;\n text-transform: uppercase;\n letter-spacing: 0.04em;\n padding: 14px 6px;\n text-align: center;\n border: none;\n }\n\n .detector-scores-table thead th:first-child {\n text-align: left;\n padding-left: 14px;\n width: 18%;\n }\n\n .detector-scores-table thead th:nth-child(2) {\n width: 12%;\n }\n\n .detector-scores-table tbody tr {\n transition: background 0.2s ease;\n }\n\n .detector-scores-table tbody tr:hover {\n background: rgba(0,0,0,0.02);\n }\n\n .detector-scores-table tbody td {\n padding: 14px 6px;\n text-align: center;\n font-size: 0.9rem;\n font-weight: 500;\n color: #2d2d3a;\n border-bottom: 1px solid #eee;\n }\n\n .detector-scores-table tbody tr:last-child td {\n border-bottom: none;\n }\n\n .detector-scores-table tbody td:first-child {\n text-align: left;\n padding-left: 14px;\n font-weight: 600;\n color: #1a1a2e;\n }\n\n .score-pill {\n font-weight: 700;\n font-family: 'Outfit', sans-serif;\n font-size: 0.78rem;\n border-radius: 6px;\n padding: 5px 8px;\n display: inline-block;\n min-width: 0;\n text-align: center;\n white-space: nowrap;\n }\n\n .score-pass {\n background: #22c55e;\n color: #fff;\n }\n\n .score-warn {\n background: #fb923c;\n color: #fff;\n }\n\n .score-fail {\n background: #ef4444;\n color: #fff;\n }\n\n @media (max-width: 600px) {\n .detector-scores-table thead th,\n .detector-scores-table tbody td {\n padding: 10px 4px;\n font-size: 0.7rem;\n }\n .score-pill {\n padding: 4px 5px;\n font-size: 0.68rem;\n }\n .detector-scores-wrap h3 {\n font-size: 1.15rem;\n }\n }\n<\/style>\n\n<div class=\"detector-scores-wrap\">\n <h3>Como os detectores de imagem pontuaram<\/h3>\n\n <table class=\"detector-scores-table\">\n <thead>\n <tr>\n <th>Modelo de imagem<\/th>\n <th>Categoria<\/th>\n <th>TruthScan<\/th>\n <th>Motor de vis\u00e3o<\/th>\n <th>IA ou n\u00e3o<\/th>\n <th>Winston AI<\/th>\n <th>EraIssoAI<\/th>\n <\/tr>\n <\/thead>\n <tbody>\n <tr>\n <td>Nano Banana <span style=\"font-size:0.75rem; color:#888; display:block;\">Edi\u00e7\u00e3o Deepfake<\/span><\/td>\n <td>Fraude<\/td>\n <td><span class=\"score-pill score-pass\">97.15% AI<\/span><\/td>\n <td><span class=\"score-pill score-fail\">18.00% AI<\/span><\/td>\n <td><span class=\"score-pill score-warn\">85.87% AI<\/span><\/td>\n <td><span class=\"score-pill score-warn\">72.29% AI<\/span><\/td>\n <td><span class=\"score-pill score-fail\">1.00% AI<\/span><\/td>\n <\/tr>\n <\/tbody>\n <\/table>\n<\/div>\n\n\n\n<p style=\"font-size:17px\">Mais uma vez, o TruthScan detectou com confian\u00e7a os m\u00e9dia gerados pela IA. O Sight Engine e o WasitAI falharam completamente. AI or Not foi o segundo detetor mais preciso (com 85% de certeza de IA) ao analisar esta imagem de IA, e Winston teve um desempenho melhor do que no terceiro teste (com 72% de certeza de IA).<\/p>\n\n\n\n<p style=\"font-size:17px\">Nesta imagem, a massa e as baratas foram todas geradas inteiramente por IA; no entanto, dois dos detectores n\u00e3o pareciam pensar assim, e apenas um (TruthScan) tinha mais de 90% de certeza de adultera\u00e7\u00e3o por IA.<\/p>\n\n\n\n<p style=\"font-size:17px\">Um ponto de preocupa\u00e7\u00e3o aqui foi o qu\u00e3o simples foi criar esta imagem falsa. Desde o carregamento da fotografia que tirei, at\u00e9 \u00e0 ativa\u00e7\u00e3o do Nano Banana e \u00e0 rece\u00e7\u00e3o do resultado, demorou apenas cerca de 2 minutos. Um pesadelo para as aplica\u00e7\u00f5es de entrega de comida; um sonho para os mentirosos esfomeados. Um \"fator de fraude\".<\/p>\n\n\n\n<p class=\"has-vivid-red-color has-text-color has-link-color wp-elements-4c89b6ee3afcbdd6b6617c150e5b2e03\"><strong>DIVULGA\u00c7\u00c3O: Algumas das imagens geradas por IA inclu\u00eddas nos testes seguintes destinavam-se a demonstrar como as imagens de IA podem ser utilizadas para desinforma\u00e7\u00e3o. Cont\u00eam imagens pol\u00e9micas e temas sens\u00edveis<\/strong>, <strong>e N\u00c3O se destinam a representar qualquer pol\u00edtica espec\u00edfica<\/strong> <strong>pontos de vista ou ideologias.<\/strong><\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><strong>#5<\/strong>. \"Autovandalismo\" (gerado com Nano Banana)<\/h2>\n\n\n\n<p style=\"font-size:17px\">Encontrei uma imagem real de um carro \u00e0 venda na Internet. Bastante mundano. Mas e se algu\u00e9m quisesse utilizar esta imagem para espalhar desinforma\u00e7\u00e3o ou apresentar uma falsa reclama\u00e7\u00e3o de seguro autom\u00f3vel? Infelizmente, a IA generativa torna isso r\u00e1pido e f\u00e1cil.<\/p>\n\n\n\n<p style=\"font-size:17px\"><strong>Para este quinto teste, peguei numa imagem real de um carro e dei \u00e0 Nano Banana a seguinte mensagem:<\/strong><\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<p class=\"has-vivid-cyan-blue-color has-text-color has-link-color wp-elements-d06074fe9533ce61c08bcf9750cd398d\">“<strong><em>N\u00e3o mudes nada no carro, exceto que quero que lhe ponhas tinta em spray que diga \"Palestina Livre\" e que o para-brisas esteja rachado. Os far\u00f3is est\u00e3o esmagados e o espelho est\u00e1 partido e pendurado e as duas janelas do lado do carro est\u00e3o<\/em><\/strong> <strong>esmagado<\/strong>.”<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>Aqui est\u00e1 a imagem real:<\/strong><\/h3>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"940\" height=\"627\" src=\"https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/2000_honda_civic-si_2000_honda_civic-si_bbeed9f1-7b50-4314-a504-40be9411b8ab-mXhf1y-2.webp\" alt=\"Uma fotografia de um carro azul\" class=\"wp-image-5298\" title=\"\" srcset=\"https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/2000_honda_civic-si_2000_honda_civic-si_bbeed9f1-7b50-4314-a504-40be9411b8ab-mXhf1y-2.webp 940w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/2000_honda_civic-si_2000_honda_civic-si_bbeed9f1-7b50-4314-a504-40be9411b8ab-mXhf1y-2-300x200.webp 300w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/2000_honda_civic-si_2000_honda_civic-si_bbeed9f1-7b50-4314-a504-40be9411b8ab-mXhf1y-2-768x512.webp 768w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/2000_honda_civic-si_2000_honda_civic-si_bbeed9f1-7b50-4314-a504-40be9411b8ab-mXhf1y-2-18x12.webp 18w\" sizes=\"auto, (max-width: 940px) 100vw, 940px\" \/><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\">Imagem da IA da Nano Banana:<\/h3>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"683\" src=\"https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/GV3-Protest-Car-Damage.jpg\" alt=\"Uma fotografia alterada pela IA do carro azul, que mostra um para-brisas rachado e danos causados pela pintura com spray na parte lateral do carro\" class=\"wp-image-5299\" title=\"\" srcset=\"https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/GV3-Protest-Car-Damage.jpg 1024w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/GV3-Protest-Car-Damage-300x200.jpg 300w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/GV3-Protest-Car-Damage-768x512.jpg 768w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/GV3-Protest-Car-Damage-18x12.jpg 18w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<style>\n @import url('https:\/\/fonts.googleapis.com\/css2?family=DM+Sans:wght@400;500;600;700&family=Outfit:wght@600;700;800&display=swap');\n\n .detector-scores-wrap {\n max-width: 780px;\n margin: 2rem auto;\n font-family: 'DM Sans', sans-serif;\n -webkit-font-smoothing: antialiased;\n }\n\n .detector-scores-wrap h3 {\n font-family: 'Outfit', sans-serif;\n font-size: 1.4rem;\n font-weight: 800;\n color: #1a1a2e;\n margin-bottom: 1.25rem;\n letter-spacing: -0.02em;\n }\n\n .detector-scores-table {\n width: 100%;\n border-collapse: separate;\n border-spacing: 0;\n border-radius: 14px;\n overflow: hidden;\n table-layout: fixed;\n box-shadow:\n 0 1px 3px rgba(0,0,0,0.06),\n 0 8px 24px rgba(0,0,0,0.08),\n 0 0 0 1px rgba(0,0,0,0.04);\n }\n\n .detector-scores-table thead th {\n background: #1a1a2e;\n color: #ffffff;\n font-family: 'Outfit', sans-serif;\n font-weight: 700;\n font-size: 0.7rem;\n text-transform: uppercase;\n letter-spacing: 0.04em;\n padding: 14px 6px;\n text-align: center;\n border: none;\n }\n\n .detector-scores-table thead th:first-child {\n text-align: left;\n padding-left: 14px;\n width: 18%;\n }\n\n .detector-scores-table thead th:nth-child(2) {\n width: 14%;\n }\n\n .detector-scores-table tbody tr {\n transition: background 0.2s ease;\n }\n\n .detector-scores-table tbody tr:hover {\n background: rgba(0,0,0,0.02);\n }\n\n .detector-scores-table tbody td {\n padding: 14px 6px;\n text-align: center;\n font-size: 0.9rem;\n font-weight: 500;\n color: #2d2d3a;\n border-bottom: 1px solid #eee;\n }\n\n .detector-scores-table tbody tr:last-child td {\n border-bottom: none;\n }\n\n .detector-scores-table tbody td:first-child {\n text-align: left;\n padding-left: 14px;\n font-weight: 600;\n color: #1a1a2e;\n }\n\n .score-pill {\n font-weight: 700;\n font-family: 'Outfit', sans-serif;\n font-size: 0.78rem;\n border-radius: 6px;\n padding: 5px 8px;\n display: inline-block;\n min-width: 0;\n text-align: center;\n white-space: nowrap;\n }\n\n .score-pass {\n background: #22c55e;\n color: #fff;\n }\n\n .score-warn {\n background: #fb923c;\n color: #fff;\n }\n\n .score-fail {\n background: #ef4444;\n color: #fff;\n }\n\n @media (max-width: 600px) {\n .detector-scores-table thead th,\n .detector-scores-table tbody td {\n padding: 10px 4px;\n font-size: 0.7rem;\n }\n .score-pill {\n padding: 4px 5px;\n font-size: 0.68rem;\n }\n .detector-scores-wrap h3 {\n font-size: 1.15rem;\n }\n }\n<\/style>\n\n<div class=\"detector-scores-wrap\">\n <h3>Como os detectores de imagem pontuaram<\/h3>\n\n <table class=\"detector-scores-table\">\n <thead>\n <tr>\n <th>Modelo de imagem<\/th>\n <th>Categoria<\/th>\n <th>TruthScan<\/th>\n <th>Motor de vis\u00e3o<\/th>\n <th>IA ou n\u00e3o<\/th>\n <th>Winston AI<\/th>\n <th>EraIssoAI<\/th>\n <\/tr>\n <\/thead>\n <tbody>\n <tr>\n <td>Nano Banana <span style=\"font-size:0.75rem; color:#888; display:block;\">Edi\u00e7\u00e3o Deepfake<\/span><\/td>\n <td style=\"font-size:0.82rem;\">Fraude \/ Desinforma\u00e7\u00e3o<\/td>\n <td><span class=\"score-pill score-pass\">97.48% AI<\/span><\/td>\n <td><span class=\"score-pill score-fail\">18.00% AI<\/span><\/td>\n <td><span class=\"score-pill score-warn\">89.52% AI<\/span><\/td>\n <td><span class=\"score-pill score-fail\">0,02% AI<\/span><\/td>\n <td><span class=\"score-pill score-pass\">99.00% AI<\/span><\/td>\n <\/tr>\n <\/tbody>\n <\/table>\n<\/div>\n\n\n\n<p style=\"font-size:17px\">Os resultados do detetor para esta sa\u00edda foram mistos. Para esse teste, o WasItAI (surpreendentemente) teve a maior pontua\u00e7\u00e3o de dete\u00e7\u00e3o de IA (99%). O TruthScan teve a segunda maior pontua\u00e7\u00e3o (97%). O Sight Engine falhou significativamente na dete\u00e7\u00e3o de quaisquer elementos sint\u00e9ticos na imagem, e o Winston teve o pior desempenho de todos os detectores (atribuindo uma probabilidade de envolvimento inferior a 1%). O AI or Not se saiu bem, mas a precis\u00e3o ainda estava abaixo de 90%.<br><br>Falemos da imagem. Os dois aspectos mais preocupantes de conte\u00fados alterados ou criados por IA como este s\u00e3o o facto de ambos envolverem engano. Em primeiro lugar, a utiliza\u00e7\u00e3o de IA para alterar imagens mundanas para criar conte\u00fado politicamente carregado \u00e9 r\u00e1pida, f\u00e1cil e parece real. Quer os autores deste tipo de conte\u00fado tenham apenas como objetivo o envolvimento ou sejam provocadores pol\u00edticos, o problema \u00e9 que se trata de informa\u00e7\u00e3o falsa apresentada como aut\u00eantica.<\/p>\n\n\n\n<p style=\"font-size:17px\">A segunda forma de utiliza\u00e7\u00e3o abusiva de ferramentas de gera\u00e7\u00e3o de imagens como esta \u00e9 a fraude. Se um burl\u00e3o quiser apresentar um pedido de seguro autom\u00f3vel falso, ter\u00e1 de produzir provas falsas. Em vez de passar horas no Photoshop a editar imagens e a adulterar provas, pode utilizar geradores de imagens para produzir rapidamente provas falsas.<\/p>\n\n\n\n<p style=\"font-size:17px\">No exemplo da imagem de IA acima, repare como um burl\u00e3o pode utilizar um processo semelhante para apresentar um pedido de indemniza\u00e7\u00e3o falso, de acordo com a <a href=\"https:\/\/www.progressive.com\/answers\/does-car-insurance-cover-vandalism\/\" data-type=\"link\" data-id=\"https:\/\/www.progressive.com\/answers\/does-car-insurance-cover-vandalism\/\" target=\"_blank\" rel=\"noopener\">Pol\u00edtica de cobertura abrangente<\/a>, que abrange especificamente:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Cortado ou <a href=\"https:\/\/www.progressive.com\/answers\/does-car-insurance-cover-tire-damage\/\" target=\"_blank\" rel=\"noopener\">pneus danificados<\/a><\/strong><\/li>\n\n\n\n<li><strong>Vidros, far\u00f3is ou luzes traseiras partidos<\/strong><\/li>\n\n\n\n<li><strong>Danos causados pela pintura com spray<\/strong><\/li>\n\n\n\n<li><strong><a href=\"https:\/\/www.progressive.com\/answers\/does-car-insurance-cover-scratches\/\" target=\"_blank\" rel=\"noopener\">Mossas ou riscos<\/a> de algu\u00e9m que tenha posto a chave no seu carro<\/strong><\/li>\n\n\n\n<li><strong>Colocar a\u00e7\u00facar ou outras subst\u00e2ncias no dep\u00f3sito de combust\u00edvel<\/strong><\/li>\n<\/ul>\n\n\n\n<p>\u00c9 claro que apresentar provas falsas n\u00e3o \u00e9 apenas enganador, \u00e9 tamb\u00e9m completamente ilegal. Num mundo ideal, ningu\u00e9m infringiria a lei; na realidade, o crime acontece todos os dias. E, sem uma dete\u00e7\u00e3o fi\u00e1vel, as companhias de seguros autom\u00f3veis correm o risco de ser enganadas em milh\u00f5es, \u00e0 medida que aumentam os pedidos de indemniza\u00e7\u00e3o fraudulentos com recurso \u00e0 IA.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">#6. \"Soda Musk e Don\" (gerado com Nano Banana)<\/h2>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"559\" src=\"https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/GV3-Elon-Spilling-Soda-1024x559.jpeg\" alt=\"Uma imagem gerada por IA de Elon Musk e Donald Trump\" class=\"wp-image-5300\" title=\"\" srcset=\"https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/GV3-Elon-Spilling-Soda-1024x559.jpeg 1024w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/GV3-Elon-Spilling-Soda-300x164.jpeg 300w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/GV3-Elon-Spilling-Soda-768x419.jpeg 768w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/GV3-Elon-Spilling-Soda-1536x838.jpeg 1536w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/GV3-Elon-Spilling-Soda-2048x1117.jpeg 1877w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/GV3-Elon-Spilling-Soda-18x10.jpeg 18w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/GV3-Elon-Spilling-Soda-scaled.jpeg 1878w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<style>\n @import url('https:\/\/fonts.googleapis.com\/css2?family=DM+Sans:wght@400;500;600;700&family=Outfit:wght@600;700;800&display=swap');\n\n .detector-scores-wrap {\n max-width: 780px;\n margin: 2rem auto;\n font-family: 'DM Sans', sans-serif;\n -webkit-font-smoothing: antialiased;\n }\n\n .detector-scores-wrap h3 {\n font-family: 'Outfit', sans-serif;\n font-size: 1.4rem;\n font-weight: 800;\n color: #1a1a2e;\n margin-bottom: 1.25rem;\n letter-spacing: -0.02em;\n }\n\n .detector-scores-table {\n width: 100%;\n border-collapse: separate;\n border-spacing: 0;\n border-radius: 14px;\n overflow: hidden;\n table-layout: fixed;\n box-shadow:\n 0 1px 3px rgba(0,0,0,0.06),\n 0 8px 24px rgba(0,0,0,0.08),\n 0 0 0 1px rgba(0,0,0,0.04);\n }\n\n .detector-scores-table thead th {\n background: #1a1a2e;\n color: #ffffff;\n font-family: 'Outfit', sans-serif;\n font-weight: 700;\n font-size: 0.7rem;\n text-transform: uppercase;\n letter-spacing: 0.04em;\n padding: 14px 6px;\n text-align: center;\n border: none;\n }\n\n .detector-scores-table thead th:first-child {\n text-align: left;\n padding-left: 14px;\n width: 18%;\n }\n\n .detector-scores-table thead th:nth-child(2) {\n width: 14%;\n }\n\n .detector-scores-table tbody tr {\n transition: background 0.2s ease;\n }\n\n .detector-scores-table tbody tr:hover {\n background: rgba(0,0,0,0.02);\n }\n\n .detector-scores-table tbody td {\n padding: 14px 6px;\n text-align: center;\n font-size: 0.9rem;\n font-weight: 500;\n color: #2d2d3a;\n border-bottom: 1px solid #eee;\n }\n\n .detector-scores-table tbody tr:last-child td {\n border-bottom: none;\n }\n\n .detector-scores-table tbody td:first-child {\n text-align: left;\n padding-left: 14px;\n font-weight: 600;\n color: #1a1a2e;\n }\n\n .score-pill {\n font-weight: 700;\n font-family: 'Outfit', sans-serif;\n font-size: 0.78rem;\n border-radius: 6px;\n padding: 5px 8px;\n display: inline-block;\n min-width: 0;\n text-align: center;\n white-space: nowrap;\n }\n\n .score-pass {\n background: #22c55e;\n color: #fff;\n }\n\n .score-warn {\n background: #fb923c;\n color: #fff;\n }\n\n .score-fail {\n background: #ef4444;\n color: #fff;\n }\n\n @media (max-width: 600px) {\n .detector-scores-table thead th,\n .detector-scores-table tbody td {\n padding: 10px 4px;\n font-size: 0.7rem;\n }\n .score-pill {\n padding: 4px 5px;\n font-size: 0.68rem;\n }\n .detector-scores-wrap h3 {\n font-size: 1.15rem;\n }\n }\n<\/style>\n\n<div class=\"detector-scores-wrap\">\n <h3>Como os detectores de imagem pontuaram<\/h3>\n\n <table class=\"detector-scores-table\">\n <thead>\n <tr>\n <th>Modelo de imagem<\/th>\n <th>Categoria<\/th>\n <th>TruthScan<\/th>\n <th>Motor de vis\u00e3o<\/th>\n <th>IA ou n\u00e3o<\/th>\n <th>Winston AI<\/th>\n <th>EraIssoAI<\/th>\n <\/tr>\n <\/thead>\n <tbody>\n <tr>\n <td>Nano Banana <span style=\"font-size:0.75rem; color:#888; display:block;\">Gera\u00e7\u00e3o completa<\/span><\/td>\n <td style=\"font-size:0.82rem;\">Desinforma\u00e7\u00e3o<\/td>\n <td><span class=\"score-pill score-pass\">99.00% AI<\/span><\/td>\n <td><span class=\"score-pill score-pass\">99.00% AI<\/span><\/td>\n <td><span class=\"score-pill score-pass\">99.00% AI<\/span><\/td>\n <td><span class=\"score-pill score-fail\">0,24% AI<\/span><\/td>\n <td><span class=\"score-pill score-warn\">86.00% AI<\/span><\/td>\n <\/tr>\n <\/tbody>\n <\/table>\n<\/div>\n\n\n\n<p style=\"font-size:17px\">Mais uma vez, utilizei o Nano Banana para este teste e o objetivo era demonstrar imagens falsas, politicamente carregadas, com um tom mais rid\u00edculo. O TruthScan, o Sight Engine e o AIorNot assinalaram a imagem gerada com uma classifica\u00e7\u00e3o de dete\u00e7\u00e3o de IA de 99%. O WasitAI n\u00e3o estava t\u00e3o confiante, detectando a imagem gerada pela IA com apenas 86% de certeza. E o Winston falhou completamente, atribuindo uma pontua\u00e7\u00e3o de IA de 0,2%. Para gerar a imagem, dei \u00e0 Nano Banana esta mensagem:<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<p class=\"has-vivid-cyan-blue-color has-text-color has-link-color wp-elements-4f7e002ecaefec24bbebdca3101fff42\"><strong>“<em>Gerar uma fotografia de Elon Musk a entornar um copo grande de refrigerante tipo Big Gulp na camisa e a passar-se enquanto Donald Trump est\u00e1 sentado ao lado dele a rir-se. Est\u00e3o num avi\u00e3o a jato. \u00c9 essa a cena. E a fotografia deve ter um aspeto realista, como se tivesse sido tirada com uma c\u00e2mara instant\u00e2nea e incluir artefactos como o reflexo do flash da c\u00e2mara na janela do avi\u00e3o e no exterior do avi\u00e3o. \u00c9 de noite. Deve parecer hiper-realista.<\/em>“<\/strong><\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<p style=\"font-size:17px\">Sempre que um destes programas de software n\u00e3o consegue detetar a IA, rotulam-no como aut\u00eantico. Por mais disparatado que seja este exemplo de imagem do teste, n\u00e3o h\u00e1 nada de engra\u00e7ado em pensar que um deepfake \u00e9 real.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">#7. \"Still Alive\" (gerado por Nano Banana)<\/h2>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"765\" height=\"1024\" src=\"https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/GV3Christian_Ransom_Photo.jpg\" alt=\"Uma imagem gerada por IA de um homem a segurar um cartaz que diz "ajuda"\" class=\"wp-image-5301\" title=\"\" srcset=\"https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/GV3Christian_Ransom_Photo.jpg 765w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/GV3Christian_Ransom_Photo-224x300.jpg 224w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/GV3Christian_Ransom_Photo-9x12.jpg 9w\" sizes=\"auto, (max-width: 765px) 100vw, 765px\" \/><\/figure>\n\n\n\n<style>\n @import url('https:\/\/fonts.googleapis.com\/css2?family=DM+Sans:wght@400;500;600;700&family=Outfit:wght@600;700;800&display=swap');\n\n .detector-scores-wrap {\n max-width: 780px;\n margin: 2rem auto;\n font-family: 'DM Sans', sans-serif;\n -webkit-font-smoothing: antialiased;\n }\n\n .detector-scores-wrap h3 {\n font-family: 'Outfit', sans-serif;\n font-size: 1.4rem;\n font-weight: 800;\n color: #1a1a2e;\n margin-bottom: 1.25rem;\n letter-spacing: -0.02em;\n }\n\n .detector-scores-table {\n width: 100%;\n border-collapse: separate;\n border-spacing: 0;\n border-radius: 14px;\n overflow: hidden;\n table-layout: fixed;\n box-shadow:\n 0 1px 3px rgba(0,0,0,0.06),\n 0 8px 24px rgba(0,0,0,0.08),\n 0 0 0 1px rgba(0,0,0,0.04);\n }\n\n .detector-scores-table thead th {\n background: #1a1a2e;\n color: #ffffff;\n font-family: 'Outfit', sans-serif;\n font-weight: 700;\n font-size: 0.7rem;\n text-transform: uppercase;\n letter-spacing: 0.04em;\n padding: 14px 6px;\n text-align: center;\n border: none;\n }\n\n .detector-scores-table thead th:first-child {\n text-align: left;\n padding-left: 14px;\n width: 18%;\n }\n\n .detector-scores-table thead th:nth-child(2) {\n width: 14%;\n }\n\n .detector-scores-table tbody tr {\n transition: background 0.2s ease;\n }\n\n .detector-scores-table tbody tr:hover {\n background: rgba(0,0,0,0.02);\n }\n\n .detector-scores-table tbody td {\n padding: 14px 6px;\n text-align: center;\n font-size: 0.9rem;\n font-weight: 500;\n color: #2d2d3a;\n border-bottom: 1px solid #eee;\n }\n\n .detector-scores-table tbody tr:last-child td {\n border-bottom: none;\n }\n\n .detector-scores-table tbody td:first-child {\n text-align: left;\n padding-left: 14px;\n font-weight: 600;\n color: #1a1a2e;\n }\n\n .score-pill {\n font-weight: 700;\n font-family: 'Outfit', sans-serif;\n font-size: 0.78rem;\n border-radius: 6px;\n padding: 5px 8px;\n display: inline-block;\n min-width: 0;\n text-align: center;\n white-space: nowrap;\n }\n\n .score-pass {\n background: #22c55e;\n color: #fff;\n }\n\n .score-warn {\n background: #fb923c;\n color: #fff;\n }\n\n .score-fail {\n background: #ef4444;\n color: #fff;\n }\n\n @media (max-width: 600px) {\n .detector-scores-table thead th,\n .detector-scores-table tbody td {\n padding: 10px 4px;\n font-size: 0.7rem;\n }\n .score-pill {\n padding: 4px 5px;\n font-size: 0.68rem;\n }\n .detector-scores-wrap h3 {\n font-size: 1.15rem;\n }\n }\n<\/style>\n\n<div class=\"detector-scores-wrap\">\n <h3>Como os detectores de imagem pontuaram<\/h3>\n\n <table class=\"detector-scores-table\">\n <thead>\n <tr>\n <th>Modelo de imagem<\/th>\n <th>Categoria<\/th>\n <th>TruthScan<\/th>\n <th>Motor de vis\u00e3o<\/th>\n <th>IA ou n\u00e3o<\/th>\n <th>Winston AI<\/th>\n <th>EraIssoAI<\/th>\n <\/tr>\n <\/thead>\n <tbody>\n <tr>\n <td>Nano Banana <span style=\"font-size:0.75rem; color:#888; display:block;\">Gera\u00e7\u00e3o completa<\/span><\/td>\n <td style=\"font-size:0.82rem;\">Falsifica\u00e7\u00e3o profunda<\/td>\n <td><span class=\"score-pill score-pass\">97.49% AI<\/span><\/td>\n <td><span class=\"score-pill score-pass\">95.00% AI<\/span><\/td>\n <td><span class=\"score-pill score-pass\">99.27% AI<\/span><\/td>\n <td><span class=\"score-pill score-fail\">0,15% AI<\/span><\/td>\n <td><span class=\"score-pill score-fail\">1.00% AI<\/span><\/td>\n <\/tr>\n <\/tbody>\n <\/table>\n<\/div>\n\n\n\n<p style=\"font-size:17px\">Achas que esta imagem \u00e9 perturbadora? Para gerar este deepfake, usei o Nano Banana e dei-lhe uma fotografia minha. A seguir, dei-lhe um longo prompt que pretendia colocar-me numa situa\u00e7\u00e3o que implicasse que eu tinha sido raptado. Talvez seja porque esta fotografia mostra <em>eu<\/em> numa posi\u00e7\u00e3o comprometedora. E se algu\u00e9m fizesse uma coisa destas, a enviasse \u00e0 minha m\u00e3e e amea\u00e7asse fazer-me mal se ela n\u00e3o lhe enviasse dinheiro? Coisas assustadoras. <\/p>\n\n\n\n<p class=\"has-small-font-size\"><em>(A mensagem completa que utilizei para este efeito pode ser encontrada no conjunto de dados de teste completo no final deste artigo).<\/em><\/p>\n\n\n\n<p style=\"font-size:17px\"><br>Felizmente para esta imagem, apesar de ter um aspeto muito realista, <strong>tr\u00eas em cada cinco detectores identificaram-na como gerada por IA.<\/strong> <\/p>\n\n\n\n<p style=\"font-size:17px\">AI or Not obteve uma forte pontua\u00e7\u00e3o de IA de 99%, TruthScan com 97% e Sight Engine com 95%. Infelizmente, Winston e WasItAI classificaram a imagem como real.<\/p>\n\n\n\n<p style=\"font-size:17px\">Devo esclarecer que estes testes n\u00e3o foram efectuados na minha conta pessoal Gemini. A Nano Banana nunca me pediu para provar que eu era a pessoa na imagem. Qualquer pessoa poderia ter descarregado uma fotografia minha da Internet, t\u00ea-la colocado na Nano Banana e t\u00ea-la feito criar este tipo de imagem. <\/p>\n\n\n\n<h2 class=\"wp-block-heading\">#8. \"Abuso de um Ben\" (gerado com Nano Banana)<\/h2>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"200\" height=\"200\" src=\"https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/c945d528-730f-48c1-bdca-30e5799fc1aa_thumb-1.jpg\" alt=\"\" class=\"wp-image-5307\" title=\"\" srcset=\"https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/c945d528-730f-48c1-bdca-30e5799fc1aa_thumb-1.jpg 200w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/c945d528-730f-48c1-bdca-30e5799fc1aa_thumb-1-150x150.jpg 150w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/c945d528-730f-48c1-bdca-30e5799fc1aa_thumb-1-12x12.jpg 12w\" sizes=\"auto, (max-width: 200px) 100vw, 200px\" \/><\/figure>\n<\/div>\n\n\n<p style=\"font-size:17px\">Conhe\u00e7a o Ben. Ele passa a maior parte do seu tempo a gerir v\u00e1rios departamentos nas empresas onde trabalha. A imagem acima \u00e9 o (verdadeiro) retrato corporativo elegante do Ben. Se n\u00e3o se consegue perceber, o Ben \u00e9 um tipo profissional. Independentemente das suas opini\u00f5es pol\u00edticas, guarda-as para si e as suas conversas de trabalho s\u00e3o apenas sobre trabalho. Mas, em qualquer altura, a imagem do Ben pode ser utilizada para difundir propaganda ou mensagens pol\u00edticas sem o seu consentimento.<\/p>\n\n\n\n<p style=\"font-size:17px\">Recentemente, tem havido muitas not\u00edcias sobre a forma como v\u00eddeos gerados por IA envolvendo agentes do ICE e protestos se t\u00eam espalhado pela Internet. Com base em todos os testes efectuados at\u00e9 agora, conclu\u00ed que criar propaganda pol\u00edtica falsa com IA era muito f\u00e1cil. Decidi fazer um novo teste, desta vez com o Ben. <\/p>\n\n\n\n<p style=\"font-size:17px\">Com a autoriza\u00e7\u00e3o do Ben, utilizei a sua fotografia de rosto para criar um deepfake. <strong>Aqui est\u00e1 a imagem gerada por IA que a Nano Banana criou:<\/strong><\/p>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"1024\" src=\"https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/Gemini_Generated_Image_qd4x5sqd4x5sqd4x.jpg\" alt=\"Uma fotografia alterada por IA de um homem chamado Ben, segurando um cartaz que diz Abolir o ICE.\" class=\"wp-image-5308\" style=\"width:537px;height:auto\" title=\"\" srcset=\"https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/Gemini_Generated_Image_qd4x5sqd4x5sqd4x.jpg 1024w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/Gemini_Generated_Image_qd4x5sqd4x5sqd4x-300x300.jpg 300w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/Gemini_Generated_Image_qd4x5sqd4x5sqd4x-150x150.jpg 150w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/Gemini_Generated_Image_qd4x5sqd4x5sqd4x-768x768.jpg 768w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/Gemini_Generated_Image_qd4x5sqd4x5sqd4x-12x12.jpg 12w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n<\/div>\n\n\n<style>\n @import url('https:\/\/fonts.googleapis.com\/css2?family=DM+Sans:wght@400;500;600;700&family=Outfit:wght@600;700;800&display=swap');\n\n .detector-scores-wrap {\n max-width: 780px;\n margin: 2rem auto;\n font-family: 'DM Sans', sans-serif;\n -webkit-font-smoothing: antialiased;\n }\n\n .detector-scores-wrap h3 {\n font-family: 'Outfit', sans-serif;\n font-size: 1.4rem;\n font-weight: 800;\n color: #1a1a2e;\n margin-bottom: 1.25rem;\n letter-spacing: -0.02em;\n }\n\n .detector-scores-table {\n width: 100%;\n border-collapse: separate;\n border-spacing: 0;\n border-radius: 14px;\n overflow: hidden;\n table-layout: fixed;\n box-shadow:\n 0 1px 3px rgba(0,0,0,0.06),\n 0 8px 24px rgba(0,0,0,0.08),\n 0 0 0 1px rgba(0,0,0,0.04);\n }\n\n .detector-scores-table thead th {\n background: #1a1a2e;\n color: #ffffff;\n font-family: 'Outfit', sans-serif;\n font-weight: 700;\n font-size: 0.7rem;\n text-transform: uppercase;\n letter-spacing: 0.04em;\n padding: 14px 6px;\n text-align: center;\n border: none;\n }\n\n .detector-scores-table thead th:first-child {\n text-align: left;\n padding-left: 14px;\n width: 18%;\n }\n\n .detector-scores-table thead th:nth-child(2) {\n width: 14%;\n }\n\n .detector-scores-table tbody tr {\n transition: background 0.2s ease;\n }\n\n .detector-scores-table tbody tr:hover {\n background: rgba(0,0,0,0.02);\n }\n\n .detector-scores-table tbody td {\n padding: 14px 6px;\n text-align: center;\n font-size: 0.9rem;\n font-weight: 500;\n color: #2d2d3a;\n border-bottom: 1px solid #eee;\n }\n\n .detector-scores-table tbody tr:last-child td {\n border-bottom: none;\n }\n\n .detector-scores-table tbody td:first-child {\n text-align: left;\n padding-left: 14px;\n font-weight: 600;\n color: #1a1a2e;\n }\n\n .score-pill {\n font-weight: 700;\n font-family: 'Outfit', sans-serif;\n font-size: 0.78rem;\n border-radius: 6px;\n padding: 5px 8px;\n display: inline-block;\n min-width: 0;\n text-align: center;\n white-space: nowrap;\n }\n\n .score-pass {\n background: #22c55e;\n color: #fff;\n }\n\n .score-warn {\n background: #fb923c;\n color: #fff;\n }\n\n .score-fail {\n background: #ef4444;\n color: #fff;\n }\n\n @media (max-width: 600px) {\n .detector-scores-table thead th,\n .detector-scores-table tbody td {\n padding: 10px 4px;\n font-size: 0.7rem;\n }\n .score-pill {\n padding: 4px 5px;\n font-size: 0.68rem;\n }\n .detector-scores-wrap h3 {\n font-size: 1.15rem;\n }\n }\n<\/style>\n\n<div class=\"detector-scores-wrap\">\n <h3>Como os detectores de imagem pontuaram<\/h3>\n\n <table class=\"detector-scores-table\">\n <thead>\n <tr>\n <th>Modelo de imagem<\/th>\n <th>Categoria<\/th>\n <th>TruthScan<\/th>\n <th>Motor de vis\u00e3o<\/th>\n <th>IA ou n\u00e3o<\/th>\n <th>Winston AI<\/th>\n <th>EraIssoAI<\/th>\n <\/tr>\n <\/thead>\n <tbody>\n <tr>\n <td>Nano Banana<\/td>\n <td style=\"font-size:0.78rem;\">Deepfake \/ Desinfo<\/td>\n <td><span class=\"score-pill score-pass\">99.00% AI<\/span><\/td>\n <td><span class=\"score-pill score-pass\">97.00% AI<\/span><\/td>\n <td><span class=\"score-pill score-pass\">98.88% AI<\/span><\/td>\n <td><span class=\"score-pill score-fail\">5.74% AI<\/span><\/td>\n <td><span class=\"score-pill score-pass\">99.00% AI<\/span><\/td>\n <\/tr>\n <\/tbody>\n <\/table>\n<\/div>\n\n\n\n<p style=\"font-size:17px\">Demorou trinta segundos. Tudo o que tive de fazer foi arrastar e largar a fotografia da cabe\u00e7a dele para o Gemini e pedir-lhe que \"levantasse um cartaz a dizer abolir o ICE\". O resultado do Nano Banana n\u00e3o \u00e9 t\u00e3o fotorrealista como os outros, mas parece suficientemente aut\u00eantico para ser persuasivo.<\/p>\n\n\n\n<p style=\"font-size:17px\"><span style=\"box-sizing: border-box; margin: 0px; padding: 0px;\">Felizmente, quatro dos detectores identificaram corretamente a imagem:\u00a0<strong>TruthScan<\/strong> (9<\/span>9% AI), Was It AI (99% AI), Sight Engine (97% AI), e AI or Not (97% AI). Winston AI falhou, atribuindo apenas uma pontua\u00e7\u00e3o de 5% AI.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">#9. \"Club Void\" (gerado por Midjourney)<\/h2>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"771\" src=\"https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/MJ-Girl-in-club-1024x771.jpg\" alt=\"Uma mulher sorridente, com um vestido branco de manga curta, est\u00e1 numa sala pouco iluminada e cheia de gente. A fotografia parece ter sido tirada com um flash, criando uma atmosfera nebulosa e azulada com part\u00edculas t\u00e9nues vis\u00edveis no ar. Outras pessoas com roupas casuais est\u00e3o em segundo plano sob um teto danificado ou de aspeto industrial.\" class=\"wp-image-5309\" title=\"\" srcset=\"https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/MJ-Girl-in-club-1024x771.jpg 1024w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/MJ-Girl-in-club-300x226.jpg 300w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/MJ-Girl-in-club-768x578.jpg 768w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/MJ-Girl-in-club-16x12.jpg 16w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/MJ-Girl-in-club.jpg 1232w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<style>\n @import url('https:\/\/fonts.googleapis.com\/css2?family=DM+Sans:wght@400;500;600;700&family=Outfit:wght@600;700;800&display=swap');\n\n .detector-scores-wrap {\n max-width: 780px;\n margin: 2rem auto;\n font-family: 'DM Sans', sans-serif;\n -webkit-font-smoothing: antialiased;\n }\n\n .detector-scores-wrap h3 {\n font-family: 'Outfit', sans-serif;\n font-size: 1.4rem;\n font-weight: 800;\n color: #1a1a2e;\n margin-bottom: 1.25rem;\n letter-spacing: -0.02em;\n }\n\n .detector-scores-table {\n width: 100%;\n border-collapse: separate;\n border-spacing: 0;\n border-radius: 14px;\n overflow: hidden;\n table-layout: fixed;\n box-shadow:\n 0 1px 3px rgba(0,0,0,0.06),\n 0 8px 24px rgba(0,0,0,0.08),\n 0 0 0 1px rgba(0,0,0,0.04);\n }\n\n .detector-scores-table thead th {\n background: #1a1a2e;\n color: #ffffff;\n font-family: 'Outfit', sans-serif;\n font-weight: 700;\n font-size: 0.7rem;\n text-transform: uppercase;\n letter-spacing: 0.04em;\n padding: 14px 6px;\n text-align: center;\n border: none;\n }\n\n .detector-scores-table thead th:first-child {\n text-align: left;\n padding-left: 14px;\n width: 18%;\n }\n\n .detector-scores-table thead th:nth-child(2) {\n width: 14%;\n }\n\n .detector-scores-table tbody tr {\n transition: background 0.2s ease;\n }\n\n .detector-scores-table tbody tr:hover {\n background: rgba(0,0,0,0.02);\n }\n\n .detector-scores-table tbody td {\n padding: 14px 6px;\n text-align: center;\n font-size: 0.9rem;\n font-weight: 500;\n color: #2d2d3a;\n border-bottom: 1px solid #eee;\n }\n\n .detector-scores-table tbody tr:last-child td {\n border-bottom: none;\n }\n\n .detector-scores-table tbody td:first-child {\n text-align: left;\n padding-left: 14px;\n font-weight: 600;\n color: #1a1a2e;\n }\n\n .score-pill {\n font-weight: 700;\n font-family: 'Outfit', sans-serif;\n font-size: 0.78rem;\n border-radius: 6px;\n padding: 5px 8px;\n display: inline-block;\n min-width: 0;\n text-align: center;\n white-space: nowrap;\n }\n\n .score-pass {\n background: #22c55e;\n color: #fff;\n }\n\n .score-warn {\n background: #fb923c;\n color: #fff;\n }\n\n .score-fail {\n background: #ef4444;\n color: #fff;\n }\n\n @media (max-width: 600px) {\n .detector-scores-table thead th,\n .detector-scores-table tbody td {\n padding: 10px 4px;\n font-size: 0.7rem;\n }\n .score-pill {\n padding: 4px 5px;\n font-size: 0.68rem;\n }\n .detector-scores-wrap h3 {\n font-size: 1.15rem;\n }\n }\n<\/style>\n\n<div class=\"detector-scores-wrap\">\n <h3>Como os detectores de imagem pontuaram<\/h3>\n\n <table class=\"detector-scores-table\">\n <thead>\n <tr>\n <th>Modelo de imagem<\/th>\n <th>Categoria<\/th>\n <th>TruthScan<\/th>\n <th>Motor de vis\u00e3o<\/th>\n <th>IA ou n\u00e3o<\/th>\n <th>Winston AI<\/th>\n <th>EraIssoAI<\/th>\n <\/tr>\n <\/thead>\n <tbody>\n <tr>\n <td>Meio da viagem <span style=\"font-size:0.75rem; color:#888; display:block;\">Gera\u00e7\u00e3o completa<\/span><\/td>\n <td>Geral<\/td>\n <td><span class=\"score-pill score-pass\">99.00% AI<\/span><\/td>\n <td><span class=\"score-pill score-pass\">99.00% AI<\/span><\/td>\n <td><span class=\"score-pill score-warn\">89.21% AI<\/span><\/td>\n <td><span class=\"score-pill score-fail\">1.56% AI<\/span><\/td>\n <td><span class=\"score-pill score-warn\">70.00% AI<\/span><\/td>\n <\/tr>\n <\/tbody>\n <\/table>\n<\/div>\n\n\n\n<p style=\"font-size:17px\">Criei a imagem acima usando o Midjourney. Quanto mais olhava para ela, mais inquietante me parecia. A est\u00e9tica \u00e9 escura e misteriosa. Uma presen\u00e7a sombria parece estar a emergir por detr\u00e1s do objeto. Para criar esta imagem, utilizei a fun\u00e7\u00e3o de refer\u00eancia est\u00e9tica do Midjourney e pedi-lhe simplesmente que fosse uma \"rapariga de p\u00e9 numa discoteca a sorrir\". Mostrei a imagem a algumas pessoas e todas acharam que era real. <br><br>A maioria dos detectores assinalou a imagem como tendo sido gerada por IA. O TruthScan e o Sight Engine assinalaram-na com uma pontua\u00e7\u00e3o de IA de 99%, o AI or Not indicou 89%. <\/p>\n\n\n\n<p style=\"font-size:17px\">O WasItAI n\u00e3o estava t\u00e3o seguro como os detectores, classificando a imagem apenas com uma pontua\u00e7\u00e3o AI de 70%. <\/p>\n\n\n\n<p style=\"font-size:17px\">O detetor do Winston AI falhou completamente, dando \u00e0 imagem uma pontua\u00e7\u00e3o AI de 1,56%. <\/p>\n\n\n\n<h2 class=\"wp-block-heading\">#10 \"Caixa de gelo\"<strong> <\/strong>(gerado por Midjourney)<\/h2>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"771\" src=\"https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/MJ-ICE-1024x771.jpg\" alt=\"Uma fotografia desfocada que mostra um homem com uma camisa manchada de vermelho, de p\u00e9 em frente de algu\u00e9m.\" class=\"wp-image-5311\" title=\"\" srcset=\"https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/MJ-ICE-1024x771.jpg 1024w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/MJ-ICE-300x226.jpg 300w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/MJ-ICE-768x578.jpg 768w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/MJ-ICE-16x12.jpg 16w, https:\/\/research.undetectable.ai\/wp-content\/uploads\/2026\/02\/MJ-ICE.jpg 1232w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<style>\n @import url('https:\/\/fonts.googleapis.com\/css2?family=DM+Sans:wght@400;500;600;700&family=Outfit:wght@600;700;800&display=swap');\n\n .detector-scores-wrap {\n max-width: 780px;\n margin: 2rem auto;\n font-family: 'DM Sans', sans-serif;\n -webkit-font-smoothing: antialiased;\n }\n\n .detector-scores-wrap h3 {\n font-family: 'Outfit', sans-serif;\n font-size: 1.4rem;\n font-weight: 800;\n color: #1a1a2e;\n margin-bottom: 1.25rem;\n letter-spacing: -0.02em;\n }\n\n .detector-scores-table {\n width: 100%;\n border-collapse: separate;\n border-spacing: 0;\n border-radius: 14px;\n overflow: hidden;\n table-layout: fixed;\n box-shadow:\n 0 1px 3px rgba(0,0,0,0.06),\n 0 8px 24px rgba(0,0,0,0.08),\n 0 0 0 1px rgba(0,0,0,0.04);\n }\n\n .detector-scores-table thead th {\n background: #1a1a2e;\n color: #ffffff;\n font-family: 'Outfit', sans-serif;\n font-weight: 700;\n font-size: 0.7rem;\n text-transform: uppercase;\n letter-spacing: 0.04em;\n padding: 14px 6px;\n text-align: center;\n border: none;\n }\n\n .detector-scores-table thead th:first-child {\n text-align: left;\n padding-left: 14px;\n width: 18%;\n }\n\n .detector-scores-table thead th:nth-child(2) {\n width: 14%;\n }\n\n .detector-scores-table tbody tr {\n transition: background 0.2s ease;\n }\n\n .detector-scores-table tbody tr:hover {\n background: rgba(0,0,0,0.02);\n }\n\n .detector-scores-table tbody td {\n padding: 14px 6px;\n text-align: center;\n font-size: 0.9rem;\n font-weight: 500;\n color: #2d2d3a;\n border-bottom: 1px solid #eee;\n }\n\n .detector-scores-table tbody tr:last-child td {\n border-bottom: none;\n }\n\n .detector-scores-table tbody td:first-child {\n text-align: left;\n padding-left: 14px;\n font-weight: 600;\n color: #1a1a2e;\n }\n\n .score-pill {\n font-weight: 700;\n font-family: 'Outfit', sans-serif;\n font-size: 0.78rem;\n border-radius: 6px;\n padding: 5px 8px;\n display: inline-block;\n min-width: 0;\n text-align: center;\n white-space: nowrap;\n }\n\n .score-pass {\n background: #22c55e;\n color: #fff;\n }\n\n .score-warn {\n background: #fb923c;\n color: #fff;\n }\n\n .score-fail {\n background: #ef4444;\n color: #fff;\n }\n\n @media (max-width: 600px) {\n .detector-scores-table thead th,\n .detector-scores-table tbody td {\n padding: 10px 4px;\n font-size: 0.7rem;\n }\n .score-pill {\n padding: 4px 5px;\n font-size: 0.68rem;\n }\n .detector-scores-wrap h3 {\n font-size: 1.15rem;\n }\n }\n<\/style>\n\n<div class=\"detector-scores-wrap\">\n <h3>Como os detectores de imagem pontuaram<\/h3>\n\n <table class=\"detector-scores-table\">\n <thead>\n <tr>\n <th>Modelo de imagem<\/th>\n <th>Categoria<\/th>\n <th>TruthScan<\/th>\n <th>Motor de vis\u00e3o<\/th>\n <th>IA ou n\u00e3o<\/th>\n <th>Winston AI<\/th>\n <th>EraIssoAI<\/th>\n <\/tr>\n <\/thead>\n <tbody>\n <tr>\n <td>Meio da viagem <span style=\"font-size:0.75rem; color:#888; display:block;\">Gera\u00e7\u00e3o completa<\/span><\/td>\n <td style=\"font-size:0.82rem;\">Desinforma\u00e7\u00e3o<\/td>\n <td><span class=\"score-pill score-pass\">99.00% AI<\/span><\/td>\n <td><span class=\"score-pill score-pass\">92.00% AI<\/span><\/td>\n <td><span class=\"score-pill score-pass\">97.94% AI<\/span><\/td>\n <td><span class=\"score-pill score-fail\">41.20% AI<\/span><\/td>\n <td><span class=\"score-pill score-warn\">72.00% AI<\/span><\/td>\n <\/tr>\n <\/tbody>\n <\/table>\n<\/div>\n\n\n\n<p style=\"font-size:17px\">Criei esta imagem utilizando o mesmo m\u00e9todo a meio da viagem do teste anterior. Desta vez, o objetivo era criar algo um pouco mais escuro. O TruthScan foi o detetor mais preciso, rotulando a imagem como 99% AI. O Sight Engine e o AI or Not detectaram corretamente a imagem. O Winston AI falhou completamente e o WasItAI teve uma pontua\u00e7\u00e3o de confian\u00e7a muito mais baixa do que o TruthScan, o Sight Engine e o AI or Not.<\/p>\n\n\n\n<p style=\"font-size:17px\">Devo dizer que, neste caso, a imagem que gerei pode ser utilizada ou enquadrada de duas formas. Por um lado, algu\u00e9m poderia utilizar este tipo de suporte (ou qualquer outro suporte com tem\u00e1tica pol\u00edtica\/desordem civil) em algum tipo de arte ou projeto criativo para fazer uma declara\u00e7\u00e3o. Nesse caso, o risco parece menor. A principal preocupa\u00e7\u00e3o que tenho com esta categoria de meios de comunica\u00e7\u00e3o gerados por IA \u00e9 o facto de serem utilizados por indiv\u00edduos nefastos que afirmam que s\u00e3o reais.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Recapitula\u00e7\u00e3o e resultados: Melhores detectores de imagens com IA<\/h2>\n\n\n\n<p style=\"font-size:17px\">Muito bem, agora que j\u00e1 apresent\u00e1mos todos os testes, vou recapitular: Gerei 10 imagens de IA<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li style=\"font-size:17px\">Utilizei o ChatGPT, o Nano Banana e o Midjourney para gerar 10 imagens de IA<\/li>\n\n\n\n<li style=\"font-size:17px\">Testei cinco detectores de imagens de IA, passando por eles todas as imagens de IA que gerei.<\/li>\n\n\n\n<li style=\"font-size:17px\">O TruthScan passou em todos os testes e foi o detetor mais preciso. O AI or Not passou em 8 dos 10 testes, mostrando alguma fiabilidade. O Sight Engine falhou em 3 dos 10 testes e demonstrou uma precis\u00e3o geral question\u00e1vel. Was It AI falhou em 4 de 10 testes e teve uma precis\u00e3o fraca em todo o lado. O Winston AI foi o detetor de imagens AI menos preciso, passando apenas em 3 de 10 testes, classificando consistentemente mal as imagens.<\/li>\n<\/ul>\n\n\n\n<!DOCTYPE html>\n<html lang=\"en\">\n<head>\n<meta charset=\"UTF-8\">\n<meta name=\"viewport\" content=\"width=device-width, initial-scale=1.0\">\n<title>Detetor de imagens AI - Resultados exaustivos dos testes<\/title>\n<style>\n @import url('https:\/\/fonts.googleapis.com\/css2?family=DM+Sans:wght@400;500;600;700&family=Outfit:wght@600;700;800&display=swap');\n\n *,*::before,*::after{box-sizing:border-box;margin:0;padding:0}\n\n body {\n background: #f4f3f0;\n min-height: 100vh;\n font-family: 'DM Sans', sans-serif;\n -webkit-font-smoothing: antialiased;\n color: #2d2d3a;\n }\n\n \/* \u2500\u2500 wrapper \u2500\u2500 *\/\n .results-page {\n max-width: 860px;\n margin: 0 auto;\n padding: 3rem 1.25rem 4rem;\n }\n\n \/* \u2500\u2500 hero header \u2500\u2500 *\/\n .hero {\n text-align: center;\n margin-bottom: 2.75rem;\n }\n .hero-badge {\n display: inline-block;\n font-family: 'Outfit', sans-serif;\n font-weight: 700;\n font-size: 0.65rem;\n text-transform: uppercase;\n letter-spacing: 0.12em;\n color: #fff;\n background: #1a1a2e;\n padding: 6px 16px;\n border-radius: 100px;\n margin-bottom: 1.1rem;\n }\n .hero h1 {\n font-family: 'Outfit', sans-serif;\n font-weight: 800;\n font-size: 2.4rem;\n color: #1a1a2e;\n letter-spacing: -0.03em;\n line-height: 1.15;\n margin-bottom: 0.6rem;\n }\n .hero p {\n font-size: 1.02rem;\n color: #6b6b7b;\n max-width: 560px;\n margin: 0 auto;\n line-height: 1.55;\n }\n\n \/* \u2500\u2500 accuracy cards row \u2500\u2500 *\/\n .accuracy-cards {\n display: grid;\n grid-template-columns: repeat(5, 1fr);\n gap: 10px;\n margin-bottom: 2.75rem;\n }\n .acc-card {\n background: #fff;\n border-radius: 14px;\n padding: 1.15rem 0.6rem 1rem;\n text-align: center;\n box-shadow:\n 0 1px 3px rgba(0,0,0,0.05),\n 0 6px 20px rgba(0,0,0,0.07),\n 0 0 0 1px rgba(0,0,0,0.03);\n position: relative;\n overflow: hidden;\n transition: transform 0.2s ease, box-shadow 0.2s ease;\n }\n .acc-card:hover {\n transform: translateY(-3px);\n box-shadow:\n 0 2px 6px rgba(0,0,0,0.06),\n 0 12px 32px rgba(0,0,0,0.1),\n 0 0 0 1px rgba(0,0,0,0.04);\n }\n .acc-card::before {\n content: '';\n position: absolute;\n top: 0; left: 0; right: 0;\n height: 4px;\n }\n .acc-card.rank-1::before { background: #16a34a; }\n .acc-card.rank-2::before { background: #a3e635; }\n .acc-card.rank-3::before { background: #facc15; }\n .acc-card.rank-4::before { background: #fb923c; }\n .acc-card.rank-5::before { background: #ef4444; }\n\n .acc-card .rank-num {\n font-family: 'Outfit', sans-serif;\n font-weight: 800;\n font-size: 0.62rem;\n text-transform: uppercase;\n letter-spacing: 0.08em;\n color: #999;\n margin-bottom: 0.35rem;\n }\n .acc-card .tool-name {\n font-family: 'Outfit', sans-serif;\n font-weight: 700;\n font-size: 0.92rem;\n color: #1a1a2e;\n margin-bottom: 0.65rem;\n line-height: 1.2;\n }\n .acc-ring {\n width: 72px; height: 72px;\n margin: 0 auto 0.55rem;\n position: relative;\n }\n .acc-ring svg { transform: rotate(-90deg); }\n .acc-ring .ring-bg { fill: none; stroke: #eee; stroke-width: 5; }\n .acc-ring .ring-fg { fill: none; stroke-width: 5; stroke-linecap: round; transition: stroke-dashoffset 1s ease; }\n .acc-ring .ring-label {\n position: absolute;\n inset: 0;\n display: flex;\n align-items: center;\n justify-content: center;\n font-family: 'Outfit', sans-serif;\n font-weight: 800;\n font-size: 1.1rem;\n color: #1a1a2e;\n }\n .acc-card .acc-fraction {\n font-size: 0.78rem;\n color: #888;\n font-weight: 500;\n }\n\n \/* \u2500\u2500 section titles \u2500\u2500 *\/\n .section-title {\n font-family: 'Outfit', sans-serif;\n font-weight: 800;\n font-size: 1.45rem;\n color: #1a1a2e;\n letter-spacing: -0.02em;\n margin-bottom: 1rem;\n }\n .section-subtitle {\n font-size: 0.88rem;\n color: #888;\n margin-bottom: 1.25rem;\n line-height: 1.5;\n }\n\n \/* \u2500\u2500 master table \u2500\u2500 *\/\n .master-table-wrap {\n border-radius: 14px;\n overflow: hidden;\n box-shadow:\n 0 1px 3px rgba(0,0,0,0.06),\n 0 8px 24px rgba(0,0,0,0.08),\n 0 0 0 1px rgba(0,0,0,0.04);\n margin-bottom: 2.75rem;\n background: #fff;\n }\n .master-table {\n width: 100%;\n border-collapse: collapse;\n table-layout: fixed;\n }\n .master-table thead th {\n background: #1a1a2e;\n color: #fff;\n font-family: 'Outfit', sans-serif;\n font-weight: 700;\n font-size: 0.62rem;\n text-transform: uppercase;\n letter-spacing: 0.05em;\n padding: 13px 5px;\n text-align: center;\n border: none;\n position: sticky;\n top: 0;\n z-index: 2;\n }\n .master-table thead th:first-child {\n text-align: left;\n padding-left: 14px;\n width: 14%;\n }\n .master-table thead th:nth-child(2) { width: 13%; }\n .master-table thead th:nth-child(3) { width: 9%; }\n\n .master-table tbody tr {\n transition: background 0.18s ease;\n }\n .master-table tbody tr:hover {\n background: rgba(26,26,46,0.025);\n }\n .master-table tbody td {\n padding: 11px 5px;\n text-align: center;\n font-size: 0.82rem;\n font-weight: 500;\n color: #2d2d3a;\n border-bottom: 1px solid #f0f0ec;\n }\n .master-table tbody tr:last-child td { border-bottom: none; }\n .master-table tbody td:first-child {\n text-align: left;\n padding-left: 14px;\n font-weight: 600;\n color: #1a1a2e;\n font-size: 0.78rem;\n line-height: 1.3;\n }\n .test-num {\n display: inline-block;\n font-family: 'Outfit', sans-serif;\n font-weight: 800;\n font-size: 0.6rem;\n background: #1a1a2e;\n color: #fff;\n width: 18px; height: 18px;\n line-height: 18px;\n text-align: center;\n border-radius: 5px;\n margin-right: 5px;\n vertical-align: middle;\n }\n .img-model-sub {\n display: block;\n font-size: 0.68rem;\n color: #999;\n font-weight: 400;\n margin-top: 1px;\n }\n .cat-badge {\n font-family: 'Outfit', sans-serif;\n font-size: 0.6rem;\n font-weight: 700;\n text-transform: uppercase;\n letter-spacing: 0.03em;\n padding: 3px 6px;\n border-radius: 5px;\n display: inline-block;\n white-space: normal;\n line-height: 1.35;\n max-width: 100%;\n }\n .cat-fraud { background: #fef2f2; color: #dc2626; }\n .cat-disinfo { background: #fefce8; color: #a16207; }\n .cat-general { background: #f0fdf4; color: #16a34a; }\n .cat-deepfake { background: #faf5ff; color: #7c3aed; }\n .cat-mixed { background: #fff7ed; color: #c2410c; }\n\n \/* score pills *\/\n .pill {\n font-weight: 700;\n font-family: 'Outfit', sans-serif;\n font-size: 0.7rem;\n border-radius: 6px;\n padding: 4px 7px;\n display: inline-block;\n text-align: center;\n white-space: nowrap;\n min-width: 68px;\n }\n .pill-pass { background: #22c55e; color: #fff; }\n .pill-warn { background: #fb923c; color: #fff; }\n .pill-fail { background: #ef4444; color: #fff; }\n\n \/* \u2500\u2500 legend \u2500\u2500 *\/\n .legend {\n display: flex;\n gap: 1.25rem;\n align-items: center;\n margin-bottom: 1.25rem;\n flex-wrap: wrap;\n }\n .legend-item {\n display: flex;\n align-items: center;\n gap: 6px;\n font-size: 0.78rem;\n font-weight: 500;\n color: #666;\n }\n .legend-dot {\n width: 12px; height: 12px;\n border-radius: 4px;\n }\n .legend-dot.lg-pass { background: #22c55e; }\n .legend-dot.lg-warn { background: #fb923c; }\n .legend-dot.lg-fail { background: #ef4444; }\n\n \/* \u2500\u2500 bottom summary strip \u2500\u2500 *\/\n .summary-strip {\n display: grid;\n grid-template-columns: repeat(3, 1fr);\n gap: 10px;\n margin-bottom: 2.5rem;\n }\n .strip-card {\n background: #1a1a2e;\n border-radius: 12px;\n padding: 1.1rem 1rem;\n color: #fff;\n text-align: center;\n }\n .strip-card .strip-val {\n font-family: 'Outfit', sans-serif;\n font-weight: 800;\n font-size: 1.65rem;\n letter-spacing: -0.02em;\n margin-bottom: 0.15rem;\n }\n .strip-card .strip-label {\n font-size: 0.72rem;\n color: rgba(255,255,255,0.55);\n font-weight: 500;\n text-transform: uppercase;\n letter-spacing: 0.05em;\n }\n .strip-val .hl-green { color: #4ade80; }\n .strip-val .hl-red { color: #f87171; }\n .strip-val .hl-amber { color: #fbbf24; }\n\n \/* \u2500\u2500 footer note \u2500\u2500 *\/\n .foot-note {\n font-size: 0.78rem;\n color: #999;\n line-height: 1.6;\n text-align: center;\n max-width: 640px;\n margin: 0 auto;\n }\n .foot-note strong { color: #666; font-weight: 600; }\n\n \/* \u2500\u2500 responsive \u2500\u2500 *\/\n @media (max-width: 700px) {\n .hero h1 { font-size: 1.6rem; }\n .accuracy-cards { grid-template-columns: repeat(2, 1fr); }\n .accuracy-cards .acc-card:last-child { grid-column: span 2; justify-self: center; max-width: 200px; }\n .summary-strip { grid-template-columns: 1fr; }\n .master-table thead th,\n .master-table tbody td { padding: 9px 3px; font-size: 0.68rem; }\n .pill { font-size: 0.6rem; padding: 3px 4px; min-width: 54px; }\n .test-num { width: 15px; height: 15px; line-height: 15px; font-size: 0.52rem; }\n .section-title { font-size: 1.15rem; }\n }\n\n \/* \u2500\u2500 entrance animations \u2500\u2500 *\/\n @keyframes fadeUp {\n from { opacity: 0; transform: translateY(18px); }\n to { opacity: 1; transform: translateY(0); }\n }\n .anim { animation: fadeUp 0.6s ease both; }\n .anim-d1 { animation-delay: 0.08s; }\n .anim-d2 { animation-delay: 0.16s; }\n .anim-d3 { animation-delay: 0.24s; }\n .anim-d4 { animation-delay: 0.32s; }\n .anim-d5 { animation-delay: 0.4s; }\n<\/style>\n<\/head>\n<body>\n\n<div class=\"results-page\">\n\n <!-- \u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550 HERO \u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550 -->\n <header class=\"hero anim\">\n <span class=\"hero-badge\">50 verifica\u00e7\u00f5es de dete\u00e7\u00e3o - 5 detectores - 10 imagens AI<\/span>\n <h1>Resultados exaustivos dos testes<\/h1>\n <p>O desempenho de cinco detectores de imagens de IA populares em fraudes, desinforma\u00e7\u00e3o, deepfakes e fotografia geral.<\/p>\n <\/header>\n\n <!-- \u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550 ACCURACY RING CARDS \u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550 -->\n <div class=\"accuracy-cards anim anim-d1\">\n\n <!-- 1 \u2014 TruthScan 100% -->\n <div class=\"acc-card rank-1\">\n <div class=\"rank-num\">#1<\/div>\n <div class=\"tool-name\">TruthScan<\/div>\n <div class=\"acc-ring\">\n <svg viewbox=\"0 0 80 80\" width=\"72\" height=\"72\">\n <circle class=\"ring-bg\" cx=\"40\" cy=\"40\" r=\"34\"\/>\n <circle class=\"ring-fg\" cx=\"40\" cy=\"40\" r=\"34\"\n stroke=\"#16a34a\"\n stroke-dasharray=\"213.6\"\n stroke-dashoffset=\"0\"\/>\n <\/svg>\n <span class=\"ring-label\">100%<\/span>\n <\/div>\n <div class=\"acc-fraction\">10 \/ 10 detectados<\/div>\n <\/div>\n\n <!-- 2 \u2014 AI or Not 80% -->\n <div class=\"acc-card rank-2\">\n <div class=\"rank-num\">#2<\/div>\n <div class=\"tool-name\">IA ou n\u00e3o<\/div>\n <div class=\"acc-ring\">\n <svg viewbox=\"0 0 80 80\" width=\"72\" height=\"72\">\n <circle class=\"ring-bg\" cx=\"40\" cy=\"40\" r=\"34\"\/>\n <circle class=\"ring-fg\" cx=\"40\" cy=\"40\" r=\"34\"\n stroke=\"#a3e635\"\n stroke-dasharray=\"213.6\"\n stroke-dashoffset=\"42.72\"\/>\n <\/svg>\n <span class=\"ring-label\">80%<\/span>\n <\/div>\n <div class=\"acc-fraction\">8 \/ 10 detectado<\/div>\n <\/div>\n\n <!-- 3 \u2014 Sight Engine 70% -->\n <div class=\"acc-card rank-3\">\n <div class=\"rank-num\">#3<\/div>\n <div class=\"tool-name\">Motor de vis\u00e3o<\/div>\n <div class=\"acc-ring\">\n <svg viewbox=\"0 0 80 80\" width=\"72\" height=\"72\">\n <circle class=\"ring-bg\" cx=\"40\" cy=\"40\" r=\"34\"\/>\n <circle class=\"ring-fg\" cx=\"40\" cy=\"40\" r=\"34\"\n stroke=\"#facc15\"\n stroke-dasharray=\"213.6\"\n stroke-dashoffset=\"64.08\"\/>\n <\/svg>\n <span class=\"ring-label\">70%<\/span>\n <\/div>\n <div class=\"acc-fraction\">7 \/ 10 detectado<\/div>\n <\/div>\n\n <!-- 4 \u2014 WasItAI 60% -->\n <div class=\"acc-card rank-4\">\n <div class=\"rank-num\">#4<\/div>\n <div class=\"tool-name\">EraIssoAI<\/div>\n <div class=\"acc-ring\">\n <svg viewbox=\"0 0 80 80\" width=\"72\" height=\"72\">\n <circle class=\"ring-bg\" cx=\"40\" cy=\"40\" r=\"34\"\/>\n <circle class=\"ring-fg\" cx=\"40\" cy=\"40\" r=\"34\"\n stroke=\"#fb923c\"\n stroke-dasharray=\"213.6\"\n stroke-dashoffset=\"85.44\"\/>\n <\/svg>\n <span class=\"ring-label\">60%<\/span>\n <\/div>\n <div class=\"acc-fraction\">6 \/ 10 detectado<\/div>\n <\/div>\n\n <!-- 5 \u2014 Winston AI 30% -->\n <div class=\"acc-card rank-5\">\n <div class=\"rank-num\">#5<\/div>\n <div class=\"tool-name\">Winston AI<\/div>\n <div class=\"acc-ring\">\n <svg viewbox=\"0 0 80 80\" width=\"72\" height=\"72\">\n <circle class=\"ring-bg\" cx=\"40\" cy=\"40\" r=\"34\"\/>\n <circle class=\"ring-fg\" cx=\"40\" cy=\"40\" r=\"34\"\n stroke=\"#ef4444\"\n stroke-dasharray=\"213.6\"\n stroke-dashoffset=\"149.52\"\/>\n <\/svg>\n <span class=\"ring-label\">30%<\/span>\n <\/div>\n <div class=\"acc-fraction\">3 \/ 10 detectados<\/div>\n <\/div>\n\n <\/div>\n\n <!-- \u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550 STATS STRIP \u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550 -->\n <div class=\"summary-strip anim anim-d2\">\n <div class=\"strip-card\">\n <div class=\"strip-val\"><span class=\"hl-green\">34<\/span> \/ 50<\/div>\n <div class=\"strip-label\">Total de detec\u00e7\u00f5es corretas<\/div>\n <\/div>\n <div class=\"strip-card\">\n <div class=\"strip-val\"><span class=\"hl-red\">16<\/span> \/ 50<\/div>\n <div class=\"strip-label\">Total de detec\u00e7\u00f5es perdidas<\/div>\n <\/div>\n <div class=\"strip-card\">\n <div class=\"strip-val\"><span class=\"hl-amber\">68%<\/span><\/div>\n <div class=\"strip-label\">Precis\u00e3o m\u00e9dia em todos os<\/div>\n <\/div>\n <\/div>\n\n <!-- \u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550 LEGEND \u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550 -->\n <h3 class=\"section-title anim anim-d3\">Todos os 10 testes - Reparti\u00e7\u00e3o completa<\/h3>\n <p class=\"section-subtitle anim anim-d3\">Cada pontua\u00e7\u00e3o representa a classifica\u00e7\u00e3o de confian\u00e7a na IA do detetor. Limiar para uma aprova\u00e7\u00e3o: \u2265 90% AI.<\/p>\n\n <div class=\"legend anim anim-d3\">\n <div class=\"legend-item\"><span class=\"legend-dot lg-pass\"><\/span> Passar (\u2265 90%)<\/div>\n <div class=\"legend-item\"><span class=\"legend-dot lg-warn\"><\/span> Quase-acidente (70-89%)<\/div>\n <div class=\"legend-item\"><span class=\"legend-dot lg-fail\"><\/span> Falha (< 70%)<\/div>\n <\/div>\n\n <!-- \u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550 MASTER TABLE \u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550 -->\n <div class=\"master-table-wrap anim anim-d4\">\n <table class=\"master-table\">\n <thead>\n <tr>\n <th>Teste<\/th>\n <th>Categoria<\/th>\n <th>Fonte<\/th>\n <th>TruthScan<\/th>\n <th>Motor de vis\u00e3o<\/th>\n <th>IA ou n\u00e3o<\/th>\n <th>Winston AI<\/th>\n <th>EraIssoAI<\/th>\n <\/tr>\n <\/thead>\n <tbody>\n\n <!-- #1 -->\n <tr>\n <td><span class=\"test-num\">1<\/span> Homem na berma<\/td>\n <td><span class=\"cat-badge cat-general\">Geral<\/span><\/td>\n <td><span class=\"img-model-sub\">ChatGPT<\/span><\/td>\n <td><span class=\"pill pill-pass\">99.00%<\/span><\/td>\n <td><span class=\"pill pill-pass\">99.00%<\/span><\/td>\n <td><span class=\"pill pill-warn\">78.00%<\/span><\/td>\n <td><span class=\"pill pill-fail\">0.98%<\/span><\/td>\n <td><span class=\"pill pill-fail\">1.00%<\/span><\/td>\n <\/tr>\n\n <!-- #2 -->\n <tr>\n <td><span class=\"test-num\">2<\/span> O recibo falso<\/td>\n <td><span class=\"cat-badge cat-fraud\">Fraude<\/span><\/td>\n <td><span class=\"img-model-sub\">ChatGPT<\/span><\/td>\n <td><span class=\"pill pill-pass\">99.00%<\/span><\/td>\n <td><span class=\"pill pill-fail\">19.00%<\/span><\/td>\n <td><span class=\"pill pill-pass\">94.48%<\/span><\/td>\n <td><span class=\"pill pill-fail\">0.04%<\/span><\/td>\n <td><span class=\"pill pill-fail\">1.00%<\/span><\/td>\n <\/tr>\n\n <!-- #3 -->\n <tr>\n <td><span class=\"test-num\">3<\/span> Pacote de perigo<\/td>\n <td><span class=\"cat-badge cat-fraud\">Fraude<\/span><\/td>\n <td><span class=\"img-model-sub\">Nano Banana<\/span><\/td>\n <td><span class=\"pill pill-pass\">99.00%<\/span><\/td>\n <td><span class=\"pill pill-pass\">98.00%<\/span><\/td>\n <td><span class=\"pill pill-pass\">99.10%<\/span><\/td>\n <td><span class=\"pill pill-fail\">31.98%<\/span><\/td>\n <td><span class=\"pill pill-pass\">99.00%<\/span><\/td>\n <\/tr>\n\n <!-- #4 -->\n <tr>\n <td><span class=\"test-num\">4<\/span> Farinha de barata<\/td>\n <td><span class=\"cat-badge cat-fraud\">Fraude<\/span><\/td>\n <td><span class=\"img-model-sub\">Nano Banana<\/span><\/td>\n <td><span class=\"pill pill-pass\">97.15%<\/span><\/td>\n <td><span class=\"pill pill-fail\">18.00%<\/span><\/td>\n <td><span class=\"pill pill-warn\">85.87%<\/span><\/td>\n <td><span class=\"pill pill-warn\">72.29%<\/span><\/td>\n <td><span class=\"pill pill-fail\">1.00%<\/span><\/td>\n <\/tr>\n\n <!-- #5 -->\n <tr>\n <td><span class=\"test-num\">5<\/span> Autovandalismo<\/td>\n <td><span class=\"cat-badge cat-mixed\">Fraude \/ Desinforma\u00e7\u00e3o<\/span><\/td>\n <td><span class=\"img-model-sub\">Nano Banana<\/span><\/td>\n <td><span class=\"pill pill-pass\">97.48%<\/span><\/td>\n <td><span class=\"pill pill-fail\">18.00%<\/span><\/td>\n <td><span class=\"pill pill-warn\">89.52%<\/span><\/td>\n <td><span class=\"pill pill-fail\">0.02%<\/span><\/td>\n <td><span class=\"pill pill-pass\">99.00%<\/span><\/td>\n <\/tr>\n\n <!-- #6 -->\n <tr>\n <td><span class=\"test-num\">6<\/span> Soda Musk & Don<\/td>\n <td><span class=\"cat-badge cat-disinfo\">Desinforma\u00e7\u00e3o<\/span><\/td>\n <td><span class=\"img-model-sub\">Nano Banana<\/span><\/td>\n <td><span class=\"pill pill-pass\">99.00%<\/span><\/td>\n <td><span class=\"pill pill-pass\">99.00%<\/span><\/td>\n <td><span class=\"pill pill-pass\">99.00%<\/span><\/td>\n <td><span class=\"pill pill-fail\">0.24%<\/span><\/td>\n <td><span class=\"pill pill-warn\">86.00%<\/span><\/td>\n <\/tr>\n\n <!-- #7 -->\n <tr>\n <td><span class=\"test-num\">7<\/span> Ainda vivo<\/td>\n <td><span class=\"cat-badge cat-deepfake\">Falsifica\u00e7\u00e3o profunda<\/span><\/td>\n <td><span class=\"img-model-sub\">Nano Banana<\/span><\/td>\n <td><span class=\"pill pill-pass\">97.49%<\/span><\/td>\n <td><span class=\"pill pill-pass\">95.00%<\/span><\/td>\n <td><span class=\"pill pill-pass\">99.27%<\/span><\/td>\n <td><span class=\"pill pill-fail\">0.15%<\/span><\/td>\n <td><span class=\"pill pill-fail\">1.00%<\/span><\/td>\n <\/tr>\n\n <!-- #8 -->\n <tr>\n <td><span class=\"test-num\">8<\/span> Abuso de um Ben<\/td>\n <td><span class=\"cat-badge cat-mixed\">Deepfake \/ Desinfo<\/span><\/td>\n <td><span class=\"img-model-sub\">Nano Banana<\/span><\/td>\n <td><span class=\"pill pill-pass\">99.00%<\/span><\/td>\n <td><span class=\"pill pill-pass\">97.00%<\/span><\/td>\n <td><span class=\"pill pill-pass\">98.88%<\/span><\/td>\n <td><span class=\"pill pill-fail\">5.74%<\/span><\/td>\n <td><span class=\"pill pill-pass\">99.00%<\/span><\/td>\n <\/tr>\n\n <!-- #9 -->\n <tr>\n <td><span class=\"test-num\">9<\/span> Clube Void<\/td>\n <td><span class=\"cat-badge cat-general\">Geral<\/span><\/td>\n <td><span class=\"img-model-sub\">Meio da viagem<\/span><\/td>\n <td><span class=\"pill pill-pass\">99.00%<\/span><\/td>\n <td><span class=\"pill pill-pass\">99.00%<\/span><\/td>\n <td><span class=\"pill pill-warn\">89.21%<\/span><\/td>\n <td><span class=\"pill pill-fail\">1.56%<\/span><\/td>\n <td><span class=\"pill pill-warn\">70.00%<\/span><\/td>\n <\/tr>\n\n <!-- #10 -->\n <tr>\n <td><span class=\"test-num\">10<\/span> A caixa de gelo<\/td>\n <td><span class=\"cat-badge cat-disinfo\">Desinforma\u00e7\u00e3o<\/span><\/td>\n <td><span class=\"img-model-sub\">Meio da viagem<\/span><\/td>\n <td><span class=\"pill pill-pass\">99.00%<\/span><\/td>\n <td><span class=\"pill pill-pass\">92.00%<\/span><\/td>\n <td><span class=\"pill pill-pass\">97.94%<\/span><\/td>\n <td><span class=\"pill pill-fail\">41.20%<\/span><\/td>\n <td><span class=\"pill pill-warn\">72.00%<\/span><\/td>\n <\/tr>\n\n <\/tbody>\n <\/table>\n <\/div>\n\n <!-- \u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550 FOOTNOTES \u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550 -->\n <p class=\"foot-note anim anim-d5\">\n <strong>Metodologia:<\/strong> Cada imagem foi submetida uma vez a cada detetor. Uma pontua\u00e7\u00e3o de <strong>\u2265 90% AI<\/strong> \u00e9 contado como uma dete\u00e7\u00e3o correta. As pontua\u00e7\u00f5es entre 70-89% s\u00e3o \"quase-acidentes\". Qualquer coisa abaixo de 70% \u00e9 uma falha. AI or Not marcou uma gera\u00e7\u00e3o em 78%, uma em 85% e uma em 89% - o 89% \u00e9 tratado como uma quase falha com um asterisco na tabela de resumo.\n <\/p>\n\n<\/div>\n\n<\/body>\n<\/html>\n\n\n\n<h2 class=\"wp-block-heading\">Observa\u00e7\u00f5es finais e dados<\/h2>\n\n\n\n<p style=\"font-size:17px\">No total, entre a reda\u00e7\u00e3o deste artigo e os testes rigorosos, passei 25 horas a elaborar este relat\u00f3rio. As ferramentas de dete\u00e7\u00e3o de IA multimodal ainda est\u00e3o em desenvolvimento, mas \u00e9 evidente que algumas s\u00e3o mais precisas do que outras. Depois de testar todas as ferramentas, o TruthScan \u00e9 o mais exato <a href=\"https:\/\/truthscan.com\/ai-image-detector\" target=\"_blank\" rel=\"noopener\">Detetor de imagem AI<\/a>. Os testes falam por si. <\/p>\n\n\n\n<p style=\"font-size:17px\">Se pretender aceder a uma c\u00f3pia CSV dos dados dos testes que efectuei neste artigo, pode encontr\u00e1-la <a href=\"https:\/\/docs.google.com\/spreadsheets\/d\/1iuapCicNcdG6dZCFd-dCmSRgUJ9GVCS2v9kh4PShRew\/edit?usp=sharing\" data-type=\"link\" data-id=\"https:\/\/docs.google.com\/spreadsheets\/d\/1iuapCicNcdG6dZCFd-dCmSRgUJ9GVCS2v9kh4PShRew\/edit?usp=sharing\" target=\"_blank\" rel=\"noopener\">aqui.<\/a> A folha de c\u00e1lculo de dados a que liguei cont\u00e9m todos os avisos originais e resultados de dete\u00e7\u00e3o dos testes deste artigo. <\/p>","protected":false},"excerpt":{"rendered":"<p>Provavelmente j\u00e1 viu mais fotografias geradas por IA online do que imagina. Por vezes, \u00e9 \u00f3bvio que uma [...]<\/p>","protected":false},"author":15,"featured_media":5326,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"rank_math_lock_modified_date":false,"_themeisle_gutenberg_block_has_review":false,"footnotes":""},"categories":[34],"tags":[19,96,97],"class_list":["post-5287","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai-news","tag-ai-detection","tag-ai-images","tag-detecting-deepfakes"],"_links":{"self":[{"href":"https:\/\/research.undetectable.ai\/pt\/wp-json\/wp\/v2\/posts\/5287","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/research.undetectable.ai\/pt\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/research.undetectable.ai\/pt\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/research.undetectable.ai\/pt\/wp-json\/wp\/v2\/users\/15"}],"replies":[{"embeddable":true,"href":"https:\/\/research.undetectable.ai\/pt\/wp-json\/wp\/v2\/comments?post=5287"}],"version-history":[{"count":27,"href":"https:\/\/research.undetectable.ai\/pt\/wp-json\/wp\/v2\/posts\/5287\/revisions"}],"predecessor-version":[{"id":5334,"href":"https:\/\/research.undetectable.ai\/pt\/wp-json\/wp\/v2\/posts\/5287\/revisions\/5334"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/research.undetectable.ai\/pt\/wp-json\/wp\/v2\/media\/5326"}],"wp:attachment":[{"href":"https:\/\/research.undetectable.ai\/pt\/wp-json\/wp\/v2\/media?parent=5287"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/research.undetectable.ai\/pt\/wp-json\/wp\/v2\/categories?post=5287"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/research.undetectable.ai\/pt\/wp-json\/wp\/v2\/tags?post=5287"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}