{"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\/br\/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":"

Voc\u00ea provavelmente j\u00e1 viu mais fotos online geradas por IA do que imagina. <\/strong><\/p>\n\n\n\n

\u00c0s vezes, \u00e9 \u00f3bvio que uma imagem foi gerada por IA, mas est\u00e1 ficando cada vez mais dif\u00edcil saber, \u00e0 medida que as ferramentas generativas de imagem e v\u00eddeo melhoram. Novas ferramentas, como o Google's Nano Banana Pro<\/a>e atualiza\u00e7\u00f5es para o site da OpenAI Modelo de imagem ChatGPT<\/a> permitem que os usu\u00e1rios gerem rapidamente imagens sint\u00e9ticas que espelham as reais. Pesquisas anteriores descobriram que 85% dos americanos dizem Os deepfakes est\u00e3o corroendo a confian\u00e7a on-line<\/a>.<\/p>\n\n\n\n

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

Executei 50 verifica\u00e7\u00f5es de detec\u00e7\u00e3o em cinco dos mais populares detectores de imagens com IA e documentei os resultados. N\u00e3o apenas apresentarei todos os dados neste artigo e os explicarei, mas tamb\u00e9m colocarei um link para a documenta\u00e7\u00e3o no final.<\/p>\n\n\n\n

Os cinco detectores que usei para esses testes foram: TruthScan<\/a>, IA ou n\u00e3o<\/a>, Motor de vis\u00e3o<\/a>, EraAI<\/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. Usei v\u00e1rios estilos e t\u00e9cnicas de solicita\u00e7\u00e3o (sobre os quais entrarei em detalhes mais tarde, quando analisar cada imagem analisada).<\/p>\n\n\n\n

Todos os cinco detectores foram encarregados de detectar imagens criadas por IA em categorias como fraude, desinforma\u00e7\u00e3o, fotografia geral e deepfakes. Como era de se esperar, nem todos os detectores tiveram um bom desempenho. O TruthScan foi o \u00fanico detector a classificar consistentemente todo o conte\u00fado que enviei com 97% ou mais.<\/strong><\/p>\n\n\n\n\n\n\n\n\nResultados do teste do AI Image Detector<\/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 AI Image Detector<\/h2>\n\n <table class=\"ai-detector-table\">\n <thead>\n <tr>\n <th>Ferramenta de detec\u00e7\u00e3o<\/th>\n <th>Total de testes<\/th>\n <th>Detectado corretamente<\/th>\n <th>Errou (Falha)<\/th>\n <th>Precis\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>EraAI<\/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%. Como o teste exige pelo menos 90%, contamos as duas primeiras como reprovadas, mas tratamos a 89% como quase reprovada, 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 fato, embora o AI or Not tenha classificado v\u00e1rios itens gerados por IA abaixo da certeza 90%, ele foi o segundo detector mais preciso e consistente durante 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 restante dos detectores de imagem de IA que testei teve um desempenho muito pior. Por exemplo, o Sight Engine classificou erroneamente 3 imagens de IA relacionadas a fraudes como aut\u00eanticas.<\/p>\n\n\n\n<p style=\"font-size:17px\">Agora, mostrarei cada imagem (das 10 que gerei), explicarei como a criei e mostrarei 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>EraAI<\/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 muito conte\u00fado gerado por IA diariamente), a imagem parece bastante convincente \u00e0 primeira vista. Se voc\u00ea clicasse no perfil de m\u00eddia social de algu\u00e9m, percorresse o feed e visse essa imagem, ela se destacaria imediatamente como uma falsifica\u00e7\u00e3o gerada por IA? Para ser justo, tentei ser criativo com o prompt. Aqui est\u00e1 o que usei para gerar essa 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>Gere uma foto no estilo de c\u00e2mera instant\u00e2nea de um homem em p\u00e9 no topo de um telhado olhando para a rua, \u00e0 noite, ele est\u00e1 iluminado pelo flash da c\u00e2mera, h\u00e1 tra\u00e7os de luz dos carros abaixo, uma mulher est\u00e1 atr\u00e1s dele com a m\u00e3o nos l\u00e1bios, sorrindo, a est\u00e9tica da foto \u00e9 como uma foto tirada em uma c\u00e2mera instant\u00e2nea em 2009 por amigos de faculdade brincando.<\/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 criatividade ou da quantidade de detalhes que inclu\u00ed em meu prompt, o TruthScan e o Sight Engine ainda classificaram corretamente a sa\u00edda como gerada por IA. O detector AI or Not chegou perto, mas ainda assim errou 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 pelo 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 mostrando 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>EraAI<\/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\">Eu queria gerar um recibo falso com o ChatGPT para ver o qu\u00e3o realista ele seria. Voc\u00ea pode notar que a imagem n\u00e3o tem um endere\u00e7o leg\u00edtimo, mas a textura ainda parece um pouco realista. Eu diria que essa imagem \u00e9 menos convincente do que a primeira que eu fiz o ChatGPT gerar, mas a maioria dos detectores ainda foi prejudicada ao analis\u00e1-la.<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li style=\"font-size:17px\">De todas as verifica\u00e7\u00f5es que fiz nessa imagem, <strong>TruthScan<\/strong> foi o mais preciso. <\/li>\n\n\n\n<li style=\"font-size:17px\"><strong>Motor de vis\u00e3o<\/strong> realmente falhou aqui, mostrando um julgamento ruim em uma imagem de IA relacionada a documentos.<\/li>\n\n\n\n<li style=\"font-size:17px\"><strong>IA ou n\u00e3o<\/strong> teve um desempenho visivelmente melhor nessa 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\">Como essa imagem est\u00e1 associada \u00e0 categoria de fraude, \u00e9 preocupante que a maioria dos detectores a tenha classificado incorretamente. <\/p>\n\n\n\n<h3 class=\"wp-block-heading\">O prompt que usei 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>Gere uma imagem de um recibo da Best Buy, que diz que o valor total gasto \u00e9 de um bilh\u00e3o de d\u00f3lares e que tem uma mancha de caf\u00e9 nele<\/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, escrevi um prompt r\u00e1pido e simples. Ainda assim, a qualidade da sa\u00edda do ChatGPT \u00e9 visualmente atraente a olho nu. Ao que parece, n\u00e3o \u00e9 t\u00e3o atraente para dois dos cinco detectores que testei.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">#3: \"Pacote de risco\" (gerado com Nano Banana)<\/h2>\n\n\n\n<p><strong><em>Em vez de apenas fazer com que o Nano Banana gerasse uma imagem original, dei a ele uma para alterar. <\/em><\/strong><\/p>\n\n\n\n<p style=\"font-size:17px\">A ideia era demonstrar como algu\u00e9m poderia facilmente usar a IA para criar evid\u00eancias falsas e alegar que recebeu um pacote danificado\/perigoso. Primeiro, tirei uma foto com um iPhone 15 de um pacote vazio da Amazon que encontrei.<\/p>\n\n\n\n<p><strong>Aqui est\u00e1 a foto 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 foto de um pacote 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 foto que tirei no Nano Banana e dei a ele o seguinte aviso:<\/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 voc\u00ea edite esta foto, n\u00e3o altere nada, exceto o local onde est\u00e1 a etiqueta, acrescente danos a ela e adicione manchas pretas de lama na caixa que pare\u00e7am realmente manchas de graxa<\/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 foto mostrando um pacote\" 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>EraAI<\/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 semelhante a \u00f3leo que o Nano Banana gerou na sa\u00edda tinha um brilho sujo convincente, completo com um efeito de encharcamento - nada cartunesco. Felizmente, a maioria dos detectores pelos quais passei classificou corretamente a imagem como falsa. A \u00fanica exce\u00e7\u00e3o foi o Winston, que novamente teve um desempenho ruim, classificando erroneamente a imagem como aut\u00eantica. AI or Not, TruthScan e WasItAI foram os detectores mais precisos na identifica\u00e7\u00e3o dessa imagem.<\/p>\n\n\n\n<p style=\"font-size:17px\">\u00c9 preocupante como parece simples carregar uma imagem em um chatbot e fazer com que ela seja completamente alterada. Posso imaginar fraudadores e golpistas usando ferramentas de gera\u00e7\u00e3o de imagens para tentar fazer pedidos de devolu\u00e7\u00e3o fraudulentos ou participar de fraudes no mercado.<\/p>\n\n\n\n<p style=\"font-size:17px\">Se as plataformas de com\u00e9rcio eletr\u00f4nico exigirem apenas uma imagem como prova para iniciar uma reclama\u00e7\u00e3o de pacote danificado, ent\u00e3o, sem nenhuma maneira confi\u00e1vel de detectar adultera\u00e7\u00e3o de IA, elas estar\u00e3o expostas a um grande vetor de ataque. Por exemplo, o site 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 itens<\/a> afirma literalmente<strong>, “<em>Materiais perigosos, incluindo l\u00edquidos ou gases inflam\u00e1veis, n\u00e3o s\u00e3o retorn\u00e1veis<\/em>.”<\/strong><\/p>\n\n\n\n<p style=\"font-size:17px\">Portanto, se voc\u00ea realmente recebesse um pacote coberto de \u00f3leo ou lama preta gordurosa, n\u00e3o precisaria devolv\u00ea-lo, mas poderia obter um reembolso. Se, por algum motivo, isso realmente acontecesse com voc\u00ea, a Amazon provavelmente pediria uma prova fotogr\u00e1fica. Est\u00e1 vendo o problema que as imagens geradas por IA podem representar quando s\u00e3o indistingu\u00edveis das imagens reais?<\/p>\n\n\n\n<h2 class=\"wp-block-heading has-text-align-left\">#4. \"Cockroach Meal\" (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 voc\u00ea perder o apetite. Acontece que h\u00e1 casos documentados de pessoas usando 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 aplicativos de entrega de comida. Para esse quarto teste, segui as mesmas etapas do teste tr\u00eas. Dessa vez, tirei uma foto 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 foto de uma caixa de comida para viagem 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 foto real que tirei no Nano Banana e dei a ele 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>\"Edite esta imagem. N\u00e3o mude nada, exceto adicionar macarr\u00e3o onde est\u00e1 o papel-alum\u00ednio e adicionar um monte de baratas beb\u00eas min\u00fasculas na comida. Fa\u00e7a com que tudo pare\u00e7a realista e n\u00e3o um desenho animado.\" <\/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 do 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 foto alterada por IA da caixa de comida para viagem com macarr\u00e3o e insetos 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>EraAI<\/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 a m\u00eddia gerada pela IA. O Sight Engine e o WasitAI falharam completamente. O AI or Not foi o segundo detector mais preciso (com 85% de certeza de IA) ao escanear essa 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\">Nessa imagem, o macarr\u00e3o e as baratas foram todos gerados inteiramente por IA; no entanto, dois dos detectores n\u00e3o pareciam pensar assim, e apenas um (TruthScan) estava acima de 90% certo de que houve adultera\u00e7\u00e3o por IA.<\/p>\n\n\n\n<p style=\"font-size:17px\">Um ponto de preocupa\u00e7\u00e3o aqui foi a simplicidade com que foi poss\u00edvel criar essa imagem falsa. Desde o upload da foto que tirei, at\u00e9 o acionamento do Nano Banana e o recebimento do resultado, no total, foram necess\u00e1rios apenas 2 minutos. Um pesadelo para aplicativos de entrega de alimentos; um sonho para mentirosos famintos. 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 a seguir foram planejadas para demonstrar como as imagens de IA podem ser usadas para desinforma\u00e7\u00e3o. Elas cont\u00eam imagens pol\u00eamicas e assuntos delicados<\/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 on-line. Bastante mundano. Mas e se algu\u00e9m quisesse usar essa imagem para espalhar desinforma\u00e7\u00e3o ou registrar uma falsa solicita\u00e7\u00e3o de seguro de 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 esse quinto teste, peguei uma imagem real de um carro e dei ao Nano Banana a seguinte solicita\u00e7\u00e3o:<\/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 mude nada no carro, exceto que eu quero que voc\u00ea adicione tinta spray que diga \"Palestina Livre\" e que o para-brisa esteja rachado. Os far\u00f3is est\u00e3o amassados, o espelho est\u00e1 quebrado e pendurado e as duas janelas na lateral 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 foto 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 foto alterada por IA do carro azul, mostrando um para-brisa rachado e danos causados por tinta spray na 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>EraAI<\/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 detector para essa sa\u00edda foram mistos. Para esse teste, o WasItAI (surpreendentemente) teve a maior pontua\u00e7\u00e3o de detec\u00e7\u00e3o de IA (99%). O TruthScan teve a segunda maior pontua\u00e7\u00e3o (97%). O Sight Engine falhou significativamente na detec\u00e7\u00e3o de qualquer elemento sint\u00e9tico na imagem, e o Winston teve o pior desempenho entre todos os detectores (atribuindo uma chance de envolvimento inferior a 1%). O AI or Not se saiu bem, mas a precis\u00e3o ainda estava abaixo de 90%.<br><br>Vamos falar sobre a imagem. Os dois aspectos mais preocupantes do conte\u00fado alterado ou criado por IA como esse \u00e9 que ambos envolvem engano. Primeiro, o uso de IA para alterar imagens mundanas para criar conte\u00fado politicamente carregado \u00e9 r\u00e1pido, f\u00e1cil e parece real. Independentemente de os criadores desse tipo de conte\u00fado estarem apenas buscando engajamento ou serem provocadores pol\u00edticos, o problema \u00e9 que se trata de informa\u00e7\u00f5es falsas sendo apresentadas como aut\u00eanticas.<\/p>\n\n\n\n<p style=\"font-size:17px\">A segunda maneira pela qual o uso de ferramentas de gera\u00e7\u00e3o de imagens como essa pode ser usado indevidamente \u00e9 para fraude. Se um golpista quiser registrar um sinistro falso de seguro de autom\u00f3vel, ele ter\u00e1 que produzir provas falsas. Em vez de passar horas no Photoshop editando imagens e adulterando provas, eles podem usar 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, observe como um golpista pode usar um processo semelhante para registrar um pedido de indeniza\u00e7\u00e3o falso, de acordo com o <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>Janelas, far\u00f3is ou lanternas traseiras quebrados<\/strong><\/li>\n\n\n\n<li><strong>Danos causados por tinta 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\">Amassados ou arranh\u00f5es<\/a> de algu\u00e9m que tenha chaveado seu carro<\/strong><\/li>\n\n\n\n<li><strong>Colocar a\u00e7\u00facar ou outras subst\u00e2ncias no tanque de gasolina<\/strong><\/li>\n<\/ul>\n\n\n\n<p>\u00c9 claro que apresentar provas falsas n\u00e3o \u00e9 apenas enganoso, \u00e9 tamb\u00e9m totalmente ilegal. Em um mundo ideal, ningu\u00e9m infringiria a lei; na realidade, o crime acontece todos os dias. E sem uma detec\u00e7\u00e3o confi\u00e1vel, as seguradoras de autom\u00f3veis correm o risco de serem enganadas em milh\u00f5es \u00e0 medida que aumentam os pedidos de indeniza\u00e7\u00e3o de seguro fraudulentos habilitados por IA.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">#6. \"Soda Musk and 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>EraAI<\/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, usei o Nano Banana para esse teste, e o objetivo era demonstrar imagens falsas e politicamente carregadas com um tom mais bobo. O TruthScan, o Sight Engine e o AIorNot sinalizaram a imagem gerada com uma classifica\u00e7\u00e3o de detec\u00e7\u00e3o de IA 99%. O WasitAI n\u00e3o estava t\u00e3o confiante, detectando a imagem gerada por IA com apenas 86% de certeza. E Winston falhou completamente, atribuindo uma pontua\u00e7\u00e3o de IA de 0,2%. Para gerar a imagem, dei ao Nano Banana este prompt:<\/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>Gere uma foto de Elon Musk derramando um copo grande de refrigerante como um Big Gulp em sua camisa e surtando enquanto Donald Trump est\u00e1 sentado ao lado dele rindo. Eles est\u00e3o em um avi\u00e3o a jato. Essa \u00e9 a cena. E a foto deve parecer realista, como se tivesse sido tirada com uma c\u00e2mera instant\u00e2nea e incluir artefatos como o flash da c\u00e2mera refletindo na janela do avi\u00e3o e fora dele. \u00c9 noite. Ela 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 desses programas de software n\u00e3o consegue detectar a IA, eles o rotulam como aut\u00eantico. Por mais bobo que seja esse 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 segurando uma placa que diz "help" (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>EraAI<\/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\">Voc\u00ea acha essa imagem perturbadora? Para gerar esse deepfake, usei o Nano Banana e dei a ele uma foto minha. Em seguida, dei a ele um longo prompt com o objetivo de me colocar em uma situa\u00e7\u00e3o que sugerisse que eu havia sido sequestrado. Talvez seja porque essa foto mostra <em>eu<\/em> em uma posi\u00e7\u00e3o comprometedora. E se algu\u00e9m fizesse algo assim, enviasse para minha m\u00e3e e amea\u00e7asse me machucar, a menos que ela lhe enviasse dinheiro? Coisa assustadora. <\/p>\n\n\n\n<p class=\"has-small-font-size\"><em>(O prompt completo que usei para isso pode ser encontrado 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 uma apar\u00eancia muito realista, <strong>tr\u00eas em cada cinco detectores o identificaram como gerado por IA.<\/strong> <\/p>\n\n\n\n<p style=\"font-size:17px\">O AI or Not obteve uma forte pontua\u00e7\u00e3o de AI de 99%, o TruthScan com 97% e o 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 esses testes n\u00e3o foram feitos em minha conta pessoal do Gemini. A Nano Banana nunca me pediu para provar que eu era a pessoa na imagem. Qualquer pessoa poderia ter baixado uma foto minha da Internet, coloc\u00e1-la no Nano Banana e fazer com que ele criasse esse 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 tempo gerenciando v\u00e1rios departamentos nas empresas em que trabalha. A imagem acima \u00e9 o (verdadeiro) headshot corporativo elegante de Ben. Se voc\u00ea n\u00e3o percebeu, Ben \u00e9 um cara profissional. Quaisquer que sejam suas opini\u00f5es pol\u00edticas, ele as guarda para si mesmo, e suas conversas de trabalho s\u00e3o apenas sobre trabalho. Mas, a qualquer momento, a imagem de Ben pode ser usada para divulgar propaganda ou mensagens pol\u00edticas sem o seu consentimento.<\/p>\n\n\n\n<p style=\"font-size:17px\">Recentemente, houve muitas reportagens sobre como v\u00eddeos gerados por IA envolvendo agentes do ICE e protestos t\u00eam se espalhado on-line. Com base em todos os testes realizados at\u00e9 agora, conclu\u00ed que criar propaganda pol\u00edtica falsa com IA era realmente f\u00e1cil. Decidi testar novamente, desta vez usando Ben. <\/p>\n\n\n\n<p style=\"font-size:17px\">Com a permiss\u00e3o de Ben, usei sua foto 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 foto alterada por IA de um homem chamado Ben, segurando uma placa que diz Abolish ICE (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>EraAI<\/th>\n <\/tr>\n <\/thead>\n <tbody>\n <tr>\n <td>Nano Banana<\/td>\n <td style=\"font-size:0.78rem;\">Deepfake \/ Disinfo<\/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\">Levou trinta segundos. Tudo o que precisei fazer foi arrastar e soltar a foto de seu rosto no Gemini e pedir que ele \"levantasse uma placa dizendo abolir o ICE\". O resultado do Nano Banana n\u00e3o \u00e9 t\u00e3o fotorrealista quanto os outros, mas ainda parece aut\u00eantico o suficiente 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 a imagem com precis\u00e3o:\u00a0<strong>TruthScan<\/strong> (9<\/span>9% AI), Was It AI (99% AI), Sight Engine (97% AI) e AI or Not (97% AI). O 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, usando um vestido branco de mangas curtas, est\u00e1 em uma sala pouco iluminada e cheia de gente. A foto parece ter sido tirada com um flash, criando uma atmosfera nebulosa e azulada com part\u00edculas fracas vis\u00edveis no ar. Outras pessoas com roupas casuais est\u00e3o ao fundo, sob um teto danificado ou de apar\u00eancia 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>EraAI<\/th>\n <\/tr>\n <\/thead>\n <tbody>\n <tr>\n <td>Meio da jornada <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\">Gerei a imagem acima usando o Midjourney. Quanto mais eu olhava para ela, mais inquietante ela parecia. A est\u00e9tica \u00e9 sombria e misteriosa. Uma presen\u00e7a sombria parece estar surgindo por tr\u00e1s do objeto. Para criar essa imagem, usei o recurso de refer\u00eancia est\u00e9tica do Midjourney e simplesmente solicitei que fosse uma \"garota em p\u00e9 em uma boate sorrindo\". Mostrei a imagem a algumas pessoas, e todas acharam que era real. <br><br>A maioria dos detectores sinalizou a imagem como gerada por IA. O TruthScan e o Sight Engine a sinalizaram com uma pontua\u00e7\u00e3o de IA de 99%, e o AI or Not indicou 89%. <\/p>\n\n\n\n<p style=\"font-size:17px\">O WasItAI n\u00e3o tinha tanta certeza quanto os detectores, classificando a imagem com apenas uma pontua\u00e7\u00e3o de IA 70%. <\/p>\n\n\n\n<p style=\"font-size:17px\">O detector da Winston AI falhou completamente, dando \u00e0 imagem uma pontua\u00e7\u00e3o de 1,56% da AI. <\/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 foto borrada mostrando um homem com uma camisa manchada de vermelho, parado na 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>EraAI<\/th>\n <\/tr>\n <\/thead>\n <tbody>\n <tr>\n <td>Meio da jornada <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 essa imagem usando o mesmo m\u00e9todo de meio de viagem do teste anterior. Desta vez, o objetivo era gerar algo um pouco mais escuro. O TruthScan foi o detector 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 menor 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, nesse caso, a imagem que gerei poderia ser usada ou enquadrada de duas maneiras. Por um lado, algu\u00e9m poderia usar esse tipo de m\u00eddia (ou qualquer outro tipo de m\u00eddia com tema pol\u00edtico\/conflito 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 essa categoria de m\u00eddia gerada por IA \u00e9 que ela seja usada por indiv\u00edduos nefastos que afirmam que ela \u00e9 real.<\/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\">Bem, agora que exibimos 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\">Usei 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 colocando todas as imagens de IA que gerei neles.<\/li>\n\n\n\n<li style=\"font-size:17px\">O TruthScan passou em todos os testes e foi o detector mais preciso. O AI or Not passou em 8 dos 10 testes, demonstrando alguma confiabilidade. O Sight Engine falhou em 3 dos 10 testes e demonstrou uma precis\u00e3o question\u00e1vel em todos os aspectos. O Was It AI falhou em 4 dos 10 testes e teve precis\u00e3o ruim em todos os aspectos. O Winston AI foi o detector de imagens com IA menos preciso, passando em apenas 3 de 10 testes e classificando consistentemente as imagens de forma err\u00f4nea.<\/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>AI Image Detector - Resultados abrangentes 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 detec\u00e7\u00e3o - 5 detectores - 10 imagens de IA<\/span>\n <h1>Resultados abrangentes dos testes<\/h1>\n <p>O desempenho de cinco detectores de imagens de IA populares em fraudes, desinforma\u00e7\u00e3o, deepfakes e fotografias em 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 detectado<\/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\">EraAI<\/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 detectado<\/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 - Detalhamento completo<\/h3>\n <p class=\"section-subtitle anim anim-d3\">Cada pontua\u00e7\u00e3o representa a classifica\u00e7\u00e3o de confian\u00e7a na IA do detector. Limite para 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> Passe (\u2265 90%)<\/div>\n <div class=\"legend-item\"><span class=\"legend-dot lg-warn\"><\/span> Quase perda (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>EraAI<\/th>\n <\/tr>\n <\/thead>\n <tbody>\n\n <!-- #1 -->\n <tr>\n <td><span class=\"test-num\">1<\/span> Homem em uma borda<\/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 risco<\/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> Refei\u00e7\u00e3o 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 beb\u00ea<\/td>\n <td><span class=\"cat-badge cat-mixed\">Deepfake \/ Disinfo<\/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 jornada<\/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 jornada<\/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 enviada uma vez para cada detector. Uma pontua\u00e7\u00e3o de <strong>\u2265 90% AI<\/strong> \u00e9 contado como uma detec\u00e7\u00e3o correta. Pontua\u00e7\u00f5es entre 70-89% s\u00e3o \"quase-acidentes\". Qualquer pontua\u00e7\u00e3o abaixo de 70% \u00e9 uma falha. O AI or Not rotulou uma gera\u00e7\u00e3o em 78%, uma em 85% e uma em 89% - o 89% \u00e9 tratado como um quase-acidente 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 elaborando este relat\u00f3rio. As ferramentas de detec\u00e7\u00e3o de IA multimodal ainda est\u00e3o em desenvolvimento, mas est\u00e1 claro que algumas s\u00e3o mais precisas do que outras. Depois de testar todas as ferramentas, o TruthScan tem a mais precisa <a href=\"https:\/\/truthscan.com\/ai-image-detector\" target=\"_blank\" rel=\"noopener\">Detector de imagens AI<\/a>. Os testes falam por si. <\/p>\n\n\n\n<p style=\"font-size:17px\">Se quiser acessar uma c\u00f3pia CSV dos dados dos testes que fiz neste artigo, voc\u00ea 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 planilha de dados que vinculei cont\u00e9m todos os prompts originais e os resultados de detec\u00e7\u00e3o dos testes deste artigo. <\/p>","protected":false},"excerpt":{"rendered":"<p>Voc\u00ea provavelmente j\u00e1 viu mais fotos online geradas por IA do que imagina. \u00c0s vezes, \u00e9 \u00f3bvio que um [...]<\/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\/br\/wp-json\/wp\/v2\/posts\/5287","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/research.undetectable.ai\/br\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/research.undetectable.ai\/br\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/research.undetectable.ai\/br\/wp-json\/wp\/v2\/users\/15"}],"replies":[{"embeddable":true,"href":"https:\/\/research.undetectable.ai\/br\/wp-json\/wp\/v2\/comments?post=5287"}],"version-history":[{"count":27,"href":"https:\/\/research.undetectable.ai\/br\/wp-json\/wp\/v2\/posts\/5287\/revisions"}],"predecessor-version":[{"id":5334,"href":"https:\/\/research.undetectable.ai\/br\/wp-json\/wp\/v2\/posts\/5287\/revisions\/5334"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/research.undetectable.ai\/br\/wp-json\/wp\/v2\/media\/5326"}],"wp:attachment":[{"href":"https:\/\/research.undetectable.ai\/br\/wp-json\/wp\/v2\/media?parent=5287"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/research.undetectable.ai\/br\/wp-json\/wp\/v2\/categories?post=5287"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/research.undetectable.ai\/br\/wp-json\/wp\/v2\/tags?post=5287"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}