Ethics and Integrity

Our first flagship product, the AI humanizer, is our most popular one. It works by using an algorithm to rewrite a piece of text the way a human would. When someone humanizes a piece of text, it becomes less likely to be detected as AI-generated.

People use our humanizer because they don't want their content to seem rigid or robotic. Naturally, there are people with both good and bad intentions. But what exactly are the ethical and permissible uses of our humanizer? See the table below for a few examples:

Every case of acceptable use of humanization should fall into "Any time the use of AI writing is not prohibited." Our humanizer should be used to improve content and communication; to build trust, not betray it.

That said, there are use cases for detection bypassing that we consider permissible.

In the first example, there was not a "gross violation of trust." In the second example, not only did a "gross violation of trust" occur, but the individual was also humanizing content "to circumvent any 'no use of AI' policies," — which is wrong.

We actually have users who, for various reasons, must create content that doesn't flag as AI-generated. They are told they can use AI, but that the content must meet human-quality standards.

The thing is, when people believe a piece of content is AI-generated (even if it isn't), they trust it less.

Research from the Nuremberg Institute for Market Decisions (NIM) found that content described as AI-made was perceived more negatively than the identical content presented as human-created. People were less inclined to click on or engage with products featured in AI-generated ads.

For text content, there are words and writing styles that the public now believes indicate something was made by AI. Examples include the em dash "—", and even words like "delve" or "realm."

It goes deeper. Without over-explaining, we've identified hundreds of patterns indicative of AI-writing.

The problem is that humans display these patterns in their writing, too. Many writers have found it easier to "humanize out" these patterns from their prose rather than completely change the way they write. Their main motivation is to maintain trust despite emerging linguistic pattern biases.

As stated on our Terms of Service, we do not allow people to "Engage in any automated use of the system, such as using scripts to send comments or messages, or using any data mining, robots, or similar data gathering and extraction tools."

Spamming is something we have banned over 7,000 users for. While Chatbot tools across various platforms use our humanization API to send messages, they are labeled as chatbots and do not impersonate humans.

Unapproved automations and automated abuse are among the easiest violations for us to protect against and enforce. And in light of agentic AI systems (like clawdbot), we have an absolutely zero tolerance policy for agentic or bot systems using our services to present themselves as real humans. We are watching.

We have never, and will never, condone cheating. Our moderation team investigates .edu email signups.

But we take personal privacy seriously. Further, we will never levy any such accusations if there is a chance we might be wrong. We know that some students will use AI to write their essays. If you're a student reading this, please know that you are only harming yourself by completely outsourcing your thought process to AI.

We sometimes hear the proposition that our software "makes it easier to cheat." The problem, though, is that cheating is a symptom of a larger problem. Our software also makes it easier for ethical users to create authentic communication.

The fact that a bad actor might try to use our software to assist with academic cheating, we feel, doesn't negate the positive uses of our humanizer.

We are always open to dialogue with members of the academic community, and one of our ongoing goals is to minimize harm as much as is feasible.

There are cases where students have gained interest in using our humanizer, not to cheat, but to protect themselves from being falsely accused by inaccurate AI detection systems (false positives). We'll discuss that further in the detector section.

We have guardrails in place to detect automated and bot content, as well as user signups from .edu email addresses. We use internal moderation checkpoints to ensure users follow our policies. The limitations all boil down to privacy.

The only instances in which privacy isn't considered are when public safety is at direct risk of physical harm, or when we are lawfully required by a United States court order to turn over personal data. Our enforcement mainly targets law-breaking or direct physical harm to the public.

We feel that AI disclosure is multifaceted. The main factor is requirement (e.g., internal governance policies/regulations/laws/TOS).

We do not believe that everyone is morally obligated to disclose their use of AI.

We support AI-governance policies and believe every company should have them. Simply put, if you do not have internal governance policies that articulate and regulate how AI is used (e.g., when AI use must be disclosed, acceptable versus unacceptable use cases, etc.), you have a blind spot in your organization.

Perhaps bad data is coming from AI tools, or your employees are exposing confidential information to AI systems that don't protect it. In any case, being able to attribute the positive and negative effects of any AI tool usage is logical. We are in a boom right now. These tools are new. Some people are using them without considering the data they're getting from these tools — or what they're sharing with them.

Our external governance policy (for how our tool can and cannot be used) is defined in our Terms of Service. Internally, we maintain AI-governance policies across all departments.

Our second-most-popular tool is our AI text detector. Like the humanizer, there are right and wrong ways to use it.

First, AI text detection, in its current state, is NOT enough to objectively rule out the use of artificial intelligence (with absolute certainty) in a piece of text. Modern text detectors (including ours) analyze the structure and syntax of written content.

The AI percentage assigned to a piece of text is a probability score, or likelihood, of AI generation. Similar to using stylometric analysis during investigations, this score should be considered a "jumping off" or "starting point" to an investigation. This is why a single AI text detection score alone is NOT enough to rule out cheating or deception.

These examples illustrate:

Hopefully, these examples make one thing clear: when using an AI text detector, it must be a supplementary tool to an investigation. As of now, a single instance of text detection flagging cannot be used as proof beyond a reasonable doubt of wrongdoing.

In academic or professional investigations, text detection should be a secondary or tertiary piece of evidence that supports an argument or conclusion of wrongdoing. A single score by itself is not enough.

In the beginning, we released what we called a "consensus-based" detection model. It was 2023. Detectors were new, and at the time were considered 'black box' tools – there wasn't a unified methodology. We created 8 models that mimicked the major detectors at the time. When users analyzed a piece of text with our detector, it showed them how each model scored it.

The limitation we discovered was that some of these detectors were skewing the overall results. Initially, if 5 of 8 models flagged the text as AI-generated, the overall consensus would conclude that the text likely was.

As we increased our Machine Learning and Research department, we removed inaccurate detector models.

Today, our detector shows only one score; however, it is still based on several detection layers. Another reason we changed this is that people began to think that the algorithm scores we modeled after other detection tools would yield the same results as those tools. The problem was that some of those detection tools would switch between beta models and display different results than our algorithm.

We didn't want people using our detector as a live attribution of other detection tools. Though we cannot publicly name anything, we have indeed identified which detection models are scientific, and the "black box" status of text detection has slowly shifted to "understandable" and "quantifiable" science.

One of the main aspects our detector looks at is linguistic entropy. That is, how unpredictable or variable a block of text is. And, where entropy measures uncertainty, we use perplexity as an evaluation metric.

Because large language models operate with token-level predictability, they often generate outputs with lower perplexity. This can vary, but LLMs generally follow a predictable pattern.

That said, the fact that tools like our humanizer exist (designed to mimic human writing habits) means text detectors can fail.

AI text detectors are a fallible tool. They are useful to have in your arsenal, but it's not an omniscient oracle. We do not condone solely using text detection to accuse someone of wrongdoing.

That all said, using a reliable AI text detector is better than not using one at all. It's not just the tool, though, but also how it's used and understood.