AI: the Problem with Detection

As social networks and publishing platforms rush to respond to artificial intelligence, many are turning to AI detection tools to decide what content is acceptable. On the surface this sounds reasonable. In practice it creates a serious problem for human writers whose natural style happens to look statistically similar to AI generated text.

There is no hidden watermark in AI text, no secret tag and no embedded marker that survives copying and pasting. What exists instead are statistical models that try to guess whether something was written by a machine based on patterns in the writing. Those guesses are often wrong.

There is no watermark in AI text

A common misconception is that AI generated text carries some kind of invisible signature. In reality, plain text produced by large language models is just text. Once it is copied into another editor, platform or document, it contains no metadata that identifies its origin.

AI detectors do not read hidden tags. They do not scan for embedded codes. They simply analyse how the writing behaves and compare it to patterns they have learned from examples of human and machine generated text.

What statistical clues actually are

When people talk about statistical clues in writing, they are referring to measurable patterns in the text. These patterns are not visible to the eye, but they can be calculated by software. Several key concepts are commonly used.

Perplexity

Perplexity is a measure of how predictable the text is. AI models tend to produce smooth, evenly paced sentences with word choices that follow common patterns. Human writers are more likely to introduce odd phrasing, unusual word combinations and small bumps in the flow. Very low perplexity can cause detectors to label text as AI.

Burstiness

Burstiness describes variation in sentence length and rhythm. Human writing often mixes short, sharp sentences with longer, more complex ones. AI writing frequently settles into a neat, consistent rhythm. Low burstiness is another clue detectors use to flag content.

Token probability patterns

Language models choose each word based on probability. Detectors can examine how likely each word is to follow the previous one. Long chains of highly probable word choices can look machine like, while humans are more prone to unexpected jumps, digressions and idiosyncratic phrasing.

Stylometric fingerprints

Stylometry is the study of writing style. AI tends to avoid typos, maintain a consistent tone, keep structure tidy and minimise abrupt topic shifts. Many human writers, especially in professional or academic contexts, also write this way. When style is very clean and consistent, detectors may treat it as AI like.

When human writers look like AI

Some people naturally write in a clear, organised and low chaos style. Their sentences are similar in length, their tone is calm and their paragraphs are structured. They avoid slang, filler and dramatic swings in emotion. This kind of writing is often valued in professional and scientific settings.

Unfortunately, it is also the kind of writing that statistical detectors associate with AI. A human author who prefers tidy structure and measured language can find their work flagged as machine generated simply because it shares these surface traits with AI output. The more disciplined the writing, the more likely it is to be caught by automated systems.

Why human writers are at risk

Human writers are at risk because AI detection tools do not understand intent, originality or authorship. They only measure patterns. A person who writes clearly and consistently can be treated as suspicious, while messy or chaotic writing is treated as more human. This creates a strange incentive where disciplined writing is punished and untidy writing is rewarded.

Writers in journalism, academia, science, government and community organisations are particularly exposed. Their work is expected to be structured and calm, which means it can be mislabelled by automated systems. False positives can lead to content removal, account penalties or reputational harm, even when the writing is entirely human.

The risk is not theoretical. Studies show that AI detectors frequently misclassify human text, especially when it is professional, technical or carefully edited. This places many legitimate writers in the path of automated moderation systems that are not designed to handle nuance.

What responsible moderation should look like

Responsible moderation should recognise that AI detection is not reliable enough to be used as the sole basis for decisions. Platforms should treat detection scores as indicators, not evidence. Human review should be required before any action is taken against a writer or publisher.

Moderation should consider context, behaviour and demonstrated authorship. A writer with a long history of original work should not be penalised because a detector finds their style too neat. Platforms should provide clear appeal processes and transparent explanations when content is flagged.

Most importantly, moderation should avoid punishing people for writing well. Clear, structured and professional writing is a strength, not a sign of automation. Systems that treat tidy writing as suspicious risk silencing the very voices that contribute most to informed public discussion.

The limits and risks of AI detection

Research into AI detection consistently shows high rates of false positives and false negatives. Clean human writing is often mislabelled as AI, while AI text that has been edited or paraphrased can slip through undetected. Different detectors disagree with each other, and none are accurate enough to serve as definitive proof of authorship.

Despite these limitations, social networks and platforms are increasingly using detection tools to moderate content. In doing so they risk punishing writers whose only offence is that they write clearly. Academic authors, journalists, technical writers and careful bloggers are particularly exposed to this problem.

Fairness, authorship and the future of moderation

The current rush to label and control AI generated content has created a climate of suspicion around writing that looks too neat. Platforms are afraid of spam, misinformation and mass produced articles, but the tools they are using to fight these problems are blunt. They do not understand intent, effort or originality. They only see patterns.

For human writers, this raises important questions about fairness and trust. A person who has spent years developing a calm, structured style can now be told that their work looks artificial. There is no marker in their text, no evidence of machine involvement, only a statistical guess. Treating that guess as fact is not a sound basis for policy.

As moderation practices evolve, there is a strong case for moving away from simple AI detection scores and towards more nuanced assessments that consider context, behaviour and demonstrated authorship. Until then, many human writers will continue to live with the risk that their best work may be mistaken for something it is not.