Is AI writing undetectable? Mostly yes, and that is why it is the wrong question

The short answer is that yes, AI writing is largely undetectable in any way you could rely on, because the tools claiming to detect it are unreliable in both directions. The longer answer is that this makes the question itself a trap, and chasing an answer to it is how people end up spending money on the wrong problem entirely.

What a detector is actually doing

An AI detector does not recognise machine writing. It estimates how predictable the text is.

Language models generate text by repeatedly choosing likely next tokens, which tends to produce prose that sits closer to the statistical middle than human writing does. Detectors measure proxies for that, and turn the result into a percentage. That percentage has the shape of a verdict and the substance of a correlation.

The consequence is two failure modes, both well documented:

False positives on human writing. Prose that is naturally regular scores as machine written. This hits non native English speakers hardest, along with anybody writing in a formal, technical or template driven register, which describes most professional and academic writing. This is measured rather than anecdotal. Liang and colleagues evaluated several widely used GPT detectors against writing from native and non-native English writers and reported that the detectors consistently misclassified the non-native samples as AI generated while identifying the native samples accurately, and concluded by cautioning against using them in evaluative or educational settings at all. Real people have lost grades and contracts over this.

False negatives on machine writing. Light editing collapses detector confidence quickly. Changing sentence lengths, adding a specific example, and cutting the connective phrases models over-use is usually enough. The same paper found that simple prompting strategies bypassed the detectors outright. The signal was never robust to editing.

A tool with meaningful error rates in both directions, producing a number with no confidence interval attached, is not a tool you can make a decision with. That is the honest state of it.

Why “undetectable” is the wrong target

Suppose you succeed completely and nothing can tell. What have you gained?

Nothing, because no search engine or assistant is running a detector on your page and penalising the score. That is not how any of this works. The thing evaluating your content is asking whether it is useful, specific, and worth citing, and those questions are answered by the content itself regardless of what produced it.

Google’s own published position is about quality and intent rather than method. Its guidance on AI generated content says appropriate use of AI is not against its guidelines and that the focus is on quality of content rather than how it was produced, and the spam policies name scaled content abuse as the violation: many pages generated for the primary purpose of manipulating rankings rather than helping users, described there as unoriginal content of little value “no matter how it’s created”.

Read carefully, that is a statement about content that adds nothing, and it applies identically to a content farm staffed by underpaid humans. The line is drawn at purpose and at scale, not at authorship.

So optimising for undetectability is optimising against a judge that is not in the room.

What actually decides whether AI assisted content performs

The failure modes of bad AI content are not exotic and they are not new. They are the failure modes of bad content, arriving faster and in greater volume.

It says what everything else says. A model trained on the web produces the consensus of the web. Ask it for an article on a competitive topic and you get an efficient synthesis of the top ten results. That page has nothing to be cited for, because everything in it is already available in ten places with more authority.

It has no facts of its own. Nothing measured, nothing tested, no number that came from somewhere the model could not have known about. This is the single largest differentiator available, and it is the one thing a model structurally cannot provide.

It is confidently wrong in specific places. Models produce plausible citations, plausible statistics and plausible product details. Publishing those unchecked is how a page acquires a factual error that outlives it. We have watched a model produce a citation to a URL that returned 404, formatted perfectly, while testing prompts for our own library — which is why every link on this page was requested before it was published rather than after.

It is produced at a volume nobody edits. Forty pages nobody read before publishing is the actual risk in AI writing, and it has nothing to do with detection.

The version that works

Use the model where it is genuinely strong: structuring an argument you already have, drafting from notes you supply, rewriting for clarity, generating the first version of something you will heavily edit, and doing the mechanical parts of research.

Then add the parts it cannot:

• Something you measured. Your own data, however small, is the most defensible content asset available. A number nobody else has is worth more than five hundred words nobody else needed. Google’s generative AI guidance makes the same point from the other side, contrasting “commodity content” that restates common knowledge with content carrying a first hand or expert view, and saying that difference will influence visibility more than anything else in the guide. Our two studies, on who blocks the AI crawlers and on what agency sites actually fail, exist for exactly this reason.
• A position. Models hedge by construction. A page willing to say “this is mostly not worth doing, and here is why” is rare enough to be memorable.
• Verification. Every fact, figure and link checked against a source, with the date it was read. This is unglamorous and it is most of the value.
• A reason for the page to exist. If you cannot say what this page offers that the existing results do not, the honest answer is not to publish it.

Do that and the detector question never comes up, because nothing about the result reads as filler. Skip it and no amount of humanising will help, because the problem was never the prose.

What about the humaniser tools

They rewrite text to defeat classifiers, usually by increasing variance in sentence structure and swapping in less predictable word choices. They work reasonably well at that narrow task and they tend to make the writing worse, because the qualities that lower a detector score are not the same as the qualities that make prose good.

More to the point, they treat a symptom of a problem you may not have. If the content is substantive, nobody is checking. If it is not, unpredictable phrasing does not fix it. We went through the category in more depth in how to humanize AI content, including the editing passes that are worth doing for the reader and happen to move a detector score as a side effect.

Our position, and it is the reason we do not sell one: detectors are unreliable in both directions, evasion is not a strategy, and the effort is better spent on the thing that actually gets pages cited, which is having something in them that could not have been generated.

If you want the mechanical half checked rather than argued about, the free audit reports whether a page carries a date, an author and a passage that can be lifted out on its own, which are the three things a machine can actually verify about provenance.

Sources

Read on 15 August 2026.

• Liang, Yuksekgonul, Mao, Wu and Zou, GPT detectors are biased against non-native English writers, arXiv 2304.02819, for the false positive finding and for prompting bypassing the detectors.
• Google, Google Search’s guidance about AI-generated content, for the position that the focus is quality rather than production method.
• Google, spam policies for Google web search, for the definition of scaled content abuse and the phrase “no matter how it’s created”.
• Google, optimizing for generative AI search, for the commodity versus non-commodity content distinction.