Pangram Labs explained how to spot AI-written texts and why detectors still make mistakes

Pangram Labs attempted to answer a question that becomes increasingly important as generative models grow: can we reliably understand that text was written not by a human, but by AI. The conversation shifts the discussion from the intuitive "it seems a bot wrote this" to a more complex reality: recognition works not like a test with a right answer, but as a probability assessment based on a set of indirect signs.

Why Text Gives Itself Away

Generative models already write more cleanly and evenly than many people. They rarely make noticeable spelling mistakes, usually maintain paragraph logic, and quickly assemble a convincing explanation on almost any topic. Therefore, the main signal has long been not correctness as such.

The problem is different: machine-generated text is increasingly good enough to pass an editor's, teacher's, or ordinary reader's first check and look perfectly normal against the backdrop of average internet content. But such writing often leaves an impression of excessive correctness. Phrases are connected too neatly, paragraph rhythm rarely breaks, and intonation seldom ventures into personal experience, doubt, or unexpected observation.

The reader cannot always formally explain this impression, but notices that the text seems to be assembled from a template without natural roughness. This effect of strange smoothness has become one of the first everyday signs of AI writing today.

In such text, there is often something slightly off.

How Detectors Work

Max Spero, CEO of Pangram Labs, describes detectors not as a magical authorship scanner, but as a pattern analysis system. Instead of one decisive marker, such tools usually gather several statistical and stylistic signals at once: how predictable the text is, how sentence length varies, whether the same logical connectives repeat, whether there are traces of natural editing, and how diverse the vocabulary distribution is. In practice, this is more of a model for assessing similarity than a technical analogue of a polygraph test.

• too even sentence and paragraph lengths
• repeating argumentation patterns
• neat, but uniform phrase structure
• low vocabulary variation despite overall coherence
• absence of minor deviations that often occur in live writing

The key point is that even a strong detector outputs a probability, not a final verdict. It does not answer with complete certainty who exactly wrote the text, but only shows how similar it is to the result of generation. This is especially important now, when authors increasingly use AI as a draft, then rewrite, shorten, and supplement it manually. As a result, the output is a mixed document where machine and human contributions are already difficult to separate by boundary.

Where Errors Will Occur

The conversation separately raises the problem of false positives and false negatives. In the first case, human text is mistakenly recognized as machine-generated, which is quite real for formal writing, student papers, or texts by authors who do not write in their native language and choose the safest constructions. In the second case, AI text, conversely, passes as human—especially if it was edited, personal details added, typical phrases removed, and given a more uneven, conversational rhythm.

From this follows an unpleasant but practical conclusion: detectors are dangerous to use as the sole tool for sanctions in education, hiring, or moderation. The cost of error is too high if a probabilistic label turns into an accusation. At the same time, there is a broader risk for the internet as a whole.

When the cost of text production almost falls to zero, the network quickly fills with a huge volume of acceptable but empty content, and trust increasingly depends not on the text itself, but on the reputation of the platform and the transparency of its origin.

What This Means

Pangram Labs formulates an important, though inconvenient conclusion for the market: recognizing AI text will become a permanent task, but a perfect test will likely not appear. For media, platforms, teachers, and users, this means transitioning from binary thinking to a probabilistic verification model. Simply put, we will have to trust automatic labels less and look more at context, editing, publication history, and quality of sources. This will gradually become the new norm of editorial hygiene.