The recap nobody watched: spotting coverage written by a machine

A recap of an episode is a strange thing to fake, because the source material is public and anyone who watched can check. That has not stopped it. A large and growing share of fan coverage is now assembled from press summaries, cast lists and earlier articles by someone, or something, that never sat through the episode.

The result reads fluently, which is the problem. Fluency used to be evidence of effort, and it no longer is. A guide to how to detect AI written articles sets out the technical markers; what follows is the version a viewer can apply while reading.

The same shift is why review platforms have moved towards verifying that a transaction occurred rather than assessing whether the text looks genuine. WebVouch checks the transaction before the review appears and weighs how steadily reviews arrived, because the star rating and the prose were the two things that stopped being hard to produce.

The tell is specificity, not style

Machine-assembled coverage is competent at summarising what happened and poor at anything that required being there.

It will tell you that a contestant confronted another contestant. It will not tell you that the confrontation happened while one of them was still holding a plate, or that the room went quiet for a beat too long before anyone spoke, or that the edit cut away at the exact moment a reaction would have been most revealing.

Those details are not decoration. They are the evidence that a person watched. A recap with none of them, across two thousand words, is a summary of a summary, and the writer's opinion in the final paragraph is a summary too.

Four markers that hold up

No timestamps or sequence errors. Events described in the order a press release listed them rather than the order they aired. Where a piece gets the running order wrong, nobody checked against the episode.

Cast names without faces. Every participant is named correctly, because the cast list was available, but nobody is described. Real coverage mentions what someone wore, or how they sounded, because those things are unavoidable when watching.

Opinion with no object. Phrases like "fans were divided" and "viewers had strong reactions" that never quote a fan, a viewer or a reaction. The claim is structurally safe and carries no information.

Uniform paragraph length. A person writing about something they watched produces uneven paragraphs, because some moments deserve four sentences and some deserve one. Assembled text tends towards regularity.

None of the four is conclusive alone. A tired writer on a deadline produces something that trips two of them. The combination is what carries the signal, and it is usually visible within the first few hundred words rather than requiring a full read.

Why this matters beyond recaps

The same production method is used far more consequentially elsewhere, and the giveaways are identical.

Product reviews written without the product. Restaurant write-ups by people who never ate there. Service reviews that describe a category rather than an experience. In each case the text performs the shape of first-hand knowledge without containing any, and in each case the missing element is the same: the small, unglamorous detail that only shows up when someone was actually present.

This is why verifying before publication solves a different problem from moderating afterwards. Verification asks whether the transaction happened. Moderation asks whether the text looks acceptable, and text has become very good at looking acceptable.

A useful consequence for readers: a verified review that the business cannot edit or remove is evidence about the transaction even when the writing is unremarkable. The prose was never the load-bearing part.

The reading habit worth building

Ask one question of any piece claiming first-hand knowledge: what is in here that could only be known by being there?

If the answer is nothing, the piece is a rearrangement of publicly available material. That does not make it false. It makes it a summary, and a summary should be read as one rather than as a report.

The same question works on a review. What did this person experience that a description of the product could not have supplied? An answer means somebody bought the thing. No answer means the text is consistent with never having seen it.

The detail does not have to be dramatic. A note about how long delivery actually took, or that the colour was slightly off from the listing photo, or that the support reply arrived on a Sunday. Small, specific, useless to invent because it advances no argument. That is exactly why it is the reliable marker.

Where this is heading

There is no reliable technical detector, and the ones that exist produce false positives on careful human writing and false negatives on careful machine writing. Relying on a tool to sort this out is going to remain unsatisfying.

What does not degrade is the specificity test, because it does not ask how the text was produced. It asks what the text knows. A writer who watched has access to things a generator does not, and the gap shows up within a paragraph or two for anyone reading with the question in mind.

The recaps are not going away, and neither is the production method behind them. Reading them as summaries rather than as reports costs nothing and puts the right weight on them.