Generative AI for content: what it does well and badly

A generative model does not know things. It produces the most plausible continuation given what came before, learned from an enormous amount of prior material.

That is not a slight. Plausible continuation is astonishingly useful: it covers drafting, rewriting, summarising, translating, and generating an image matching a description.

But it explains the whole shape of what follows. The model optimises for plausible, not for true, not for original, and not for good. Those often coincide, because true things are usually plausible. When they diverge, the model follows plausible, every time.

So you get output that reads exactly like a confident expert, because it learned the shape of confident expert writing, and that shape is entirely independent of whether the content is correct. Fluency is not evidence of accuracy. They are separate properties that a language model couples only by accident.

There is a pattern to where these tools earn their place: tasks where you can instantly tell if the output is right, and where the work was mechanical rather than creative.

• Transcription and subtitles. You can read them and see if they match. Formerly an hour; now a button.
• Translation and dubbing. Verifiable if you or someone you trust speaks the language.
• Reformatting. Turning a long video into clips, resizing for another aspect ratio, restructuring a script into bullets.
• Variations on a theme. Thirty title options to choose from. You are the filter, so the model's mistakes cost nothing.
• Getting past the blank page. A bad first draft you rewrite beats staring at nothing.
• Editing mechanics. Background removal, object cleanup, colour matching.

The common thread is that verification is cheap and you stay the judge. The model proposes at volume; you dispose. That division of labour plays to the strength, plausible output, and neutralises the weakness, no idea if it is true.

The failures are not random. They follow directly from optimising for plausibility.

Confident fabrication. Ask for a statistic, a study, or a quote and you may get one that does not exist, formatted perfectly, with a plausible source. This is the single most dangerous failure for anyone posting factual content, because the output gives you no signal that it happened.

Regression to the average. Trained on everything, it produces the middle of everything. Ask for a hook and you get the hook that has been written ten thousand times. On a platform where the ranker is actively penalising sameness, generic is not a neutral outcome, it is a losing one.

No taste. It cannot tell which of its thirty titles is good, because good is not a property it models.

Confident about your domain. Its fluency is uniform, but its reliability is not. It sounds identical whether it is on solid ground or improvising, and you only notice the difference in the areas you personally know well.

One rule covers most of the risk, and it follows from the failure modes rather than from caution.

Never publish a factual claim from a model that you have not checked against a source.

SAFE            you can verify instantly, or errors are harmless
  subtitles       you read them
  30 hook ideas   you pick
  translation     a speaker checks
  a rough draft   you rewrite it

DANGEROUS       plausible and wrong looks identical to plausible and right
  statistics      may be invented
  study citations may not exist
  quotes          may be misattributed
  dates, prices   may be confidently wrong
  medical, legal, financial claims

The asymmetry is what makes this workable. In the safe column, a mistake costs you a moment. In the dangerous column, a mistake is published under your name, to an audience, and you had no signal it was coming.

Notice this is not "do not use AI". It is: use it where being wrong is cheap and visible.

Labelling AI content has moved from etiquette to obligation, on two separate tracks.

Platform rules. The major platforms require you to disclose realistic synthetic media, and they apply their own labels when their detection fires. Undisclosed synthetic content that their systems catch can be labelled or demoted regardless of your intent.

Law. The EU AI Act sets transparency obligations for generated content, with a particular focus on material that could mislead: synthetic media depicting real people or events generally has to be disclosed. Obligations phase in over time, and the details differ by use, so treat this lesson as orientation and check the current text for anything commercial.

The practical line is about deception, not tooling. Nobody expects a disclosure because you used AI to remove a background or generate subtitles. The obligation attaches when a reasonable viewer would be misled about whether something is real: a synthetic voice of a real person, a photorealistic event that did not happen, a fabricated endorsement.

Teams that get value from these tools converge on roughly the same division of labour.

Read the pattern: the model does volume and mechanics, you do judgement and facts. Every failure mode of these tools lands in the rows marked you, which is why those rows cannot be delegated.

Invert this, letting the model decide what to say while you tidy the output, and you get exactly the generic, occasionally-false content that the word slop describes.

Here is the part the tooling pitches leave out. If a tool lets you produce content with no effort, it grants everyone else the same power at the same moment. A capability everybody has is not an advantage. It is a new floor.

So the effect of cheap generation is not that AI users beat non-users. It is that the volume of adequate content explodes, and adequate stops being enough to earn attention.

Which raises the value of exactly what cannot be generated: a genuine point of view, real expertise, a specific experience, actual access, a distinctive voice. A model trained on everything published cannot produce the thing you learned last week that nobody has written down.

The uncomfortable synthesis: use these tools for everything that is not the reason someone watches you, and never for the thing that is. The model can make your subtitles. It cannot have your opinion, and an opinion is generally what the audience came for.

One question routes almost every task: can you check the output cheaply? If yes, delegate freely. If no, the model's confident-but-wrong failure mode lands on you.