Internet Is Fun 
“Slop” is the term that stuck for mass-produced generated content: articles, images, videos and reviews made cheaply, at volume, to fill space and collect attention. It is not the same thing as “made with AI” β plenty of useful work uses these tools. Slop is specifically the stuff made because generating it is nearly free.
The tells are worth learning, because the volume is not going down and search results are not filtering it out reliably.
Text: the structural tells
Stylistic tells age badly β the giveaway phrases of eighteen months ago have been trained out. The structural ones hold up much better:
- It never gets specific. Real expertise produces odd particulars: a version number, an exception, a thing that went wrong. Slop stays at the level of the general because the general is all it has.
- Perfect symmetry. Five sections, each with three bullets, each bullet the same length. Human writing is lumpy β people spend four paragraphs on the interesting bit and one sentence on the dull bit.
- It restates the question. A long preamble about why the topic matters, then a summary of what will be covered, then the content, then a summary of what was covered.
- No position. It lists considerations and declines to conclude. An author who has done the thing has an opinion about it.
- Confident about the checkable, vague about the rest. Facts that are easy to verify are correct; anything requiring judgement dissolves into hedging.
Images: what to look at
Hands and text used to be the standard advice and both have improved a great deal. More durable tells:
- Background incoherence. The subject is fine; the things behind it do not resolve into objects. Railings that do not connect, text on signs that is nearly-letters.
- Texture uniformity. Everything has the same surface quality and the same soft ambient light, regardless of what it is made of.
- Too tidy. Real photographs contain junk β a cable, a bin, an awkward crop. Generated ones are composed with nothing out of place.
- Repeating motifs. The same face, the same window, the same tree at slightly different scales.
Treat all of these as probabilistic. None of them is proof, and confidently accusing a photographer of generating their own work is its own problem.
Where slop concentrates
It clusters where the reward is automatic and the quality bar is low: product roundups, recipe sites, news aggregation, “best X for Y” listicles, and comment sections. Anywhere the traffic converts without anyone reading carefully.
It is rarer in places with a named author who has to face the same audience next week β which is, not coincidentally, the same property that made forums work. See the rise and fall of the internet forum.
Practical defences
Prefer sources with a person attached. A named author with a history is the strongest available signal, and it is cheap to check.
Check whether the site existed two years ago. Archives make this trivial, and a great deal of slop is published on domains with no past. How web archives actually work covers the method.
Read the specific parts first. Skip to the section that would require real knowledge. If it is thin, the rest is decoration.
Use sources that are not ranked by engagement. Curated link collections, RSS from people you chose, and communities with actual moderation. RSS is unglamorous and works.
A caveat worth keeping
The point of this is not to play detective. Plenty of good writing is edited with these tools and plenty of bad writing is entirely human. The useful question is not “was a model involved” but “did anyone who knew anything check this” β and that question has always been the right one.
For the other side of the same subject, see what AI is genuinely bad at and is this AI tool useful or just a demo.
