For a while, “slop” was a word for dismissing anything made with AI. If you were a believer, you learned to tune out the word and whoever said it.

That no longer works. The believers are drowning in it too.

Everyone now sits on both sides of the problem. An associate drafts a diligence summary in minutes, and a partner spends the evening working out which parts are true. The hour one person saves lands on someone else’s desk.

Sometimes nobody reads it at all. A document arrives with more detail than anyone can absorb, so the recipient has AI analyze it and draft the reply. The reply comes back longer. So does the next one. Each round adds detail and removes understanding, and the work slows down while looking busier than ever. The AI is talking to the AI. The humans are just hitting send. We run into it in our own work, and of all the problems firms bring to us, it is the one we hear about most.

Developers are already documenting the rise of slop. In a 2026 study of 1,154 Reddit and Hacker News posts, researchers found a tragedy of the commons: individual speed paid for by reviewers and maintainers. By the end of 2025, Merriam-Webster had made “slop” its word of the year.

Bad work is not new. Bad work arriving faster than anyone can check it is.

Where it comes from

Effort used to be the filter. When building something took weeks, the weeks forced a question first: is this worth making? The tools removed the effort, and the question went with it. Knowing what is worth making is still the hard part. Skilled people, not just novices, now build things nobody decided should exist, because nothing made them decide.

The tools rarely tell you they have failed. Ask for something that cannot be done and you will usually get something anyway. Ask for something poorly specified and you will get a confident answer to a question you did not mean to ask. The failure does not announce itself. It arrives looking finished, and nobody has a reason to doubt it. That is a much harder problem than an error message.

Generation outpaces oversight. The producing side scaled. The checking side did not. In the study, that gap between generating and reviewing was a dominant theme. Every unreviewed thing that ships is a bet that it was fine, and organizations place that bet many times a day without noticing they are betting.

Then the damage spreads. The study found that slop erodes trust. Once reviewers expect slop, everything looks suspect, and good work gets skimmed, discounted, or sent back alongside the bad. That is the real tragedy of the commons. The careful believer pays for the careless one.

What actually helps

Someone skilled has to decide what gets built. Not review it after the fact. Decide it in advance. That takes a stated vision of what the system is for and what belongs in it, held by people with the judgment and authority to say no. Better tools do not supply that. Everyone is buying cameras. Nobody is hiring a director. Without direction, output volume becomes the measure of progress, because it is the only thing left that is easy to measure.

Evaluation standards have to be strict enough to reject things. A standard that has never stopped anything is not a standard. It is a formality. The test is not whether the output looks good. It is whether it would damage the system, mislead someone, or cause a problem further down the line. We hold our own work to this, and it is uncomfortable by design, because the alternative is finding out later in front of a client.

Oversight has to be aimed, not merely available. This is the easiest one to get wrong, because failure here looks like success. Audit trails get built. Approval steps get added. Then volume rises and the trail becomes something nobody opens. Traceability that exists but goes unexercised provides the feeling of control without the control.

More diligence will not close the gap, because review capacity did not scale. Review has to go where being wrong is expensive, and everything else ships knowing nobody checked it. The bet from earlier still gets placed. It just gets placed on purpose. When we built our own platform, we made every output traceable to its source. That matters because the trail gets followed where it counts.

Not an argument for slowing down

None of this is a case for using less AI. Volume is not a problem on its own. A firm that responds to slop by generating less in total will be slower and no better.

The problem is volume without judgment attached. Judgment is now the scarce input, and that changes the job of anyone running a team. The job is no longer getting more produced. It is deciding what should be produced, setting a bar that rejects things, and choosing where review goes.

The organizations drowning in slop are not the ones using AI too much. They are the ones that scaled production and left everything else where it was.

References:
Baltes, S., Cheong, M., & Treude, C. (2026). “An endless stream of AI slop”: How developers discuss the burden of AI-assisted software development. arXiv:2603.27249. https://arxiv.org/abs/2603.27249
Feldman, E. (2025, December 17). Merriam-Webster’s word of the year for 2025 is “slop,” the A.I.-generated junk that fills our social media feeds. Smithsonian Magazine. https://www.smithsonianmag.com/smart-news/merriam-websters-word-of-the-year-for-2025-is-slop-the-ai-generated-junk-that-fills-our-social-media-feeds-180987887/