The Sloppiest Thing About AI — oxblog
There is a fashionable new way to avoid thinking. Notice an em dash, a tidy<br>list or a sentence with suspiciously polished edges. Announce that AI was<br>involved. Dismiss the whole thing.
The post may be correct. The product may work. The patch may fix the bug. None<br>of that matters once the detector in someone’s head has beeped. The work has<br>been placed in the AI bucket, the bucket has been labelled slop, and<br>judgment can stop before it becomes tiring.
AI did not create this habit. It merely brought it to the surface. Plenty of<br>people have always cared more about the signals surrounding an idea than the<br>idea itself. Tell them a person they admire said something and they lean<br>forward. Tell them the same words came from someone they despise and they stop<br>listening. Now authorship by a machine provides an even cheaper excuse.
Provenance is not quality
Where something came from can matter. It can expose a conflict of interest,<br>establish authority or tell us who is responsible. What it cannot do is spare<br>us from examining the thing itself.
Is the claim true? Is the reasoning sound? Does the product solve the stated<br>problem? Can the contributor explain and defend the patch? Those questions<br>measure substance. “Was AI involved?” does not answer any of them.
A beautifully written falsehood remains false. An awkwardly written insight<br>remains an insight. A human can produce empty, derivative nonsense with great<br>confidence; a machine can help express a useful idea with unusual clarity.<br>The origin does not reverse the result.
Anyone who changes an opinion merely because they discovered who — or what —<br>formed the sentence was not evaluating the sentence. They were evaluating a<br>social cue and mistaking the sensation for thought.
The sloppiest classifier
AI slop is real. Cheap generation has made it possible to flood every inbox,<br>issue tracker and publishing system with plausible-looking material that no<br>one cared enough to verify. The review cost is real too: ten seconds of<br>generation can impose ten minutes of work on someone else.
Reject that behaviour without hesitation. Reject fabricated claims,<br>repetition, empty summaries, untested patches, evasive authors and submissions<br>whose creators cannot answer basic questions about them. Rate-limit the flood.<br>Close the gate when there are not enough people to operate it. None of this<br>requires pretending that every use of AI produces the same result.
Calling all AI-assisted work “slop” is itself the sloppiest possible act of<br>classification: putting unlike things into one bucket because teasing out the<br>differences requires effort. It is intellectual laziness dressed as taste.<br>Sometimes it is worse — a performance of discernment by people unwilling to<br>do any discerning.
Blanket bans are surrender
Projects, publications and communities that ban AI-generated contributions<br>usually present the ban as a defence of quality. It is an admission that they<br>have chosen not to measure quality.
Good rules describe the failure they are designed to prevent. Require<br>evidence. Require tests. Require disclosure where it is relevant. Require the<br>submitter to understand, revise and take responsibility for every word or line<br>they send. Reject duplication and mass submission. These rules work whether<br>the offending material came from a model, a careless human or a human using a<br>model carelessly.
A provenance ban does the opposite. It rejects honest contributors while<br>rewarding anyone willing to conceal their tools. It removes people who have<br>something useful to contribute but lack the time, confidence or writing skill<br>to package it in the approved manner. And it guarantees that some good work<br>will be discarded for no defect present in the work.
This does not mean a maintainer must accept every contribution. No volunteer<br>owes the internet an unlimited review queue. But a thousand hours of unpaid<br>work buys gratitude, not infallibility. Scarcity of maintainer attention is a<br>sound reason to control volume. It is not evidence that an entire class of<br>tools is incapable of producing value.
Design for the world that exists
Generated material is not a passing current one can nobly swim against. It is<br>a collapse in the cost of producing words, images and code. That change will<br>not be reversed by a rule in a contribution guide.
The serious response is to build better filters around the properties that<br>matter: truth, originality, usefulness, accountability and cost imposed on<br>others. The unserious response is to prohibit the new tool and congratulate<br>oneself for preserving standards. One adapts the system to reality. The other<br>tries to preserve the appearance of the old world while the ground moves<br>underneath it.
This is what makes blanket bans so contemptible. They take the easiest<br>available action — say no to a category — and leave the difficult work of<br>preserving quality undone. They protect a familiar surface at the expense of<br>the substance they claim to...