The Sloppiest Thing About AI

Mansour — Fri, 21 Aug 2026 04:32:22 +0000

Originally published on oxblog.

There is a fashionable new way to avoid thinking. Notice an em dash, a tidy
list or a sentence with suspiciously polished edges. Announce that AI was
involved. Dismiss the whole thing.

The article may be correct. The product may work. The patch may fix the bug.
None of that matters once the detector in someone’s head has beeped. The work
enters the AI bucket, the bucket is labelled slop, and judgment can stop
before it becomes tiring.

AI did not create this habit. It merely brought it to the surface.

People have always cared more about the signals surrounding an idea than the
idea itself. Tell them someone they admire said something and they lean
forward. Tell them the same words came from someone they despise and they stop
listening. AI now provides an even cheaper excuse.

Provenance is not quality

Is the claim true? Is the reasoning sound? Does the product solve the stated
problem? Can the contributor explain and defend the patch?

Those questions measure substance. “Was AI involved?” does not answer any of
them.

A beautifully written falsehood remains false. An awkwardly written insight
remains an insight. A human can produce derivative nonsense; a machine can
help express a useful idea clearly. The origin does not reverse the result.

The sloppiest classifier

AI slop is real. Cheap generation makes it possible to flood inboxes, issue
trackers and publishing systems with plausible-looking material that nobody
cared enough to verify.

Reject fabricated claims, repetition, untested patches, evasive authors and
submissions whose creators cannot answer basic questions about them.
Rate-limit the flood. Control the review burden.

None of that requires pretending every use of AI produces the same result.

Calling all AI-assisted work “slop” is itself the sloppiest possible
classification: putting unlike things into one bucket because teasing out the
differences requires effort.

It is intellectual laziness dressed as taste.

Blanket bans are surrender

Good rules describe the failure they are intended to prevent.

Require evidence. Require tests. Require disclosure where relevant. Require
submitters to understand, revise and accept responsibility for every word or
line they send.

These rules work whether the offending material came from a model, a careless
human or a human using a model carelessly.

Blanket bans do the opposite. They reject honest contributors while rewarding
anyone willing to conceal their tools. They exclude people with useful
contributions who lack the time, confidence or writing ability to package them
attractively.

They also guarantee that some good work will be discarded for no defect
present in the work.

A thousand hours of unpaid maintenance buys gratitude, not infallibility.
Scarcity of maintainer attention is a sound reason to control volume. It is
not evidence that an entire class of tools cannot produce value.

Responsibility is the boundary

Using AI is not an exemption from standards.

If you submit the work, you own it. You own its errors, omissions, invented
citations, insecure code and tedious prose. “The model did it” is no more
acceptable than “my editor did it” or “my compiler allowed it.”

A tool can assist production. It cannot inherit responsibility.

Ban unaccountable work. Ban abuse. Ban deception where disclosure is
required.

Do not ban useful work because the person behind it used leverage.

The disclosure

Every sentence in this article was generated by AI from my argument,
direction and requested tone.

If that fact created an urge to dismiss what you just read, examine the urge.
Nothing in the argument changed. No claim became less true. You merely learned
something about the process and felt tempted to substitute it for an
assessment of the result.

That is precisely the point.

Judge the work. Demand evidence. Demand accountability. Reject rubbish
without apology.