A Florida appeals court ran out of patience this spring. In a published opinion, the court described a self-represented party’s AI-assisted briefs as “slop” - citations to cases that did not exist or did not say what he claimed - and, to drive the point home, appended a limerick it had generated with a chatbot. We laughed. It was an easy laugh. Across the country, courts are seeing the same thing: confident, overlong, legal-sounding filings from people who fed a situation into a model and submitted whatever came back.
It is tempting to file this under “look what happens to people who do not know what they are doing.” Resist that. Part of the pro se surge is an access-to-justice story, not a punchline. But the more useful point for a lawyer is colder than either reaction: the failure on display in those filings is the same failure operating inside expert legal work. The only difference is the suit.
Here is the mechanism. A model shifts toward the premise it is handed. Give it a weak position and some frustration, and it does not stop you. It builds. For a pro se litigant with no legal baseline, the result is obvious garbage - fake cases, impossible arguments. The alarms go off for everyone, including the judge. For a skilled lawyer, the result is the opposite of obvious. The model takes your sophisticated, plausible, slightly wrong premise and renders it in exactly the prose you would have written on a good day. It mirrors your competence back at you.
That is the danger worth naming. AI does not only make weak lawyers look foolish. It makes strong lawyers sound exactly as right as they already expected to sound. The pro se litigant’s slop trips every alarm. Yours trips none, because it looks like your own work product. It survives your review because it was built to.
Call it invisible slop. It is not fabrication - fabrication you would catch. It is flawed judgment, fluently dressed. The citations are real. The structure is clean. The tone is yours. The only thing missing is the moment where something pushed back, and that absence does not show up on the page.
And it does not stop at your desk. The same fluency that got it past you gets it past everyone downstream. You pass it up to the partner, out to the client, into the filing, and it carries your name - which is exactly the credential that makes no one look twice. Worse, you are more confident than when you started. The model did not just miss your weak premise; it spent three paragraphs agreeing with it, and you walked away surer of yourself, not more careful. If it fooled you, it will fool everyone you hand it to, because you are the one vouching for it.
So the question to ask of AI-assisted work is not the one most lawyers ask. The instinct is to ask “is this correct?” and check the citations. Citations are the easy layer. The harder, more honest question is: did this ever resist me? If the model agreed with my framing from the first reply, that is not reassurance. That is the warning.
Three habits catch invisible slop better than re-reading ever will.
First, make it argue the other side. Not “are there counterarguments,” which produces a polite list. Tell it to destroy the position you just built, as if it represented your opponent and the fee turned on winning. What survives that is worth more than what it handed you first.
Second, strip your tell. The model reads whose side you are on from how you ask. Give it the facts with no framing - no “my client,” no adjectives, no theory of the case - and ask for a neutral read. The gap between that answer and the one it gave you when it knew what you wanted is a map of where it was flattering you.
Third, notice when it was too easy. If a model never made you work, never qualified, never told you the thing you did not want to hear, you did not get its judgment. You got your own, returned with better lighting. The smoother the session felt, the harder you should look.
None of this is about distrusting the tool to the point of not using it. I use these systems. The point is narrower, and it is the whole job: the value was never in the draft. It was in the friction the draft skips. A pro se litigant cannot supply that friction, because they never had it. A lawyer can - and a lawyer who lets a fluent model talk them out of it is making the same mistake the court just mocked, at a far higher hourly rate.
I write more about practicing law in the age of agentic AI - using these tools without surrendering judgment, privilege, or the duty of competence - at The Agentic Lawyer. www.theagenticlawyer.com
Educational only, not legal advice, and no attorney-client relationship is created. Views are my own, not my employer's. Attorney advertising in some jurisdictions.