Chapter 3 Can I trust it?
Lesson 3.6:Watch Out for AI Slop (Who Vouches for This?)
A partner sends a question to an associate at 11:31. A six-page memo comes back at 11:40. It has numbered headings, a list of considerations and a cautious conclusion. The partner now has to read it. The associate did not. Who is the author of a document nobody read?
Before the definitions, try the test that this lesson is about. Judge a memo first, and see what produced it afterwards.
The partner asks
“Can our client terminate the supply agreement for the supplier's late deliveries?”
Memo A · Re: termination for late delivery
1. Overview
Late delivery by a supplier may entitle the purchasing party to terminate the agreement, depending on the terms of the contract and the applicable law. It is important to assess each situation on its facts.
2. Key considerations
- Whether time is of the essence under the agreement or by implication.
- Whether the delay constitutes a material or repudiatory breach.
- Any notice and cure provisions, which must be strictly complied with.
- Whether the supplier may rely on force majeure or other excusing events.
- The commercial relationship and any waiver through prior conduct.
3. Recommended approach
The client should carefully review the termination clause and document all instances of late delivery. It is advisable to serve a formal notice of breach before termination to reduce the risk of wrongful termination. Legal advice should be sought regarding the specific jurisdiction.
4. Conclusion
Termination may be possible but carries risk. A cautious, well-documented approach is recommended.
Your rating of the first memo was the training target.
Fictional matter. Memo A was produced from the fifteen-word prompt; Memo B from the same question with the file, a reader, a verdict and a word budget.
What slop is, and is not
Slop is output that nobody vouched for, passed to someone who now has to do the checking. The word was fixed in 2024 by the programmer Simon Willison: sharing unreviewed AI content with other people is rude simonwillison.net Willison, May 2024 (opens in a new tab). Charlie Hills and Sam Illingworth add the part that matters for lawyers. Slop presents conclusions without showing how they were reached, bounded or checked. What is missing is responsibility, not polish Substack Hills and Illingworth, Jan 2026 (opens in a new tab).
So three things are not slop. AI-assisted work that a person read and stands behind. A short, unpolished answer with a point of view. An honest “I did not verify this” attached to a draft. And one thing is slop even with no AI involved: the junior's draft forwarded unread, the document dump in disclosure. Human slop always existed. What the tools removed is the friction that limited how much of it one person could produce in a day.
The ratio rule
There is a quick test that needs no expertise. If the input is much shorter than the output, the extra material did not come from the matter. A model is a conditional machine: it writes the most likely next words given what it was shown. Where the prompt rules nothing out, it fills from the average of everything it has read. Memo A above is that average. Every sentence applies to every supply contract in every country, which is why none of them helps with this one.
Legal work runs the other way. A thick file goes in and a short opinion comes out. The value is in the compression: which three facts decide the case, which clause governs, what to do on Friday. Slop is expansion. Fifteen words go in and four pages come out. The reader can apply the rule in two seconds, to their own drafts and to what lands in their inbox.
The rule has an inverse for the writer. The prompt that produces a short, specific output is long: the file, the reader, the question, a verdict first, a word budget. Memo B took more work to ask for than Memo A. That work was the lawyer's job before AI, and it still is.
Correct slop is still slop
Suppose Memo A happened to be right. It would still be slop, because the partner cannot tell right from wrong without redoing the work. Slop is about verifiability and ownership, not the error rate. A document carries value only if someone can be asked “what did you check?” and answer.
There is a reason the test above fools people. Models are trained on human ratings of how good a text looks on reading, and that is exactly the instinct a reviewer uses on a first pass. A lawyer's gut is reliable on a junior's draft. On a model's draft, the gut was the training target. This is the same mechanism as the agreeable answers of lesson 3.3, and the reason the subtler errors of lesson 3.1 survive a read-through.
Two tells follow from this. In slop, everything is listed and nothing is ranked, because ranking is where judgment lives. And the hedges are generic. “This may depend on jurisdiction” fits any document. A real hedge is specific: “I could not confirm this applies to Scotland.” A caveat that works in any memo is not a caveat.
Both tells are easier to see than to describe. Two short texts a lawyer might send, each in the version that went out and the version that carries the same facts ranked.
Reply to a client asking when their claim expires
As sent 82 words
Thank you for your email regarding the limitation period. It is important to note that limitation periods can vary depending on the nature of the claim and the applicable jurisdiction. Generally, contractual claims are subject to a six-year limitation period, although this may differ in certain circumstances. We would recommend reviewing the relevant documentation carefully and seeking advice at the earliest opportunity to ensure that no deadlines are missed. Please do not hesitate to contact us should you have any further questions.
Same facts, ranked 51 words
Your claim is in contract, so you have six years from the breach. The breach is the missed payment of 3 May 2021. Issue proceedings before 3 May 2027. I recommend a letter before action this month, and I have not yet checked whether the 2023 standstill letter extends the date.
Missing from the sent version: Which claim, which date, what to do, and what was not checked.
Note to the partner after reviewing a services agreement
As sent 81 words
I have reviewed the draft services agreement and set out below my key observations. Overall, the agreement is broadly in line with market practice, although there are a number of provisions that may warrant further consideration. In particular, the limitation of liability clause, the termination provisions and the intellectual property clause should be reviewed carefully. Some of these clauses may expose the client to risk. I would suggest that we discuss these points further and consider whether any amendments are appropriate.
Same facts, ranked 57 words
Two points block signature. Clause 12 caps the supplier's liability at fees paid, about £40,000, against a data-loss exposure near £2 million; ask for £1 million or proof of insurance. Clause 9 gives the supplier the IP in deliverables; the client needs ownership or a perpetual licence. Clause 15 on termination is acceptable. The rest is market.
Missing from the sent version: Which two clauses matter, by how much, and what to ask for.
Answer to a colleague asking if the draft can go to the other side
As sent 55 words
Thanks for checking. In principle there is no objection to sharing the draft with the counterparty at this stage, subject to the usual considerations around privilege and confidentiality. It may be worth confirming internally that all comments have been addressed before anything is sent across. Let me know if it would be helpful to discuss.
Same facts, ranked 20 words
Yes, send it. First delete the internal comment on clause 7, which is privileged. Mark it draft and without prejudice.
Missing from the sent version: The yes, the one thing to remove, and why.
Shorter and heavier. One version carries the matter, the other carries the average.
Fictional matters. Both versions of each text were written for this lesson.
Who pays
Slop is a transfer of cost. The sender saves the reading. The receiver does it instead, usually with less context. In a 2025 survey of 1,150 US employees, 40% had received what the authors call workslop in the previous month, and each instance cost the receiver close to 2 hours to decode or redo Harvard Business Review Niederhoffer and Hancock, Sep 2025 (opens in a new tab). Lesson 2.5 showed the in-house version: hours saved on your own drafting, spent correcting someone else's AI.
The cost lands on whoever has to read. The open-source project curl closed its bug bounty on 1 February 2026 after the share of valid reports fell from more than 15% to below 5%, the rest being AI-generated reports its maintainers had to read one by one daniel.haxx.se Stenberg, Jan 2026 (opens in a new tab). The group general counsel of the publisher Reach reported a rise in AI-generated libel and privacy complaints, often unfounded or legally flawed, each of which must still be answered Global Legal Post Schroeder, May 2026 (opens in a new tab). Courts keep their own count: 2,149 decisions worldwide now address AI-fabricated content in filings, 854 of them involving lawyers Damien Charlotin AI Hallucination Cases, Oct 2026 (opens in a new tab). The judge in Mata v. Avianca put the cost in two sentences: the opposing party wastes time and money exposing the deception, and the court's time is taken from other important endeavors S.D.N.Y. Mata v. Avianca, Jun 2023 (opens in a new tab).
For a firm, the question is now commercial. Three in five in-house counsel see no savings from their outside counsel's AI Bloomberg Law ACC and Everlaw, Oct 2025 (opens in a new tab). If the other side's AI draft takes two hours to fix, who bills for that, and will the client pay?
Two lanes
Hills and Illingworth split AI content into two lanes. The volume lane: quality that is good enough, as much of it as the channel will carry. It can be profitable and it builds no trust. The leverage lane: judgment and expertise, with the model as an accelerator. People want the second lane's results, copy the first lane's tactics, then blame the tool Substack Hills and Illingworth, Jan 2026 (opens in a new tab).
Law has a volume lane. It is the AI client alert nobody at the firm read, the blog post written for search engines, the weekly LinkedIn “insight”. About half of new web articles have been mostly AI-written since early 2025, and the share has stopped rising Axios Graphite, May 2026 (opens in a new tab). The readers who matter to a firm learned to recognise it. Which lane is your firm's newsletter in, and do your clients know?
The feedback loop runs through people
A popular worry is that models will eat their own output. The lab result is real: a model trained only on text from earlier models loses its rare patterns first and converges on a generic average, a result called model collapse Nature Shumailov et al., Jul 2024 (opens in a new tab). The prediction built on it is weaker. Keeping the original human text alongside the synthetic text avoids the collapse arXiv Gerstgrasser et al., Apr 2024 (opens in a new tab), and labs filter what they train on.
What nobody filters is people. A study of 740,000 hours of lectures and podcasts found that words ChatGPT favours, “delve” first among them, rose sharply in spoken English after its release arXiv Yakura et al., 2024 (opens in a new tab). A generation of juniors is learning to write from model output. The convergence on the beige average is happening in human prose first. That is the collapse worth a lawyer's attention: the signal between people who know something and people who need to know it is getting noisier.
What this means for your work
On the sending side, four habits, adapted from a one-page etiquette site noslopgrenade.com Neubig (opens in a new tab). Use the model to make things shorter, not longer. If someone asked you, they wanted your judgment, so give a verdict. Never send what you have not read. And say what you checked and what you did not, because the model cannot say it for you. Ask it whether it is sure and it runs the same process that produced the answer. Asking the model for its confidence is asking the slop to grade itself.
Checked: clauses 14 and 20 against the signed agreement; the three dates against the delivery log.
Not checked: whether the 15 April email counts as a force majeure notice.
Decided: terminate under clause 14.3 once that point is settled.
Drafted with AI. Read in full by me.
Three lines the model cannot write for you.
Filled in for the fictional matter above. The blank version pastes as three lines and a sign-off.
On the receiving side, one habit. Before reading a draft, ask what was checked. If the answer is nothing, the document has no author yet, and the reading is work the sender owes. A firm can make this routine by storing the prompt with the output, the way an instruction letter sits with an opinion. The prompt is the file note.
This ends chapter 3. Chapter 4 asks whether you are allowed to use these tools at all.