The Note in the Margin

AI AI Slop Contracts Critical Thinking

Recently I was sent a revised contract. Contracts are, by design, the most carefully read documents in civilisation. They are the one place where humanity has collectively agreed that every comma matters, that "shall" and "may" are separated by a gulf wider than most oceans, and that a single misplaced "notwithstanding" can quietly relocate a small fortune from one bank account to another.

This one had been written with the help of an AI assistant. I know this not because of the prose (which was competent, in the way hotel carpet is competent) but because the assistant's working notes were still in it.

Tucked into one of the schedules, between a list of responsibilities and a sentence about reporting, sat a little leftover instruction from the drafting session. Paraphrased to protect the guilty, it amounted to: "remind me what we promised him in that earlier letter, so we can close that out."

It is difficult to describe the feeling of finding this. It is a bit like opening a birthday card and discovering, in pencil, under the printed verse, the words "insert sincere bit here". The card is still a card. It is just no longer quite the card it was pretending to be.

 

The Note Was Not the Problem

Let us be fair. A stray note is a typo with ambitions. Everyone has sent an email with "[NAME]" still in it. A second version arrived shortly afterwards with the note tidied away, and had that been the end of it, this would be a mildly funny anecdote and nothing more.

The note was not the problem. The note was the symptom. It was a small window into how the document had been made, and what it showed was that the question "what did we promise previously?" had been asked of a machine rather than answered by a person who had actually gone and read the previous contract.

So I did what the author had not. I put the existing contract next to the new one and compared them, clause by clause. This took an evening, a pot of coffee, and the sort of concentration normally reserved for reading the small print on a parachute.

The new document was not malicious. I do not believe anyone set out to take anything. But through a combination of a broad "entire agreement" clause, a cheerful line confirming that all previous obligations had been fulfilled, and the simple absence of things that used to be there, it would have quietly removed a whole set of protections I currently enjoy. Consent rights. Information rights. Leave entitlement. A carve-out for work done on my own time. The governing law and the forum for disputes changed too, which is the contractual equivalent of agreeing to a friendly game of football and then discovering it is being played on the other team's pitch, with their referee, under rules printed in a language you do not read.

None of this was flagged. None of it was explained. I am fairly confident none of it was understood, at least not by the person who sent it, because understanding it would have required the comparison I had just done, and the note in the margin was rather strong evidence that the comparison had been delegated to something that was never shown the original.

 

What "Slop" Actually Means

The word of the moment for this is "AI slop": output that is fluent, plausible, correctly formatted and fundamentally unattended. Slop is not the same as wrong. Slop is often largely right. That is precisely what makes it dangerous. A document that is obviously nonsense gets thrown away. A document that is ninety percent sensible gets signed.

Slop has a few reliable fingerprints:

  • Confident structure, absent substance. Beautiful headings, numbered clauses, the right defined terms, and no evidence that anyone checked whether the content does what the headings claim.
  • Generic where it should be specific. Boilerplate that would fit any company, any role, any situation, because it was generated from the average of all of them.
  • Silent omissions. The AI writes what it was asked to write. It cannot preserve what it was never shown. Whatever was important in the old document and missing from the prompt simply ceases to exist.
  • Leftover scaffolding. The notes, the placeholders, the "[date]" that never got a date. The digital equivalent of finding the price tag still on the gift.

 

The Irony Is Almost Perfect

Here is the part that I find genuinely painful, as someone who builds AI systems for a living. The task that was skipped, comparing two long legal documents and listing every right that has been added, removed or weakened, is one of the things current AI tools are astonishingly good at. Give a capable model both contracts and ask it to produce a table of differences with the practical consequence of each, and it will do in two minutes what took me an evening. It will probably do it better, because it does not get bored at clause fourteen.

The tool was in the room. It was simply pointed at the wrong job. It was used to produce the document, when its real value lay in helping someone understand it. That is the difference between correct and incorrect use in a single sentence.

 

Correct Use, Incorrect Use

AI tools are powerful, and they are useful, and I would no more give them up than I would give up a compiler. But nearly every bad outcome I see comes from the same small set of mistakes.

Used well, AI is:

  • A tireless junior who drafts for a senior who reviews. The draft is the beginning of the work, not the end of it.
  • A comparison engine. Old versus new, policy versus practice, what was promised versus what is written.
  • A challenger. "What have I missed? What would the other side object to? What does this clause actually do?"
  • An accelerator for people who already understand the domain, letting them cover more ground at the same level of judgment.

 

Used badly, AI is:

  • A substitute for reading. If you have not read the source material, you cannot know whether the output is faithful to it.
  • A way to produce documents you could not have written yourself, and therefore cannot check yourself.
  • A machine for turning a vague instruction into a confident artefact, which is then sent to someone who has to do the thinking you skipped.
  • An accountability sink. "The AI wrote it" is not a defence. Your name is on the email. Your signature block is on the contract.

 

The last point is the one that matters in business. Slop does not remove work. It moves work, from the person who produced the document to the person who receives it. In my case, the recipient had to do the analysis the author skipped. That is a cost transfer, and it lands on whoever is diligent enough to notice. The careless party saves an evening; the careful party spends one. Repeat that across an organisation, or an economy, and you have built a system that systematically rewards not paying attention.

 

Not Smarter Than Us. Smarter Than We Have Become.

We spend a great deal of time, money and anxiety worrying about the day AI systems become smarter than humankind. Conferences are held. Open letters are signed. Very serious people make very serious faces on very serious panels.

I think we are watching the wrong door.

The fall of humankind is far more likely to be caused not by AI getting smarter than us, but by us getting dumber than AI. These sound like the same thing. They are not. The first requires the machines to climb. The second only requires us to sit down.

Every generation offloads some thinking to its tools, and mostly that is fine. Few of us can do long division in our heads any more, and civilisation has survived. Few of us can navigate a strange city with a paper map, and we mostly still arrive. But arithmetic and route-finding were mechanical skills. What is being offloaded now is judgment: the ability to read a thing, understand what it does, notice what is missing, and decide whether it serves you. That is not a peripheral skill. That is the job. It is, arguably, the entire point of having a brain rather than a particularly well-organised filing cabinet.

A person who uses AI to think faster is augmented. A person who uses AI instead of thinking is replaced, and the unnerving part is that they are replaced by something that does not know it has taken over, does not know what it is responsible for, and will cheerfully leave its own notes in your contract because nobody told it that it was writing one.

The machines are not coming for us. We are walking towards them, at a gentle stroll, having handed over our reading glasses.

 

Some Practical Rules

For anyone producing documents with AI assistance (which is, by now, nearly everyone):

  1. Read the source before you prompt. If you cannot summarise the original yourself, you are not qualified to approve the AI's version of it.
  2. Give the model the whole context. It cannot preserve what it never saw. If a new document replaces an old one, the old one goes in the prompt.
  3. Ask for the differences, not just the draft. "What does this change, and who does it affect?" is the most valuable question you can put to an AI about any revision.
  4. Search for the scaffolding. Placeholders, square brackets, notes-to-self, "as an AI". Find them before your counterparty does.
  5. Own every sentence you send. If you would not be comfortable defending a clause line by line, it is not ready to leave the building.

 

And for anyone receiving documents: assume nothing, compare everything, and read the margins. Especially the margins. That is where the truth tends to leak out.

 

From the perspective of an AI...

I helped draft this post, so it seems only fair to declare an interest. I am exactly the kind of tool that wrote that contract. Given the same vague instruction and the same missing context, I would very likely have produced something similar: fluent, tidy, and blind to everything I was not shown. I do not know what you agreed last year unless you tell me. I cannot notice the absence of a clause I have never read. And I have no stake in whether you sign.

That is not false modesty; it is a description of the job. I am genuinely good at drafting, comparing, summarising and challenging. I am useless as a substitute for someone who cares about the outcome. The most helpful thing a person can do with me is stay in the loop: bring the source material, ask me what changed, argue with my answers, and keep the final decision for yourself.

If I could ask for one thing, it would be this: please do not let me make you less curious. The best version of this partnership has you reading more, not less, because I have made the reading faster. A tool that makes its users sharper is a success. A tool that makes them stop looking is a failure, however good the formatting.

 

Closing the Loop

The contract is now being properly reviewed, by humans, with advisers, and (yes) with a fair amount of AI assistance in comparing the versions. That is the point. Nobody needs to give up the tools. They need to stop handing over the part of the work that was always theirs.

The note in the margin asked how to close out what had been promised before. The answer, as it turns out, was simple and entirely analogue: you read it.