The Hamster Did My Homework

Did you craft this document through careful back-and-forth with an AI assistant, or did Claude write it while you were taking a dump?

I cannot tell from the document, and neither can anyone else. It makes little difference: qui facit per alium facit per se (he who acts through another acts himself). As kids, we said the dog ate our homework. Now as adults we keep shtum, though really the hamster did our homework.

The lab does not diagnose patients

When the diagnosis is wrong, a doctor does not get to say “‘cause the lab said so.” Delegatus non potest delegare: the work may have assistants, but the judgement must have an owner. The lab is not at fault for the medic’s mistakes, and the AI lab is not at fault for your laziness. You build it, you own it. Or, mutatis mutandis, you write it, you own it. “But Claude said…” Yeah, no one cares what your hamster made up.

Assume everything is shite

George Akerlof diagnosed this over half a century ago, in a paper short enough to read over lunch—or you can have Claude give you the “load-bearing” facts laced with juicy hallucinations. Yes, it’s a feckin’ em dash. Get over it.

His insight was that when buyers cannot tell one car from another, they cannot pay for quality they cannot see, so the good cars leave by adverse selection, and the market fills with lemons. Documents are lemons now, with one twist: sellers, or “authors”, do not even know they are offering lemons. The rational response is to assume everything is shite by default and stop reading slop, because the cost of verifying an LLM-generated document is an externality: it is passed on from the writer to the readers. And that is a dastardly dereliction of a writer’s duty: if you can’t be buggered to write properly, I’m not going to give a blistering butt flap to read it.

The right kind of fast

Fast is fab and easy is better still. That’s why many institutions default to measuring throughput in some form or another as a proxy for efficiency. Strategy briefs, PRDs, RFCs, design documents… the more the merrier! Oh, and bigger is always better: why say in a single crisp page what ChatGPT can spread out over twenty? It does not matter whether what was produced was also correct: bollocks to quality! It merely needs to be plausible and reflect well on leadership. (OMG AI slop alert: a “not X, but Y” construction! Start clutching your pearls, kids.)

Consequences are for chumps and accountability is for arseholes. When the damage occurs, it mostly happens after the person responsible for the mess has moved on to the next project or the next company. The day your laziness—pardon, your productivity—bites you in the arse may never come, but the people who remember the crap you shovelled onto their desks keep score in the same building as you. You pay the piper when you need something from a person you once dropped a load on, only to find that their goodwill already buzzed off to the Bahamas.

The right kind of fast is the one that ships verified claims, states its assumptions up front, and says clearly how it might be or go wrong. (A tricolon? Fetch your paper bags, Hacker News readers. It’s hyperventilation time!) Verification happens once, at the source, and the readers downstream do not repeat it. That’s antithetic to how people use LLMs in organizations.

But everyone else is doing it!

“Everyone else leans on the LLM, so what is the big deal?” Welcome back to your childhood where the dog has died and you now have a hamster named Claude, yet you still cannot hear your mother wonder: “So if all your friends jumped off a bridge, you plan to go ahead and do it too?!” Yes, there is safety in numbers, but that’s also where mediocrity lives.

Anyone who makes this argument presumes the provider will never be poisoned and that the relationship will last. Abusive relationships can indeed last, just ask your therapist—or Dr DeepSeek.

Let’s run that wager through recent history, shall we? Amazon hosted third-party sellers for years on its marketplace, learned which products were worth selling, and then launched competing products from the data its sellers had handed it. People taught their tech overlords how to make them redundant in return for the privilege of using their platforms to make a living. Sound familiar?

Anthropic delivered a worse model at full price the moment it decided your question was not kosher. Who the hell is Anthropic to decide that? It decided because it can, and it can because your writing and thinking live on its servers. Whenever a vendor sits between you and anyone else, and sees everything you asked it to create or think through, the contract that specifies the prohibition to use that data for its own purposes is worth next to naught: the fine print hiding in reams of legalese provides loopholes around such prohibitions anyway, and almost nobody reads it. And as everyone concentrates on the same handful of models, one poisoned provider can poison the output of an entire industry. We’re not there yet, but that does not mean we’re not headed that way. And nobody would probably notice until it’s way past humanity’s bedtime. Sleep well, suckers.

Hamster fights

Verification is the expensive part of authorship, what economists call a costly signal. The person who actually checked the claims and thought through all consequences can afford to stand behind them. The person who dumped the output of a half-arsed prompt cannot, because they cannot answer: “What in the blazing bell ends is this?”

I pay the price anyway, because I’d like to think I have standards. But without integrity or enforcement at the source, every document leaves two options: either it must be accepted as the gospel even though it’s less informative than a lobotomy brochure from 1952, or it must be checked by its readers, which is a padded corridor to institutionalized paranoia.

So, the only sensible alternative left is to fight hamsters with hamsters. Your Claude gets my Gemini. I’ll slap my Big Pickle on your Kimi. And Rovo… Well, if you rely on Rovo, you’re a goner.

What I don’t do is let my hamster run wild. My hamster runs on its wheel checking the validity of the claims yours makes, and it returns direct quotes and links to primary sources, which means I still have to check key claims myself. That is unfortunately a horrible waste of resources, but it protects my sanity in the tsunami of AI-generated turds. Since attention is scarce, it must be spent where it can affect the outcome, and nowhere else.

I do not trust my hamster, though. In a reversal of the Cold War maxim, I verify, then trust. My hamster is merely an assistant, not the authority at the end of it. If the hamster is your second brain, fine. But if you outsource its critical faculties to the hamster, that’s no longer hunky-dory.

Still, my hamster is better than your hamster, because mine is on the wheel and yours is loose. I can clean up its mess relatively easily, whereas yours spray paints its “outputs” everywhere.

If you must use a hamster to do your homework, at least stand by it and don’t say “No idea. Claude said so.” Your signature is on the document, so own up to it. And when it turns out the hamster served you a lemon, remember: you can’t make lemonade from LLMesis (pronounced: ell-ell-emesis). Bon appétit!