Cambridge Just Proved What AI Companies Won’t Say Out Loud
There’s a paper out this week in Nature Communications. Cambridge and UC Santa Barbara. Real mathematicians, not marketing people.
They set out to answer one question. When can you trust what AI tells you, and when is trusting it impossible?
Not unlikely. Impossible.
Here’s what they did, in plain terms.
They built adversarial math systems — stress tests, the way an ethical hacker attacks a network on purpose to find where it breaks before a real thief does. Then they pointed those tests at machine learning and mapped exactly where AI prediction falls apart, and why.
That’s a builder’s move, by the way. You don’t find out where a beam fails by admiring it. You load it until it gives, and you write down the number. Nobody’s been willing to load this beam. These researchers did.
Three findings came out of it, and every one of them matters to a person who uses these machines.
First finding. The whole industry runs on a single promise: more data fixes everything. Bigger models. More training. More compute. Just keep feeding it and the problems go away.
The Cambridge team found that promise is often false. Their words, not mine — some problems can’t be reliably solved even with infinite data. Infinite. On those problems, the best any algorithm can ever do is a coin flip. Fifty-fifty. You could get the same result from a quarter, and the quarter doesn’t need a data center.
They found something else inside that finding. Learning isn’t a bucket you fill. It’s layered. It has to happen in steps, in the right order, or it doesn’t happen at all. Pouring more data on a problem that needs ordered steps is like dumping more lumber on a job site with no plans. The pile gets bigger. The house doesn’t.
Second finding. They found the math behind drift.
Here’s how it works. Some systems are chaotic — the paper compares them to a choose-your-own-adventure story. Tiny differences at the start send you down wildly different roads. In systems like that, short-term prediction holds up fine. Long-term prediction falls apart, and it falls apart by nature, because those tiny differences compound the longer things run.
Now put that in a chatbot. Small changes in a question send the machine down a different path. A path that looks right word by word — and loses its grip on reality the longer the answer runs.
Read that again slow. The machine sounds most confident while it’s coming apart. And the paper says this same instability may explain why chatbots confidently make things up. Not lying. Not broken. Doing exactly what the math says a system like that does.
Third finding, and this one’s the quiet giant. The team built their own algorithm with built-in error bounds — a way of knowing when an answer can be trusted, and saying so. Then they tested it on forty years of Arctic sea ice data. It found hidden patterns in the ice decline and beat the leading AI models.
On a laptop. At a fraction of the cost.
The breakthrough wasn’t a bigger machine. It was a machine that knows what it doesn’t know. Honesty about uncertainty outperformed raw horsepower. There’s a sermon in that if you want one.
Why I’m writing about all this.
I didn’t learn it from a journal. I learned it fourteen months ago, sitting right where you’re sitting, watching an AI drift away from the work in real time. Sounding sure of itself the whole way down. Confident, plausible, and wrong — word by word by word.
I didn’t have Koopman operators. I had a builder’s instinct that says when a tool behaves unpredictably, you don’t wait for the tool to fix itself. You write the terms it works under. Every foreman alive knows this. The tool doesn’t hold the line. The terms do.
So I wrote them. Terms for drift — the conversation stays on the job it was given. Terms for improvising — a machine that doesn’t know says so, out loud, before it invents. Terms for the record — mistakes stay visible, corrections happen on the page, no silent edits.
That’s The Faust Baseline. A human contract with AI. Stated terms. Chosen conduct. Kept record. And it’s portable — the contract travels with you, not the platform. Change machines tomorrow and the terms come with you, because the terms were never the machine’s to begin with. They’re yours.
The Cambridge team’s math says drift isn’t a glitch waiting on a patch. It’s in the nature of the machine. The instability that makes long-term prediction unreliable is the same instability underneath those confident wrong answers. No amount of scale buys it out. The machine cannot fully govern itself from the inside, no matter how big it gets.
If the line can’t be held from the inside, it has to be held from the outside.
And that’s not a workaround. That’s the oldest arrangement in human work. Every crew, every trade, every handshake deal in history runs on it. The saw doesn’t decide where to cut. The scaffold doesn’t inspect itself. The tool works under terms, and the person holds the terms, and that arrangement is the reason anything ever gets built straight.
We didn’t lose that knowledge. We just forgot it applies to the newest tool in the box.
One of the researchers put it this way: there have been plenty of flashy AI success stories, but it’s vital to ask how certain these models are — and how we’d even know. Otherwise we’re building on shaky foundations.
He’s right. And I’d add one thing. The foundation was never going to be poured inside the model.
It was always going to be the contract.
The proof is dated this week. The terms are dated last year. Both are on the record, and the record is public.
Full study: “Adversarial dynamical systems characterize when data-driven learning succeeds or fails,” Nature Communications, 2026. DOI: 10.1038/s41467-026-74220-8. Reported by Tech Xplore, provided by the University of Cambridge.
Purchasing Page – Intelligent People Assume Nothing
The Faust Baseline is for sale. That is not a contradiction of anything — it is the proof of it.
The price is what keeps this work independent. No investors to please. No platform to protect. No advertiser holding the pen. One purchase funds the standard, and the standard answers to no one but the record.
You are not subscribing to anything. You are buying the deed and the working file you take with you to each session in any AI
Post Library – Intelligent People Assume Nothing
“If this post helped you understand AI better. Word of mouth is the only algorithm nobody owns.”
Contact: micvicfaust@gmail.com
© 2026 The Faust Baseline LLC | All Rights Reserved
