You Can’t Eliminate AI Hallucination. You Can Absolutely Reduce It.

There’s a particular kind of argument doing the rounds in AI circles right now, and it’s hard to resist because it flatters you for following it. The pitch goes like this: hallucinations aren’t really about missing facts or weak grounding at all. They’re something deeper, a “geometric” failure, the model wandering into the low-density regions of its own activation space where its learned structure breaks down and it’s forced to emit a token because the architecture has no way to say nothing. Grounding can’t fix it. Retrieval can’t fix it. The problem is baked into the shape of the model itself.

It’s elegant. It’s confident. It sounds like the opening of a physics paper. And it’s becoming a cottage industry. For anyone responsible for putting these systems into production, it’s worth understanding why that’s exactly the problem.

Rigour and the costume of rigour aren’t the same thing

I’ve spent enough years around security vendors to develop an allergy. The tell isn’t whether something sounds technical; it’s whether the technical language is doing work or doing theatre. A lot of what’s circulating about hallucination right now fails that test. Try to find a single claim in it you could actually test. What counts as one of these “coherence manifolds”? How would you measure whether a model’s internal state has “entered a region where its learned structure fails”? At what point does the grand-sounding language become something you could prove wrong?

You usually can’t, because the vocabulary isn’t there to be tested. It’s there to sound inevitable. That’s a rhetorical move, not a scientific one, and it’s the same move a certain kind of vendor makes when they wrap an ordinary product in the language of an immutable law of nature: you can’t argue with physics, so stop trying and buy the thing.

And there is almost always a thing. Follow these arguments to the end and they tend to arrive at the same destination, the author’s own framework, method, or repo, positioned as the bridge across a gap they’ve just spent a thousand words declaring uncrossable. Once you see the structure you can’t unsee it: define a problem as architecturally unsolvable, then present yourself as the one person holding the solution. It’s a good business model. It’s bad engineering advice.

The real explanation isn’t one lab’s theory. It’s a consensus.

Here’s what makes the “it’s structural, give up” story more than just unhelpful. It’s almost certainly wrong, and not because one rival camp says so. Because over the past two years the major labs have converged on the same answer from completely different directions, and none of those directions is “topology we can’t fix.”

Start with the top-down account. In September 2025 a team led by Adam Kalai at OpenAI published a paper on why language models hallucinate, and it does the thing the doom-merchants refuse to do: it explains the behaviour in terms you can test and act on. Hallucinations, they prove, are ordinary statistical errors, the generative cousin of misclassification in any supervised model. They show mathematically that even a model trained on perfectly clean data will still hallucinate, and they tie the floor to something concrete: the fraction of facts that appear exactly once in training. If a fifth of the birthdays in your data show up only once, a well-calibrated model will miss at least a fifth of birthday questions. That’s not mysticism. That’s countable.

Then the behaviour persists, for a depressingly mundane reason: the way we grade these models rewards confident guessing over honest uncertainty. Most benchmarks score a model on whether it got the answer right. Say “I don’t know” and you score zero. Take a wild guess and you might land it. Run that incentive across millions of training signals and you get exactly what you’d expect: a system optimised to be a good test-taker, one that has learned bluffing beats abstaining. The same group had already shown, a year earlier, that a perfectly calibrated model is mathematically forced to hallucinate at a floor rate. This isn’t a hand-wave. It’s a proof.

Now the part that should settle it. Anthropic came at the same question from the opposite end, bottom-up, by tracing the actual circuits inside one of their models. They found something that turns the popular story on its head: the model’s default is to refuse. There’s a circuit sitting on by default that says, in effect, “I don’t have enough to answer this.” When the model recognises a name it knows well, a competing “known answer” feature fires and switches that refusal off. A hallucination is that switch misfiring, the model recognises a name, assumes it knows the rest, and suppresses its own “I can’t answer” reflex. The researchers could trigger it deliberately, making the model confidently invent a biography for a person who doesn’t exist, by hand-activating that one feature.

Sit with what that means. One lab proved it from the statistics looking in; another watched it happen at the wiring level looking out. Google found that models often encode internally whether they’re about to lie, but don’t act on the signal. Meta found that both fine-tuning and reinforcement learning actively push models toward confident fabrication. Four labs, four methods, one answer. None of them is “it’s geometric and you’re stuck.” All of them point at training and evaluation incentives, the things we control.

That’s not a counsel of despair. It’s a to-do list.

Why enterprises deploying AI solutions should care which story is true

This isn’t an academic spat. The two explanations hand you opposite instructions, and if you’re the one signing off on an AI system in production, you have to pick.

The “it’s geometric, it’s unfixable” story tells you to give up. Hallucination is structural, grounding is a misdiagnosis, so retrieval and verification are wasted effort and your only real hope is a new architecture and, conveniently, whatever the person making the argument happens to be selling. Believe that and you stop doing the very things the whole field just demonstrated actually work.

The consensus story tells you to get to work, and everything on its list is something you can do now:

• Reward the model for saying “I don’t know.” When you evaluate or fine-tune, stop scoring abstention as failure. Penalise confident wrong answers harder than honest uncertainty. OpenAI reported that grading this way cut one model’s hallucination rate on a test set from roughly three-quarters to around a quarter. You are not at the mercy of the architecture here; you are at the mercy of your own grading.
• Ground it anyway. Retrieval doesn’t cure hallucination, nothing does completely, but the claim that it can’t help is contradicted by the labs’ own results: Meta nearly doubled factual accuracy on a domain by grounding answers in a real source of record. Tie the answer to a citable document and you give the system somewhere true to stand.
• Instrument for confidence, not just correctness. Capture when the model hedges, when it refuses, when it’s operating outside its grounding. Google’s finding is the lever here: the model often already knows it’s on thin ice, so surface that signal instead of throwing it away. A system that knows its limits and tells you is worth more in a regulated workflow than a more “accurate” one that bluffs with total composure.

Notice that none of this waits of conceptual topology to arrive. It’s available now.

The discipline the doom-sellers are missing

I’ll grant the structural crowd one thing, and the labs back them on it: hallucination really is, in a narrow sense, permanent. Every major team says the same, you cannot drive it to zero in a base model, and anyone selling you one that “doesn’t hallucinate” is the vendor you should trust least. On that, the doom and the data agree.

But “you can’t eliminate it” and “you can’t reduce it, so don’t try” are very different sentences, and the gap between them is where every serious AI governance decision actually lives. The job was never to find an architecture that can’t be wrong. It’s to build systems that know when they might be, tell you, and leave a trail you can check. That’s not topology. It’s engineering, and it’s the unglamorous, testable, do-it-this-quarter kind.

So be careful with any theory that explains why a hard problem is unsolvable and then, in the same breath, offers to sell you the way across. The most useful explanations rarely arrive that way. They come with a to-do list instead of a sales pitch, and the items on it work whether or not the grand theory behind them ever pans out.