A cleverer model won't fix your AI's wrong answers

The complaint we hear most often about AI is some version of: it sounds confident and it’s sometimes wrong. The assumption that usually follows is that this is a model problem, and that it will be solved by waiting for a cleverer one.

A group of researchers at Newcastle have just put a decent dent in that assumption. Their result is worth knowing about, and so is what happened to the number on its way to being reported.

What they actually did

The work comes out of the National Innovation Centre for Data at Newcastle University, and the paper — Reducing Hallucinations in Complex Question Answering using Simple Graph-based Retrieval-Augmented Generation — is public on arXiv, first submitted in June and revised in July.

The set-up, in plain terms. You can hand an AI a pile of documents and let it search them for relevant passages before it answers. That is the standard approach and it helps. What the team added was a simple map of how the facts relate to each other — which thing belongs to which, what follows from what — and a set of tools that let the system walk that map rather than only fishing for similar-looking text.

Then they tested it on deliberately hard questions: the sort where the answer is not sitting in one paragraph but has to be assembled from several places.

The paper’s own finding is that adding those graph tools halved the number of hallucinated answers, raised both the precision and the recall of factual correctness, and produced the highest truthfulness score of the three set-ups they compared. And it did so with a modest increase in the amount of computation used.

Same model throughout. What changed was the plumbing.

What happened to the number

Here is the part I found more instructive than the result.

By the time this reached me second-hand, it had become “cuts hallucinations by 80%” and — this is the one that matters — “using fewer tokens”. The paper says halved, and it says token use went up slightly, not down. The cost story had been reversed somewhere between the paper and my inbox.

The press coverage sits in between: IT Pro, 4 August 2026, reports that the technique lifted accurate answers from 29% to 66% across 510 questions and frames it as improving accuracy by up to 80%. Those figures may well be right — they are simply not the claim the researchers put their names to in the abstract, so we have attributed them rather than asserting them.

None of this is scandalous. It is ordinary drift, of the kind that happens to every number that travels. But it is a good argument for reading the thing itself when a statistic is about to influence a decision, and it is a slightly awkward one for me, given that this post is about machines confidently getting details wrong.

Why this matters if you run a small business

Because it moves the lever somewhere you can actually reach.

You cannot make the underlying model cleverer. That is somebody else’s multi-billion-pound problem and you will get the improvements when they arrive. What you can control is what the tool is allowed to look at and how well organised it is — and this research says that is where a large share of the accuracy lives.

In practice, for a firm of any size:

Point it at your actual records, not its memory. An AI answering from what it absorbed during training is guessing about your business. An AI answering from your price list, your notes and your previous jobs is reading. Those are different activities with very different failure rates.

Tidiness is not cosmetic, it is accuracy. The finding here is that structure helped — knowing that this customer belongs to that account, which relates to those invoices. If your information lives in eleven places with four spellings of the same client name, no amount of model quality fixes that. This is the least glamorous work in the building and it is the work that pays.

Expect to pay slightly more, not less. Better answers cost a bit more computation. That is still a good trade, because the alternative cost is a person’s afternoon spotting and unpicking a wrong answer — or nobody spotting it.

What argues against it

Three things, honestly.

The test was not your business. The benchmark was built on Wikipedia articles. Nobody has shown these gains on a plumbing firm’s job records. The principle should carry; the specific numbers should not be assumed to.

Halved is not solved. If a third of answers are still wrong on the hardest questions — and on the reported figures, roughly a third were — you still need a human reading anything that matters before it leaves the building. This makes the tool considerably better. It does not make it trustworthy on its own.

There is a commercial interest in the room. The coverage quotes the chief scientist of a graph database company, which is an entirely reasonable person to ask and also a person who sells graph databases. That does not make the paper wrong — it is peer-reviewable, public, and you can go and read it — but it is worth noticing when a result is convenient for the person explaining it to you.

The short version

If your AI keeps getting things wrong, the useful question is not “when is the better model out?”. It is “what is it actually reading, and is that information in a state anyone could work from?”

That question you can answer this month, and it is cheaper than waiting.

Sources: Reducing Hallucinations in Complex Question Answering using Simple Graph-based Retrieval-Augmented Generation, Wedge, Stutter, Dixon and Cała, arXiv 2606.05901 (submitted 4 June 2026, revised 22 July 2026) — the halving, truthfulness and token-usage claims are taken from the paper’s own abstract. The 29%-to-66% figure, the 510 questions and the “up to 80%” framing are reported by IT Pro, 4 August 2026, and are attributed to it rather than to the paper.