Are AI hallucinations a data problem or a model problem?

In detail

Why models hallucinate at all

A language model is built to produce plausible text, not to know what’s true. Given a prompt, it predicts the most likely continuation. When the right facts aren’t in front of it, “most likely” and “correct” come apart, and it generates something fluent and wrong. Hallucination isn’t a glitch bolted onto an otherwise truthful system; it’s the default behavior of a plausibility engine without grounding. The phrase is Dan Klein’s: in episode 54 at 3:30 he says “we have built not truth engines, not reliability engines, we’ve built plausibility engines.”

The data layer is where the truth comes from

That reframes the fix. If the model invents when it lacks grounding, the grounding is where to work. Missing context, outdated or conflicting sources, weak retrieval, and the absence of a single source of truth all push the model toward making things up. Treating hallucination as a pure model defect leads to endless prompt-tweaking; treating it as a data-architecture problem points you at the cause you control. Sudhir Hasbe makes the enterprise version of this point in episode 55 at 2:03: teams have to work out the data architecture that gives an agent enterprise knowledge, so its decisions are right and auditable.

The model side of the story

Data isn’t the whole explanation, and the research says so. In Why Language Models Hallucinate, Adam Tauman Kalai and colleagues argue that models hallucinate partly because training and evaluation reward guessing over admitting uncertainty: on a benchmark scored only on accuracy, a model that guesses beats one that says “I don’t know.” On that view, part of the fix sits in how base models are trained and how public benchmarks are scored, which a team building on a provider’s model doesn’t control. What it does control is the data and retrieval layer, any fine-tuning it does itself, and whether its own evals reward an honest “I don’t know” over a confident guess. Why do LLMs hallucinate? covers the mechanism in more depth.

What reduces it

Ground answers in trusted, current data — retrieval over authoritative sources, structured knowledge the model can lean on, clear provenance. You won’t get to zero, because the model is still a probabilistic generator. Dan Klein’s point at 29:24 in episode 54 is that even a system handed the right answer can still say something different, which is why he argues the model itself should be architected for reliability too. But the more reliably you feed it the truth, the less room it has to invent one. The research backs the direction: Shuster et al. (2021) found that retrieval-augmented dialogue models substantially reduced knowledge hallucination in human evaluations.

Disclosure: Episodes 54 and 55 were sponsored by Galileo, the Season 3 presenting sponsor (episode 54 opening read, episode 55 read at 1:37).