AI Hallucinations: a human problem with a human solution

The hallucination problem isn't really about hallucinations. It's about framing.

I keep noticing how strange it is that we talk about AI getting things wrong as if it's some unique flaw. We don't dismiss a junior for needing direction. We don't write off an advisor for getting something wrong. But with AI, it feels like we're over-indexing the hallucination problem, to the point where some people are using it as the validation of their opinion that AI isn't good.

I think the reason runs deeper than the technology. We've spent decades with computers being binary - type a calculation, get the right answer. The trust has always sat with the tool, not the user. So when AI behaves differently, our reflex is that the tool's failed.

We're using a human-shaped tool with a computer-shaped mental model. And we're surprised when it doesn't quite fit.

Why we keep blaming the tool

The reflex is everywhere once you start looking. Every time someone uses AI badly and gets a bad result, the headlines tend to blame AI rather than the person. Lawyers cited fake cases in court filings, got fined, and the story became "AI hallucinates" rather than "the lawyer didn't read what they were filing before submitting it." Nobody blames a typewriter for typos. With AI, somehow we do.

That's the bit that doesn't quite fit. We've trained ourselves on decades of software where the tool is reliable and the user is the variable. Excel doesn't guess. Calculators don't have opinions. If something's wrong it's a bug in the software and it's broken.

AI flips that. It produces something more like an opinion than an answer. Sometimes that opinion's brilliant, sometimes it's wrong, but it's never the kind of thing you should accept without thought. The cost of using it well is staying in the loop yourself - reading the output, applying judgement, owning the result. Most of the people getting tripped up have skipped that part, and then blamed the model for the mistake.

We're using the wrong category for it.

Stop thinking of AI as a computer

I think the mental shift is harder than the technical one.

Now before I say this I'm not suggesting that AI is conscious or that we should be treating AI as a living sentient being.

But I do think we should treat the output from an AI system in the same way we deal with a human's opinion, instead of how we treat a computer's output. Stop thinking of AI as a computer. Treat its output like a document from a human.

There are two different framings I often think about when working with AI. Sometimes I treat it like a junior - someone I brief on a task, expect to check the work, set up a halfway check-in for, build trust with over time. I don't expect them to nail it first time. I don't fire them when they need redirection. I give them the work, I review what comes back, and I build a working relationship around that.

Other times I treat it more like an advisor. Someone I ask for an opinion, or some expertise, or help thinking something through. I don't blindly accept what an advisor tells me. I also don't refuse to ask just because they're sometimes wrong. I take what's useful, discard what isn't, and apply my own judgement to land on something. People have always operated this way with the trusted people in their inner circles. Advisors give me input. I decide.

Either framing works. Both are better than the calculator one. The unifying point is that the judgement comes from you, not from the tool. Once you accept that, the question stops being "is this AI right?" and becomes "what's useful in what it's given me, and what do I want to do with it?"

That's a different question. It also happens to be the question you'd ask of any genuinely useful collaborator.

What this looks like in practice

Once you treat AI as a collaborator rather than a calculator, the question of how to work with it gets simpler. The answer is to use the same kind of guardrails you'd put around any work you care about, sized to the risk of the output.

For low-stakes stuff the guardrails can be light. For higher-stakes work, you build more in. A style guide. A tone-of-voice check. A list of words or phrases never to use. A fact-checking pass that requires sources for every number. A manual review of anything that's going public.

Here's a concrete example. When I go from a voice note to something written for me - an email, a proposal, a presentation, anything - I'll always ask the tool to check its own output against the original transcript afterwards, looking for anything that's wrong or wasn't actually in the source. That one extra step, applied to almost any voice-to-written workflow, takes the error rate from something like 5% down to 0.5%. It costs almost nothing. Almost no one does it.

That's the gap really. The people I know who've quietly figured this out are getting the required accuracy and getting all of that productivity gain. The people refusing to use the tool are getting nothing.

And I guess this is why I get frustrated with the hallucination pushback, with a small amount of effort you can mitigate the majority of the problem. It's not obvious because we're used to thinking about software in a certain way.

Rather than refuse to use it, build process around the tool you've actually got, not the one you wish you had.

The models might not get fixed

The model-makers themselves have stopped saying this is something they're about to fix. An OpenAI paper from September showed hallucinations come from how today's models are trained and how we evaluate them - structural features, not a bug in the code that's going to get fixed at some point. So if you're waiting for AI to stop getting things wrong before you'll use it, you're probably waiting for something that isn't coming.

The good news is you don't need to. The framing is the thing. Treat AI like a human collaborator from the start - sometimes a junior, sometimes an advisor - and the hallucination "problem" stops looking like a problem. It looks like a feature of working with a powerful, imperfect, highly capable tool that can increase the amount of output and impact you can have on your business. Which, if you've ever managed a team, is something you already know how to do.