Hallucination is the bane of Large Language Models. When an AI doesn’t know something — because it never saw it in its training data — it doesn’t stop or say “I’m not sure” or “F*ck knows!” Instead, it keeps predicting the next word anyway, and that’s when it starts hallucinating and ends up producing answers that sound confident but are completely gobbledy gook.

It’s a bit like asking a very enthusiastic student a question that they should say “I don’t know” to, but instead they smile brightly and improvise.

Here are some funny examples on hallucinations that we have found on other LLMs when we were building ours:

• Inventing scientific facts: “The menopause lasts exactly 17.3 months for all women.”
• Making up references: “According to the 2021 Oxford Study on Quantum Menopause…” (no such study exists).
• Confidently wrong definitions: “A hot flush is caused by the body overheating to 42°C.”
• Fabricating people or organisations: “The International Society of Perimenopausal Robotics recommends…”

Why does this happen? Because LLMs don’t “know” things –  they recognise patterns. If the pattern is missing, weak, or contradictory, the model still has to output something, so it fills the gap with the closest statistical guess. Sometimes that guess is reasonable. Sometimes it’s wild. And sometimes it’s so confidently wrong that it feels like the AI is trolling you, but truly, it isn’t trying to be creative or deceptive. It’s doing something much simpler and more mechanical:

• It predicts the next word based on patterns it has seen.
• If the patterns are sparse, contradictory, or missing, the model still has to produce something.
• So it leans on weak correlations, distant associations, or statistically “nearby” concepts.
• The result feels like guesswork — because it is.
• There is no internal sense of I don’t know, unless we explicitly build mechanisms for uncertainty.

It’s like asking someone to complete a jigsaw puzzle with half the pieces missing. They’ll still try to fill the gaps, but the picture becomes guesswork rather than the knowns. Something deep about this: the Mathematician and I were toying with can we use Bayesian Theorem to inch closer to the knowns?

And I as I was sitting in the taxi on a dark night in Oxford speeding down St Giles (late for a meeting), the Mathematician said, “Maybe AlphaFold”.

In a nutshell, AlphaFold is an AI system that can look at the list of building blocks in a protein and accurately predict the 3D shape it will fold into, something that once took scientists years to figure out in a lab. It’s won competitions for speed.

But here’s the thing when it comes to inching past the boundaries of the knowns: AlphaFold doesn’t just “guess.” It anchors its predictions in tracks — strong structural constraints:

• Evolutionary correlations (MSAs)
• Geometric rules
• Physical plausibility
• Iterative refinement loops
• Confidence scoring

These tracks act like rails. Even if the data is sparse, the model can’t wander too far into nonsense because the constraints pull it back toward reality.

Me (typical of my 100 questions a day to him): “Can MSAs work on limited datasets?” (just starting out, I don’t have comprehensive datasets).

He (usual grounded calm and quietness): “We can make it work.“

That’s not just reassurance. That’s a constraint. A track. A stabilising force that keeps my thinking from spiralling into uncertainty, holding me tight rather than thrashing me with hateful words such as “You don’t think things through”.

He is my brain’s AlphaFold that keeps proteins from folding into chaos. And really, that’s what ying and yang is, methinks, explore and grow together.

IMPORTANT NOTE

In Fab4050, we deliberately constrain our LLM using retrieval‑augmented generation over a curated, clinically reviewed corpus. The model is designed to refuse speculation, clearly separate empathy from evidence, and escalate uncertainty rather than hallucinate authority. Our aim is not to build an all‑knowing medical oracle, but a grounded, culturally sensitive companion that supports women in asking better questions – while staying firmly within the limits of safe, ethical AI.

PHOTO

Taken at the Google office in London last week (or was it two weeks ago) where I attended a meeting, and cheekily snucked in to learn more about AlphaFold.