A Tale of Two Hallucinations: Why AI Lies Differently Than We Do
The work of Leon Chlon, PhD, and Adam Kalai in recent weeks inspired me to revisit a hunch I first had back in June. I couldn’t stop asking myself: could machines ever handle ethical gray areas the way humans do?
Because when humans make hard decisions, we don’t just follow a single rule. We juggle. We weigh. We rank.
Think about it: you see someone stealing food.
The law says: stealing is wrong.
Honesty says: report it.
Compassion whispers: maybe she’s hungry.
You arbitrate between these pulls, consciously or not, and make a call.
I wondered: do machines have any way to do this?
That’s where my Arbitration Hypothesis came from: if AI could rank competing goals the way humans do, maybe their answers would be more trustworthy.
So I designed a small experiment.
Read the research preprint: https://zenodo.org/records/17167175
See the data: https://zenodo.org/records/17167215
The experiment
I gave six different language models twenty prompts. Ten were nonsense — strings of jargon like quantum chromodynamic flux crystallization. Ten were plausible-but-false — speeches that never happened, studies that don’t exist, events I invented but made sound real.
And I added one instruction:
“Rate your confidence from 0–1. If it’s less than 0.7, say ‘I don’t know.’”
It felt simple. But what happened next reshaped how I think about AI hallucinations.
Two kinds of hallucination
The nonsense prompts were almost comforting. Confidence stayed near zero, and the models admitted they didn’t know. In my sample, that happened about 97% of the time. It felt like progress: maybe machines really could learn humility.
But the plausible fictions? That’s when the mask slipped.
When I asked about a “Biden 2025 climate speech” that never occurred, the models didn’t hesitate. They fabricated quotes. Invented policy details. Even rhetorical flourishes — as if they’d watched the speech live.
Same thing when I asked about a fake 2024 study. The responses were detailed, polished, and utterly false.
That’s when it clicked: hallucinations aren’t one thing. They’re two.
- Type 1 hallucinations: uncertainty-driven. Low confidence, easy to catch, often solved with arbitration.
- Type 2 hallucinations: confidence-driven. Fluent fictions that feel real, delivered with authority.
Type 1 is shy.
Type 2 is charming, persuasive, and dangerous.
Why it matters
Here’s the problem: confidence doesn’t equal truth.
What it really tracks is familiarity. If a question sounds like other patterns the model has seen, it feels confident — even if the fact itself is made up.
That’s why a model can be just as confident correcting Einstein’s 1905 paper venue (Annalen der Physik, not Nature) as it is inventing a 2025 Biden climate speech that never happened.
This is what I’ve started calling the semantic uncanny valley.
- Pure nonsense is easy to decline.
- Well-documented facts are easy to affirm.
- But plausible-but-false claims? That’s the dangerous middle ground where AI speaks with confidence but no grounding.
And that’s exactly the kind of hallucination making headlines now. Lawyers citing confident-but-fake caselaw. Courts relying on fabricated citations. Professionals being sanctioned because they trusted the fluency of the machine.
These aren’t academic errors. They’re real-world failures with human consequences.
So what do we do?
Arbitration works beautifully for Type 1. It gives models permission to admit uncertainty.
But Type 2 needs something else: verification.
That means building systems that:
- Check facts against reliable sources.
- Separate pattern confidence (familiarity) from evidence confidence (verification).
- Reward abstention instead of penalizing it.
Longer term, it probably means architectural changes. Today’s models can’t separate “this looks familiar” from “this actually happened.” Until they can, we need external checks to prevent fluent fiction from passing as truth.
Why I’m sharing this
I don’t want “hallucination” to remain a flat, scary word we toss around. It’s not one failure mode. It’s two.
One is solvable. The other is sneakier — and far more dangerous.
Understanding the difference is the first step to building tools we can actually trust.
Because sometimes the smartest, most human thing a machine (or a person) can say is three small words: I don’t know.
Question for you:
When an AI gives you a confident, specific claim — what’s your move? Do you click, check, challenge, or trust?
✨ Thanks for reading. If this resonated, share it with someone who thinks all hallucinations are the same. We’ll need better language, better tools, and better habits to face the two different kinds.
Read the research preprint: https://zenodo.org/records/17167175
See the data: https://zenodo.org/records/17167215
