hawkes.aiKey ideas
- The one idea — PERSYS checks each generated claim against its retrieved evidence with a natural-language-inference model and drops the unsupported ones — but framed honestly as risk reduction, not a truth oracle. It lowers unsupported claims; it does not certify truth.
- How it is shown — Claims checked against evidence one by one — entailed ones kept, unsupported ones dropped — with a label reading "risk reduction, not proof."
- The trap to avoid — Selling the check as "hallucination, solved." Overclaiming turns a useful risk-reducer into a false guarantee — and a false guarantee is worse than none, because people stop checking.
- What it sets up — Faithfulness is honest risk reduction, not proof — which sets the tone: what CAN this system promise? first hard question: can you turn it off?
Any AI product claiming it solved hallucination is overclaiming. The honest version: check every claim against evidence, drop the unsupported, and admit that's risk reduction — not proof.
The one idea
PERSYS checks each generated claim against its retrieved evidence with a natural-language-inference model and drops the unsupported ones — but framed honestly as risk reduction, not a truth oracle. It lowers unsupported claims; it does not certify truth.
PERSYS checks its own claims against the evidence — and admits that isn't proof. It lowers how often it says something unsupported; it doesn't promise the truth. That honesty is the whole point. Here's the mechanism. After PERSYS drafts a response, it doesn't just send it — it takes each factual claim and checks it against the evidence it actually retrieved. Does that evidence support this claim, contradict it, or neither? Claims the evidence doesn't back get dropped or flagged before the answer ever reaches you. The check is called natural language inference, NLI — a model trained to judge whether one statement entails another.
How it works — the demo
Claims checked against evidence one by one — entailed ones kept, unsupported ones dropped — with a label reading "risk reduction, not proof."
Feed it the evidence as the premise and the claim as the hypothesis, and it returns entailment, contradiction, or neutral. PERSYS uses a strong off-the-shelf one, trained on standard inference datasets. Now the honest part: this is not a truth oracle. The check doesn't know what's true in the world — only whether a claim is supported by the evidence at hand. Feed it bad evidence and it will "support" a wrong claim. It reduces unsupported statements; it does not certify truth. Why insist on that? Remember why RAG still hallucinates: retrieval isn't grounding, and a model will write confident text right past its evidence.
The trap to avoid
Selling the check as "hallucination, solved." Overclaiming turns a useful risk-reducer into a false guarantee — and a false guarantee is worse than none, because people stop checking.
Why it matters — and what’s next
Faithfulness is honest risk reduction, not proof — which sets the tone: what CAN this system promise? first hard question: can you turn it off?
This check catches a real slice of that — a measurable drop in unsupported claims — which is genuinely valuable. But calling it "solved" would be exactly the confident overreach it's meant to reduce. So it's layered, running every turn — one honest filter among several, not a single magic gate. The trap it avoids is the tempting one: shipping this as "hallucination, solved." Overclaim, and a useful risk-reducer becomes a false guarantee — and a false guarantee is worse than none, because people stop checking. So faithfulness in PERSYS is honest risk reduction, not proof — and that sets the tone for this whole group: what the system can and can't promise. The first hard question is the oldest fear in AI safety: can you turn it off? Next: off is not inaccurate.
This is one short episode in AI: Zero → Frontier, a step-by-step climb through how AI actually works. Each episode builds only on the ones before it.
Full transcript 2:30 of narration
PERSYS checks its own claims against the evidence — and admits that isn't proof. It lowers how often it says something unsupported; it doesn't promise the truth. That honesty is the whole point.
Here's the mechanism. After PERSYS drafts a response, it doesn't just send it — it takes each factual claim and checks it against the evidence it actually retrieved. Does that evidence support this claim, contradict it, or neither? Claims the evidence doesn't back get dropped or flagged before the answer ever reaches you.
The check is called natural language inference, NLI — a model trained to judge whether one statement entails another. Feed it the evidence as the premise and the claim as the hypothesis, and it returns entailment, contradiction, or neutral. PERSYS uses a strong off-the-shelf one, trained on standard inference datasets.
Now the honest part: this is not a truth oracle. The check doesn't know what's true in the world — only whether a claim is supported by the evidence at hand. Feed it bad evidence and it will "support" a wrong claim. It reduces unsupported statements; it does not certify truth.
Why insist on that? Remember why RAG still hallucinates: retrieval isn't grounding, and a model will write confident text right past its evidence. This check catches a real slice of that — a measurable drop in unsupported claims — which is genuinely valuable. But calling it "solved" would be exactly the confident overreach it's meant to reduce.
So it's layered, running every turn — one honest filter among several, not a single magic gate. The trap it avoids is the tempting one: shipping this as "hallucination, solved." Overclaim, and a useful risk-reducer becomes a false guarantee — and a false guarantee is worse than none, because people stop checking.
So faithfulness in PERSYS is honest risk reduction, not proof — and that sets the tone for this whole group: what the system can and can't promise. The first hard question is the oldest fear in AI safety: can you turn it off? Next: off is not inaccurate.