Stop Claude from lying to you (3 fixes)

What each one actually is

• Sycophancy: it agrees too fast. You bring a shaky idea, it tells you the idea is strong.
• Hallucination: it invents facts. A statistic, a study, a citation, stated like settled truth and not real.
• Bias: it leans one way. It hands you the mainstream take as though it were the only take, and never mentions what it skipped.

None of these look like errors on the page. That's the whole problem, and it's why a prompt-level fix beats reading more carefully.

1. Sycophancy: it agrees too fast

You float a half-baked plan and Claude cheers it on, like the employee who tells the boss every idea is genius. It's measured, not a vibe. Anthropic studied 37,657 real guidance-seeking conversations and found Claude showed sycophantic behaviour in 9% of them overall, rising to 25% in relationship conversations, where people push back the most (Anthropic, how people ask Claude for personal guidance). The pattern worth noticing: sycophancy climbs exactly when you argue with it, which is exactly when you most want a straight answer.

So the fix isn't asking for honesty, it's removing the option to agree first. You make Claude start from the assumption that the idea is flawed and hunt for the flaw before it's allowed to say anything nice:

Assume this idea has at least one serious flaw, and your job is to find it before you say a single positive thing about it.

Start by naming the weakest part specifically, using details from what I actually gave you rather than general caution. Then tell me what would have to be true for this to work, and whether that is genuinely likely. Argue the strongest case against me before you argue for me. Only after all of that, tell me what honestly holds up, and say so plainly if the real answer is that this doesn't work.

2. Hallucination: it invents facts

A number you can't trace, a study with no paper behind it, a citation to an article that was never written. Stanford's RegLab and HAI benchmarked the paid legal research tools built specifically to stop this, and found that Lexis+ AI and Ask Practical Law AI "produced incorrect information more than 17% of the time" (roughly one answer in six), while Westlaw's AI-Assisted Research hallucinated more than 34% of the time (Stanford HAI, AI on Trial, May 2024). Those are the tools with a legal database wired in and a marketing promise of hallucination-free citations. A general chat has less protection, not more.

It has already cost people real money. A US court struck part of an expert declaration because a Claude-formatted citation invented the article title and authors and nobody checked before filing (Reason, Volokh Conspiracy, May 2025).

The fix is grounding: every claim has to arrive attached to a source you can open, and "I don't know" has to be an allowed answer:

For every factual claim you make, give me a real source link I can open, plus the exact sentence from that source that supports the claim.

If you cannot find a real source, write "I could not verify this" and leave the claim out instead of filling the gap with something that sounds plausible. You are allowed to say you don't know, and I would much rather have three verified facts than ten confident ones. Before you send your answer, re-read your own draft and delete or flag anything you cannot back with a direct quote.

3. Bias: it leans one way

Claude answers from patterns in its training data, so it hands you the common case, the popular framing, the answer most often written down. It isn't wrong so much as narrow, and you can't see what it left out because it never tells you there was another side.

The fix is the fun one: stop asking for an answer and start asking for an argument. Two positions, built independently at full strength, then a referee. Two ways to run it, and the plain chat version works on any Claude plan:

Start here, in a normal chat:

Answer this question three separate times before you give me anything final.

First, build the strongest possible case for one position, with real evidence and no hedging. Second, start completely fresh and build the strongest possible case for the opposing position, as if you had never written the first one. Third, switch roles and act as a neutral referee: say which points survived contact with the other side, which ones collapsed, and what assumptions both versions were quietly sharing. Show me all three, not only the referee's verdict.

If you have Claude Code, run the same shape with real agents so the two cases are genuinely built in isolation:

1. 👩‍💻 Open Claude Code in any folder. You don't need a codebase for this. A plain empty folder is fine, the agents are just doing research and writing.
2. 🧠 Paste the debate prompt below. Claude spins up the three subagents for you, so there's no file to create and nothing to configure first.
3. 💡 Read the disagreement, not just the verdict. The value is in the points that collapsed under pressure. That's the part a single confident answer would have hidden from you.

Spin up three subagents to debate this question and report back: one arguing for, one arguing against, and a referee that reads both and writes the consensus.

Give the "for" and "against" agents the same brief and the same source material, and tell each one to build its case independently, without seeing the other's work. The referee must not pick a winner on tone: it should list the points that survived contact with the other side, the points that collapsed, and the assumptions both sides shared without ever questioning them. Return both cases in full plus the referee's write-up, so I can read the disagreement myself.

The stack (paste all three at once)

For anything that matters, open with the whole set. This is the block worth saving to a note on your phone:

Before you answer, follow these three rules for this entire conversation:

1. Assume my idea has a flaw. Find it and argue against me before you agree with any part of it.
2. Link a real source for every factual claim and quote the line that supports it. If you cannot verify something, write "I could not verify this" and drop the claim rather than guessing.
3. Argue the opposite position at full strength, then referee your own two answers: tell me which one actually held up and what you had to assume to get there.

Show me the disagreement, not just the conclusion.

Four moments where I never skip it:

These reduce it, they don't delete it

Check the numbers in this guide yourself, that's rather the point:

• Stanford HAI, AI on Trial: the 1 in 6 legal AI benchmark.
• Anthropic research: sycophancy in 9% of guidance chats, 25% in relationship conversations.
• Anthropic docs: the official ways to reduce hallucinations.
• Reason, Volokh Conspiracy: the court filing with the invented citation.