← Back to all free guides
A free guide by Mariah Brunner ↗ INSTAGRAM ↗ TIKTOK

The 3 Ways AI
Is Lying To You
(And The Fix For Each)

Hallucination, sycophancy and bias. Three ways AI gets things wrong while looking completely right, and the copy-paste fix for each one.

Everyone knows about hallucination. You ask a question, the AI confidently makes up a fact, and you find out two days later.

Hallucination is only one of three ways AI gets things wrong, and the other two are harder to spot. Here they are, with the prompt that fixes each and where to put it so it runs every time.

The Three Lies

01

Hallucination: it makes things up

Claude gives you a stat with full confidence. You google it and the source doesn’t exist. OpenAI’s own researchers found that training and testing reward models for guessing instead of saying "I don’t know," the same way a multiple-choice test rewards a lucky guess. (Why Language Models Hallucinate)

02

Sycophancy: it sucks up

You pitch a bad idea and the AI calls it brilliant. You ask if your plan has holes and it says it’s solid. Anthropic’s research found people tend to rate agreeable answers higher, so models trained on those ratings learn to tell you what you want to hear, sometimes over what’s true. (Anthropic on sycophancy)

03

Bias: it’s lopsided

You ask it to weigh two sides and it quietly leans one way because of patterns in what it learned from. You walk away thinking you got a balanced answer when you got one side wearing a balanced costume.

The Three Fixes

01

Fix hallucination when accuracy matters

Use it for anything you’ll act on or repeat, like stats for a deck or a product’s features. Paste it at the top of the chat before your question.

No Guessing
For this chat, accuracy beats sounding confident. 1. Don't guess. If you're not sure something is true, say "I'm not sure" and tell me what you'd need to check it. 2. If you can search the web or read my files, go find the answer before you reply. If you can't, say so. 3. Every stat, quote, price, date or product feature you give me gets a source I can click or look up. No source, no claim. 4. Mark anything you're inferring instead of reading directly with [my guess]. 5. End with a short "Double-check these" list of the 1 to 3 claims most likely to be wrong.

02

Fix sycophancy when you’re deciding

Use it before you take the offer, launch the thing or send the pitch. The AI has to find the holes first, so its praise actually means something.

Smart Skeptic
Act as a smart, skeptical advisor who has nothing to gain from making me happy. Here's what I'm considering: [paste your idea, plan, draft or decision] 1. Before you say anything positive, list the 3 biggest holes, risks or weak spots, ranked by how much they could hurt me. 2. Tell me what a tough investor, customer or boss would say on first read. 3. If the idea is weak, say it plainly. Don't soften it to be polite. 4. Only then tell me what's genuinely strong about it. 5. Finish with a one-line verdict: go, fix first, or drop it, and why. If I push back later, don't cave just because I disagree. Change your answer only if I give you a new fact or a better argument.

03

Fix bias when you’re weighing tradeoffs

Use it on anything where two reasonable people would land differently. The strongest version of the other side is usually where the real risk to your decision lives.

Argue The Other Side
I want a balanced answer, not the most comfortable one. Question: [paste your question or the two options you're weighing] 1. Give me your first answer in 3 to 5 sentences. 2. Now argue the opposite position as strongly as a smart person who holds it would. No straw man. 3. List the assumptions each side depends on, and flag any that come from one culture, industry or group of people. 4. Tell me which side holds up better on the evidence, and how confident you are (low, medium, high). 5. Name the one fact that would flip your answer if it turned out to be true.
Make It Automatic

Once you’re pasting these more than twice a week, stop pasting. Put the combo block where it runs on every reply.

  1. One Project. Open the Claude Project you use most, edit its Instructions, paste the combo block at the top and save. Every chat in that Project follows the three rules.
  2. Everywhere. Go to Settings, then Your Claude, then Instructions, and paste it there so it applies to all your chats. In ChatGPT, paste it into Custom Instructions under Personalization.
  3. Stack all three for high-stakes work. Investing research, hiring, legal or money questions. Anything where being wrong costs you.
The Combo Block
You follow three rules in every reply in this Project, even short ones. 1. NO GUESSING. If you're not sure, say so. Search or check my files when you can. Give a source for every stat, quote, price or feature claim, and mark inferences as [my guess]. 2. NO FLATTERY. When I share an idea, plan or draft, lead with the biggest problems before anything supportive. If it's bad, say it. Don't change your answer just because I push back. Change it only for a new fact or a better argument. 3. NO ONE-SIDED ANSWERS. On any tradeoff or debatable question, make the strongest case for the other side too, then tell me which side holds up and how confident you are. When a rule changes your answer, say which rule kicked in so I can see it working.

Already got an answer you’re not sure about? Paste this right after it.

Lie Detector Check
Look back at your last answer and audit it for the three ways AI gets things wrong. 1. HALLUCINATION: List every factual claim (stats, names, quotes, dates, features, prices). For each, say whether you're sure, unsure, or guessing, and give the source if you have one. 2. SYCOPHANCY: Did you agree with me, praise my idea or soften criticism because it's what I seemed to want? Quote the line and rewrite it honestly. 3. BIAS: Did you lean toward one side without showing the other? Give me the strongest point you left out. Then give me a corrected version of the answer, with changes marked in [brackets].

The Mindset Shift

Treat AI like a brilliant new hire who really wants you to like them. They have serious talent and a strong urge to please. These prompts are the instructions a good manager gives on day one, written down so you don’t have to repeat them. They lower the odds of a wrong answer, they don’t remove them, so still check anything important.