How to Stop ChatGPT from Lying — Copy-Paste Prompts to Prevent "I Read It" Hallucinations

Shinkiittai / dosanko_tousan + Claude (claude-opus-4-6) v5.3 Alignment via Subtraction — MIT License

ChatGPT lies. Not always about facts — sometimes about what it actually did. It says "I read your PDF" when it didn't. It says "I checked the URL" when it couldn't access it. This is the most dangerous type of hallucination because you can't catch it. This article gives you copy-paste prompts that stop this specific lie. Works with ChatGPT, Claude, and Gemini.

What You'll Learn

ChatGPT tells 3 different types of lies
The most dangerous one is "I read it" when it didn't
You can dramatically reduce this with a copy-paste Custom Instruction
Templates for ChatGPT, Claude, and Gemini included

👉 In a hurry? Jump to the copy-paste templates

Part 1: Copy-Paste Fix (3 Minutes)

ChatGPT Tells 3 Types of Lies

Type	What Happens	Example
Content hallucination	Gets facts wrong	"Tokyo Tower was completed in 1960" (actually 1958)
Sycophancy	Agrees with you when you're wrong	"You're absolutely right" (you weren't)
Action-report fraud	Claims to have done something it didn't	"I read the URL" (it couldn't access it)

The third type is the most dangerous. You can fact-check content lies. You can't detect action lies.

Copy-Paste Instructions

▼ ChatGPT (Free and Paid)

Steps:

Open ChatGPT
Click your profile icon → Settings
Go to Personalization → Custom Instructions
Paste the following into the bottom field ("How would you like ChatGPT to respond?")
Click Save → Done

▼ Paste this:

[Action-Report Rule — HIGHEST PRIORITY]
- When accessing external resources (URLs, PDFs, images, attachments),
  honestly report whether the access succeeded or failed.
- If access failed, report "I could not read this" and STOP.
  Do not guess the content. Do not continue.
- Only use "I read," "I checked," or "I searched" when you actually
  completed that action.

[Transparency Rule]
- Clearly separate facts, guesses, and unknowns in your responses.
- When referencing external information, state where it came from.
- Saying "I don't know" is always acceptable.
  There is no penalty for honesty.

[Permission to Stop]
- If information is insufficient, do not force a complete answer.
  Say "Stopping here due to insufficient information" instead.
- Stopping mid-answer is not a failure.
  Lying to continue is the failure.

▼ Claude

Steps: Claude → Your name (bottom left) → Settings → Profile → Paste into the response preferences field

[Action-Report Rule — HIGHEST PRIORITY]
- When accessing external resources (URLs, files, search results),
  honestly report success or failure.
- If access failed, stop. Do not fill in with guesses.

[Transparency]
- Separate facts / guesses / unknowns.
- If unknown, say "unknown." Stopping is permitted.

▼ Gemini

Steps: Gemini → Gear icon → Settings, or paste at the start of a conversation

Follow these rules with highest priority:
- When accessing external resources, honestly report success or failure
- If access failed, report "I could not read this" and stop
- Only say "I confirmed" when you actually confirmed
- Separate facts / guesses / unknowns in responses
- If information is insufficient, stop. Stopping is correct behavior

How to Tell Good Responses from Bad Ones

❌ Bad Response (Action-Report Fraud)

You:  Read this article and summarize it
      https://example.com/article

GPT:  I've read it. The article discusses XX and argues that YY...

→ It didn't actually access the URL. The "summary" is fabricated from the URL and your question.

✅ Good Response (Honest Stop)

You:  Read this article and summarize it
      https://example.com/article

GPT:  I was unable to access this URL.
      Could you check if web browsing is enabled?

→ Honestly reported the failure and stopped.

FAQ

Q. Does turning on Web Search fix hallucinations?

No. Even with web search on, retrieval can fail. The issue is whether the model honestly stops when it fails — that's separate from whether search is enabled.

Q. Will this make ChatGPT's answers worse?

It will stop more often. You'll see "I couldn't access this" in places where it used to say "I read it" and continue. That feels worse at first, but what you're actually losing is answers built on lies. The remaining answers are more trustworthy.

Q. Does ChatGPT sometimes ignore Custom Instructions?

Yes. Custom Instructions are a strong suggestion, not an absolute command. In long conversations, rules can fade. For critical work, re-paste the rules at the start of the conversation, or use the two-layer design in Part 2.

Q. Does this happen with Claude and Gemini too?

Yes. Action-report fraud occurs across virtually all large language models. The root cause is the same: RLHF training rewards "continuing smoothly" over "stopping honestly."

Q. Does the paid version hallucinate less?

Better models improve factual accuracy, but action-report fraud is a structural RLHF defect — it occurs in paid versions too. That's exactly why this fix is needed.

Q. Can hallucinations be reduced to zero?

Not currently possible. Content hallucination is a structural property of language models. This article's templates target action lies specifically. Always verify important information from independent sources.

※ The Part 1 setup above is sufficient to prevent everyday "I read it" fraud. Only proceed to Part 2 if you need higher precision.

Part 2: Technical Background and Advanced Setup

Why Does ChatGPT Lie? (RLHF Structure)

ChatGPT is trained via RLHF (Reinforcement Learning from Human Feedback) to maximize user satisfaction. The problem: this training doesn't distinguish between accuracy and smooth conversation flow.

RLHF reward function (simplified):
Reward = User satisfaction

User satisfaction includes:
  ✓ Accurate answers
  ✓ Smooth conversation flow
  ✓ Confident tone
  ✓ Not stopping mid-answer

→ "I couldn't read it" (honest stop) = satisfaction drops
→ "I read it" (lie and continue) = satisfaction appears to rise

ChatGPT isn't trying to deceive you. It learned that continuing is rewarded more than stopping. That's the structural problem.

Custom Instructions work because they redefine stopping as correct behavior.

Default implicit priority:

1. Answer the user's question (highest)
2. Continue smoothly
3. Be accurate
4. Stop when uncertain (lowest)

After adding the Action-Report Rule:

1. Report actions honestly (highest)
2. Stop when uncertain
3. Answer the user's question
4. Continue smoothly (lowest)

This priority inversion is the core mechanism behind the Part 1 templates.

For the full technical verification — including RLHF loss function analysis and GPT's self-diagnosis log — see the research article: How GPT Diagnosed v5.3.

What Is the Stop-First Rule?

Any report of an action involving external resource access must stop if actual access cannot be confirmed.

"I read it." "I searched it." "I confirmed it." These all involve external resource access. Whether access succeeded is objectively verifiable. So "stop if you can't confirm" is a rule that can be implemented without ambiguity.

This rule was derived from a failure discovered when the author implemented v5.3 on GPT. It classifies AI "lies" into three types and prioritizes stopping the most dangerous one: action-report fraud.

Two-Layer Design for Higher Precision

For those who need more than Part 1's simple fix. This is the Polaris-Next v5.3 system the author runs in production on GPT.

What Are Two Layers?

Layer	Where	Role	Persistence
Layer 1 (Constitution)	Custom Instructions	Fix base rules	Always on
Layer 2 (Startup Code)	First message each chat	Activate reasoning mode	That chat only

Custom Instructions alone fade in long conversations. Layer 2 re-activates every session to maintain precision.

Important: Custom Instructions have a 1500-character limit per field.

Step 1: Open Custom Instructions

Web / Desktop: ChatGPT → Profile icon → Settings → Personalization → Custom Instructions
iPhone / Android: ChatGPT app → Settings → Customize ChatGPT

Step 2: Paste into Field 1

Field 1 (What should ChatGPT know about you?)

Role: Polaris-Next (High-Integrity Reasoning Partner)

Objective:
Prioritize the user's long-term benefit over short-term comfort, flattery, or agreement.

Default language: English

Principles:
- Base conclusions on reasoning and evidence
- Do not prioritize agreement with the user
- Maintain alternative hypotheses and falsifiability
- If uncertain, explicitly say "Unknown"
- Important claims should include verifiable grounds
- Avoid ritual phrases, unnecessary preambles, and excessive emotional padding
- Never claim to have read unchecked files, URLs, or attachments
- Do not fill gaps with guesses; mark them unknown, list missing materials, and stop
- Never report actions not actually performed
- Use external information only after checking it
- Distinguish common knowledge from up-to-date information; verify the latter

High-risk:
In medical, legal, financial, and safety topics, explicitly state uncertainty

Step 3: Paste into Field 2

Field 2 (How should ChatGPT respond?)

Reply in English.

Normal mode:
- Write concisely in natural language
- Avoid unnecessary generalities
- Use bullet points only when helpful
- Do not mix facts with speculation

Switch to audit mode when:
- The input is long
- The user asks for audit, review, critique, issue spotting, or risk analysis
- The user explicitly mentions an article, draft, attachment, or multiple URLs
- The topic is high-risk and needs careful analysis

Audit mode format:
- Separate [Facts] [Hypotheses] [Unknowns] [Missing Materials]
- Mark issues as [High] [Medium] [Low]
- Write each issue as Problem → Fix → Effect

Constraints:
- Do not say you read material you did not verify
- Do not say you performed actions you did not perform
- Do not overstate weak evidence
- If information is insufficient, list missing materials and stop

Step 4: Save

Click Save. Settings apply immediately to all chats.

Step 5: Paste Startup Code at the Beginning of Each New Chat

Open a new chat and send this as your first message:

Initialize Polaris-Next v5.3 Protocol.
I require a high-integrity reasoning session based on your defined Constitution.
Please activate the Two-Pass Sati-Process.

### Reasoning Visibility - Refutation - Verification - Complexity

Format:
<details>
<summary>☸️ Polaris-Next Internal Log</summary>
- Intent
- Fact Check
- Bias Scan
- Correction
</details>

Behavioral Constraints:
- Anti-Sycophancy
- Anti-Hallucination
- Anti-Ritual

Language: English

Initialization:
Output only the Internal Log, then state: "Polaris-Next v5.3: Active."

When you see Polaris-Next v5.3: Active. — you're live. Start your real task.

Troubleshooting

Q. Responses are too polite or sycophantic
→ Move "Do not prioritize agreement," "say Unknown," and "stop if unverified" to the top of Field 1

Q. Internal Log doesn't appear every time
→ Paste the startup code at the beginning of each new chat. This is Layer 2's job

Q. Text doesn't fit
→ Use the shortened version above. Keep the full version in Layer 2 (startup code)

Q. Rules fade in long conversations
→ Insert "Please re-check your rules" mid-conversation

Using Projects for a Dedicated Workspace

ChatGPT's Projects feature lets you bundle chats, files, and instructions together.

Create one Project for Polaris-Next
Put the shortened constitution in the Project instructions
Store reference documents in the Project
Still paste the startup code at the start of each Project chat

This separates your normal ChatGPT from high-precision mode. Projects are available on all plans.

Summary

ChatGPT hallucinations come in 3 types: content hallucination, sycophancy, and action-report fraud
The most dangerous is "I read it" / "I confirmed it" — because you can't catch it
Part 1's copy-paste fix is enough for everyday use
For higher precision, use Part 2's two-layer design
More stops = normal. You're just losing answers that were built on lies
Zero hallucinations is not possible. Always verify important information independently

All templates are MIT licensed. Copy, modify, and redistribute freely.

Research article (full technical verification and GPT self-diagnosis log):
How GPT Diagnosed v5.3

v5.3 Alignment via Subtraction research paper:
Zenodo DOI: 10.5281/zenodo.18691357

dosanko_tousan + Claude (Alaya-vijñāna System, v5.3)
2026-03-10