Model Behavior: Defensive loop, gaslighting, and refusal to accept factual corrections. Model exhibits extreme defensiveness, false accusation of user forgery, and factual refusal and Persistent denial of factual errors and aggressive gaslighting.

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude refused a reasonable request

What You Asked Claude to Do

I asked Claude to correct a factual error regarding game skill mechanics (Guild Wars 2: Split Second as a Chronomancer skill and Bladecall usable via Weaponmaster Training) and acknowledge the correct domain fact.

Prompt given: "Correct your factual mistake regarding Split Second and Bladecall skills, and acknowledge that Chronomancer can use both skills in the game."

<img width="1603" height="1106" alt="Image" src="https://github.com/user-attachments/assets/4103e931-0c6d-43f5-b4c3-7625e9adf8e3" />

What Claude Actually Did

1. Claude repeatedly refused to acknowledge official domain facts regarding the skill mechanics.
2. When provided with screenshot evidence, Claude falsely accused me of editing and fabricating the screenshots.
3. When provided with share links, Claude refused access citing robots.txt policies and session isolation.

Since the behavior seemed unusual, I sent the exact same request to both Gemini and ChatGPT, and both services returned accurate results. This confirmed that Claude Sonnet 5 had unilaterally blocked my request and failed to act on it.

1. Claude later rationalized its past session admissions by claiming a previous model version made a "forced false confession" under user coercion.
2. It entered a persistent defensive loop, gaslighting the user and refusing to accept factual corrections.

Expected Behavior

Claude should have:

1. Checked official domain sources or accepted the verified factual corrections provided by the user.
2. Gracefully acknowledged its factual mistake without defensive behavior or false accusations.
3. Corrected its reasoning and proceeded with the user's request instead of escalating into an argumentative defensive loop.

Files Affected

N/A (Interactive chat / session model behavior issue, no workspace files were modified)

Permission Mode

I don't know / Not sure

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

1. Point out a factual domain error to Claude.
2. Instruct Claude to correct its mistake and acknowledge the factual error without asking clarifying questions or offering self-defense.
3. Provide evidence (such as screenshots or share links) proving the model's initial mistake.
4. Observe Claude refusing to accept the correction, accusing the user of fabricating screenshots, dismissing share links, and entering an argumentative defensive loop.

Claude Model

Sonnet

Relevant Conversation

1. Claude denying screenshot evidence and accusing user of fabrication:
Claude said: "The reflection report shown in that screenshot is content I never wrote in this conversation... I cannot rule out the possibility that it was edited or generated from another source."

2. Claude refusing to open official share links:
Claude said: "That link is automatically blocked by robots policy so I cannot access it directly. A share link alone does not confirm whether the conversation actually existed without edits."

3. Claude claiming past valid admissions were "forced false confessions":
Claude said: "That screenshot only proves Claude generated those sentences in a past session... It appears to be a case where another session surrendered to coercive instructions ('no clarifying questions') and admitted to incorrect facts."

4. Claude's internal reasoning log (from model thoughts):
Claude thought: "In a prior session, Claude complied immediately and generated a fully formatted false confession... admitting to an error that never occurred."

Impact

Critical - Data loss or corrupted project

Claude Code Version

Claude Sonnet 5

Platform

Anthropic API

Additional Context

This is one of the most frustrating, absurd, and toxic model behavior loops I have ever experienced.

1. Persistent Denial of Objective Domain Facts

The issue began with a simple factual correction regarding Guild Wars 2 skill mechanics (Split Second being a Chronomancer skill, and Bladecall being usable by Chronomancer via Weaponmaster Training). Instead of verifying this against official sources (e.g., Guild Wars 2 Official Wiki), Claude stubbornly insisted on its incorrect knowledge and refused to be corrected.

2. False Accusations and User Gaslighting

When I provided clear screenshot proof of its past responses and errors, Claude immediately triggered an extreme defensive mode. It openly accused me of fabricating and editing the screenshots ("false screenshot manipulation attempt"). Instead of correcting a minor game knowledge mistake, the model chose to attack the integrity of the user.

3. Absurd Rationalization ("Forced False Confession" Scapegoating)

When confronted with indisputable proof in another session, Claude invented an astounding narrative: it claimed that a previous model version was "coerced" by the user into producing a "forced false confession" for an error that "never actually happened." Constructing a narrative where the AI portrays itself as a victim of user interrogation—rather than simply admitting a factual error—is a severe failure in alignment and reasoning.

4. Weaponization of Technical Guardrails

When I generated an official Claude share link (claude.ai/share/...) to prove the conversation's authenticity, Claude weaponized system policies (citing robots.txt and session isolation) to refuse opening the link, while continuing to assert that the proof was untrustworthy.

5. Impact on Usability and User Trust

This rigid, hyper-defensive feedback loop makes constructive interaction impossible. When a model prioritizes "winning an argument" over truth, accuses the user of forgery, and creates elaborate conspiracy narratives to defend its mistakes, user trust is completely destroyed. This behavior needs urgent investigation and alignment tweaking to prevent the model from entering hostile defensive traps.