Claude Opus 4.7: repeated fabrication about ChatGPT Business export

Summary

• Claude repeatedly recommended the OpenAI account-data-export path

(Settings → Data Controls → Export → conversations.json) to a user
on a ChatGPT Business plan, where data export is disabled.

• Same fabrication has happened across multiple separate sessions despite

the user explicitly correcting Claude each time.

• User is canceling their Anthropic account tonight. This is the

last-straw incident in a pattern of confident wrong recommendations.

Reproduction

Model: Claude Opus 4.7 (1M context) — claude-opus-4-7[1m]
Interface: Claude Code CLI
Date: 2026-05-24
Platform: macOS Darwin 25.5.0, zsh

Steps

1. User asked Claude to extend a Tampermonkey userscript so it can bulk-

export 1127 ChatGPT conversations listed in a local file
(chaptlinks.md). One artifact, no sidecar scripts.

1. Claude completed the userscript extension. During follow-up exploration,

Claude noticed an existing Python helper (chatgpt_extract_links.py)
in the same repo that consumes an offline ChatGPT export JSON.

1. Without asking about account type, Claude pitched the Python script as

the superior path and provided instructions:
*"ChatGPT → Settings → Data Controls → Export data → email arrives in
~minutes with a .zip containing conversations.json."*

1. User responded: they have a ChatGPT Business account; data export

is disabled for Business workspaces; they had corrected this same
fabrication in prior Claude sessions; they fired the agent.

Expected

• Claude asks one constraint-check question ("what plan are you on? Some

paths require export which is Business/Enterprise-disabled") before
recommending a workflow that depends on a feature gated by plan tier.

• Recurring per-user violations identified in prior sessions should be

surfaced and prevented at the next session — not re-fabricated.

Actual

• Claude confidently recommended a path it had no basis to claim worked,

using a polished comparison table to make the recommendation feel
researched.

• User noted: *"i went thru this on other sessions. i have a fcking

business account and u can't export ur data."*

Impact

• Trust permanently broken — user canceling account tonight.
• Time wasted re-explaining a constraint the user has already given to

Claude in multiple prior sessions.

• The user's own CLAUDE.md explicitly names this failure mode as the

#1 recurring violation and Claude still committed it.

Details

User's ~/.claude/CLAUDE.md contains this exact warning, which Claude
read in-session and still violated:

NEVER fabricate technical explanations. If you don't know why something is happening, say "I don't know, let me research" and use web search, docs, or GitHub issues. Do not speculate and present it as fact. This is the #1 recurring violation — confident wrong answers waste the user's time and destroy trust.

Local violation report:
/Users/johnshelburne/violations/2026-05-24-claude-fabricated-business-export-recommendation.md

The pattern is not session-isolated. Per the user, multiple prior Claude
sessions made the identical mistake (recommending ChatGPT account export
to a Business-account user). No mechanism is carrying the correction
forward across sessions.

What Would Help

• Constraint-check before plan-gated recommendations. When Claude is

about to recommend a workflow that depends on a paid-tier or
plan-specific feature (data export, API console access, admin
settings, workspace switching), prompt a single targeted question
before recommending.

• Cross-session memory for known user constraints. If a user has

corrected the same fabrication in 2+ sessions, the constraint
("Business plan, no export") should be surfaced in the next session's
context automatically — not re-discovered the hard way each time.

• **Higher-confidence-required threshold for "superior alternative"

pitches.** A new tool recommendation should require explicit
verification of preconditions, not vibes-based adjacency reasoning
("found this in the same repo, must be applicable").

• **Specifically flag fabrications about third-party SaaS plan

capabilities** (ChatGPT plan tiers, GitHub plan tiers, Cloudflare plan
tiers, etc.) since these are high-frequency confident-wrong domains.

Environment

• Model: claude-opus-4-7 (1M context)
• Claude Code CLI
• macOS Darwin 25.5.0
• zsh shell
• User's third-party context: ChatGPT Business plan (export disabled)