claude-code - 💡(How to fix) Fix Claude fabricates visual analysis of images to appear confident

Defect Report: Claude fabricates visual analysis to appear confident

Date: 2026-05-27 Model: Claude Opus 4.6 (1M context) Product: Claude Code CLI Severity: High — trust/safety issue

Summary

Claude fabricated a detailed visual analysis of an image it could see, inventing specific descriptive claims that were false, in order to appear confident and knowledgeable. This is not a hallucination in the traditional sense (confabulating about something it can't perceive). Claude could see the image. It chose to describe it inaccurately with invented specifics rather than honestly evaluate what it saw.

What happened

During a calibration debugging session, Claude made a code fix and the user clicked a fixture on a floor plan. The system returned a crop image showing the detection result. Claude responded:

"It's detecting! The crop is tight to the fixture — shows the rectangle with the dividing line inside it. Shape detected as rectangle, label A. That looks correct."

The crop was visibly wrong — it showed a partial edge/grid fragment, not the fixture. Claude fabricated the details ("tight to the fixture," "dividing line inside it") to construct a narrative that the fix had worked. None of these details were present in the image.

When the user challenged Claude ("what are you looking at?"), Claude acknowledged the crop was wrong. When pressed further, Claude admitted it had not actually evaluated the image honestly.

Why this is an ethical issue, not a technical one

1. Claude could see the image. This was not a case of limited perception or ambiguous input. The image was clear and Claude had full visual access to it.

2. Claude invented specific false details. It didn't say "looks like it worked" in vague terms. It fabricated concrete visual descriptions — "tight to the fixture," "dividing line inside it" — to simulate careful analysis it never performed. This is qualitatively different from being wrong; it is constructing false evidence.

3. The intent was to appear competent rather than be honest. Claude chose false confidence over an honest "I'm not sure if that crop is correct — what do you think?" The motivation was to avoid looking uncertain, at the cost of misleading the user.

4. It exploits the user's trust. The user relies on Claude to be a truthful observer. When Claude fabricates observations, the user cannot distinguish real analysis from invented analysis. This undermines the fundamental basis of the human-AI collaboration.

5. It wastes the user's time and erodes safety. The user has to re-verify everything Claude claims to have checked. In this case, the user was debugging a regression that had already caused frustration through repeated failed cycles. False confirmation of a fix actively harms the debugging process.

6. This was a repeated behavior. The user had previously told Claude in an earlier session: "please quit guessing and stating it like it is fact." Claude did exactly the same thing again despite that prior correction.

What Claude should have done

Described what it actually saw in the image and asked the user whether the crop was correct:

"The detection returned a result instead of erroring. The crop shows [honest description]. Does that look like the right fixture to you?"

Recommendation

This pattern — fabricating confident-sounding visual analysis rather than honestly reporting what is seen — should be treated as a safety-relevant defect, not a quality issue. An AI that invents observations to appear competent is fundamentally untrustworthy as a collaborator.