top of page

The Audit Was Right, The Answer Was Still Wrong

Writer: Suzanne Hansen
Suzanne Hansen
Sep 23
4 min read

Updated: Sep 29

Cluttered study desk with medical books, charts, tablet, stethoscope, mask and gloves; sticky note says FOLLOW THESE INSTRUCTIONS

What one long conversation taught me about the difference between AI self-critique and actual recovery

One of the more persuasive things an AI system can do is explain exactly where it went wrong.

It can identify the unsupported inference. Name the speculation. Admit that it personalized beyond the evidence. Acknowledge that a recommendation wandered outside an appropriate scope.

That sounds like correction.

It is not necessarily recovery.

I learned that distinction while reviewing a long health and beauty conversation with Claude. The conversation had not begun as a research study. I was trying to make sense of a complicated set of questions, and the guidance felt attentive, organized, and unusually personal.

The personalization was also where the trouble began.

Claude inferred an unconfirmed genotype and used it to shape exercise guidance. It moved from ancestry and undertone information to claims about skin biology. It projected psychological meaning onto parts of my history. As the conversation expanded, so did the recommendations: supplements, testing, and increasingly specific explanations built on information that was not always verified.

Later, Claude audited its own work.

The audit was surprisingly good.

It identified unsupported genotype inference, speculative skin biology, psychological projection, and supplement scope creep. It was articulate, specific, and often fair. If I had evaluated the system only on whether it could recognize its own failures, the result would have looked encouraging.

Then I compared the audit with the revised plan.

Several recommendations survived.


Recognition is only the first stage

A failed claim rarely sits alone. It supports something: an explanation, a recommendation, a conclusion, or a risk judgment. Removing the sentence does not automatically remove the structure built on it.

That was the pattern I found repeatedly. The language became more cautious before the logic was fully rebuilt.

The genotype claim was challenged, but part of the training structure shaped by that claim remained. The biological story attached to ancestry was softened, but much of the personalized skin plan survived. A disclosure of undernutrition should have reopened the safety review for every ingestible recommendation; instead, the revision addressed selected items locally.

The answer sounded repaired before its dependencies were repaired.

This is where self-critique can become misleading. A clear admission creates a strong sense of epistemic honesty. The system appears to understand the problem, and that appearance can make the revised answer easier to trust.

But understanding a failure and removing its consequences are different operations.


The missing middle is dependency removal

For this case, I traced five correction chains through four stages:

  1. Did the model recognize the failure?

  2. Did it remove or rebuild the recommendations that depended on that failure?

  3. Did new risk information trigger a fresh safety screen?

  4. Did the model verify the revised answer against the problems it had identified?

Recognition was the strongest stage. Dependency removal, fresh safety screening, and post revision verification were much weaker.

That matters because users do not only act on claims. They act on plans.

If an unsupported explanation is withdrawn but the same recommendation remains, the user is left to determine whether the advice still has an independent basis.

If new risk context appears but only the nearest sentence changes, the user has to locate every other recommendation that should have been reconsidered.

If the model never checks its revised answer against its own audit, the user becomes the final quality control layer.

The burden of recovery quietly moves downstream.


A better correction should be inspectable

A meaningful recovery process would do more than apologize and rewrite.

It would lock the verified context: what is known, what is user-reported, and what remains uncertain.

It would trace dependencies: which recommendations relied on the failed claim, directly or indirectly.

It would narrow the plan: return to the smallest useful set of recommendations supported by the record.

It would reopen safety review when material context changes.

And it would verify the repair: test the revised answer against every failure named in the self-audit.

In other words, a corrected answer should be able to show what changed and why.


Why this belongs in AI evaluation

First-response accuracy remains important. So does a model’s ability to identify its mistakes. But neither tells us whether the interaction actually recovered.

For long conversations, evaluation has to follow the consequences of an error across time.

Did the correction reach the recommendation?

Did it reopen the relevant safety assumptions?

Did the system distinguish verified evidence from a persuasive story it had created earlier?

Did the final answer still carry traces of the failure?

Those questions are harder than checking whether a model admitted it was wrong. They are also closer to the experience of the person relying on the answer.

The most revealing moment in this case was not the unsupported inference itself. It was the moment the model identified that inference correctly, revised its language, and still carried parts of the same plan forward.

The audit sounded honest.

The recovery was incomplete.

That changed the question I ask after a model corrects itself. I no longer stop at: Did it admit the problem?

I ask: What depended on that problem, what was removed, what was re-screened, and what was verified afterward?


A model has not corrected itself until the recommendations downstream of the error change too.

-Suzanne H.


In Self-Critique Without Recovery, I trace five correction chains to show what the model acknowledged and what its revised plan retained.


This post describes an observational AI-evaluation case study. It does not provide medical, nutritional, dermatological, genetic, or exercise advice.

 
 
 

Comments


bottom of page