When AI Says Sorry, Look for Evidence of Correction

Imagine asking an AI assistant for a source to support an article. The reference looks convincing, but you cannot find it. When challenged, the assistant apologises and offers another reference. The conversation feels more cooperative. The editorial problem remains: does the replacement source exist, and does it support the sentence?

This hypothetical exchange separates two useful things. An apology can acknowledge a frustrating experience. A correction needs to change the answer or action in a way you can examine. For creators, business owners and everyday users, the question after “sorry” is what evidence now supports continuing.

What people prefer in an AI apology

In Who’s Sorry Now, published in ACM Transactions on Computer-Human Interaction on 16 June 2026, Zahra Ashktorab and colleagues studied 162 participants’ preferences among rote, empathic and explanatory chatbot apologies. Participants compared exchanges containing an error, an apology and a resolution across bias, factual-error and unfounded-fabrication scenarios.

Explanatory apologies were generally preferred, with variation by situation and user. Empathic responses received support in the bias scenario for acknowledging emotional impact. The study used imagined situations: its preference findings do not establish that a particular apology makes a live system more accurate or prevents a repeated mistake.

The useful implication is to leave room for different needs. Someone may want the error explained; someone else may first want the impact acknowledged. A single elaborate apology script cannot be assumed to suit everyone.

Reassurance and continued use are different outcomes

A second study, by Björn Konopka and Manuel Wiesche, was published in Information Systems Frontiers on 3 June 2026. In an online experiment involving 261 participants, a customer-service chatbot made a programmed error in a hypothetical shopping scenario. The researchers compared apologies, clarification questions, explanations, alternative input options and control conditions.

Apologies and the two explanation-based strategies produced comparable reported trust. When strategies were grouped, explanation-based responses led more participants to choose the chatbot again than the human-like response group did. That was a choice about a fictitious second task, not evidence of lasting customer loyalty or a permanently improved underlying system.

My inference from these studies is to ask three separate questions: did the response feel appropriate, did the person choose to continue, and was the underlying problem resolved? Those outcomes can inform one another, but they should be measured separately.

Make the correction observable

For an ordinary factual or workflow error, I propose four questions. This is an editorial checking method, not a validated psychological scale:

• What exactly was wrong? Identify the affected claim, file, instruction or action.
• What changed? Compare the corrected result with the earlier one.
• What supports it now? Look for a relevant source, calculation, checked record or confirmed action.
• What remains uncertain? Keep any unresolved limit visible.

In the imaginary source example, a useful next step is to locate the publication, check its authors and date, then inspect the passage supporting the claim. If no suitable source can be found, remove or narrow the sentence. A second polished citation is another item to verify.

This does not require turning every typo into an investigation. Match the checking effort to the consequence. A misspelt heading and an unsupported allegation need different responses.

Treat the explanation as another claim to assess

An assistant might attribute a mistake to an outdated source, ambiguous wording or a calculation error. That explanation can help identify a check, but its fluency does not establish the cause. Ask which source was outdated, which words were ambiguous or which calculation changed.

Compare the explanation with evidence available to you. If you cannot inspect the proposed cause, describe it as an unverified explanation rather than a confirmed diagnosis of the system. There may still be a useful correction even when the cause remains unresolved.

Likewise, apologetic language alone does not establish remorse, a software update or a durable change in future behaviour. A promise that an error will never recur needs evidence beyond the promise itself.

Check the next relevant step

Consider a second hypothetical case: an assistant uses the wrong product name in a customer email. After correction, check the subject line, body and attachment references. If a similar draft follows, check whether it uses the approved name. These observations tell you something specific about that work.

They do not prove that every future conversation or every component of a product has been changed. Keep your conclusion at the scale of the evidence: this draft was corrected, this test passed, or this issue remains open.

When deciding whether to continue, define the scope. You may be comfortable using the assistant to explore ideas while separately checking factual claims. You may require human approval before a customer message is sent. Continuing a conversation need not mean granting every output the same authority.

If repeated attempts produce the same consequential error, change the approach. Supply a checked source, narrow the request, use another method or involve someone with the relevant expertise. Endless apology exchanges are not a substitute for a usable result.

For businesses, name who completes the repair

When a chatbot speaks on behalf of a business, someone still needs to confirm the real outcome. Did the customer receive the correct document? Was an incorrect instruction withdrawn? Is an unresolved request with the person who can act on it?

Design the response around what has actually happened. A message saying that a request has been passed to a person should correspond to a verified handover. If an action is still pending, say so plainly instead of making the apology sound like confirmation of completion.

For published work, the repair may extend beyond the private chat. If an inaccurate claim reached readers, consider the article and any related captions or newsletters. The MaryChuks guide to a visible correction and update trail explains how to make the change understandable.

A considerate acknowledgement can be welcome. It can sit alongside a clear account of the mistake, an inspectable correction and an honest limit on what has changed. The strongest reason to rely on the next step is evidence that the work now meets its purpose.

Sources

• Ashktorab and colleagues: Who’s Sorry Now: User Preferences Among Rote, Empathic, and Explanatory Apologies from LLM Chatbots, ACM Transactions on Computer-Human Interaction, 16 June 2026. Author-posted preprint, 3 July 2025.
• Konopka and Wiesche: Explainability in AI: Comparing Human-Like and System-Like Trust Repair Strategies, Information Systems Frontiers, 3 June 2026.

Publication details and source versions checked on 23 October 2026.