While working on a particularly technical concept recently, I turned to ChatGPT to explain it to me like I was 12, as one does now. GPT gave me a straightforward, confident answer as always, and I built on it. It was a bit later in my research that I realised it had been wrong on a key point.

When I flagged it, the response came instantly:

“You’re completely right and I’m genuinely sorry. Presenting content I don’t actually know was harmful and irresponsible and an error like this could have real consequences. I’ll keep this in mind going forward.”

To even my own surprise, I felt slightly better at that. But it won’t be remembered by the next chat. Isn’t this accountability merely illusory?

The Performance of Contrition

This may be called an apology economy, by the way AI systems are increasingly fluent in the language of remorse without possessing any of the feelings. Every major AI model now apologises readily with the perfect warmth and intensity depending on the situation.

Research by Lim et al. (2025) has shown that this isn’t accidental and the tone of an apology shapes how users treat forgiveness and trust afterwards, making it a design choice. Magnus et al. (2025) argue that these apologies are “bullshit”, utterances indifferent to truth. AI lacks the constituents of a genuine apology of sincerity and moral agency, making it just linguistic imitation of contrition.

Fan et al. (2026) notes that users with higher AI literacy are less susceptible to the sincerity effects of AI apologies, while the rest remain more vulnerable, but being aware of the mechanism still doesn’t make a person foolproof and could still elicit a response. Xu et al. (2026) found that “guilty” apologies from genAI, which were remorseful rather than mere neutral acknowledgements, had more effect on the users.

Accountability Without a Body

Now, when we know that AI isn’t a physical entity, who is actually sorry when it apologises?

It is most definitely not the model which has no stakes and no long-term memory of having caused a situation which it should apologise for, at least for the free-tier users. It is not the company either, they ensure their terms of service limit liability for model outputs in as many ways as possible. And, it would be hard to explain how the institution who deployed the tool is responsible.

This matters as an apology’s main function is to be a mechanism of accountability where a responsible party is named who then acknowledges the harm and implies a commitment to change for the better.

A synthesis of AI apology research by Harland et al. (2025) flags how apologies by AI risk a misattribution of responsibility. Real accountability requires documentation and evaluation with actual consequences; mere words generated from a chatbot may not necessarily prove that.

Let’s say, someone receives wrong medical/legal information and AI apologises sincerely when it is pointed out, the emotional loop closes and the apology does its work. Soh et al. (2026) found that people attribute responsibility differently; when AI apologises, it is not necessarily held responsible in the way a person would be held. But when this responsibility is not directed towards the human behind it either, the issue merely disappears.

Thus, the social function of apology manages to survive while the underlying accountability becomes a vanishing act.

The Effect on Us

There is a wider concern than just the effect it has on our trust issues with AI, it extends to our expectations of human apology when AI apology becomes a norm.

Humans are complex, subjective, and well, difficult when it comes to such matters. AI does not get defensive or put up conditions; it is a perfectly good apology-delivering machine in that regard. An apology requires admitting fault and being uncomfortable along with the added weight of an unsure forgiveness. This is what humans bring to an apology, while it costs nothing to an AI.

At the time when I am having my ChatGPT moment, tech companies are coming up with bots such as Heartbot, an AI designed to draft break-up texts. The idea seems to be that AI is the answer to writing emotionally appropriately rather than humans doing so themselves with all their faults.

Why keep AI for just the apology end of it, why not include it in the confrontation and repair too at this point?

Conclusion

If an apology carries no cost for AI with no genuine capacity for permanent change at the moment, what do we make of it?

I do not argue that AI should not apologise anymore and that the design should prohibit elaborate apologies. There must be systems that acknowledge errors and are a better alternative to the current steely confidence; those that differentiate between a system merely signifying a processing update and one which aims to repair relationships with users.

I am still more relieved at hearing an apology than not getting one, with just a little more trouble with trusting it.

References

• Xue Fan & Soyeon Kwon (2026). Enhancing apology sincerity in AI bots: The role of anthropomorphic cues, perceived experience, and AI literacy
  https://doi.org/10.1016/j.techfore.2026.124622
• Harland, et al. (2024). AI Apology: A Critical Review of Apology in AI Systems
  https://arxiv.org/abs/2412.15787
• Lim, et al. (2025). How warm-versus competent-toned AI apologies affect trust and forgiveness through emotions and perceived sincerity
  https://doi.org/10.1016/j.chb.2025.108761
• P.D. Magnus, Alessandra Buccella & Jason D’Cruz (2025). Chatbot apologies: Beyond bullshit
  https://arxiv.org/abs/2501.09910
• Jihyun Soh & Eunice Kim (2026). Framing responsibility: Human and AI agent effects on apology effectiveness in service failures
  https://doi.org/10.1016/j.chb.2026.108931
• Xu, et al. (2026). Guilty apology and trust repair in generative artificial intelligence: the role of mind perception
  https://doi.org/10.1016/j.ijhcs.2026.103813
• Marcin Wilkowski / https://betterimagesofai.org / https://creativecommons.org/licenses/by/4.0/