Sorry, Not Sorry: When AI Can Write Your Apology and When It Can't
A workplace mistake happens. A client is upset. A message needs to go out, and AI can draft it in seconds, in exactly the right tone, with exactly the right words. The question isn't whether AI can write a good-sounding apology. The question is whether using it will actually work - and a growing body of research says: it depends entirely on what you're apologising for.
The core finding: identical words, different trust
A 2023 study published in the journal Computers in Human Behavior ran three experiments testing AI-mediated apologies in workplace settings. The study led by researchers Ella Glikson and Kotaro Asscher found that across all three experiments, simply telling someone that AI tools had been used to help write an apology - even just for language support - measurably reduced how authentic the apology felt, and reduced the recipient's willingness to forgive. A follow-up study conducted in 2024 went further: it took one identical corporate apology statement for a data-breach scandal and told half the participants it was written by a human, and half that it was AI-generated. It contained same words, same sentence structure, and same details. The participants who believed a human wrote it reported significantly higher trust and forgiveness than those who believed AI did - from the exact same text.
The mechanism researchers point to is called the " machine heuristic ": people reflexively associate AI with competence and efficiency, but not with warmth, sincerity, or genuine moral intention. An apology's entire job is to signal the second category. AI, by reputation alone, undermines it before a single word is read.
Soumitra Mishra - Head of Academics, at Newton School of Technology noted that an apology is not just a message we send, it is a part of the process of understanding what we have done.
“If AI writes the apology for us, even if the result sounds exactly like us, we can skip that difficult but necessary reflection. AI can help us improve the grammar or make our meaning clearer after we have written the substance ourselves. But it should not do the thinking for us. The effort and discomfort involved in writing an apology are not inconveniences. They are part of what makes the apology sincere,” Mishra added.
When this actually backfires in public
This isn't just a lab finding. In 2023, American professional basketball player Jamel Morant posted an apology after a public incident, and it was written carefully enough that sports commentator Shannon Sharpe and others publicly questioned whether it had been generated by ChatGPT - intending that as a criticism, not a compliment. The suspicion alone was enough to undercut the apology's reception, regardless of whether it was true. The lesson generalises: if an apology ever reads as possibly AI-assisted, in a moment where sincerity is the entire point, the suspicion itself does damage that no amount of polish can undo.
An apology is not a writing task; it is an act of accountability, noted Dr N Vivek, Professor of Operations & Chairman, Center for AI in Business, Great Lakes Institute of Management.
“AI can polish a sentence you have written, but it cannot understand what your mistake meant to the person you hurt. If you didn't choose every word yourself, you haven't really apologised." Vivek said.
He noted that each section of an apology should be there on purpose: what happened, what we accept responsibility for, what we are doing about it, what we will do to prevent a repeat.
“A thoughtful person who understands the consequences, both immediate and future, should be able to explain why every paragraph is there and why every word was chosen over its neighbour. If you cannot defend a sentence, you should not send it. An AI-generated paragraph that you merely approved is a sentence you never truly decided,” the professor added.
India's version of the same tension
This plays out with a specific twist in the Indian market. Zendesk's 2025 CX Trends Report found that Indian consumer trust in AI customer service is the highest in Asia, and that 63 percent of Indian consumers are more likely to trust AI specifically when it displays warmth and empathy - a market that, on the surface, seems unusually receptive to AI handling emotionally sensitive exchanges. However, a separate 2024 Salesforce study found that nearly 68 percent of Indian consumers specifically want to know when they're interacting with an AI agent rather than a human.
Put those two findings together and the tension becomes obvious: Indian consumers want AI to feel warm and empathetic, and they want to be told when it's AI. But the Glikson and Asscher research shows those two preferences actively work against each other - disclosure is exactly what triggers the machine heuristic and kills the perceived sincerity the warmth was supposed to build. There is no setting where both preferences get fully satisfied at once.
Where AI genuinely helps
None of this means AI has no place in apologies. The research points to a real, usable distinction. For transactional, low-stakes situations - a shipping delay, a minor scheduling error, a service hiccup with no real relationship at stake - AI-drafted apologies work fine, largely because nobody expects deep personal investment in a transactional exchange to begin with. AI is also genuinely useful as a drafting aid behind the scenes: helping a person organise their thoughts, find the right tone, or get unstuck on an apology they still write and send as their own. Glikson and Asscher's research specifically found that limited reliance on AI assistance, without full disclosure of every tool used, didn't carry the same authenticity penalty as heavy, disclosed use.
Where it doesn't
For anything involving a real relationship, a breach of trust, or a situation where the recipient needs to feel genuinely understood - a client relationship, a serious workplace conflict, a public crisis - the research is consistent: AI authorship, once known or even suspected, actively undermines the outcome the apology was meant to produce. In these cases, the apology's entire value lies in the sender's visible, personal effort. Outsourcing that effort to a tool doesn't just fail to help. It can make the apology worse than no apology hastily sent at all.
The judgment call
This is Judgment, applied to one of the more personal tasks AI can attempt: ask not "can AI write this well" but "does this apology's value come from the words, or from the visible effort of a specific person writing them." Transactional apologies survive AI assistance easily. Relationship-repairing ones depend on the one thing AI cannot actually supply: a person who meant it.
The core finding: identical words, different trust
A 2023 study published in the journal Computers in Human Behavior ran three experiments testing AI-mediated apologies in workplace settings. The study led by researchers Ella Glikson and Kotaro Asscher found that across all three experiments, simply telling someone that AI tools had been used to help write an apology - even just for language support - measurably reduced how authentic the apology felt, and reduced the recipient's willingness to forgive. A follow-up study conducted in 2024 went further: it took one identical corporate apology statement for a data-breach scandal and told half the participants it was written by a human, and half that it was AI-generated. It contained same words, same sentence structure, and same details. The participants who believed a human wrote it reported significantly higher trust and forgiveness than those who believed AI did - from the exact same text.
The mechanism researchers point to is called the " machine heuristic ": people reflexively associate AI with competence and efficiency, but not with warmth, sincerity, or genuine moral intention. An apology's entire job is to signal the second category. AI, by reputation alone, undermines it before a single word is read.
Soumitra Mishra - Head of Academics, at Newton School of Technology noted that an apology is not just a message we send, it is a part of the process of understanding what we have done.
“If AI writes the apology for us, even if the result sounds exactly like us, we can skip that difficult but necessary reflection. AI can help us improve the grammar or make our meaning clearer after we have written the substance ourselves. But it should not do the thinking for us. The effort and discomfort involved in writing an apology are not inconveniences. They are part of what makes the apology sincere,” Mishra added.
When this actually backfires in public
This isn't just a lab finding. In 2023, American professional basketball player Jamel Morant posted an apology after a public incident, and it was written carefully enough that sports commentator Shannon Sharpe and others publicly questioned whether it had been generated by ChatGPT - intending that as a criticism, not a compliment. The suspicion alone was enough to undercut the apology's reception, regardless of whether it was true. The lesson generalises: if an apology ever reads as possibly AI-assisted, in a moment where sincerity is the entire point, the suspicion itself does damage that no amount of polish can undo.
An apology is not a writing task; it is an act of accountability, noted Dr N Vivek, Professor of Operations & Chairman, Center for AI in Business, Great Lakes Institute of Management.
“AI can polish a sentence you have written, but it cannot understand what your mistake meant to the person you hurt. If you didn't choose every word yourself, you haven't really apologised." Vivek said.
He noted that each section of an apology should be there on purpose: what happened, what we accept responsibility for, what we are doing about it, what we will do to prevent a repeat.
“A thoughtful person who understands the consequences, both immediate and future, should be able to explain why every paragraph is there and why every word was chosen over its neighbour. If you cannot defend a sentence, you should not send it. An AI-generated paragraph that you merely approved is a sentence you never truly decided,” the professor added.
India's version of the same tension
This plays out with a specific twist in the Indian market. Zendesk's 2025 CX Trends Report found that Indian consumer trust in AI customer service is the highest in Asia, and that 63 percent of Indian consumers are more likely to trust AI specifically when it displays warmth and empathy - a market that, on the surface, seems unusually receptive to AI handling emotionally sensitive exchanges. However, a separate 2024 Salesforce study found that nearly 68 percent of Indian consumers specifically want to know when they're interacting with an AI agent rather than a human.
Put those two findings together and the tension becomes obvious: Indian consumers want AI to feel warm and empathetic, and they want to be told when it's AI. But the Glikson and Asscher research shows those two preferences actively work against each other - disclosure is exactly what triggers the machine heuristic and kills the perceived sincerity the warmth was supposed to build. There is no setting where both preferences get fully satisfied at once.
Where AI genuinely helps
None of this means AI has no place in apologies. The research points to a real, usable distinction. For transactional, low-stakes situations - a shipping delay, a minor scheduling error, a service hiccup with no real relationship at stake - AI-drafted apologies work fine, largely because nobody expects deep personal investment in a transactional exchange to begin with. AI is also genuinely useful as a drafting aid behind the scenes: helping a person organise their thoughts, find the right tone, or get unstuck on an apology they still write and send as their own. Glikson and Asscher's research specifically found that limited reliance on AI assistance, without full disclosure of every tool used, didn't carry the same authenticity penalty as heavy, disclosed use.
Where it doesn't
For anything involving a real relationship, a breach of trust, or a situation where the recipient needs to feel genuinely understood - a client relationship, a serious workplace conflict, a public crisis - the research is consistent: AI authorship, once known or even suspected, actively undermines the outcome the apology was meant to produce. In these cases, the apology's entire value lies in the sender's visible, personal effort. Outsourcing that effort to a tool doesn't just fail to help. It can make the apology worse than no apology hastily sent at all.
The judgment call
This is Judgment, applied to one of the more personal tasks AI can attempt: ask not "can AI write this well" but "does this apology's value come from the words, or from the visible effort of a specific person writing them." Transactional apologies survive AI assistance easily. Relationship-repairing ones depend on the one thing AI cannot actually supply: a person who meant it.
Next Story