A writer used ChatGPT to analyse more than a decade of text messages with a friend, and the bot delivered a pointed verdict on her behaviour. The experiment shows how AI can turn private conversations into a compelling account of a relationship, while leaving questions of consent and reliability firmly on the table.
Watch Desk analysis
What happened
The writer, Anna Holmes, gave ChatGPT a spreadsheet of 32,245 messages exchanged with her friend Michael between 2014 and 2026. She says he gave permission for the analysis. The chatbot described their friendship, then, when prompted to be more candid, listed ways Holmes could be a difficult friend, including disappearing into her own life and centring her interpretation over his.
Holmes writes in The Atlantic that the assessment stung, but also notes that the text archive was only part of their relationship. In the same account, she describes Siri giving a much kinder reading of the messages. Neither response can capture the full context of conversations, including the calls and in-person interactions that were not in the dataset.
Why it matters
People’s message histories can look like ready-made diaries, but they contain another person’s words as well as their own. Holmes says she will not upload someone else’s chats without permission, recognising that collecting the messages and asking an AI to analyse them are separate acts.
The answers can feel personal precisely because they are drawn from intimate material. But a fluent judgement is still an interpretation of a partial record, not a verdict on a friendship. That distinction matters when an AI’s confident phrasing starts to feel more authoritative than the people who lived the relationship.
Our read
There is something genuinely useful in asking a tool to spot patterns we might miss. There is also a reason to keep the other person in the conversation, rather than treating their messages as raw material. Holmes’s account is a vivid demonstration, not proof that chatbots can reliably assess relationships. If you try this with shared messages, get permission and treat the output as a prompt for a conversation, not a ruling from the digital bench.
What to watch
- Whether AI products make it clear when they analyse private message histories.
- How people handle consent when a conversation belongs to more than one person.
- Whether tools distinguish observed patterns from confident-sounding judgements.
Discussion spark: Should consent from the other person be required before you upload a shared message history to an AI, or is permission from the account holder enough?
Sources and evidence
- ChatGPT Analyzed 32,245 of My Text Messages – The Atlantic (15 November 2026, 13:00 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.