> ## Content Index > Fetch the complete content index at: https://katecarruthers.com/llms.txt > Use this file to discover other available public pages before exploring further. # When my AI reads the email your AI wrote - URL: https://katecarruthers.com/when-my-ai-reads-the-email-your-ai-wrote/ - Published: 2026-11-05T19:40:00.000Z - Updated: 2026-11-05T19:39:59.000Z - Description: When your AI writes the email and mine reads it, what are we actually communicating? AI can help with the words, but it cannot do the listening or take responsibility for what we agree. - Author: Kate Carruthers - Tags: artificial intelligence, human communication, attention, ai agents, ai governance, future of work, digital relationships If your AI writes an email, and my AI reads it and drafts the reply, what exactly are we doing? A meeting might get booked. An invoice might get paid. That seems a reasonable use of automation. But if the message is supposed to be an apology, a difficult conversation or an expression of concern, I am less convinced. Something has been exchanged. Whether we have actually ***communicated*** is another matter. A recent [Malwarebytes article on artificial intimacy](https://www.malwarebytes.com/blog/ai/2026/10/artificial-intimacy-love-and-lies-in-the-age-of-ai?ref=katecarruthers.com) looks at how AI is becoming involved in our relationships, including through the messages we send and the advice we seek. It got me thinking less about people forming relationships with chatbots and more about what happens when chatbots start conducting our relationships for us. We do not need to stop talking altogether. We can keep sending messages while gradually removing ourselves from the exchange. ## Attention is all you need The irony is that the 2017 paper that introduced the transformer architecture was called [*Attention Is All You Need*](https://arxiv.org/abs/1706.03762?ref=katecarruthers.com). That architecture went on to underpin much of the current generation of large language models, including the systems we now use to write and summarise our correspondence. The paper was describing a mathematical mechanism for weighting relationships between elements of an input. Machine attention is not human care, interest or understanding. Nevertheless, there is something worth considering here. We are using technology built around machine attention to avoid giving each other our own. I have no objection to using AI to help find the words. Assistance can make communication possible, particularly when someone is struggling to express themselves or working in another language. There is no virtue in making that harder than it needs to be. But helping me express what I think is different from deciding what I think. And helping me read a message is different from responding so I do not have to read it. When I write to someone about something important, I generally want them to consider what I have said. I do not necessarily need an elegant response. I would rather have a clumsy reply from someone who has thought about it than a polished one from a system they have left to deal with me. The scarce resource is **attention**, not words. ## Turning bullet points into emails and back again Consider a fairly ordinary workplace exchange. - I have three points to communicate. My AI turns them into a six-paragraph email with a friendly opening, some background and an offer to discuss further. - Your AI reduces it to three points. - You dictate a short response. Your AI turns that into another email. Mine summarises it. What have we gained? Perhaps the tools have helped us communicate clearly. Or perhaps we have built an elaborate process for expanding and compressing text because nobody wants to question why we are producing so much of it. If the machines are the intended readers, why make them exchange simulated human correspondence? We could use structured requests, with clear information about what is needed, by when and who can approve it. If a person needs to consider the matter, then we need to allow time for that. Generating a longer message and summarising it at the other end does not establish that anyone has understood the issue. I am wary of productivity claims that measure how quickly we can write and process emails without asking why we are sending them. I think we need to spend more time considering the purpose of ordinary things we do, like emailing each other. When my AI writes the email and your AI reads it, what is that particular act of communication for? We might be completing a useful transaction. But if the point was for us to discuss something, understand each other or make a decision, have we actually done that? Automating unnecessary correspondence does not make it necessary. ## What are we delegating? An apology is a useful example because its purpose is not simply to transmit information. The words might say that I did something, that it hurt you and that I regret it. But you might reasonably want to know whether I understand what happened and intend to do anything differently. An AI can help me draft the apology. It cannot establish that I have accepted responsibility. The same applies to a manager responding to a staff member who is struggling. A considerate message might be helpful, but it is not evidence that the manager has listened, understood the problem or decided to act. I can delegate the drafting. I cannot delegate that work and still claim to have done it. Research offers an interesting complication here. In [two experiments published in Scientific Reports](https://www.nature.com/articles/s41598-023-30938-9?ref=katecarruthers.com), AI-generated suggested replies made communication faster and increased positive emotional language. Actual use could improve perceptions of a conversation partner, while believing that the other person was using more AI-generated replies was associated with less favourable judgements of them. These were suggested replies, not agents conducting entire conversations. But the findings point to something worth taking seriously: a more pleasant exchange does not necessarily mean a stronger connection. How the words came to be written can matter to the person receiving them. ## Who actually agreed? There is also the risk that neither of us notices when the meaning changes. You ask your AI to make an email more diplomatic. In the process, it waters down your concerns. My AI summarises the message and tells me you are happy to proceed. A reply goes back confirming the work. You think you raised an objection. I think I have your approval. Neither of us has read what the systems actually exchanged. That is a hypothetical, but it raises a straightforward accountability question. Which part of the exchange represents a decision either of us actually made? An automatically generated acknowledgement should not become evidence that someone considered a risk or accepted responsibility. Nor should we assume that sending a message means its contents have been understood. If we delegate correspondence, we need to be clear about what authority goes with it. Scheduling a meeting is one thing. Committing someone to a course of action is another. There still needs to be a person who knows what has been agreed and is prepared to stand behind it. ## Some of the effort matters I understand the attraction of handing over difficult communication. Most of us have messages we would rather not write and conversations we would rather avoid. But sometimes working out what to say is part of working out what we think. Trying to explain a disagreement can reveal that my argument is not particularly good. Writing an apology can make me confront what I did. Reading a long, frustrated message from a colleague can reveal a problem that a summary would leave out. None of this means we should preserve every tedious email or perform every administrative task by hand. It means we should be careful about treating all the effort involved in communication as waste. If we measure success by how little time people spend reading and responding, we may end up with a very efficient system for making sure nobody has to listen. I am happy for an AI to sort out a meeting time. I am less happy with the idea that it can do the listening on my behalf. If you are writing to me about something that matters, reading it is part of my responsibility. So is thinking about what you said before I reply. I can get help with the words. Paying attention is still my job.