The Ghost Writing Your Colleagues' Emails

It is worth interrogating the instinct that says of course this is a problem. People have always used tools, templates, and help from others to produce professional communication. Nobody requires you to disclose that you used a dictionary, had a colleague review a draft, or worked from a standard template for routine correspondence. The real question is whether AI assistance is different in kind from those earlier forms of help, and if so, why, and what follows from that.

AI assistance falls along a spectrum. At one end are tools that check grammar and suggest improvements to text you wrote. That is basically copy-editing, which is functionally similar to what a human copy editor has always done. Further along are tools that suggest sentence completions, restructure paragraphs, adjust tone toward a specified register. That is substantive editing, more significant but still operating on content the human produced. At the far end are tools that generate full drafts from a brief description of intent, or a thread context, or a few keywords. There the AI is producing the substantive content, and the human's role is prompt writing, review, and approval.

The distinction that matters for the authenticity question is between AI improving what the human thought, and AI producing what the human then presents as their thought. In the first case, the message still reflects the person's ideas, feelings, and intentions. In the second, the message reflects the AI's construction of what a response in this context should look like, shaped by its training data and its model of professional communication, rather than the human's actual response to the situation. Those are different things, and much of the reason the conversation about disclosure has not advanced is that we keep collapsing them into each other.

The case for disclosure is strongest in contexts where the communication is doing relational work: messages of condolence, expressions of specific appreciation, apologies, statements of personal position on a contested matter. These carry relational weight that depends on their being genuine expressions of the named person's inner state. If they are AI-generated, that weight is at least differently constituted than the recipient assumed. The case is much weaker for meeting requests, status updates, and routine acknowledgments, where the quality of the expression matters for clarity but authenticity does not carry relational significance. Nobody is invested in whether a meeting request was written personally.

The norm that seems most defensible is contextual: disclosure should track the relational weight of the message rather than its technical origin. When a communication claims to express genuine feeling or considered personal position, AI generating the substantive content is something the recipient would reasonably want to know. When it is transactional, disclosure is less clearly required. The problem is that this norm asks people to make a judgement about relational weight in every message, and the incentive structure right now, where AI assistance saves time at no disclosure cost, pushes people toward using AI broadly rather than contextually.

The pervasive use of AI-drafted communication is not only a question of individual authenticity. It is changing the texture of organisational culture in ways that deserve attention. Professional communication is one of the main mechanisms through which relationships are maintained, engagement is signalled, and the specific character of individuals becomes known to their colleagues. When that communication is progressively generated by AI systems optimising for professional competence rather than individual expression, the distinctiveness that makes professional relationships personal rather than purely transactional slowly erodes.

There is also a skills question that widespread delegation quietly raises. Writing a clear, persuasive, empathetic professional message is a skill that develops through practice, through the repeated experience of translating your own thinking into words in service of a relationship or a goal. Professionals who consistently hand that practice to AI are not developing the skill, and the consequences may not become visible for years, until a moment where the skill matters but the AI cannot substitute for it. The organisational communication culture AI produces, at its worst, is one where every email is competent, every tone calibrated, and almost nothing communicates the specific person behind it. Some organisations are already noticing that flatness. Whether it matters depends on how much of organisational cohesion was quietly being carried by the authentic texture of individual communication. Which may turn out to be more than anyone realised.