What AI is teaching us about human communication, by Ruth Oji

A colleague recently asked me to review an email she had drafted using an AI writing assistant. The email was impeccable; every sentence was grammatically correct. The structure was logical, and the tone was professional. Yet something felt off. She was writing to a long-time client about a project delay, and the email, while perfectly polite, read as if it had been written by someone who had never met the client and had no sense of their relationship. It said all the right things and none of the necessary things. This is happening more frequently as AI writing tools become ubiquitous in professional life. People use them to draft emails, prepare presentations, generate reports, and refine their prose. The tools are remarkably capable. They can take rough ideas and turn them into polished text. They can adjust tone, fix grammar, and structure arguments coherently. For many routine communication tasks, they are genuinely useful. But their usefulness reveals something important about what they cannot do, and in revealing that, they teach us something about what human communication actually is.

AI excels at the surface features of language. It can generate sentences that are clear, correct, and well-structured. It understands grammar in a way that most humans do not, having been trained on vast amounts of text. It can mimic different styles, adjust formality levels, and produce prose that sounds fluent and professional. If communication were simply a matter of encoding information in grammatically correct sentences, AI would be nearly perfect at it.

The difficulty is that communication is not simply encoding information. It is navigating relationships, managing context, and making choices about what to say and what to leave unsaid based on who you are talking to and what you are trying to accomplish. These are precisely the things AI cannot do, because they require understanding that goes beyond pattern recognition in text.

Consider what happens when you ask an AI to write an apology email. It will produce something that contains all the conventional elements of an apology. It will acknowledge the problem, express regret, and offer to make things right. The language will be appropriate. But it will not know whether this is the first time you have disappointed this person or the fifth. It will not know whether your relationship can withstand a straightforward apology or whether you need to acknowledge a deeper pattern. It will not know whether the person you are apologizing to values directness or prefers more elaborate expressions of contrition. All of these things matter enormously to whether the apology will be received well, and none of them are accessible to the AI. This is not a limitation that better training data will solve. It is a fundamental constraint on what AI can know. When you communicate with another person, you are drawing on shared history, mutual understanding of context, and knowledge of what matters to them. You are making judgments about what needs to be said explicitly and what can be implied. You are reading signals about how your communication is being received and adjusting accordingly. AI has access to none of this.

I notice this particularly in AI-generated professional communication. A manager might use AI to draft an email to their team about a policy change. The AI will produce clear, well-organized prose that explains the change and its rationale. What it will not do is anticipate which aspects of the change will concern which team members, or acknowledge the unspoken anxieties that the change might trigger, or signal awareness of how this change relates to previous changes that did not go well. A human manager who knows their team would include these elements naturally, not because they are following a template, but because they understand the relational and historical context in which the communication is happening.

The same limitation appears in AI’s handling of tone. AI can adjust for formal versus informal registers. It can make text sound more friendly or more authoritative. But tone in human communication is not just about word choice. It is about signaling your relationship to the person you are addressing and to the content you are discussing. When a colleague writes, “I think we might want to reconsider this approach,” the hedging signals something about their confidence level, their relationship to you, and how directly they feel they can challenge your thinking. AI can produce hedged language, but it cannot know when hedging is appropriate and when it undermines your message, because that depends on context it does not have access to.

This raises an interesting question about what we are actually doing when we communicate well. We tend to think of good communication as clear expression of ideas. AI is quite good at that. But human communication involves something more. It involves understanding what your listener already knows and does not know, what they care about, what will concern them, and how they are likely to interpret what you say. It involves making choices about what to emphasize and what to downplay based on your goals and your relationship. It involves reading responses and adjusting your approach accordingly.

These are pragmatic dimensions of communication. Pragmatics, as a field, examines how context shapes meaning, how the same words can mean different things in different situations, and how much of what we communicate is implied rather than stated. From a pragmatic perspective, meaning does not reside in sentences themselves. It emerges in the interaction between speaker and listener, shaped by everything they bring to the exchange.

AI operates at the semantic level. It understands what words mean in a general sense. It can combine them into grammatically correct, logically coherent sentences. But it does not operate at the pragmatic level. It does not understand what these particular words will mean to this particular person in this particular context. It cannot know what is appropriate to say given the relationship and history between speaker and listener. It cannot sense when something that is technically accurate will nonetheless be heard as dismissive or tone-deaf.

I have seen this play out in various professional contexts. A consultant uses AI to draft a report for a client. The report is comprehensive and well-written. But it includes recommendations that, while logical, ignore political realities the consultant knows about but the AI does not. A team member uses AI to respond to a sensitive email from a colleague. The response is polite and addresses the explicit content of the email, but it misses the emotional subtext entirely, leaving the colleague feeling unheard. A leader uses AI to draft remarks for a town hall meeting. The remarks are articulate and cover all the necessary points, but they lack the personal touch that would signal genuine engagement with employee concerns.

In each case, the AI has done what it was designed to do. It has produced clear, correct, well-structured text. What it has not done is communicate in the full sense of what communication requires. It has not navigated the relationship. It has not read the context. It has not made the subtle judgments about what to say and how to say it that would make the communication effective for this particular audience in this particular situation.

This is not an argument against using AI for communication tasks. AI can be genuinely helpful for drafting, editing, and refining prose. It can save time and improve clarity. The difficulty arises when we treat AI-generated text as if it were complete communication, rather than as a starting point that requires human judgment to make it appropriate for the actual context in which it will be received.

What AI’s limitations reveal is how much of human communication happens at the level of pragmatics rather than semantics. When we communicate well, we are not simply encoding information accurately. We are making choices about how to frame that information given what we know about our audience. We are signaling our relationship to them and to the content. We are managing face, both theirs and ours. We are reading their responses and adjusting our approach. 

We are deciding what to say explicitly and what to leave implied. We are attending to emotional subtext and relational dynamics.

These are not optional extras that make communication more pleasant. They are fundamental to whether communication achieves its purpose. An email that is perfectly clear but ignores relational context can damage a relationship even as it conveys information accurately. A presentation that is logically structured but does not address the audience’s actual concerns will fail to persuade, no matter how well-crafted the arguments are. A policy announcement that is comprehensive but tone-deaf to employee anxieties will create resistance rather than buy-in. AI cannot do these things because they require understanding that extends beyond text. They require knowing the people involved, the history of their interactions, the organizational or social context, and the goals that are at stake. They require real-time reading of how communication is being received and flexibility to adjust. They require judgment about what matters in this particular situation, which depends on values and priorities that cannot be extracted from language patterns alone.

This has practical implications for how we use AI in communication. AI is most useful for tasks where context is minimal or standardized. It can help with grammar and clarity. It can generate first drafts that you then revise with your knowledge of context and relationship. It can handle routine communication where the pragmatic demands are low. What it cannot do is replace the human judgment that makes communication effective in contexts that matter. Perhaps the most valuable thing AI teaches us about communication is how much of it we were doing without realizing it. When you write an email to a colleague, you are not just encoding information. You are drawing on everything you know about them, your relationship, and the situation to make choices about tone, framing, what to include, and what to leave out. When you present to a group, you are reading the room, noticing where people seem engaged or confused, and adjusting your approach accordingly. When you have a difficult conversation, you are managing multiple layers of meaning simultaneously, attending to both content and relationship.

These are sophisticated cognitive and social skills. We exercise them constantly, often without conscious awareness. AI’s inability to do them reveals how complex they actually are. And in revealing that complexity, AI helps us see what we should be paying attention to when we communicate. Not just whether our sentences are clear, but whether we are attending to context, relationship, and the pragmatic dimensions that determine whether our communication will actually work. In the end, AI is a powerful tool for certain aspects of communication. But it is not a substitute for human judgment about what to say, how to say it, and when to say nothing at all. Those judgments require understanding that lives in relationships and contexts, not in text. And that understanding is what makes communication human.