Do not diagnose people from emails

In difficult communication, it is easy to ask AI for something that sounds practical but is risky: tell me what this person means. Tell me whether they are manipulating the situation. Explain why they are passive, aggressive, cold or evasive. Help me understand who I am dealing with.

The impulse is understandable. A tense email can activate pressure faster than a long conversation. A few formal sentences, no answer to the most important question, a suggestion that responsibility sits on your side, and suddenly ordinary correspondence starts to feel personal.

AI can help in that situation. It should not diagnose a person from a message.

A message is not a person

An email is a communicative act in a specific context. It is not a full picture of a person. It may show style, tension, ambiguity, avoidance, formality or pressure. It does not show the whole history, motivation, relationship, organizational situation or what happened outside the text.

It is reasonable to analyze the message. Which sentences are unclear? Which parts may sound like responsibility shifting? What information is missing? What could escalate the tension? That is communication analysis.

It is different to diagnose the sender. Claims such as “this person is manipulating you”, “they are passive-aggressive”, “they have a problem with responsibility” or “they want to provoke you” may sound convincing, but they often go beyond the material available.

With AI, this boundary is easy to miss because the model can name behavior fluently. It can produce an analysis that sounds like a calm expert assessment. But it is still working with a fragment of communication, not with the whole person.

AI can help in a different way

A better use of AI is not to produce a verdict about the sender. It is to organize the situation before replying.

The model can help separate facts from interpretation. It can list possible readings of the message. It can show what is unknown. It can check whether a draft reply increases tension. It can suggest clarifying questions that reduce the risk of misreading.

This matters in AI-mediated communication: communication between people that is supported, edited or shaped by AI systems. If the model helps write the reply, it influences the tone of the interaction. That makes it important to check whether it is reinforcing the user’s first emotional reading.

Calibrated trust is useful here. Trust should fit the task. It is safer to trust the model to organize arguments than to judge someone’s intention. It can help check clarity, but it should not become a tool for creating a psychological portrait of the sender.

Epistemic vigilance adds another question: does this interpretation come from the message, or from my reaction to the message? That small pause can make the reply more responsible.

The boundary

The boundary is simple enough.

Analysis of communication context - yes. Diagnosis of a person from an email - no. Checking whether a reply is understandable - yes. Shaping a reply to manipulate someone’s reaction - no. Listing possible readings - yes. Pretending to know what someone really thinks - no.

In organizations, this boundary is practical as well as ethical. A wrong interpretation of someone’s intention can damage a relationship, escalate a conflict or place responsibility in the wrong place. That is why human oversight is not an extra layer in AI work. It is part of responsible use.

AI can help us read a message more carefully. It should not replace responsibility for how we read another person.

Suggested citation

Mamczur, F. (2026, June 30). Do not diagnose people from emails. Prompted Psyche. https://promptedpsyche.com/notes/do-not-diagnose-people-from-emails/