No AI in Human Communication#
This manifesto draws a clear boundary between human-to-human communication and work artifacts.
Here, human communication means a message presented as one person speaking to another: a conversation, email, chat message, comment, feedback, disagreement, appreciation, apology, support, or leadership communication.
Those messages are written by humans, without AI generating, rewriting, or polishing them.
Reports, specifications, analyses, documentation, and similar work artifacts are different. AI may assist with them, but its involvement is made visible.
Individuals and teams who sign this manifesto commit to preserving that distinction.
Human communication is human-written#
Behind every name, face, or avatar should be a person speaking for themselves - not AI masquerading as one.
Human communication is more than information transfer. It carries attention, effort, judgment, personality, uncertainty, care, and intent. Our wording and imperfections are part of how others know us.
When a message is presented as coming from us, we write it ourselves. We do not ask AI to generate, rewrite, polish, soften, intensify, or impersonate our voice.
Authenticity does not require perfect writing. It requires the person to be present.
Research increasingly suggests that AI-mediated interpersonal messages can reduce perceived effort, warmth, authenticity, reciprocity, and willingness to forgive. See the evidence and its limits.
Do not normalize AI masquerading as people#
If one person delegates a message to AI and the other delegates the response, the conversation may become AIs talking to each other, masquerading as us.
Signatories do not reward that behavior by continuing the exchange as though it were human communication.
When a message is clearly AI-generated and presented as personal communication, ask the sender to restate it in their own words, or pause the exchange until they do.
This is not hostility toward the person, and it is not a test of writing style. Do not accuse someone based on suspicion alone.
It is a form of social resistance: refusing to normalize hidden machines as participants in human relationships.
AI-assisted work artifacts are visibly marked#
Technical reports, specifications, analyses, research summaries, documentation, meeting notes, and similar work artifacts are different from personal messages. AI may assist or generate them.
Material AI involvement must be visible before the reader relies on the artifact.
Use clear labels:
- AI-assisted: A human created the substance or first draft, and AI materially revised, expanded, reorganized, or clarified it.
- AI-generated: AI produced most of the artifact, and a human with enough relevant expertise to judge its material claims reviewed and approved it.
The label belongs near the title or at the beginning - not buried in metadata or a footnote.
A human-written message may share a labeled AI-assisted artifact. The artifact and the interpersonal message are different layers; the message itself remains human-written.
Real human expertise matters most#
AI can make almost anyone sound competent in almost any field. It cannot give them the years of practice, judgment, failure modes, tacit knowledge, and accountability that real expertise provides.
Do not throw AI-generated work artifacts at other people and make them audit the model for you.
If you cannot independently judge every material claim, assumption, source, tradeoff, and recommendation, you are not ready to present the artifact as reliable work.
AI may extend real expertise. It does not manufacture it.
If you cannot tell when AI is wrong, AI is not making you an expert.
Use AI most confidently where you already understand the field well enough to challenge its output, recognize missing context, and reject plausible-sounding mistakes.
Respect professional boundaries#
Cross-functional collaboration is valuable. Using AI to bypass the people whose expertise and responsibility cover a domain is not.
Do not use AI to silently expand into another person's or team's professional domain and then present generated output as equivalent to their expertise.
For example:
- A web developer should not generate infrastructure and route around the operations or platform team responsible for it.
- A bug report should primarily describe the observed behavior, impact, and reproduction. An AI-generated diagnosis or pointer to where the problem supposedly lives may be shared as a clearly labeled hypothesis - not as an authoritative technical explanation.
- A feature request should explain the need, context, and desired outcome. Do not attach an AI-generated implementation and present it as a vetted solution unless you genuinely understand the relevant system and tradeoffs.
- Do not use AI to play expert in a field where you lack mature experience, especially to correct or overrule people whose profession and accountability are in that field.
Professional boundaries are not an argument for silos. They are a reminder that expertise, context, and accountability matter.
Use AI to collaborate with experts, not to impersonate them or route around them.
Writing is thinking#
When a work artifact is meant to represent human analysis, judgment, or expertise, write the first version yourself.
Writing forces you to decide what you know, what you believe, how the reasoning connects, and where uncertainty remains.
Use AI afterward to challenge the reasoning, identify gaps, reorganize, shorten, proofread, or improve clarity.
If AI supplies the substance and reasoning, mark the artifact as AI-generated. Do not present model fluency as your own understanding.
Disclosure does not remove responsibility#
If you publish or circulate an artifact, you own its consequences even when AI helped create it.
Verify the facts, sources, calculations, and citations. Understand the terminology and claims. Be able to explain the reasoning, its limits, and the first reasonable follow-up without asking AI to reconstruct the argument for you.
AI-generated confidence is not expertise.
Respect human attention#
AI makes text cheap to produce. Reading still costs human time.
Do the compression before you publish. Remove noise. Include relevant context. Do not save your own time by transferring the thinking work to every reader.
AI-generated volume is cheap. Human attention isn't.
Before you send or publish#
For interpersonal communication:
- Did I write this message myself?
- Is the other person hearing me, or a model acting through me?
- Am I using AI to simulate care, confidence, empathy, or expertise?
For a work artifact:
- Is material AI involvement clearly labeled?
- Did I write and think first where the artifact represents my judgment?
- Do I have enough real expertise to judge every material claim?
- Am I respecting the professional boundaries and responsibilities of the people or teams who own this domain?
- Do I understand and stand behind every important claim?
- Can I answer a reasonable follow-up myself?
- Did I respect the reader's time?
The commitment#
Human messages are human-written.
AI-assisted work is visibly marked.
Real expertise is not replaced by generated fluency.
Professional boundaries are respected.
Human responsibility is never delegated.
Speak to people yourself. Mark AI in the work.