A leader’s field guide
Before you communicate about AI, answer six questions.
Nobody has this figured out completely. Trust does not require certainty; it requires clear boundaries, visible human accountability, and an honest way for people to respond.
Start here
Build your message with six CALMER questions
Use these before writing a policy, staff note, family letter, public statement, or project disclosure.
Built on the research-informed CALMER framework: Confirm · Acknowledge · Link · Make usable · Engage · Return. Read the research review behind the framework.
A useful test: after reading your message, can someone distinguish what is known from what is uncertain, understand why AI is involved, see what remains human, act or respond, and know when they will hear from you again?
See the structure at work
A director’s weekly newsletter, with the guidance made visible
This is how a school director might describe their own practice. The CALMER labels show the trust-building work each sentence is doing.
Scenario: Weekly school-community newsletter · AI-assisted organization · Human-written final message
When I draft the weekly newsletter, I may use an approved AI tool to help organize my notes, spot repetition, or suggest a first-pass structure.
I know some readers may worry that AI has replaced my judgment or voice; it has not. I do not use AI to decide what our community needs to hear, invent stories or quotations, or write about individual students, families, or staff.
I use AI only for organization; I choose the priorities, verify every detail, and rewrite and approve the final message in my own voice before it is sent.
The newsletter stays in our familiar format and plain language. I do not put confidential, identifying, or unpublished information into the tool, or generate or alter anyone’s voice or likeness.
If you want to know how AI contributed—or think I have missed something—reply to the newsletter; your question will come directly to me.
I will revisit this practice each term and change it if our tools, evidence, or community expectations change.
Adapt the pattern
The same guidance in different situations
These are illustrative statements to adapt, not universal language to copy without context.
Preparing a board update
I may use an approved AI tool to organize public reports and identify recurring themes as I prepare a board update. I verify every source, decide what matters, and write the final interpretation. Confidential meeting material and personally identifying information are not entered into the tool.
Notice: AI helps organize source material; the leader retains interpretation, emphasis, verification, and authorship.
A high-stakes decision
We do not use AI to make final decisions about hiring, admissions, assessment, discipline, or access to services. If an approved tool supports administrative work, a named decision-maker reviews the relevant information and remains accountable. Anyone affected may request human review.
Notice: Human oversight is translated into a decision right and an appeal route.
Synthetic images, audio, or video
This illustration was generated with AI to explain a concept; it does not depict a real person or event. We do not clone anyone’s voice or likeness without explicit consent. A human editor reviewed the material for accuracy, bias, and context.
Notice: Disclosure explains what is synthetic, while consent protects the person behind the voice or likeness.
Five useful outliers
“Human voice” can mean five different things
Each organization offers a different answer to the same question: what must remain human? Use the comparison to decide what your own message needs to preserve.
Long-form articles, newsletters, and scripts stay human-written; AI is permitted mainly for preparation and summarization. Borrow this when the author’s expertise is part of the promise.
Human oversight is described through curiosity, critical thinking, and empathy, with public notice when a conversational service is not human. Borrow this when “human review” feels too vague.
AI may be built into assignments so students practise prompting, checking, and improving its output rather than pretending it does not exist. Borrow this when capability—not avoidance—is the goal.
AI tools address production gaps such as missing shots, backgrounds, and lighting while filmmakers retain creative intent and final judgment. Borrow this when technology should extend craft, not replace authorship.
It translates text and speech in tribal languages, digitizes oral traditions, and widens access to health and government information using community-curated data and expert validation. Borrow this when AI should carry a community’s voice farther without claiming it.
The comparison: MIT protects who writes. The UK protects how people judge. Tartu builds capability through critique. Netflix keeps human judgment at the heart of creative work. Adi Vaani uses AI to preserve and extend community voice.
Before you send
Run the trust check
A strong AI message should let you answer yes to each question.
- Is the purpose specific?
- Are the boundaries visible?
- Is a person clearly accountable?
- Can the audience question, appeal, or opt out where appropriate?
- Have you named what remains uncertain?
- Is there a date for review?
Explore the evidence behind this guide
What fifteen policies consistently show
Fifteen policies at a glance
| Policy | Released | Category | Stated plainly | Left open | Reason given |
|---|---|---|---|---|---|
| WIRED | Mar 2023 | Output guardrails | Headlines, social posts and idea generation are fine; published text is not | Not stated | Undisclosed AI text would amount to plagiarism |
| University of Tartu | Apr 2023, updated 2024 and 2025 | Permission with conditions | AI use is encouraged and described openly; detection software is off the table | Practice varies by faculty, and citation expectations may change again | Detection cannot reliably identify AI writing, so it cannot be applied fairly |
| Associated Press | Aug 2023, updated Jul 2026 | Output guardrails | AI output is treated as raw, unchecked source material; photos, video and audio are never altered | How to be sure incoming material is AI-free, given unreliable detectors | Accuracy, fairness and speed, with the journalist still gathering and judging |
| Ringier | 2023 | Permission with conditions | Label AI-generated content; check, verify and supplement every output | Where “only an aid” ends and a label begins | Staff judgment is what makes the output usable |
| Singapore GovTech Pair | 2023, page updated Jul 2025 | Permission with conditions | Public officers use a secure, government-contextualized assistant for writing, research, summaries, coding, and repetitive tasks | How individual agencies set limits for higher-risk work | Free officers from routine work so they can focus on meaningful public service |
| MIT News Office | Dec 2023 | Output guardrails | Articles, newsletters and scripts are human-written; AI helps with headlines, captions and alt text | What other MIT communicators do on their own channels | Inaccuracy is well-documented, and audience trust is what is at stake |
| Ukraine news media guidelines | 2024 | Rules about rules | Risk assessment, transparency, audience awareness, human oversight | Whether the voluntary stage becomes binding, and when | Putting Council of Europe principles into national practice |
| UK Government Communications | May 2024 internally, public Sep 2024 | Permission with conditions | First drafts, summaries, accessibility audio; no public communication without human oversight | The whole approach, openly described as evolving | Accuracy, inclusivity and mitigating bias |
| Ukraine advertising and marketing guidelines | Aug 2024 | Audience rights | Label AI content, keep human control, protect audience data | Everything, technically: the guidance is voluntary | Preparing businesses now for regulation expected later |
| California AB 3030 | Signed Sep 2024, in force Jan 2025 | Audience rights | AI-written clinical messages carry a disclaimer and a way to reach a human | Anything a provider reviews; scheduling and billing sit outside the law | Patients should know what they are reading and be able to reach a person |
| College of Physicians and Surgeons of Ontario | 2025 | Permission with conditions | Scribes are encouraged; patients consent before AI records them | How far the tools can be trusted in a given clinical moment | The physician stays accountable for what the record says |
| Czech media decalogue | Aug 2025 | Rules about rules | Every signatory will have internal rules for working with AI tools | The rules themselves, and the implementation memorandum still to come | Meeting the European Media Freedom Act through self-regulation rather than law |
| PRSA | Oct 2025 | Audience rights | Disclose meaningfully, case by case, rather than with one site-wide notice | Exactly when a use crosses into needing disclosure | Audiences are owed the knowledge of who is speaking |
| Charles University | Page last updated Sep 2026 | Permission with conditions | Every thesis carries a declaration of AI use; students may decline AI entirely | Faculties set their own limits; a whole tier of use requires a conversation with a supervisor | No single rule can fit every discipline |
| University of Cambridge | Not dated | Output guardrails | Research and ideas yes; nothing published that is fully AI-written, no cloned voices or faces | Not stated | Not stated |
Every linked policy was read at its source in September 2026. The comparison records what each document states plainly, leaves open, and gives as its reason. Cambridge’s page carries no publication date.
Turn your AI principles into language people can understand and respond to.