Field guide · AI in practice
Prompt to Publish
A field guide to AI for university communicators. What these tools are, how to direct them, how to verify them, and how to know when to put them down.
If you read nothing else
The whole guide on one page
Copy it. Pin it up. Everything after this section explains why it works.
Build every prompt from four parts
Who is the AI writing as?
Role
Sets register, vocabulary and audience in one move.
What can it not guess?
Context
Names, facts and materials it has no way to know.
What is the deliverable?
Task
Format, length and components.
What are the boundaries?
Constraints
Style rules, prohibitions, what it must not invent.
Verify in three steps, every time
- Source check. No openable source, no confirmed claim.
- Claim isolation. Highlight every number, name, date, title and quote.
- Independent confirmation. Check against sources the AI did not give you.
The more confident it sounds, the more it deserves checking. Polish is not proof.
The eight principles
- Use AI to support judgment, not replace it.
- Never confuse fluency with accuracy.
- Protect private and personally identifying information.
- Do not fabricate institutional reality.
- Verify claims according to their potential harm.
- Disclose synthetic content when audiences could be misled.
- Preserve authentic voice and representation.
- The person publishing the work owns the result.
Direction going in. Doubt coming out. Everything in between is technique.
Part one
Understanding AI
You cannot calibrate trust in a tool you cannot describe.
What these tools actually are
Artificial intelligence is not one thing. It is a family of techniques for getting computers to do tasks that normally require human judgment. The AI in your phone's autocorrect, the AI that flags a fraudulent charge, and the AI drafting your story are different technologies with different failure modes.
This guide concerns the newest branch: generative AI, systems that produce new text, images, audio and video rather than sorting or scoring what already exists. Inside it sit large language models, the engine behind ChatGPT, Claude, Gemini and Copilot.
large language model (LLM) A system trained on massive text collections to predict plausible language.
The system that writes your draft has no concept of true. It has a concept of likely.
The mental model that matters: prediction, not retrieval
01
A large language model is a prediction engine. It learns the statistical patterns of language, then generates the most plausible next words given everything that came before.
02
It does not inherently retrieve verified facts. It generates from a compressed statistical impression of its training data, unless it is connected to search, uploaded documents or another retrieval system.
03
That is why the same system can draft a beautiful paragraph and invent a citation with identical confidence. Fluency and accuracy come from the same mechanism, so one cannot vouch for the other.
Seventy years in one page
The overnight revolution was seventy years in the making, with two winters in the middle.
| 1950 | Alan Turing asks whether machines can think and proposes the imitation game. |
| 1956 | The Dartmouth workshop coins the term artificial intelligence. |
| 1966 | ELIZA, the first chatbot, fools users with pattern-matched therapy scripts. |
| 1970s | The first AI winter: promises outrun hardware and funding collapses. |
| 1980s | Expert systems boom in industry, then bust. The second winter. |
| 1997 | IBM's Deep Blue defeats world chess champion Garry Kasparov. |
| 2012 | AlexNet starts the deep learning era. Neural networks finally have the data and compute they need. |
| 2016 | AlphaGo defeats Lee Sedol at Go, a game long thought safe from machines. |
| 2017 | The transformer architecture is published. It becomes the T in GPT. |
| 2022 | ChatGPT reaches 100 million users in two months, a record for consumer applications at the time. |
| 2024 to 2026 | Multimodal systems, AI Overviews and early agentic AI move the question from what it can make to what we should let it do. |
Two lessons hide in this timeline. Capability arrives in sudden jumps after long quiet periods, and every boom has carried predictions that aged badly. Hold both.
What comes next, honestly
Predictions about AI age like milk. Open questions age like wine. These are the tensions worth tracking.
Agentic AI
Systems that take multi-step actions rather than answering single prompts. Who checks work that no person saw midstream? Picture an assistant that drafts the pitch, assembles the source list and files the photo request as one chain, with a person reviewing only the end product.
Search is being rewritten
Generative engine optimization is emerging alongside SEO as AI Overviews absorb search traffic. When a parent asks a chatbot which universities offer a program, whether yours appears depends on how clearly your pages state it.
Regulation is unsettled
As of 2026, the EU regulates AI through a risk-based framework while the prevailing US federal posture emphasizes acceleration. Institutional policy cannot wait for the law to settle.
The authenticity counterweight
Provenance is becoming a technical discipline: edit-history logging, metadata and rights signals, file-level tagging. Newer professional cameras can attach a tamper-evident capture and edit log to every frame.
Content creation is decentralizing
Faculty and departments can now generate podcasts, videos and articles without contacting a communications office. Off-the-shelf tools can turn a research paper into a passable two-voice podcast in an afternoon.
Lead with standards, or inherit the cleanup.
Part two
Why this matters now
The tools arrived before the training did.
University communications teams adopted generative AI quickly, unevenly and without a shared playbook. Staff use it daily to draft stories, caption images and repurpose content, while the standards for that use still live mostly in individual judgment.
Three conditions make this urgent
Adoption runs ahead of governance
Verification practices vary desk to desk.
University content carries institutional liability
A hallucinated statistic in an enrollment story is a different kind of error than one in a dinner recipe.
Trust grows through structure
Trust in AI grows through structured, hands-on development, not unstructured exposure.
Standards that live in individual judgment are not standards yet.
What practitioners are asking
This guide was shaped by a professional panel on AI and university photography at the UPAA Symposium in June 2026, with colleagues from Michigan, Missouri S&T, Arizona State and commercial freelance practice. The questions that came up again and again:
- Will AI support our work, or be used to eliminate positions?
- How do we protect authenticity in stories about real students and faculty?
- When does AI use require disclosure?
- Who is accountable when AI-generated content is wrong?
What this guide teaches
- Construct structured prompts with the four-part Prompt Architecture Blueprint, measured by a simple before-and-after revision count.
- Evaluate AI-generated content for accuracy with a three-step verification protocol before anything reaches publication.
- Identify the credibility illusion by comparing responses to neutral, directive and emotionally framed prompts.
- Assess ethical, legal, authenticity and disclosure risks, and decide when AI should assist, when human review is required, and when AI should not be used.
How this guide was made. In the spirit of the disclosure standards it teaches: the Prompt Architecture Blueprint began as a working tool I built for colleagues in a university communications office, and this guide is its first formal write-up, first produced as graduate coursework at Kent State. Portions were drafted with AI assistance, then verified, restructured and rewritten by me.
Part three
The craft
The working model
- These systems predict plausible language. They do not retrieve verified facts.
- Fluency and accuracy are produced independently. Confidence is not accuracy.
- Hallucination is not a defect one patch removes. It comes from the same predictive mechanism that produces fluent language, though retrieval, stronger models and verification can reduce the risk.
Failure mode one
Overtrust
It sounds right, so it is.
Failure mode two
Undertrust
It is dangerous, so never.
The goal is calibrated trust: use AI where it creates real value, verify according to risk, and decline to use it when authenticity, privacy or institutional credibility would be compromised.
The Prompt Architecture Blueprint
Four parts, every prompt. The framework also works as a standardized prompt intake form for a team.
Who is the AI writing as?
Role
A role sets register, vocabulary and audience assumptions in one move.
What can it not guess?
Context
Names, facts, relationships and materials the model has no way to know. Context is where hallucination gets starved.
What exactly is the deliverable?
Task
Format, length, components. Vague tasks return vague drafts.
What are the boundaries?
Constraints
Style rules, prohibitions, things it must not invent. Constraints are where institutional risk gets fenced.
The more precise the constraints, the fewer revision cycles the draft needs. Precision up front is also lighter on the planet, a point Part five returns to.
Role: [who the AI writes as, and for which audience] Context: [names, facts and materials it cannot know] Task: [the exact deliverable: format, length, parts] Constraints: [style rules, what it must not invent, what to flag]
The gallery: weak and strong, side by side
Frameworks tell. Examples teach. Five real communications tasks, each run both ways.
1 · The feature story
WEAK
Write a story about the aquatic ecology lab.
Generic prose, invented atmosphere, a fabricated quote from an unnamed student, no sense of audience. Twenty minutes of cleanup.
STRONG
You are a senior writer for a university news site writing for alumni and prospective families. Context: the university's aquatic ecology lab hosts undergraduate researchers; the story subject is a junior biology major; the piece runs alongside three photos. Task: a 400-word feature draft with a headline and two subhead options. Constraints: no superlatives about the university, no invented quotes, AP style, do not state research findings the source material does not contain.
The draft arrives in register, at length, and quote-free where no quotes were supplied. The remaining work is voice, not triage.
2 · The caption set
WEAK
Write a fun Instagram caption for this rocket photo.
Exclamation points, a rocket emoji, and a claim the launch "broke records," which nobody said.
STRONG
Role: social media writer for a university account with a warm, wry voice. Context: photo shows the liquid-fueled rocketry team conducting a cold-weather field test; students pictured have signed releases; no launch occurred, this is instrumentation work. Task: three caption options under 125 characters plus five hashtag suggestions. Constraints: no achievement claims, no emoji, do not name students.
The model cannot invent a triumph, because the constraint forbids claims and the context says what actually happened.
3 · The summary
WEAK
Summarize this press release.
A competent summary that quietly promotes the study's hoped-for outcome into a finding.
STRONG
Role: newsletter editor for internal university staff. Context: the attached announcement describes a grant awarded to study X; results do not exist yet. Task: a 150-word item, one quote retained verbatim from the provided text. Constraints: describe the research as beginning, not concluding; every number must appear in the source; flag anything you were unsure about below the draft.
Requiring every number to trace to the source turns the model from author into summarizer. The flag request makes uncertainty visible instead of smoothing it over.
4 · The shot list
WEAK
Give me a shot list for an accreditation story.
A plausible, generic list that misses everything specific to the visit.
STRONG
Role: photo editor planning documentary coverage. Context: three-day site visit; the college requests emphasis on student-faculty interaction in the new simulation lab; two reviewers have declined to be photographed. Task: a shot list of 12 setups ordered by schedule, each with location, subjects and a one-line rationale. Constraints: nothing staged to look like instruction that is not occurring; note where releases are required.
The no-staging constraint is the authenticity standard from Part five, written into the prompt itself.
5 · The image generation request
WEAK
Make an image of diverse students in a science lab.
Stock-adjacent synthetic people, representation flattened to a template, and an image one forward away from being mistaken for documentation.
STRONG
Role: illustrator creating clearly stylized concept art for an internal slide, not for external use. Context: the slide discusses a proposed lab renovation. Task: one illustration in a flat, graphic, obviously non-photographic style. Constraints: no photorealism, no recognizable individuals, label output as AI-generated in the file name.
Style constraints keep synthetic imagery visibly synthetic. The file-name label makes the disclosure travel with the asset.
Ethics can be written into the prompt.
How does AI bluff?
No detector catches everything. Pattern recognition catches a lot.
01
Unverifiable precision
Suspiciously specific statistics with no source attached. Precision is not provenance.
02
The invented citation
Real journal, real-sounding authors, page numbers and all. Always resolve the DOI.
03
Uniform confidence
Human experts hedge unevenly; models hedge evenly or not at all. A draft with no uncertainty anywhere did not weigh any.
04
Studies-show syndrome
"Research consistently shows" followed by nothing checkable.
05
The agreeable mirror
The output adopts your framing, including your errors. Ask for the strongest case against your own premise as a stress test.
06
Stale facts, fresh tone
Confident present-tense claims about roles, prices and policies that change. Anything time-sensitive gets verified by default.
The credibility illusion
The heart of this guide is a finding from my own course research. Three prompts on the same topic, one neutral, one directive and one emotionally framed, went to the same AI model in separate conversations. The results inverted expectation: the neutral prompt produced the most confident, least qualified response, the emotionally framed prompt produced the most careful one, and none of the three cited a single source.
| Prompt framing | Sounds authoritative | Epistemic care | Sources cited |
|---|---|---|---|
| Neutral | Highest | Lowest | 0 |
| Directive | Middle | Middle | 0 |
| Emotionally framed | Lowest | Highest | 0 |
Checking against institutional guidance and survey research surfaced what even the most careful response missed, including 2025 survey data on 760 adult learners showing 56 percent were more likely to trust institutions cited in Google AI Overviews. That complicates the simple story that AI erodes trust.
credibility illusion The gap between how authoritative an AI response sounds and how much epistemic work it actually did.
The test takes ten minutes with three prompts and one model, and anyone on a team can replicate it.
The three-step verification protocol
- Source check. Ask where each factual claim came from. If the AI cannot point to a source you can open, treat the claim as unconfirmed.
- Claim isolation. Pull every checkable statement out of the draft: numbers, names, dates, titles, rankings, quotes. Highlight them. A claim you have not isolated is a claim you have not checked.
- Independent confirmation. Confirm against sources that did not come from the AI. Rank them: institutional guidance and survey data outrank commentary. Two AI tools agreeing is not confirmation.
The more authoritative an output sounds, the more verification it deserves.
Part four
The workshop
Ninety minutes, most of it in participants' hands. Hands-on and discussion blocks fill 65 of the 90 minutes.
The hook
Two AI answers to the same question, one confident and wrong, one hedged and right. The room votes before knowing which is which.
How these tools work
Prediction versus retrieval, calibrated trust, and the credibility illusion reveal.
Hands-on one: rebuild your prompt
Participants bring one current draft or task and rebuild its prompt with the Blueprint, comparing before and after.
Hands-on two: the seeded-error drill
Participants hunt for three planted errors, obvious, plausible and subtle, in an AI-drafted university story.
The spectrum and scenarios
Small groups place tasks on the risk spectrum, then work through one publish-or-don't scenario.
Reflection and handoff
Exit ticket naming one workflow change for this week. The one-page references go home with everyone.
The AI use spectrum
- Brainstorming headlines
- Building shot lists
- Reformatting approved text
- Generating interview questions
- Drafting story copy
- Summarizing source material
- Drafting captions
- Repurposing content
- Generating fictional students
- Fabricating campus scenes
- Altering documentary reality
- Publishing unverified claims
As representational or institutional risk rises, verification, approval and disclosure requirements rise with it.
Three scenarios for small groups
The fabricated caption
An AI caption states a student's major, hometown and program. Two details were inferred, not provided. What must be verified? Who owns the correction? Is it publishable after revision? Does it need disclosure?
The synthetic campus image
A department requests a generated photo of diverse students in a lab because no photography exists. Could viewers believe it is real? Does a label fix it? Is synthetic imagery appropriate for recruitment at all?
The efficiency request
A manager proposes generating large volumes of social content to save staff time. What does efficient mean here? Which tasks automate safely? Which still need a person?
How learning gets measured
- Pre and post survey: five items tracking confidence and knowledge.
- Performance artifact: every participant leaves with one rebuilt prompt and one verified draft from their real workload, plus a before-and-after revision count.
- Exit ticket: one named workflow change within the week.
- 30-day follow-up: actual adoption, which feeds the case for scaling.
Want a deeper session on one skill? Checking AI-generated quotes is a companion 90-minute workshop built entirely around verification.
Part five
Judgment
Privacy and student stories
Three layers, kept honest:
- Law. FERPA governs education records and does not always reach marketing content.
- Policy. Whatever your office formally adopts, which this guide is meant to help build.
- This guide's recommendation. Do not enter student PII, health information, protected academic-record details or other restricted institutional data into an AI tool unless the use is authorized, the tool is approved for that data, and any required consent has been obtained. Consent to the story itself remains a separate ethical expectation.
Know which layer you are standing on before anything gets pasted.
Accuracy and institutional liability
Rankings, research findings, accreditation and outcomes are claims with legal and reputational weight. Because of the credibility illusion, the most dangerous errors arrive in the most convincing packaging.
Disclosure and authenticity are different questions
Disclosure asks
Have we told the audience how this was made?
Authenticity asks
Even with disclosure, is this an appropriate representation of the institution?
A proposed working standard: AI-assisted drafting that a person has verified and rewritten needs no public disclosure. AI-generated content published substantially unedited must be disclosed. AI-generated imagery is always disclosed. This is a starting position for discussion, not settled policy.
Disclosure is not a magic ethical disinfectant. A photograph carries an implied claim that something existed in front of the camera. University photography often works as documentation, and prospective students read our images as evidence. The field's cautionary examples are real: a university billboard whose diverse student cast attended other schools, and an alumni newsletter that ran a real photograph through AI, cartooned everyone in it, and reached 77,000 inboxes.
provenance The documented origin and edit history of a piece of content. It establishes how something was made, not that an image depicts reality truthfully. In photography it is increasingly machine-readable.
Provenance is becoming infrastructure. Cameras that log capture and edit history, file-level metadata, and rights signals are turning authenticity from a value into a documentable property, where platforms choose to honor those signals. Communicators who adopt provenance practices early will hold the receipts everyone soon wants.
When not to use AI
The most important AI skill nobody teaches. Some assignments are wrong for these tools, full stop.
Documentary reality
News photography, event coverage, anything presented as a record of what happened. Generation and heavy synthetic editing end the claim to documentation.
Unconsented personal stories
A student's health, family circumstances or academic record do not enter a model unless the use is authorized, the tool is approved for that data, and every required consent, including consent to the story itself, has been obtained.
Crisis communications
Speed, pressure, institutional stakes and hallucination risk are the worst possible combination. AI should not originate crisis facts, official positions or safety instructions. Accountable people write and approve the final language.
Legal, accreditation and compliance language
Text with regulatory consequences gets drafted and reviewed by the people accountable for it.
Anything you cannot verify
If no source exists to check a claim against, the claim does not publish. This is not an AI rule. AI just industrialized the temptation.
The honest cost of saying no
It would be dishonest to pretend declining is free. A freelancer on the UPAA panel put it plainly: refuse a client's AI request once, and you may never be hired again. At many institutions, pressure to adopt AI increasingly comes from senior leadership. So the skill is not blanket refusal. It is knowing which uses are worth spending professional capital to push back on, and having the verification and authenticity language to make that case credibly.
The judgment call happens at a single desk, where one person decides what publishes.
Equity and representation
Generative systems trained mostly on Western, English-language data flatten the voices least represented in that data. Default AI voice is not neutral voice. Representation is a verification item alongside facts.
Environmental cost
By common estimates, a generative AI query draws roughly five times the electricity of a conventional web search, and data center cooling uses about two liters of water per kilowatt-hour. Estimates vary by model and infrastructure, but the direction is not in dispute. The argument is not against use. It is for intentional use, which the Blueprint already encourages: better outputs in fewer attempts.
Part six
Building a policy
This is not a policy. It is the load-bearing questions any good one has to answer.
Start from stated values
Where institutional leadership has articulated a stance on AI, anchor the policy to it. A policy that puts leadership's own language into practice is a policy leadership can adopt. Arizona State offers one model, having named its computing school for augmented intelligence. Where no stated position exists yet, the policy process itself is the right occasion to ask leadership for one.
Eight questions a working policy must answer
01
Scope
Which tools, which teams, which content types? A policy that covers everything governs nothing.
02
Risk-to-review mapping
Adopt the spectrum. As representational and institutional risk rises, so do verification, approval and disclosure. Write the approval path for each tier, with names attached.
03
Disclosure tiers
When is AI assistance silent, when is it credited, when is content labeled synthetic? The working standard in Part five is a starting draft.
04
Data and tool vetting
Which tools retain pasted content for training? Who maintains the approved list? What never gets pasted anywhere, regardless of tool?
05
Provenance and metadata standards
Decide now how imagery is tagged: capture history, edit logging, AI-training opt-outs. Cheap to adopt early, expensive to retrofit.
06
Training and literacy
Policy without training is shelfware. The workshop in Part four is a ready-made onboarding unit, repeatable team by team.
07
Review cadence and ownership
Name an owner. Set a review date measured in months, not years. The tools will lap an annual policy.
08
The escalation path
Where does a staff member take a judgment call the policy does not cover? If the answer is nowhere, the policy will be worked around.
Policy sets the floor. Literacy sets the ceiling.
Rolling it out
| Format | A 90-minute hybrid workshop, repeatable for each content team. |
| Pilot | One team first, revise, then scale. |
| Materials | The workshop deck, a handout, the verification protocol, the Blueprint template and the principles as standalone one-pagers. |
| Cost | Minimal. Tools are already licensed or free, and facilitation and follow-up use existing staff time. |
| Follow-up | A 30-day adoption survey and optional office hours. |
| Next in the series | Image generation standards, then AI-assisted digital asset management. |
From prompt to publish
The easy version of AI training teaches people where the buttons are. This guide aims at something more durable: a shared standard of practice that lets a communications team move faster without publishing carelessly, built on a framework its members already recognize and a finding they can test themselves in ten minutes.
Prompting is only the beginning. Verification, disclosure, privacy, authenticity, provenance and human accountability are the skills that carry an AI-assisted draft safely from prompt to publish.
AI can help us work faster. Our responsibility is to make sure that speed does not outrun truth, authenticity or judgment.
Reference
Glossary
- agentic AI
- Systems that plan and carry out multi-step tasks with limited human involvement between steps.
- AI Overviews
- Generated summaries at the top of search results, increasingly the first thing prospects read about an institution.
- artificial intelligence
- The family of techniques for performing tasks that normally require human judgment.
- calibrated trust
- Trust proportioned to evidence and risk. The corrective to overtrust and undertrust alike.
- credibility illusion
- The gap between how authoritative an AI response sounds and how much epistemic work it did.
- epistemic
- Relating to knowledge and how carefully it is handled.
- fine-tuning
- Additional training that specializes a general model for a narrower task or voice.
- generative AI
- Systems that produce new text, images, audio or video rather than classifying existing content.
- GEO
- Generative engine optimization: shaping content so AI systems cite and summarize it accurately.
- hallucination
- A fluent, plausible statement generated without factual basis.
- large language model (LLM)
- A system trained on massive text collections to predict plausible language.
- machine learning
- Techniques that learn patterns from data rather than following hand-written rules.
- multimodal
- A model that works across text, images, audio and video in one system.
- prompt framing
- The tone and stance a request carries. Framing changes how carefully the answer handles knowledge.
- provenance
- The documented origin and edit history of content. In photography, increasingly machine-readable.
- token
- The word-fragment unit a language model actually reads and predicts.
- training data
- The text and images a model learned its patterns from. The source of both its fluency and its blind spots.
- transformer
- The 2017 architecture behind modern language models.
Reference
References
- Birhane, A. (2020). Algorithmic colonization of Africa. SCRIPTed, 17(2), 389 to 409.
- Daud, R. (2026). Prompt framing and the credibility illusion: A comparison analysis. MDJ 61001 course study, Kent State University.
- European Parliament. (2024). EU Artificial Intelligence Act.
- Fernández-Batanero, J. M., et al. (2022). Digital competences for teacher professional development. European Journal of Teacher Education, 45(4), 513 to 531.
- Guedon, C. (2025, October 13). AI search and student trust. Search Influence.
- Nazaretsky, T., et al. (2022). Teachers' trust in AI-powered educational technology. British Journal of Educational Technology, 53(4), 914 to 931.
- NIST. (2023). AI Risk Management Framework. (2024). Generative AI Profile.
- Perry, E., & Booth, J. (2024). The practices of professional development facilitators. Professional Development in Education, 50(1), 144 to 156.
- UNESCO. (2021). Recommendation on the Ethics of Artificial Intelligence.
- UPCEA. (2025, October 24). Adult learner survey on AI search behavior.
- White House. (2025). America's AI Action Plan.
- Zewe, A. (2025). Explained: Generative AI's environmental impact. MIT News.
- Professional context: UPAA Symposium panel, June 2026, Ann Arbor, Michigan.