She is four minutes into the pitch when the question lands. The reporter has been nodding along to the backgrounder — a tight, well-sourced two-pager the team turned around overnight — and then, almost casually, between sips of coffee: “Did you write this, or did AI?” It is not a hostile question. It is a routine one now. And in the half-second before she answers, the strategist realizes she does not actually have a policy. She has a reflex, and the reflex is to say “I wrote it,” which is mostly true, except for the part where she asked the Assistant to summarize three filings and tighten the lede.
That half-second is where most of the PR profession lives right now. AI is in the workflow — in the research, the first drafts, the message-testing, the summarization — and the disclosure norms have not caught up to the practice. The instinct is to treat the question as binary: did you use AI, yes or no, confess or deny. That framing is wrong, and it is the source of most of the anxiety. The honest professional answer is not a yes or a no. It is a question about materiality, and it has a structure.
In 2025 the Public Relations Society of America gave that structure a name. Its updated ethics guidance, “Promise & Pitfalls: The Ethical Use of AI for PR Practitioners”, does not tell practitioners to disclose everything, and it does not tell them to disclose nothing. It tells them to calibrate — to disclose in proportion to how much the AI shaped the substance the audience is relying on. This post turns that principle into a working decision framework: what PRSA actually says, a decision tree for the material-versus-assistive line, how the conversation differs with clients versus journalists, the FTC overlay nobody in PR can ignore, and what to do in the gray zone where most real work lives.
What PRSA actually says
The 2025 guidance is grounded in the same place the PRSA Code of Ethics has always been grounded: honesty, disclosure, and the free flow of accurate information. What changed is the application. The document treats generative AI not as a novel ethical category but as a new surface for old obligations — the duty not to deceive, the duty to verify, the duty to protect confidential information. The headline for practitioners, as PRNews summarized the update, is a disclosure protocol keyed to material use.
Three principles do most of the load-bearing work, and they are worth stating plainly because they cut against the reflexes.
Principle 01 · Disclosure scales with materiality
You disclose when AI materially shaped the substance — not every time you touch a tool.
Using AI to fix a typo or reformat a table is not a material use any reasonable audience would expect to be told about. Using it to generate the strategic argument, the data analysis, or the quoted-as-yours content is. The test is whether disclosure would change how the audience weighs what they are reading.
Principle 02 · Accuracy is non-delegable
The practitioner owns every fact, regardless of which tool produced it.
“The AI said so” is not a defense. PRSA's guidance assumes a human verifies the output before it goes anywhere — which means a disclosed AI assist is never a substitute for checking the work. Disclosure and verification are separate obligations, and you owe both.
Principle 03 · Confidentiality is a hard perimeter
Client and embargoed material does not go into tools that could expose or train on it.
This is the obligation that has nothing to do with the audience and everything to do with the client. It runs underneath every disclosure decision: before you decide whether to tell anyone you used AI, you have to have used it on infrastructure that honored the duty of confidence in the first place.
Read together, those principles dissolve the binary. The question is never “did you use AI.” The question is “did AI shape something the audience is relying on, in a way they would not expect, such that not saying so would mislead them.” That is a materiality test, and materiality tests have decision trees.
A disclosure decision tree
Plot any AI-assisted deliverable on two axes. On the horizontal: materiality — how much the AI shaped the substance, from cosmetic assistance to authoring the load-bearing content. On the vertical: what the audience reasonably expects, from “assumes a human wrote every word” to “knows and accepts AI is in the loop.” Four zones emerge, and each carries a different obligation.
The disclosure map
Materiality × Audience expectation
Rose · Disclose
AI authored the argument, the analysis, or quoted-as-human content the audience assumes a person produced.
Op-eds, bylined thought leadership, executive quotes, testimonial-style content. Non-disclosure here misleads.
Amber · Note it
AI shaped substance, but the audience already operates with AI in the room.
Client deliverables under an AI-use agreement, internal research decks. A standing note covers it — no per-document confession needed.
Slate · Judgment call
Light AI assist, but the audience assumes pure human authorship.
Editing, restructuring, tightening a human-written draft. Lean toward a norm, not a per-instance disclosure. The gray zone.
Emerald · No need
Cosmetic or back-office assist, and AI use is expected or invisible to the audience.
Spell-check, formatting, transcription, internal first-pass summaries. No reasonable audience expects to be told.
The diagonal is where the discipline lives. The top-right and bottom-left corners are easy: if AI authored the argument and the audience expects a human, you disclose; if AI fixed a typo and nobody expects otherwise, you do not. The interesting cases run along the other diagonal — high materiality but an audience that already accepts AI (handle it with a standing norm rather than a fresh disclosure each time), and low materiality but an audience that assumes pure human authorship (where most of the genuine gray-zone judgment calls cluster). The mistake practitioners make is treating the amber-and-slate cases like the rose case — over-disclosing in a way that reads as either anxious or performative — or treating them like the emerald case and saying nothing when a standing norm would have built trust.
Run the deliverable through three questions in order, and the zone resolves on its own: Did AI shape the substance, or only the surface? Does the audience already know AI is in the loop? And would a reasonable person feel misled to learn how it was made? If the answer to the last question is yes, you are in rose no matter what the first two said.
Telling clients: make it a norm, not a confession
With clients, disclosure should never be a surprise, because by the time a deliverable is in their hands the AI-use conversation should already be closed. The professional move is to make it standing policy, negotiated up front and written into the engagement, rather than a per-project disclosure that arrives feeling like an admission.
Practically, that means an AI-use clause in the scope or master services agreement that does three things. It states that the firm uses AI in its workflow — for research, drafting, summarization, and analysis. It states the perimeter: that confidential and embargoed client material is handled only on infrastructure that does not expose it or train external models on it. And it states the guarantee that matters most to a nervous client: that a human reviews and is accountable for every deliverable, so AI never ships unsupervised into their name. Once that clause exists, the disclosure question with clients mostly disappears. The amber zone — substantive AI use with an audience that already accepts it — is covered by the standing agreement.
What the clause does not cover, and where judgment returns, is the rose-zone content the client will put their own name on: the CEO's bylined op-ed, the founder's LinkedIn post, the quote attributed to a real executive. Here the obligation runs to the client's own integrity. If an executive is going to be quoted saying something an AI drafted, the executive needs to know, approve, and own it — because the disclosure exposure is theirs, and the relationship cost of a surprise is yours.
In Hone Studio
Disclosure decisions get easier when you can show your work. When the Assistant runs with the Knowledge Base in scope, every answer is grounded in your own approved materials — prior coverage, fact sheets, the client's own positioning — with inline citations that trace each claim back to a source passage. So an AI-assisted draft is traceable to firm-owned material rather than opaque generation, and the human-in-the-loop standard holds: every output is a draft a person reviews and signs off on, never something that ships on its own. That is precisely the posture PRSA's guidance assumes.
Telling journalists: credibility is the whole asset
With journalists the calculus inverts. Clients are in a contractual relationship where AI use can be normalized in advance. Journalists are not, and the currency you are spending with them is credibility, which is expensive to earn and cheap to lose. The reporter's coffee-cup question in the opening scene is not really about AI. It is about whether they can trust what you hand them.
The principle is the same — disclose material use — but the threshold is lower, because the asymmetric cost of being caught is so much higher. A backgrounder that AI summarized from public filings is emerald-to-slate: the substance is verifiable, the facts are the facts, and the journalist will check them anyway. You do not need to announce the workflow. But a “quote” from an executive that an AI actually composed, a statistic the AI produced that you did not trace to a primary source, a piece of analysis presented as your firm's original thinking that came wholesale from a model — those are rose, and a journalist who later discovers them does not recalibrate one story. They recalibrate you.
The operating rule with media is simple: never let a journalist learn from someone else how something was made. If there is any chance the AI's role in a piece is material enough to matter to the story, say so first, plainly, and move on. Volunteered disclosure reads as confidence. Discovered non-disclosure reads as deception, and in this profession deception is the one unrecoverable error.
The FTC overlay
For anything consumer-facing, PRSA's ethics are the floor, not the ceiling — the FTC's endorsement guidance turns some disclosure questions into legal ones. Endorsements and testimonials must reflect the honest opinions of a real person; a review, quote, or “customer” voice fabricated by AI is a deceptive practice, not a gray-zone judgment call. Material connections must be disclosed, and AI-generated personas presented as authentic people cross a line the Society's ethics and the government's enforcement agree on. When the audience is consumers and the content is endorsement-shaped, the disclosure decision is no longer yours to calibrate. It is the law's.
The gray zone: editing, summarizing, ideation
Most working PR practitioners do not spend their day in the rose corner generating fake quotes. They spend it in slate: using AI to tighten a paragraph they wrote, to summarize a forty-page filing into a usable brief, to surface five framings of a message so they can pick the strongest. These are the cases the binary framing handles worst, because the honest answer to “did you use AI” is “a little, on the parts that don't change what you're relying on” — which is true, defensible, and impossible to say in a single yes or no.
The way through the gray zone is to ask what the AI actually touched relative to what the audience is relying on. Three cases, three answers.
Gray-zone case · Editing
AI tightened a draft you wrote and reasoned through.
The thinking is yours; AI worked the prose. The substance the audience relies on did not change. This lands in slate — covered by a general professional norm of AI assistance, not a per-document disclosure. The line moves to rose only if the editing quietly rewrote the argument rather than the sentences.
Gray-zone case · Summarizing
AI compressed source documents into a brief you then verified.
The facts trace to the originals and you checked them, so accuracy is intact and the work is verifiable on demand. Emerald-to-slate: no disclosure obligation to a journalist, because what they rely on is the underlying source, not your workflow. The obligation that survives is the verification — a summary you did not check is a different problem entirely.
Gray-zone case · Ideation
AI generated message framings or angles you selected among.
The model proposed; you chose, refined, and own the result. The human judgment that matters — which framing is true, defensible, and on-strategy — stayed with you. Slate. The output is scaffolding for your decision, not the decision, and disclosure tracks the decision.
The unifying test across all three is the same one PRSA's materiality principle implies: did the AI change the substance the audience is relying on, or only the surface around it? Editing prose, summarizing sources you verify, and generating options you choose among leave the load-bearing substance in human hands. That is why they rarely require disclosure — not because they are hidden, but because there is nothing material to hide. The moment any of them tips into the AI authoring the argument, the statistic, or the quote, it slides into rose, and the obligation slides with it.
Building the norm before you need it
The strategist in the opening scene was caught flat-footed not because she had done anything wrong, but because she was making an ethics decision in real time, under social pressure, with no framework loaded. That is the avoidable failure. Every other professional judgment in PR — embargo handling, conflict checks, correction policy — is decided in advance and applied in the moment. AI disclosure should be no different. A firm that has settled its disclosure posture before the reporter asks does not have a half-second of panic. It has an answer.
That answer is not a single rule. It is a calibration: an AI-use clause with clients, a default-disclose reflex for material use with journalists, a hard confidentiality perimeter underneath both, and a clear-eyed read of the gray zone that neither over-confesses the cosmetic nor under-discloses the substantive. The firms that get this right will treat the framework the way they treat their style guide — written down, taught to every account lead, and revisited as the tools and the audience expectations move.
In Hone Studio
A disclosure norm is only as credible as the infrastructure behind it. Because grounded outputs trace back to your own approved source material rather than opaque generation, the firm can stand behind “AI-assisted, fully verified, human-approved” as a true statement rather than a hopeful one. And the confidentiality principle that runs underneath every disclosure decision is structural here, not aspirational: every client gets their own database, their own infrastructure, their own deployment, and your documents never touch another client's system — and your data is never used to train AI models, per every provider's commercial API terms. The posture you disclose to clients and journalists is the posture the platform actually enforces.
Disclosure, in the end, is not a confession. It is the opposite. The practitioner who can say exactly how a deliverable was made — which parts were AI-assisted, which judgment was human, where the facts came from, who signed off — is demonstrating control, not admitting fault. The anxiety in that half-second before the answer comes from imagining disclosure as an admission of having cut a corner. Reframe it. The firm that can show its work is the firm worth trusting with the work in the first place. In a profession whose entire asset is credibility, being able to answer the disclosure question well is not a liability to manage. It is the competitive advantage.