AI-generated marketing content needs a written disclosure policy
AI-assisted research. This note was prepared with the assistance of AI tools. It may contain errors and is not professional advice. See our Legal Disclosures.
As agents produce more marketing content, teams face a question that barely existed a few years ago: when to tell people that AI was involved. The right call depends on the type of content, the channel, the audience, and the rules that apply where you operate.
Those rules are still changing. Some platforms have their own labeling requirements, especially for images, video, and content that depicts real people, and some jurisdictions are adding rules of their own. Without a written policy, each writer, reviewer, and agent makes the call differently. In our research, we treat disclosure as a policy decision made once and applied consistently, rather than a judgment made piece by piece.
This post is not legal advice. It describes how we think about building a sensible internal policy, which you should then review with your legal team before relying on it.
Why disclosure matters
Disclosure decisions affect three things a marketing team cares about, and each points in a slightly different direction.
Trust. Audiences who later discover undisclosed AI involvement in content they assumed was personal or human-written may feel misled. That reaction tends to attach to the brand, not the piece.
Compliance. Some jurisdictions and platforms require labels for certain kinds of AI-generated or altered content, particularly synthetic media and endorsements. What applies to you depends on where you publish and what you publish, so confirm the specifics with counsel rather than relying on a general summary.
Accountability. Clear internal records of AI involvement make it easier to answer questions from customers, partners, or regulators. Even where no public label is needed, the internal record usually is.
A framework for deciding
We find it helpful to consider two factors together. The first is how much AI shaped the content. We use three levels.
- Assisted: a person wrote it, and AI helped with editing or ideas.
- Drafted: AI produced a draft that a person substantially reviewed and edited.
- Generated: AI produced the content with light or no human editing.
The second factor is how personal or consequential the content is. Again, three levels cover most cases.
- Low: general informational content and internal material.
- Medium: marketing emails, social posts, and product descriptions.
- High: content presented as a personal message, testimonials, synthetic images or video of people, or content making important claims.
Disclosure becomes more important as both factors increase. Assisted, low-consequence content rarely raises the question. Generated content presented as a personal message from a named individual is the clearest case for disclosure, or for not producing it at all.
Applying the framework
Most real decisions fall in the middle, so it helps to work through one. Consider a hypothetical retailer whose agents draft product descriptions and a monthly email signed by the founder.
The product descriptions are drafted by an agent and edited by a merchandiser. They sit at drafted and medium. The team decides that a general statement on its site about how it uses AI is enough for these. It records AI involvement for each description in its content system and checks platform rules for any marketplace listings.
The founder's email is harder. If an agent writes it and the founder lightly approves it, the email is generated content presented as a personal message, at the high end of both factors. The team has two honest options.
The founder can write the email with AI help, which moves it to assisted. Or the email can carry a clear note on how it was produced. Sending it as written, under the founder's name, without either change is the option we recommend against. The issue is less the use of AI than the gap between what the reader assumes and what happened.
Lines we recommend not crossing
Some uses carry enough risk to trust that we suggest ruling them out entirely, whatever the disclosure. We recommend writing these into the policy as fixed limits.
- No fake social proof: never create testimonials, reviews, or customer quotes.
- No synthetic people without consent: never present synthetic images or video of real people without their consent and clear labeling.
- No hidden impersonation: never have agents impersonate a specific employee in direct conversations without disclosure.
- No unverified claims: never publish AI-generated claims a person has not checked.
Several of these limits can be enforced mechanically. A check that blocks any content containing a quotation attributed to a customer, unless it links to an approved source, is simple to build. Our post on guardrails as code describes the approach, and confirming the wording of any required label with counsel before encoding it saves rework.
Internal practices
A policy works only if the team can follow it without stopping to interpret it. We suggest five practices.
- Record involvement: log the level of AI involvement for each published piece in your content system.
- Follow platform rules: apply each platform's labeling requirements for images and video.
- Keep a standard statement: write a short, plain-language disclosure your team can use where appropriate.
- Tag at the source: have agents attach their involvement level to every draft, so the record does not depend on memory.
- Review twice a year: revisit the policy at least twice a year, since rules and norms are changing.
The involvement record becomes part of your audit trail. If a customer asks how a piece was produced, the answer should come from a record, not a recollection. See audit trails that satisfy legal, brand, and finance for how to structure it, and our quarterly governance checklist for where the policy review fits.
What to measure
Disclosure policies are easy to write and easy to drift from. We suggest two checks. First, sample published pieces each quarter and confirm that each has a recorded involvement level that matches how it was made. Second, track how often reviewers flag a piece as needing a disclosure decision the policy does not cover.
The second measure shows where the policy is thin. A steady stream of uncovered cases in one content type, such as short-form video, is a sign to extend the framework there.
The open question we keep returning to is how audiences read disclosure statements. A label meant to build trust may be read as a warning, or ignored entirely. We plan to study whether the wording and placement of a plain disclosure change how readers judge the content it accompanies.
How to apply this: draft a one-page disclosure policy using the two factors above, list the lines you will not cross, and have agents tag their involvement on every draft. Review the policy with legal counsel before your agents produce more customer-facing content.
Related research
Roles and permissions for a mixed human and agent team: name who can act, approve, and stop
When agents join the team, someone has to decide who can do what. Clear roles and permissions keep people and agents accountable and stop access drifting.
Escalation paths: agents should know when to stop and ask
Good agents know their limits. Clear escalation rules, and a fast way for people to resolve what agents raise, prevent many of the mistakes that damage trust.
Red-teaming marketing agents: find the failure modes before customers do
Deliberately trying to make your agents fail is one of the fastest ways to make them safer. Here is a practical red-teaming process for marketing teams.
Put this research to work
Marketix turns these patterns into software: agents, automated workflows, and a hub that keeps people in control. Sign up for early access.
