Artificial intelligence in employee advocacy programs – what to consider under the AI Act?

An employee prepares a post with the help of AI, marketing provides materials, and the company encourages advocates to publish regularly. At first glance, this looks like a natural evolution of an employee advocacy program. But questions quickly arise: does such a post need to be labelled? Who is responsible for its content? Is it enough for the employee to click “approve”? The AI Act does not prohibit the use of artificial intelligence to support employee advocates, nor does it introduce a single mandatory label for every publication. It does, however, require organizations to distinguish between different ways AI is used. For the person running the program, the key is therefore to determine what is created with the help of the tool, who reviews it, and what the audience should know about the publication.
Who is this article for?
This article will be particularly useful if you:
- run an employee advocacy or social selling program;
- are responsible for marketing, communications, HR, or employer branding;
- have advocates who use AI to create posts and visuals;
- are considering video content featuring an employee’s digital likeness or voice;
- need rules that can be applied in day-to-day content work.
Start with the use case, then determine the obligations
Discussions about the AI Act often begin with the question: “Is our tool compliant?” It is a necessary question, but from the perspective of an employee advocacy program, it is not enough. The same system might help edit a post, generate a statement delivered by a digital advocate, or analyze employee behavior. Each use case requires a separate assessment.
The AI Act is an EU regulation governing certain uses of artificial intelligence based on the risks they create. As a rule, using AI to draft a conference post does not in itself mean that a high-risk AI system is being used. The situation may be different when AI is used to assess employees or make decisions about their employment. This distinction follows from the risk-based approach described by the European Commission.
For employee advocacy, this gives a simple starting point: describe the actual use cases before creating one broad “AI policy.” The list should include both features provided by the company and tools that advocates use independently when working on company-related content.
A useful question to ask is:
Do we know at which point AI is merely assisting the author, and at which point it starts creating content or an assessment that someone may treat as credible?
An employee’s personal profile does not determine responsibility
Employee advocacy has a distinctive feature: the company organizes the program, but content is published under employees’ names. It is easy, then, to assume that because the profile is personal, the use of AI is also the author’s private matter.
Who owns the LinkedIn account does not, by itself, determine whether the rules apply. The European Commission distinguishes between personal, non-professional use and professional activity. It also explains that when employees use a system under an organization’s instructions and control, the organization is the AI deployer; the employees do not thereby become separate deployers.
The program should therefore define the respective scope of action of the company and the advocate. If marketing provides the materials, selects the tool, and defines how it should be used, the discussion of responsibility should not end with “everyone posts at their own risk.” By contrast, an employee’s independent professional activity outside the program may require a different assessment.
The division of roles should be translated into day-to-day practice:
- Who prepares the source material?
- Who verifies the data?
- Who approves the use of a person’s likeness?
- Who receives reports of errors?
Advocates should know the answers to these questions before their first publication.
Does every AI-assisted post need to be labelled?
There is no general requirement to add “generated by AI” to every text created with the help of artificial intelligence. Article 50(4) applies, among other things, to AI-generated or AI-manipulated text published for the purpose of informing the public on matters of public interest. It provides an exception where the content has undergone human review or editorial control and a natural or legal person holds editorial responsibility for the publication. These two elements need to be considered together.
Context matters for employee advocates. A post sharing a personal observation about organizing meetings is not the same as an analysis of how new legislation affects employee rights. In the latter case, you need to assess whether the text is intended to inform the public about a matter of public interest. Calling it “marketing content” is not enough to close the assessment.
Transparency is also discussed in industry media. In its analysis of the AI Act in media, advertising, and PR, Wirtualne Media highlights the risk that synthetic content may mislead audiences. In employee advocacy, this issue is particularly relevant to the trust attached to an author’s name: readers may assume that the expert has verified the information they publish.
Human review must cover the substance
The exception for reviewed text may look like the simplest solution: add an approval button and the problem disappears. But approving the material alone does not show what was actually checked.
In its current guidance, the Commission indicates that a superficial review limited to spelling or grammar is not sufficient. Verification must cover the substance, while editorial responsibility means ultimate responsibility for the publication.
Imagine an advocate preparing a post about changes in employment law. AI produces fluent copy but invents an obligation that the law does not impose. The author edits the first sentence and publishes the whole piece. The post has a human author attached to the profile, but the incorrect information still reaches the audience.
The program should therefore define verification as including:
- checking figures, dates, quotes, and sources;
- assessing whether the conclusions are actually supported by the material;
- removing experiences and examples that the author cannot confirm;
- checking whether the subject falls within the reviewer’s area of competence;
- the option to reject the publication or refer it for additional review.
Not every post needs to go through the legal department. However, a sensible organizational approach is to identify topics that require additional review, such as interpretations of law, health information, or unpublished company results. This is a process recommendation, not a statutory list of posts requiring approval.
A digital advocate is a separate case
With visual content, the question does not end with whether someone checked the script. Imagine a video in which a digital version of an employee talks about a project. The voice sounds natural, the lip movement matches the speech, and the audience may take the material for a genuine recording.
If the material meets the definition of a deep fake, the AI deployer must disclose that it has been artificially generated or manipulated. The exception for editorially reviewed text is not a general exemption for such videos. The definition covers image, audio, and video content that may falsely appear to be authentic.
In an employee advocacy program, particular attention should be paid to:
- cloning an expert’s voice;
- a realistic avatar delivering prepared text;
- generated images suggesting that an employee attended an event;
- modifications to a recording that change the meaning of the person’s statement.
These examples need to be assessed in context. An abstract illustration accompanying a post creates different audience expectations than a realistic “conference photo” from an event the author never attended.
Program rules should also define, together with the employee, the scope in which their likeness or voice may be used. Consent to appear in a recording and disclosure to the audience that AI was used to address different issues. Neither should be treated as a substitute for the other.
How should AI-generated materials be labelled in practice?
The program owner should distinguish between two layers: technical marking of outputs by the system provider and information visible to the audience where the deployer is required to provide it.
An example disclosure for synthetic video could read:
Video generated using AI. The digital likeness and voice were used with the consent of [name].
This is suggested wording and should be adapted to the specific material. The sentence about consent should only be included if it is factually accurate.
Where disclosure is required under Article 50, the information must be clear and distinguishable and available no later than the recipient’s first exposure to the content. Metadata alone does not replace a visible or audible notice to the audience of a deep fake.
In practice, it is worth checking how the disclosure appears in the mobile feed, whether it remains readable after the material is shared, and whether it disappears when the description is expanded or truncated. A note buried in the program rules does not help someone watching the video on LinkedIn.
AI literacy: what should employee advocates learn?
AI literacy means having the skills needed to use artificial intelligence knowingly and responsibly. In an employee advocacy program, this translates into the ability to assess content, sources, and the limitations of the tool.
The AI literacy obligation has applied since February 2, 2025, but its wording was amended in 2026. Current Commission information points to taking measures to support the development of the competencies of people using AI on behalf of an organization. It does not require every individual to achieve one specific, defined level. This change is worth taking into account when relying on older guidance.
A sensible starting point for advocates is a workshop based on materials similar to what they actually publish. You can prepare a draft citing a nonexistent source, a text that invents an experience for the author, and an image suggesting a real event. Participants then identify the problem and propose a correction.
The training program should focus on several skills:
- identifying information that needs verification;
- comparing generated text against source material;
- distinguishing language editing from substantive review;
- assessing when disclosure of AI use is required;
- choosing safely what information can be entered into the tool.
Program measurement should not quietly turn into employee evaluation
Data on advocates’ activity can help identify which topics generate conversations and where support is needed. A report showing the number of posts does not in itself mean that a high-risk AI system is being used.
The situation needs to be reassessed, however, if AI is used to analyze activity in order to evaluate people’s work or behavior, influence promotions, or allocate tasks. Annex III to the AI Act covers certain uses of AI in employment and worker management, including monitoring and evaluating workers’ performance or behavior. Classification depends on the intended purpose of the system and the rules in Article 6.
For the program owner, this is an important organizational boundary. It is worth defining in advance what advocates’ data is used for, who can access it, and what conclusions should not be drawn from it. A low number of posts may result from project workload, the nature of the role, or a deliberate choice of topics. It should not automatically become a measure of loyalty to the company.
According to the Commission’s current timetable, the rules for high-risk systems in the areas covered by Annex III apply from December 2, 2027. This does not postpone the transparency obligations relating to published materials.
The AI Act does not replace confidentiality and data protection rules
An advocate may prepare a post that does not require an AI label while pasting confidential client information into a generator. They may also correctly label an image while using material for which the company does not have the necessary rights.
Assessing compliance with the AI Act is only one part of the publication process. You still need to take into account personal data protection, rights to the materials used, and confidentiality obligations. Wirtualne Media’s analysis for the communications industry also highlights these parallel risks.
A practical program rule may require advocates to use approved tools and materials intended for external communication. If an advocate wants to base a post on a client conversation or an internal presentation, they should first check whether it can be used. Simply removing the company name may not solve the problem: project details may still make the organization identifiable.
A simple workflow for an employee advocacy program
Rules are easiest to apply when they are part of the normal publication process. The proposed workflow can be divided into five stages.
1. The author provides the source material
The advocate starts with their own observation or an approved source. They know what information can be entered into the tool. They do not ask AI to invent experiences, results, or client statements.
2. AI helps prepare a draft
The tool structures or shortens the text, or suggests alternatives. The program owner makes it clear that an AI-generated draft must be reviewed before publication.
3. A human reviews the substance
The author or another designated person verifies the facts and the meaning of the statement. For more complex topics, the material is referred to a subject-matter expert. For texts covered by Article 50, the conditions of the exception relating to review and editorial responsibility also need to be considered.
4. The team assesses whether disclosure is required
Before publication, the team checks the type of material, how AI was used, and the requirements of the channel. For realistic visual and audio content, the assessment includes the risk that recipients may take it for authentic.
5. The advocate publishes and receives post-publication support
The program should identify a person to whom an error, concern, or audience question can be reported. If a problem arises, the team determines the correction and how to communicate the change. Pre-publication approval alone does not close the work of maintaining content credibility.
This process can also be used when advocates work with the mAIa assistant in Sharebee. Embedding AI in the program delivers the most value when it is accompanied by clear rules for working with content. The program owner should combine tool support with author education and clear allocation of responsibility. Access to AI features alone does not demonstrate that the entire program complies with the law.
AI Act and employee advocacy – summary
The AI Act should be translated into decisions made about specific content: does the publication require disclosure of AI use, who actually reviewed it, and who is responsible for its content? Advocates also need the right competencies and an awareness that analyzing their activity may have different consequences than using AI to help write a post.
A good program gives employees a clear way of working. Authors know which materials they can use, how to verify a draft, and when to ask for help. This allows them to publish more efficiently while retaining control over what appears under their name.
FAQ – AI Act and employee advocacy
1. Does the AI Act prohibit writing posts with the help of AI?
No. AI can be used to help prepare content. The scope of obligations depends on how the system is used and on the type of material. An employee advocacy program should include publication controls and clear rules for advocates.
2. Does every LinkedIn post need to be labelled?
There is no single requirement to label all AI-assisted text. For texts intended to inform the public on matters of public interest, the conditions in Article 50 are relevant, including the exception relating to human review or editorial control and responsibility for publication. The rules of the platform and the company should also be checked separately.
3. Is an employee’s approval of a post enough?
An approval button alone does not demonstrate that verification has taken place. The review should cover the substance. If the author is not qualified to assess the information, they need support from someone with the appropriate expertise.
4. Does an employee’s consent remove the need to label a digital avatar?
You should not assume so. If the material is a deep fake subject to a disclosure obligation, agreement on the use of the person’s likeness does not replace information that must be provided to the audience.
5. When did the transparency rules start to apply?
Article 50 has applied since August 2, 2026. The limited transition period for earlier systems concerns technical marking and detectability of outputs on the provider side; it is not a general deferral of the obligations that apply to publishers.



