Czas czytania: 10 minut
Are hyperrealistic AI avatars worth adding to your corporate training?
Available research does not show that a more realistic avatar teaches more than a standard or stylized one, and since 2 August 2026 a video featuring a realistic replica of a person must disclose, from the start, that it was generated with AI.
The health and safety manager at an industrial company has an idea for the annual safety course: have the plant director present it. Not in person, since they have no time to record, but through a hyperrealistic avatar built from a few minutes of their video. The workforce knows them, respects them and listens to them. It looks like the best way to get the message across.
The idea makes sense, but it raises three questions worth answering before ordering anything: whether realism actually improves learning, what legal obligations it brings, and when it is worth the effort compared with a stock avatar or no presenter at all.
In this article we review what studies say about avatar realism, what the EU AI Act requires since August 2026, and four situations where hyperrealism is worth it.
Not all AI-generated presenters are the same, and the difference matters both for learning and for the law. These are the four options on the market today:
| Type of presenter | What it is | What it takes to create | Could it be a deep fake? |
|---|---|---|---|
| Replica of a real person | Hyperrealistic avatar built from a specific employee's image and voice | Between 20 seconds and 5 minutes of video, depending on the vendor, plus their consent | Very likely yes |
| Realistic stock avatar | Photorealistic synthetic person who doesn't correspond to anyone known | Nothing: you pick it in the tool | It can be, if it looks credible as a real person |
| Stylized avatar | 2D, 3D or cartoon-style character | Nothing: you pick it in the tool | In principle, no |
| No presenter | Narrated screens with text, images and voice | Nothing | No |
The right-hand column is this year's novelty. We explain it in detail below, but the summary is simple: the more the presenter resembles a real person, the more likely the video needs an AI-generated content label.
The question has more of an answer than it seems, because research on "pedagogical agents", characters that explain content on screen, goes back more than a decade. These are the most solid findings:
The pattern is consistent: realism does not add learning on its own; what it can add is presence, recognition and trust. Those are real effects, but it's worth knowing that this is what you pay for, not better retention. The studies also have clear limits: small samples, short videos and theoretical content, not operational procedures.
To dig into when an avatar is worth it compared with filming people, see our comparison of AI avatars vs traditional recording.
Article 50 of the AI Act has applied since 2 August 2026, and the Digital Omnibus adopted in July did not delay the obligation that affects whoever uses the videos.⁵ ⁶ It has two parts worth separating:
The question is whether a training video featuring an employee's replica is a deep fake. The regulation defines it as AI-generated content that resembles existing persons and would falsely appear authentic to the viewer. The European Commission's guidelines on Article 50, published in July 2026, largely settle the question with a very concrete example. Among the cases that are deep fakes, they list a video featuring a realistic synthetic avatar of a company CEO congratulating employees on the year's results.⁷
The same guidelines add three criteria that directly affect training:⁷
The practical consequence is that a course presented by a hyperrealistic replica of the plant director should carry a disclosure at the start of the video. This is not a minor formality: the regulation provides for fines of up to 15 million euros or 3% of worldwide turnover,⁵ and Spain's draft AI governance law, currently in parliament, classifies failure to disclose as a serious infringement.⁸
The disclosure does not replace other obligations: the person's consent, data protection and image rights still apply separately.⁷ We review the training obligations the same regulation brings in our article on Article 4 of the AI Act after the Digital Omnibus.
If realism doesn't improve retention and also brings obligations, when is it worth it? In the training projects we support, a hyperrealistic avatar of a real person makes sense in these four cases.
A safety policy presented by management, a culture change explained by the CEO, or a welcome message from the site manager. Here the value is not in the information but in the backing of someone with recognizable authority. That is exactly the social presence effect the studies do detect.
When the workforce knows and respects the technician who best masters a process, seeing them explain it strengthens trust in the procedure. The avatar lets that person "explain" without spending hours recording every time something changes.
In companies with sites in several countries, a replica that keeps the same person's face and voice in each language gives consistency to a corporate message. Recording it the traditional way would require that person to speak all those languages.
If the executive or expert has little availability and the content is updated several times a year, the replica avoids depending on their calendar for every change.
Outside these cases, a stock avatar, a stylized one or narrated screens with no presenter usually deliver the same learning at lower cost and with fewer obligations. That includes most operational procedures, where what people need to see is the machine, the step or the screen, not the person explaining it.
Before ordering the avatar, settle four points:
We cover quality risks and managing the avatar over time in detail in our guide to AI avatar risks and best practices.
A hyperrealistic avatar doesn't make people learn more. What it can do is deliver a specific message with the voice and face of someone the workforce trusts. That is the reason to use it, and also the measure of when it isn't needed.
For the safety manager in the example, the answer is probably yes, but only for part of the course: the plant director's introduction, with their consent and a clear disclosure at the start. The rest, the procedures, the risks and the measures, works just as well with a stock presenter or no presenter. Reserving hyperrealism for the moments when who speaks matters is the simplest way to benefit from it without overpaying or taking on unnecessary risk.
It depends on how realistic it is. The Commission's guidelines include realistic AI-generated human avatars within the concept of "person", and note that photorealism makes it more likely that content is a deep fake.⁷ With a photorealistic avatar, the prudent choice is to disclose. With a clearly stylized or cartoon character, in principle it isn't needed.
At the start of the video, clearly and distinguishably, and meeting accessibility requirements.⁵ ⁷ The voluntary EU code of practice on AI-generated content proposes common icons companies can use as labels.¹³
Both have different obligations. The vendor must technically mark generated content. The company publishing the video in its training is the one that must inform viewers, even if specific employees created it with the tool.⁵ ⁷
Yes. Spain's Organic Law 1/1982 states that consent to use one's image can be revoked at any time, with compensation for any damage caused.⁹ That's why the agreement should cover what happens to videos already published.
Studies published so far find similar learning outcomes, with a slight advantage for the recorded person in experience and social presence.² ³ ⁴ Most use short videos and theoretical content, so it's worth testing with your own case before generalizing.
¹ Castro-Alonso, Wong, Adesope and Paas (2021). Effectiveness of multimedia pedagogical agents predicted by diverse theories: a meta-analysis - Educational Psychology Review ² Leiker et al. (2023). Generative AI for learning: investigating the potential of synthetic learning videos - arXiv ³ Netland et al. (2025), ETH Zurich - Computers & Education ⁴ Xu et al. (2025) - British Journal of Educational Technology ⁵ Regulation (EU) 2024/1689 on Artificial Intelligence - EUR-Lex ⁶ The AI Act implementation timeline under the AI Omnibus - Future of Privacy Forum ⁷ Guidelines on the transparency obligations of Article 50 of the AI Act (July 2026) - European Commission ⁸ Draft Organic Law on the proper use and governance of artificial intelligence - Spanish Parliament Official Gazette ⁹ Organic Law 1/1982 on the civil protection of the right to honour, personal and family privacy and one's own image - BOE ¹⁰ The use of third-party images in AI systems and their visible and invisible risks - AEPD ¹¹ How do I create my personal avatar - Synthesia Help Center ¹² How to create a digital twin - HeyGen Academy ¹³ Code of practice on AI-generated content - European Commission
Program wspierania zatrudnienia na czas nieokreślony wykwalifikowanych osób młodych w ramach Krajowego Systemu Gwarancji dla Młodzieży. Vidext otrzymał dotację w wysokości 25.401€ przyznaną przez LABORA (Walencka Służba Zatrudnienia i Szkoleń) na zatrudnienie na czas nieokreślony w 2024 r. wykwalifikowanych osób młodych zarejestrowanych w Krajowym Systemie Gwarancji dla Młodzieży; działanie kwalifikuje się do współfinansowania z Europejskiego Funduszu Społecznego Plus (EFS+) 2021-2027 lub innego funduszu Unii Europejskiej. Expediente ECOGJU/2024/550/46. Niniejsza publikacja jest dokonywana w wypełnieniu obowiązków przejrzystości określonych w ustawie 19/2013 z dnia 9 grudnia.
