A plain language look at what the EU AI Act and California’s AI Transparency Act actually require of the courses, avatars, and voiceovers you’re building this year, and what they don’t.
A member in the #IgniteLearning community sent me a message in our community membership that when I received it, I thought, “Oh, this is such a timely topic. Everyone in our industry should be talking about this.” She had just finished building a full onboarding series for a client, using an AI avatar to deliver the narration instead of hiring a video crew she did not have budget for. Her question was simple. “Do I need to put a label on this saying it’s AI?”
I did not have a confident answer for her on the spot, and I have been building courses for a long time. That bothered me enough to go find out properly, including sitting in on a Synthesia webinar session last week that walked through exactly this. Here is what I learned, translated into plain language for people who build courses, not people who write legislation.
This is general educational information based on publicly available guidance, not legal advice. If your training content reaches learners in the EU or California, or touches a regulated industry, loop in your own legal counsel before you finalize your approach.
Two laws quietly went live this year
The European Union’s AI Act, specifically Article 50, and California’s AI Transparency Act, known as SB 942, both became enforceable on August 2, 2026. If you build courses, facilitator videos, or any learner-facing content with AI tools, and there is any chance a learner in the EU or California sees it, this applies to you whether anyone told you or not.
Quick sidenote. If you read an article about SB 942 that says it took effect January 1, 2026, that was the original date. An amendment called AB 853 pushed the operative date to August 2, 2026, specifically to line up with the EU timeline. A lot of the guidance still floating around online has not caught up to that change.
The question that actually matters
Both laws share one underlying idea, and it is simpler than the legal language makes it sound. The question is not whether you used AI. Nearly everyone building courses right now uses AI somewhere in the process. The question is whether what your learner sees or hears could reasonably be mistaken for something real, a real person, a real voice, a real event, or a real result, when it is not.
AI assistance and AI deception are not the same category, and both laws are drawing that line more precisely than most of us expected.
What the EU AI Act requires
Article 50 puts two separate duties on two separate roles. The company that builds the AI system, Synthesia for example, is the provider, and it must mark its AI-generated content in a machine-readable way so detection tools can identify it. You, the person who takes that content and publishes it for your learners, are the deployer, and you carry a different duty: a visible disclosure that a person can see, shown at the latest when they first encounter the content.
That visible disclosure duty applies specifically to what the EU calls deepfakes, meaning AI-generated or manipulated image, audio, or video that resembles a real person, object, place, or event closely enough that someone could believe it is authentic. It also applies to AI-generated text published on matters of public interest, like public health or safety, when that text has not gone through meaningful human review.
Meaningful review is the operative phrase. Correcting typos does not count. Actually reading the content, verifying it, editing it, and taking responsibility for it as the person who published it does. If you draft a facilitator guide with AI help and then genuinely work through it, you are covered by that exemption.
Content that is evidently artistic, fictional, or clearly a simulation gets lighter treatment under the same framework. Penalties for noncompliance can run up to €15 million or 3% of global turnover, which tells you the EU is treating this seriously, even though most of us are nowhere near that scale of exposure.
What California’s law requires
SB 942 works differently. It places its obligations on covered providers, meaning companies like Synthesia that produce a generative AI system with more than 1 million monthly users (although revisions to the bill are actively still happening, so keep a good eye on this one because it is EVER-CHANGING) or visitors in California. It does not place a direct legal obligation on you as the person generating a video. Covered providers have to offer a free detection tool, embed machine-readable provenance information in what they generate, and make a visible label option available for you to use if you choose to. The civil penalty for a covered provider is $5,000 per violation, per day.
In practice, this means the tool companies are doing heavy lifting on the technical side. Your job is deciding when to turn on the visible disclosure option they give you, based on your own audience and your own judgment about what is fair to show them.
What this looks like inside a real course build (as of today, 9/10/2026)
Your facilitator guide draft is safe. If you ask an AI tool to help you draft an outline, write learning objectives, or build a facilitator script, and then you genuinely edit it, verify it, add your own expertise, and take responsibility for what gets published, that is meaningful human review. Under the EU framework, that specifically exempts you from the text disclosure requirement.
Your AI-narrated eLearning module deserves a second look. If you are using a synthetic voice to narrate a module, most learners assume a person recorded that. A short line in your course introduction, something like narration created using an AI voice, costs you nothing and builds trust instead of risking it if a learner ever notices.
Your realistic AI avatar is the one to think through carefully. This is exactly what happened to the member who messaged me. A photorealistic avatar delivering training as though they are a live facilitator, reaching an EU or California audience, puts the visible disclosure responsibility on you as the deployer, not on the platform. Synthesia and tools like it handle the machine-readable marking automatically. The on-screen label a learner can see is yours to add.
Cloning your own voice is a disclosure, not a confession. A growing number of us are using an AI clone of our own voice to keep up with production timelines. A short disclosure line does not diminish your expertise. It tells your learner two true things: this was not a live recording, and you stand behind every word of it.
A stylized cartoon avatar is a different situation entirely. A Vyond-style animated character walking a learner through a compliance scenario is not going to be mistaken for a real person by anyone. Content that is clearly stylized or artistic gets lighter treatment under these frameworks, because it was never trying to pass as real in the first place.
Do not build a course promo around a success story that did not happen. If you use AI to generate a learner testifying that your program changed their career, and that learner does not exist, you have built a false testimonial, not an AI disclosure problem. The FTC rules against this predate the current AI conversation entirely, and no disclosure label fixes a claim that was never true.
Your branching scenario characters are usually fine as they are. The frustrated caller in a customer service module or the difficult stakeholder in a leadership scenario is understood by the learner to be constructed for the exercise. This falls closer to the fictional and instructional content both frameworks treat more lightly. A line noting that scenario characters are simulated is still good instructional design practice, even when the law does not require it.
Public-facing compliance content written entirely by AI is worth a second pass. If you are producing something that will be published externally on a topic like workplace safety, and it goes out with no meaningful human review behind it, that sits closer to the category the EU framework is actually concerned about. This is a narrow case for most in-house work, but worth knowing if you ever build anything client facing that touches public safety information.
Language you can actually use
- Narration: “Narration for this module was created using an AI voice.”
- Avatar: “This video features an AI-generated presenter.”
- Your own voice clone: “This narration uses an AI-generated version of my voice. The script was written and approved by me.”
- Simulated scenario: “Characters in this scenario are simulated for training purposes.”
- Course-level statement: “This course was developed with AI support for drafting, narration, or visual assets. All content was reviewed by me as the designer.” (You can see my disclaimer at the bottom of this blog).
Where the responsibility sits
Let’s be honest, most of us are not going to read the full text of an EU regulation before our next project deadline. What matters practically is this: if you are building for a client on a platform or tool they chose, you are very likely still the deployer of what gets published, even if you did not pick the tool. Ask where questionable content came from before you publish it under your name or a client’s name. That habit protects you regardless of which specific law applies to a given project.
What I am doing with my own work
I use AI extensively in my own course development, from drafting to storyboarding to narration. I am not moving away from any of that. What has changed is that I now think through one extra question before I publish: could a learner reasonably believe this is something it is not? If the answer is yes, a simple line of disclosure goes in. If the answer is no, I keep building the way I already was.
That is really the whole approach. Use AI to help you create more efficiently. Do not let it manufacture a reality your learner did not actually get.
About Dani Watkins
Dani is the founder of #IgniteLearning and the owner of Zenith Performance Solutions. She’s an instructional designer, trainer, and eLearning developer who creates practical resources for in-house L&D professionals, including how AI fits into responsible course development in a field that keeps moving.
Where to go for more
- Synthesia, “What does the EU AI Act Article 50 mean for my Synthesia videos?” https://help.synthesia.io/en/articles/16046624-what-does-the-eu-ai-act-article-50-mean-for-my-synthesia-videos
- California SB 942 bill text, California Legislative Information: https://leginfo.legislature.ca.gov/faces/billNavClient.xhtml?bill_id=202320240SB942
- European Commission, EU icons for labelling AI-generated content: https://digital-strategy.ec.europa.eu/en/policies/eu-icons-labelling-ai-generated-content
- European Commission, Code of Practice on Transparency of AI-generated Content: https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content
- European Commission, Guidelines on transparency obligations for providers and deployers of AI systems: https://digital-strategy.ec.europa.eu/en/library/guidelines-transparency-obligations-providers-and-deployers-ai-systems
- EU AI Act Service Desk, Article 50: https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-50
✍️ This article is general guidance, not legal advice. ZPS cannot advise on your specific obligations. For anything beyond the general position, refer to the official sources linked in this article and consult your own legal or compliance team.
Portions of this article were drafted, outlined, or edited with the assistance of artificial intelligence tools. The final content has been thoroughly reviewed, verified, and approved by human editors, who assume full editorial responsibility for its publication.