In this article8 sections
  1. Synthetic Data Is Not a Compliance-Free Zone
  2. What the AI Act Actually Covers
  3. Scenario 1: Marketing Creative and Synthetic Imagery
  4. Scenario 2: Building Training Datasets with Generative AI
  5. Scenario 3: AI-Generated Customer Support Scripts
  6. The Article 4 Obligation You Cannot Skip
  7. Your Compliance Checklist for Synthetic Data
  8. References

Synthetic Data Is Not a Compliance-Free Zone

The ai act synthetic data question lands on compliance desks constantly: if we generate fake data instead of using real personal data, are we off the hook? The answer is no, not automatically. The EU AI Act imposes obligations on deployers based on what the AI system does, not on whether the output looks like real data. Generating synthetic images, text, audio, or video with a generative AI tool can trigger Article 50 disclosure requirements, and in some cases pulls your organisation into the high-risk category under Annex III.

This post walks through the three most common synthetic data scenarios for SMEs, maps each to the relevant AI Act articles, and gives you a plain-language checklist for each.

What the AI Act Actually Covers

The EU AI Act, which entered into force on 1 August 2024, applies to deployers: organisations that use an AI system in a professional context within the EU. Article 3 defines a deployer as any natural or legal person who uses an AI system under their own authority, other than for personal non-professional purposes.

Synthetic data generation tools, including large language models used to write customer scripts, image generators used to produce marketing visuals, and audio cloning tools used to create training samples, all qualify as AI systems under Article 3(1). The moment your organisation uses one of these tools for a business purpose, you are a deployer with obligations attached.

The specific obligations depend on the risk tier. Most synthetic data use cases sit in the minimal-risk category. Two duties can apply whatever the risk level: Article 4 (AI literacy measures) and, where the use fits one of its situations, Article 50 transparency. Article 26 applies only if the system is high-risk, from 2 December 2027.

Scenario 1: Marketing Creative and Synthetic Imagery

Your marketing team uses an AI image generator to produce lifestyle photos for a campaign. The models in the images are fully synthetic, no real person photographed.

Article 50(4) is the article to know here. It requires you to disclose deep fakes: AI images, audio or video that resemble real people, objects, places or events and could falsely pass as real. Your generator's provider must also mark its output in a machine-readable way (Article 50(2)). Fully synthetic lifestyle photos of people who look real can fall under the deep fake rule; labelling other AI visuals is good practice.

Practically, this means:

  • Add a visible disclosure label to AI-generated campaign images ("Created with AI" is sufficient in most cases).

  • Retain an internal record showing which assets were AI-generated.

  • If your campaign will appear on social platforms, check that platform-level disclosure rules align with Article 50 (most major platforms now require this independently).

For evidently artistic, creative, satirical or fictional work, Article 50(4) allows a lighter disclosure, but it is still required: you disclose that AI-generated content is present in a way that does not spoil the work. Default to disclosing.

Scenario 2: Building Training Datasets with Generative AI

Your data science team uses a large language model to generate thousands of synthetic customer support queries. The goal is to train an internal classification model without using real customer data.

This scenario layers two AI systems: the generative model used to produce the dataset, and the downstream model trained on it. Your organisation is a deployer of the first system. Because you build the second yourself, you are its provider, and also its deployer when you use it.

Here the key question is whether the downstream model qualifies as high-risk under Annex III. If the classification model is used to make or assist decisions about natural persons in areas like employment, credit, or benefits, it very likely does, and the high-risk rules then apply from 2 December 2027. Training it on synthetic data does not reduce its risk classification. Risk is assessed by the model's function and context of deployment, not by the provenance of its training data.

The generative step itself is normally not high-risk, so Article 26 of the EU AI Act does not apply to it. Good practice still borrows from it:

  • Use AI tools in line with the provider's instructions.

  • Assign someone to review outputs.

  • Keep a record of use.

If your organisation generates synthetic training data at scale, build a short internal log that captures: the tool used, the date range, the volume of outputs, the intended downstream use, and who reviewed the outputs for quality and bias. This log is your internal record, and it helps if the downstream model turns out to be high-risk.

AVG (GDPR) interaction is also relevant here. Even synthetic data derived from real customer interactions may retain statistical patterns that constitute personal data under Dutch and Belgian supervisory authority guidance. Check whether a DPIA is required before you start generation.

Scenario 3: AI-Generated Customer Support Scripts

Your operations team uses a generative AI tool to draft scripts that human agents read aloud during customer calls. The voice is human, but the words are AI-written.

This is where Article 50 surprises many compliance officers.

Article 50(1) requires that when an AI system interacts directly with natural persons, those persons must be informed that they are interacting with an AI. The exception: when it is obvious from context. A human agent reading an AI script is not an AI interaction in the Article 50(1) sense, so that specific disclosure is not triggered.

However, if the script is spoken by a synthetic voice agent, the people it talks to must be told it is an AI (Article 50(1)); the voice tool's provider must also mark its audio output (Article 50(2)). Many organisations run pilot programmes with synthetic voice agents without realising this.

Practical steps for this scenario:

  • Audit your customer contact tooling. Identify every touchpoint where AI generates text or voice that reaches a customer.

  • For AI-scripted human calls: no Article 50(1) disclosure required, but document the process internally.

  • For synthetic voice agents: add an opening statement identifying the caller as an AI system. Article 50 has applied since 2 August 2026.

  • Train customer-facing staff on what counts as AI interaction under the Act. This overlaps with your Article 4 AI literacy obligation, which applies from 2 February 2025.

The Article 4 Obligation You Cannot Skip

Article 4 requires providers and deployers to take measures to support the AI literacy of their staff. No specific level is required for any individual, but it is a binding obligation that has applied since 2 February 2025.

For synthetic data workflows, AI literacy means your team understands:

  • Which tools are in use and what risk tier they occupy.

  • What disclosure obligations attach to the outputs.

  • How to escalate when an AI output is uncertain or potentially harmful.

A one-page process guide per tool, combined with a short recorded briefing, is a reasonable measure for most SME contexts. Keep attendance records: they are how you show the measures you took.

Your Compliance Checklist for Synthetic Data

Inventory first. List every AI tool used to generate text, image, audio, or video outputs in your organisation. This is the foundation for knowing which duties apply.

Classify the output. Does it interact with or depict natural persons? Does it feed a downstream model that makes decisions about people? Each yes adds obligations.

Apply Article 50 disclosure. Deep fakes and AI-written text published to inform the public on matters of public interest need a disclosure (unless a human has reviewed the text and someone holds editorial responsibility). For other content, disclosure is good practice.

Check the downstream model's risk tier. If synthetic data trains a model used in HR, credit, or customer triage, run an Annex III check. High-risk classification means conformity assessments, human oversight systems and, for Annex III systems, registration in the EU database, from 2 December 2027 (Annex III) or 2 August 2028 (AI in Annex I products).

Document everything. A simple spreadsheet covering tool, use case, volume, oversight assigned, and disclosure method is your internal record. Start it today.

Run a DPIA if in doubt. If synthetic data derives from real personal data or is used to build models that affect individuals, coordinate with your DPO and the AVG process before you begin.

If you are unsure which of your synthetic data workflows already triggers obligations, the fastest first step is a structured compliance check. Run the free 2-minute AI Act compliance check at comply.khairos.ai to get a prioritised view of where your organisation stands and which articles apply to your specific use cases.

References

  1. Article 3

  2. Article 50(4)

  3. Annex III

  4. EU AI Act

  5. Run the free 2-minute AI Act compliance check at comply.khairos.ai

Not sure where your company stands?

Answer 15 questions about how your team uses AI. You get a compliance score and the gaps to close first, as a PDF.

Am I compliant? Check in 2 minHow Khairos Comply works