Regulation (EU) 2024/1689 · Artificial Intelligence Act ● Annex III: — days

Is ChatGPT High Risk Under the EU AI Act? A Use-Case-by-Use-Case Answer

ChatGPT" is not a risk tier — the use case is. This guide maps ten common workplace uses of general-purpose AI to their tier, the article that applies and the obligation that follows.

Renomira Team · 10 min read

People ask whether a tool is high risk. The Act never asks that. It asks what the system is used for in your specific context — and then it points at an article.

Tiles spelling AI CHAT — the tool is not the tier, the use is
Photo: Jernej Furman from Slovenia / CC BY 2.0 via Wikimedia Commons

A general-purpose assistant is not classified. The application of it inside your company is. The same model can be prohibited, high-risk, limited or minimal within a single working day, depending on which prompt a colleague runs.

The four tiers, and where they come from

Tier Trigger Your obligation
Prohibited Article 5 — eight banned practices Withdraw the use case. No compliance route exists.
High risk Article 6 + Annex III (or Article 6(1) product safety) Documentation, oversight, logging, conformity assessment, registration.
Limited risk Article 50 transparency duties Disclose, mark synthetic content, retain provenance.
Minimal risk Everything else AI literacy under Article 4 is the only standing duty.

The classification decision itself has to be documented with the reasoning — “we assessed this against Annex III and concluded X because Y” — because the reasoning is what an inspector reads, not the label.

Ten common uses, mapped

1. Drafting emails and documents — minimal. No decision about a person, no synthetic media published externally. Article 4 literacy and your data-handling policy cover it.

2. Summarising meetings — minimal, with a privacy caveat. The AI Act tier stays minimal, but recording participants implicates GDPR: a lawful basis, a notice, and works-council consultation in Germany and the Netherlands. Two regimes, one tool.

3. Code generation in internal tools — minimal. Standard caution on secrets and licence provenance. No Annex III connection.

4. Customer-facing support chatbot — limited (Article 50(1)). Disclose that it is an AI. If it escalates to a human, say so. If it makes decisions about accounts — refunds denied, service terminated — you are no longer in limited territory.

5. AI-generated images in marketing — limited (Article 50(2) and 50(4)). Mark synthetic media, retain machine-readable provenance, label deepfakes.

6. Screening or ranking job applicants — high risk (Annex III point 4). This is the one that catches people out. A hiring manager pasting CVs into a chatbot has created a high-risk AI system inside your company, with the full Article 26 deployer duties.

7. Scoring or ranking candidates, promotions or performance — high risk (Annex III point 4). Employment and worker management covers recruitment, task allocation, promotion and termination decisions.

8. Assessing creditworthiness or setting insurance prices — high risk (Annex III point 5(b), 5(c)). Also triggers the fundamental-rights impact assessment under Article 27 for the deployer.

9. Inferring emotion from faces or voice at work — prohibited (Article 5(1)(f)). Not a disclosure issue. The system must be removed.

10. Using AI to decide who gets access to essential public or private services — high risk (Annex III point 5). Benefits, eligibility, emergency-call triage: the tier follows the consequence to the individual.

The rule that resolves most arguments

Ask two questions in order:

  1. Does it make or inform a decision about a natural person that affects them? If yes, look at Annex III — it is almost certainly high risk, and possibly prohibited.
  2. Does it talk to people or generate media they might mistake for real? If yes, Article 50 applies.

If both answers are no, you are in minimal-risk territory with an Article 4 literacy obligation and your own policy. Most internal productivity use falls here. Most departmental uses — HR, credit, insurance, essential services — do not.

The “shadow” version of the problem

The tier is set by the use, so nobody has to decide to build a high-risk system for one to exist. It happens when a recruiter pastes CVs into a consumer chatbot, when a line manager asks a model to rank a team for redundancy, or when a procurement tool quietly adds lead scoring.

You cannot classify what you cannot see, which is why discovery runs ahead of classification: spend analysis, SSO logs, an anonymous disclosure survey and a 120-tool catalog will surface more use cases than an architecture review ever will.

What actually changes when a use is high risk

If one of your uses lands in Annex III, the deployer duties under Article 26 follow:

  • Operate the system strictly per the provider’s instructions.
  • Assign human oversight — a named person with the authority and competence to intervene.
  • Ensure input data is relevant and sufficiently representative.
  • Keep logs automatically generated by the system, for at least six months.
  • Inform workers before the system is put into service at the workplace.
  • Cooperate with the market surveillance authority.
  • Where GDPR applies, run a data protection impact assessment.

None of that is optional, and none of it starts in December 2027 — the register, the human-oversight assignment and the worker consultation are the parts you can complete now.

Where to start

Run the free exposure scan: twelve questions about what AI your company actually uses and how you govern it. It returns your exposure score, the ranked exposures with the article that fired, and your first three moves. Confirm your email and the full report unlocks.

Frequently asked questions

Is using ChatGPT at work illegal under the EU AI Act?

No. Using a general-purpose assistant at work is legal and, in most internal uses, falls into the minimal-risk category. The risk tier attaches to the application — screening CVs with it is high risk; drafting an internal memo with it is not.

Are we a provider if our employees use a third-party chatbot?

No. Using a third-party system under your own authority makes you a deployer. You become a provider if you put your name on the system, substantially modify it, or generalise it across a market — for example, building a customer-facing product on top of a model and branding it as yours.

Does a Copilot-style coding assistant count as high risk?

Not for internal software development. High-risk classification follows Annex III and the product-safety route in Article 6(1). Code assistance for your own products sits outside both, unless the output feeds a decision that is itself in Annex III — such as automated screening of contractor applications.

What if one chatbot handles both support tickets and refund denials?

Classify each use separately and record both. The support conversation is limited risk with an Article 50 disclosure; the automated refusal of a refund is a decision about a person and needs its own classification, oversight and, where it materially affects them, an explanation under Article 86.

Where does GPAI model compliance fit — is that our job?

If you only use a general-purpose AI model, the Article 53 obligations sit with the model provider: documentation, copyright policy, training-data summary. Your obligations are deployer-side — Article 26 where the use is high risk, Article 50 where it is transparent, Article 4 always.

Next step

Put this guide to work on your own estate.

Run the free exposure scan to see which of these articles actually apply to your company — 12 questions, four minutes, no card. Confirm your email and the full report unlocks.

  • Free forever plan
  • No credit card
  • EU-hosted · GDPR-native
Run a free scan Sign up free

Build a live AI register in minutes — upgrade only when you need the audit pack.

is ChatGPT high risk under EU AI ActChatGPT EU AI Act complianceis Copilot high risk AI ActAI Act risk classification chatbotgenerative AI EU AI Act obligationsChatGPT Article 50 disclosureusing ChatGPT at work complianceEU AI Act minimal risk definition

Keep reading