EU AI Act High-Risk Checker
Is your AI system high-risk under the EU AI Act? Work through Annex I, Annex III, the Article 6(3) filter and the general-purpose AI tests to see which regime likely applies — and from which date.
The EU AI Act (Regulation (EU) 2024/1689) reserves its heaviest regime for high-risk AI systems. A system gets there by one of two routes in Article 6. Under Article 6(1), it is a safety component of — or is itself — a product covered by the Annex I harmonisation legislation and needing third-party conformity assessment: medical devices, machinery, toys, lifts, radio equipment and their neighbours. Under Article 6(2), it is used for one of the eight Annex III areas: biometrics, critical infrastructure, education, employment, essential services, law enforcement, migration and border control, and justice and democratic processes. Providers of high-risk systems carry the Chapter III build duties — risk management, data governance, Annex IV technical documentation, logging, human oversight, a quality management system, conformity assessment, CE marking and registration — while deployers carry the Article 26 use duties and, in defined cases, an Article 27 fundamental rights impact assessment.
What this checks
Most AI systems are not high-risk under the AI Act. The label sticks in two main ways: your system is tied to a regulated product listed in Annex I, or it is used for one of the eight areas listed in Annex III — biometrics, critical infrastructure, education, jobs, essential services, law enforcement, migration, and justice or elections. A separate regime covers general-purpose AI models. This checker tests those routes in order, then applies the Article 6(3) filter that can take an Annex III system back out again.
Everything runs in your browser. Your answers are never sent to a server; they are encoded into a link only if you choose to copy a result link.
- Regulation (EU) 2024/1689 (Artificial Intelligence Act) — full text on EUR-Lex (Articles 2, 3, 6, 16, 17, 26, 27, 43, 47–49, 51–55, 72, 73, 99, 101, 111, 113; Annexes I, III, IV, XI, XII)
- Regulation (EU) 2026/1744 (Digital Omnibus on AI) — the amendment that moved the high-risk application dates in Article 113
Last verified: 15 November 2026, against the consolidated EUR-Lex text 02024R1689-20260727.
Quick answers
When is an AI system high-risk under the EU AI Act?
In two ways: it is a safety component of, or is itself, an Annex I regulated product needing third-party conformity assessment (Article 6(1)) — or it is used in one of the eight Annex III areas such as biometrics, hiring, credit scoring or law enforcement (Article 6(2)). A system that profiles people in an Annex III area is always high-risk.
Annex III systems can escape through the Article 6(3) filter if they pose no significant risk, do not materially influence decisions, and only do a narrow, procedural, preparatory or review-support task — and the provider documents that assessment in advance. The high-risk duties apply from 2 December 2027 for Annex III systems and 2 August 2028 for Annex I systems, after Regulation (EU) 2026/1744 moved the original dates.
What duties does a provider of a high-risk AI system have?
The Chapter III build and process duties: risk management, data governance, Annex IV technical documentation, record-keeping, transparency to deployers, human oversight, accuracy and cybersecurity (Articles 9–15), plus a quality management system (Article 17), conformity assessment (Article 43), EU declaration of conformity (Article 47), CE marking (Article 48), registration (Article 49), post-market monitoring (Article 72) and serious-incident reporting (Article 73).
Deployers have a shorter, different list under Article 26: use the system as instructed, competent human oversight, relevant input data, monitoring, keeping logs for at least 6 months, and informing workers. Public bodies and deployers of public-benefit or credit-scoring systems also owe a fundamental rights impact assessment under Article 27.
What counts as a general-purpose AI model with systemic risk?
A general-purpose AI model is presumed to have high impact capabilities — and so systemic risk — when the cumulative compute used for its training passes 10^25 FLOPs (Article 51(2)). The Commission can also designate a model as systemic-risk on other capability grounds.
All GPAI model providers owe the Article 53 duties. Systemic-risk models add Article 55: evaluation and adversarial testing, systemic-risk mitigation, serious-incident tracking and reporting, and cybersecurity — plus notifying the Commission within 2 weeks of meeting the criterion (Article 52). GPAI duties have applied since 2 August 2025; older models must comply by 2 August 2027.
How to use
- Question 1: say honestly whether the system, the model or their output could reach the EU — a public website or app open to EU visitors counts as “possibly”.
- Questions 2–4: place your role, say whether the system is tied to an Annex I regulated product, and tick every Annex III area the system is actually used in.
- Question 5 appears only if you ticked an Annex III area: it runs the Article 6(3) filter. Question 6 then tests the general-purpose AI model regime separately.
- Read your result tier and its duty list, check the dates table below, and copy the result link if you need to share the answers with counsel.
About this tool
Two escape routes and one trap matter in practice. The trap is profiling: an Annex III system that profiles natural persons is always high-risk. The first escape is Article 6(3): an Annex III system that poses no significant risk to health, safety or fundamental rights — including by not materially influencing decision outcomes — and that only performs a narrow procedural task, preparatory work, improves a completed human result, or flags patterns for human review, falls outside the high-risk regime. The provider must document that assessment before placing the system on the market, so the escape only works on paper you wrote in advance. The second separate regime is Chapter V: providers of general-purpose AI models owe the Article 53 documentation, copyright-policy and training-summary duties, and models with systemic risk — presumed once cumulative training compute passes 10^25 FLOPs — owe the extra Article 55 duties and must notify the Commission within 2 weeks.
Timing is where most summaries online are now wrong. The Digital Omnibus, Regulation (EU) 2026/1744, amended Article 113: the Chapter III high-risk duties apply from 2 December 2027 for Annex III systems and from 2 August 2028 for Annex I systems. The original dates — 2 August 2026 and 2 August 2027 — are void, and any checklist still using them is out of date. General-purpose AI duties have applied since 2 August 2025, with models placed on the market before that date given until 2 August 2027 to comply. This checker runs entirely in your browser, stores nothing, and covers the high-risk and GPAI scope questions only — the Article 5 prohibitions, Article 50 transparency duties and the GDPR are separate questions.
How it works
The mechanism is a list, not a risk score. Article 6 does not ask how dangerous your system feels; it asks which list it is on. Annex I borrows the existing product-safety world — if the underlying product already needs a notified body to assess it, the AI inside it inherits the high-risk label. Annex III lists uses where a wrong output lands on a person’s job, loan, education, benefits or liberty. That is why the same model can be unregulated in one product and high-risk in another: the label follows the use and the product, not the technology. The transferable signal is to classify the deployment, not the model — every new customer, sector or feature is a fresh classification question.
The Article 6(3) filter is the pressure valve, and its design tells you how regulators will test it. It has a negative condition (no significant risk, no material influence on outcomes), a positive condition (only a narrow, procedural, preparatory, result-improving or pattern-flagging task), an absolute carve-out (profiling is always high-risk), and a paperwork condition (the provider documents the assessment before market placement). Read together, the filter rewards systems that genuinely leave the decision with a human and can prove it. The judgement signal: if removing the human from your workflow would change the outcome for the person affected, your system materially influences decisions — whatever the marketing page calls it.
GPAI runs on a different logic: capability, measured crudely by compute. The 10^25 FLOPs line in Article 51(2) is a presumption, not a measurement of harm — it exists because a model’s downstream uses cannot be listed in advance the way Annex III lists system uses. So the Act splits the work: model providers document and disclose down the chain (Annexes XI and XII), and system providers classify their own use at the top of the chain. If you sit in the middle — fine-tuning or rebranding a model — watch the definitions, not your contract label: putting a model or system on the market under your own name is what makes you a provider, with the duties to match.
Worked examples
Hiring platform that ranks candidates: A vendor sells software that filters job applications and ranks candidates for employers. That is Annex III point 4 — employment and recruitment. The ranking shapes who gets interviewed, so it materially influences decisions and the Article 6(3) filter does not save it. Result pattern: likely high-risk, Annex III route, duties from 2 December 2027.
AI component inside a medical device: A startup builds the imaging model inside a diagnostic device that needs third-party conformity assessment under the medical devices rules. That is the Article 6(1) Annex I route, whatever the model does. Result pattern: likely high-risk, Annex I route, duties from 2 August 2028.
Spell-check style drafting helper for caseworkers: A tool tidies up text a caseworker has already written and flags possible inconsistencies for the caseworker to check. It does not score the applicant, does not profile anyone, and the decision stays with the human. That is the Article 6(3) shape — if the provider documented the assessment before launch. Result pattern: needs legal review, because the result turns on that documented assessment.
Foundation model trained above the compute line: A lab releases a general-purpose model trained with cumulative compute over 10^25 FLOPs. Article 53 duties apply, Article 55 systemic-risk duties are added, and the Commission must be notified within 2 weeks. Result pattern: likely general-purpose AI model duties.
Quick reference
| Date / threshold | What it means |
|---|---|
| 1 August 2024 | Regulation (EU) 2024/1689 enters into force |
| 2 February 2025 | Chapters I and II apply — including the Article 5 prohibitions |
| 2 August 2025 | General-purpose AI model duties apply (Chapter V) |
| 2 August 2027 | GPAI models placed on the market before 2 August 2025 must comply (Article 111(2)) |
| 2 December 2027 | High-risk duties apply for Annex III (Article 6(2)) systems — Article 113 as amended by Regulation (EU) 2026/1744 |
| 2 August 2028 | High-risk duties apply for Annex I (Article 6(1)) systems — Article 113 as amended |
| 10^25 FLOPs | Cumulative training compute above which a GPAI model is presumed to have systemic risk (Article 51(2)) |
| EUR 15,000,000 or 3% | Top fine for breaching high-risk operator obligations (Article 99(4)); EUR 35,000,000 or 7% for prohibited practices |
Sources: Regulation (EU) 2024/1689 on EUR-Lex; Regulation (EU) 2026/1744 (Digital Omnibus on AI). Last verified: 15 November 2026.
Frequently asked questions
Is this legal advice?
No. This checker is general information built from the published text of Regulation (EU) 2024/1689 as amended by Regulation (EU) 2026/1744. It cannot weigh your product design, your contracts, or how a market surveillance authority would read your facts. Treat a “likely” result as the start of a conversation with EU AI Act counsel, not the end of one.
Why does this page say December 2027 when other sites say August 2026?
Because the law changed. Regulation (EU) 2026/1744 (the Digital Omnibus on AI) amended Article 113: Chapter III Sections 1–3 now apply from 2 December 2027 for Annex III high-risk systems and from 2 August 2028 for Annex I high-risk systems. The original 2 August 2026 and 2 August 2027 dates are void. Pages still quoting them are working from the unamended 2024 text.
We only flag applications for a human recruiter. Are we high-risk?
Possibly not — that is exactly the shape the Article 6(3) filter was written for: pattern detection that supports human review without replacing or improperly influencing it, with no material influence on the outcome. But two things can pull you back in. If the system profiles candidates, it is always high-risk. And the filter only protects you if the provider documented the assessment before the system went on the market. “The human decides” said after the fact, with no paper, is a weak position.
What are the fines?
For breaching the high-risk operator obligations, up to EUR 15,000,000 or 3% of total worldwide annual turnover, whichever is higher (Article 99(4)). Prohibited practices under Article 5 sit in the top tier: up to EUR 35,000,000 or 7%. Supplying incorrect information to authorities runs up to EUR 7,500,000 or 1%. For SMEs, the lower of the fixed amount and the percentage applies. GPAI model fines under Article 101 mirror the EUR 15,000,000 or 3% tier.
We build on top of a general-purpose AI model. Whose duties are they?
Both sets can apply at once, to different people. The model provider owes the Chapter V duties for the model. If you build a system on that model and put it to an Annex III use — hiring, credit, education — your system can be high-risk in its own right, and the Chapter III duties land on whoever provides and deploys the system. A GPAI model underneath does not absorb the system-level duties.
Does “likely not high-risk” mean we have no AI Act duties?
No. The Article 5 prohibitions apply to everyone, the Article 50 transparency duties apply to a much wider set of systems than the high-risk regime, and Article 4 requires providers and deployers to take measures for staff AI literacy. High-risk is one chapter of the Act, not the whole book.
Sources and verification
Built from Regulation (EU) 2024/1689 on EUR-Lex (consolidated text 02024R1689-20260727), including the Article 113 application dates as amended by Regulation (EU) 2026/1744, and Articles 6, 9–17, 26, 27, 43, 47–49, 51–55, 72, 73, 99 and 101 with Annexes I, III, IV, XI and XII. Last verified: 15 November 2026.
First time embedding? Read the step-by-step guide →
More free tools
NYC Local Law 144 Checker
Do you use an automated employment decision tool for a New York City job? Work through the AEDT definition in Local Law 144 — the bias audit, the published summary and the 10-business-day notice — and see what likely applies.
GPSR Product Safety Checker
Does the EU General Product Safety Regulation cover your product, and what does your role require? Check the responsible person rule, distance-sale listing information and marketplace duties.
NIS2 Scope Checker
Are you an essential or important entity under NIS2? Check your sector against Annex I and Annex II, the size floor, the regardless-of-size types and Member State designation.
European Accessibility Act Checker
Does the European Accessibility Act cover your product or service? Check the Annex product and service lists, the 28 June 2025 cut-off, the services-only microenterprise exemption and the burden defence.
EU AI Act Article 50 Checker
Four Article 50 duty blocks — interaction disclosure, machine-readable marking, emotion-recognition notice, deep fake and public-interest text disclosure — sorted by whether you are the provider, the deployer, or both.
Invoice Generator
Create professional invoices in seconds — download as PDF, no sign-up.