EU AI Act for developers: what applies to you, and when
The EU AI Act for developers: risk tiers, provider vs deployer, Article 50 transparency, GPAI rules and the AI Omnibus dates, with a Python triage script.

The EU AI Act (Regulation (EU) 2024/1689) is the European Union’s law on AI systems, and for a developer it comes down to three questions: which risk tier your use case falls in, whether you’re the provider or the deployer, and which date applies. Most LLM features (a chatbot, a content generator) land in the transparency tier and owe disclosure and labeling duties, which apply from August 2, 2026. Use cases on the high-risk list, such as CV screening or credit scoring, owe much more, and after the AI Omnibus amendment those rules apply from December 2, 2027.
This is an engineer’s reading of the regulation text, to help you decide what to ask your lawyer. It is not legal advice.
Does the EU AI Act apply to me if I’m not in the EU?
Often, yes. Article 2 covers providers placing AI systems or general-purpose models on the EU market “irrespective of whether those providers are established or located within the Union or in a third country”, and providers and deployers in a third country where the system’s output “is used in the Union.” A team in Surat or San Francisco shipping an AI feature to EU customers is in scope.
Some things are out of scope: individuals using AI for “purely personal non-professional activity”, AI developed for the “sole purpose of scientific research and development”, and most systems released under free and open-source licences, unless they’re high-risk, prohibited, or covered by the Article 50 transparency rules. Pre-release testing is excluded too, but “testing in real world conditions” is not.
What are the risk tiers in the EU AI Act?
The European Commission’s summary uses four:
- Unacceptable risk, banned (Article 5). Examples include social scoring, untargeted scraping of internet or CCTV images to build facial recognition databases, and emotion recognition in workplaces and schools.
- High risk, strict obligations. Either AI that’s a safety component of a regulated product (Annex I: machinery, medical devices and so on) or a use case listed in Annex III. Annex III has eight areas: biometrics, critical infrastructure, education, employment, access to essential services, law enforcement, migration, and justice and democratic processes. The two developers hit most often are 4(a), AI used for recruitment or selection, including targeted job ads, filtering applications and evaluating candidates, and 5(b), AI that evaluates creditworthiness or sets credit scores (fraud detection is excluded).
- Transparency risk (Article 50). Chatbots must tell people they’re talking to a machine; generated content must be marked.
- Minimal or no risk. Spam filters, AI in video games. No specific obligations.
General-purpose AI models (the LLMs themselves) sit on a separate track with their own rules, covered below.
Am I a provider or a deployer?
This decides most of your obligations, so get it right early. From Article 3:
- A provider develops an AI system or general-purpose model and places it on the market or puts it into service “under its own name or trademark.”
- A deployer is anyone “using an AI system under its authority”, other than for personal non-professional use.
If you build a product on top of a model API and sell it under your brand, you’re the provider of that AI system, even though someone else provides the model. If your company buys an HR screening tool and uses it, your company is a deployer.
The trap is Article 25. A deployer becomes a provider of a high-risk system if it puts its own name on one, makes “a substantial modification” to one, or modifies “the intended purpose” of a system, including a general-purpose one, so that it becomes high-risk. Wrapping a general chatbot API in a “rank these candidates” feature is exactly that last case.
What does Article 50 require from an LLM app?
Article 50 is the part most LLM products touch, and it has applied since August 2, 2026.
- 50(1), disclosure. Providers of systems “intended to interact directly with natural persons” must design them so people know they’re dealing with an AI, unless that’s obvious to a reasonably well-informed person. A support chatbot should say it’s a bot.
- 50(2), marking. Providers of systems generating synthetic audio, image, video or text must ensure outputs “are marked in a machine-readable format and detectable as artificially generated or manipulated.” There are exceptions for assistive editing and for output that doesn’t substantially alter the input.
- 50(3), emotion recognition and biometric categorisation. Deployers must inform the people exposed.
- 50(4), deep fakes and public-interest text. Deployers must disclose that deep fake image, audio or video “has been artificially generated or manipulated.” AI-generated text published to inform the public on matters of public interest needs the same disclosure, unless it went through human review with someone holding editorial responsibility.
The Omnibus gave one grace period here: providers of systems placed on the market before August 2, 2026 must comply with the 50(2) marking duty by December 2, 2026 (AI Omnibus text).
What changed with the AI Omnibus?
The Commission proposed a “digital omnibus” to simplify and delay parts of the Act, and it passed in 2026 as Regulation (EU) 2026/1744. Per the Future of Privacy Forum’s timeline, Parliament adopted it on June 16, 2026, the Council on June 29, and it was published in the Official Journal on July 24, 2026. The changes developers should know:
- Annex III high-risk obligations moved from August 2, 2026 to December 2, 2027.
- Annex I high-risk (AI in regulated products) moved from August 2, 2027 to August 2, 2028.
- Article 50(2) marking for systems already on the market: December 2, 2026.
- A new prohibition on AI systems that generate child sexual abuse material and non-consensual intimate imagery applies from December 2, 2026.
- AI literacy (Article 4) was softened from a duty to ensure staff literacy to a duty to support it.
- Simplified documentation and proportionate fines for small and medium companies.
The prohibitions (in force since February 2, 2025) and the GPAI rules (since August 2, 2025) were not delayed.
What do the rules for general-purpose AI models require?
These apply to whoever trains and releases the large language model or other general-purpose model, so most application developers are on the receiving end. Article 53 requires GPAI providers to keep technical documentation, give downstream providers information about the model’s capabilities and limitations, have a copyright compliance policy, and publish a summary of training content. Open-source models with public weights are exempt from the first two, but not the last two. Models trained with more than 10^25 floating point operations are presumed to have systemic risk (Article 51) and carry extra duties, and the open-source exemption doesn’t apply to them. The Commission published a GPAI Code of Practice in July 2025 as a way to show compliance.
For an app builder, the useful part is 53(1)(b): your model provider owes you documentation of capabilities and limitations. Ask for it, and keep it in your own technical file.
What are the fines?
Article 99 sets the caps, whichever is higher for large companies and whichever is lower for SMEs and start-ups:
- Prohibited practices: up to 35 million euros or 7% of worldwide annual turnover.
- Most other obligations of providers, deployers, importers, distributors and notified bodies: up to 15 million euros or 3%.
- Incorrect or misleading information to authorities: up to 7.5 million euros or 1%.
How do I triage my AI features against the Act?
Here’s a first pass I’d run over a product’s feature list before talking to counsel. It sorts each use case into a tier and prints the date that applies as of today. The lists are deliberately short; the real Annex III has many more entries.
# ai_act_triage.py: a first-pass triage of AI use cases against the EU AI Act's tiers.
# Python 3 stdlib only. Dates are Regulation (EU) 2024/1689 as amended by the AI Omnibus,
# Regulation (EU) 2026/1744. A sorting aid for engineers, not legal advice.
from datetime import date
TODAY = date(2026, 10, 31)
PROHIBITED = {"social scoring", "workplace emotion recognition", "untargeted face scraping"}
ANNEX_III = {"cv screening": "4(a) employment", "credit scoring": "5(b) creditworthiness",
"exam scoring": "3 education"}
DATES = {"prohibited": date(2025, 2, 2), "annex iii": date(2027, 12, 2),
"art 50": date(2026, 8, 2), "art 50(2) marking, pre-Aug-2026 systems": date(2026, 12, 2)}
def triage(use, generates_content=False, talks_to_people=False, on_market_before_aug_2026=False):
if use in PROHIBITED:
return "PROHIBITED (Art. 5)", DATES["prohibited"]
if use in ANNEX_III:
return f"HIGH-RISK (Annex III {ANNEX_III[use]})", DATES["annex iii"]
if generates_content and on_market_before_aug_2026:
return "TRANSPARENCY (Art. 50(2) marking)", DATES["art 50(2) marking, pre-Aug-2026 systems"]
if generates_content or talks_to_people:
return "TRANSPARENCY (Art. 50)", DATES["art 50"]
return "MINIMAL (no specific obligations)", None
cases = [("cv screening", {}), ("support chatbot", {"talks_to_people": True}),
("image generator, new", {"generates_content": True}),
("image generator, live since 2025", {"generates_content": True, "on_market_before_aug_2026": True}),
("workplace emotion recognition", {}), ("spam filter", {})]
for use, flags in cases:
tier, when = triage(use, **flags)
status = "-" if when is None else ("in force" if when <= TODAY else f"applies {when}, {(when - TODAY).days} days away")
print(f"{use:33} {tier:38} {status}")
Output (Python 3.13.11):
cv screening HIGH-RISK (Annex III 4(a) employment) applies 2027-12-02, 397 days away
support chatbot TRANSPARENCY (Art. 50) in force
image generator, new TRANSPARENCY (Art. 50) in force
image generator, live since 2025 TRANSPARENCY (Art. 50(2) marking) applies 2026-12-02, 32 days away
workplace emotion recognition PROHIBITED (Art. 5) in force
spam filter MINIMAL (no specific obligations) -
The value isn’t the code; it’s the forced conversation. Writing the feature list as rows makes someone decide whether “rank inbound job applications” is a feature you have, and the answer moves it from one row to another with a 397-day clock.
What I would do
- Inventory every AI feature, including internal tools, with its purpose in one sentence. The intended purpose is what the Act classifies, so write it the way a regulator would read it.
- For each, record provider or deployer, and check Article 25. Any feature that ranks, scores or filters people for jobs, credit, education or public services gets escalated now, even though the high-risk date is December 2027. Those obligations (risk management, data governance, logging, human oversight, technical documentation) take longer to build than a year.
- Ship Article 50 now if you haven’t (and add a red team pass for the chatbot while you’re there): a visible “you’re chatting with an AI” in chat UIs, and machine-readable marking on generated media. For images, that means provenance metadata and, where you can, watermarking.
- Keep the model provider’s documentation in your file, and log enough to reconstruct what the system did. From the data-platform and privacy seat I sat in at Walmart, this part feels familiar: it’s the same discipline as GDPR records of processing, applied to models. Teams that already map where personal data flows will find the AI inventory much easier.
- Watch the guidance. The Commission owes high-risk classification guidelines and more by 2027, and those will settle many edge cases.
For agent products in particular, the human-in-the-loop and guardrails work you’d do for safety reasons maps closely onto the Act’s human oversight and logging duties. Do it once, and document it for both.
FAQ
When does the EU AI Act come into force?
It entered into force on August 1, 2024, and applies in stages: prohibitions from February 2, 2025, general-purpose model rules from August 2, 2025, most other rules including Article 50 transparency from August 2, 2026, Annex III high-risk rules from December 2, 2027, and Annex I high-risk rules from August 2, 2028, after the AI Omnibus amendment.
Does the EU AI Act apply to ChatGPT-style chatbots?
The underlying model falls under the general-purpose AI rules, owed by its provider. A chatbot product built on it owes Article 50 disclosure, and it becomes high-risk if it’s used for an Annex III purpose such as screening job candidates.
Is open-source AI exempt from the EU AI Act?
Partly. AI systems under free and open-source licences are mostly excluded, unless they’re prohibited, high-risk or covered by Article 50. Open-source general-purpose models are exempt from some documentation duties but must still publish a training content summary and a copyright policy, and the exemption doesn’t apply to models with systemic risk.
What counts as high-risk AI under the EU AI Act?
AI that’s a safety component of products covered by EU product law (Annex I), and AI used in the Annex III areas: biometrics, critical infrastructure, education, employment, essential services such as credit, law enforcement, migration, and justice and democratic processes.
What are the penalties under the EU AI Act?
Up to 35 million euros or 7% of worldwide turnover for prohibited practices, up to 15 million or 3% for most other violations, and up to 7.5 million or 1% for misleading information to authorities.
Sources
- Regulation (EU) 2024/1689, the AI Act, via the AI Act Explorer: Article 2, Article 3, Article 5, Article 25, Article 50, Article 51, Article 53, Article 99, Annex III
- Regulation (EU) 2026/1744, the Digital Omnibus on AI, adopted July 8, 2026
- European Commission, AI Act: regulatory framework for AI
- Future of Privacy Forum, The AI Act implementation timeline: what changes under the AI Omnibus?