The EU AI Act does not change when your website needs cookie consent. It governs the AI systems that use the data afterwards. Cookie collection stays under Europe's ePrivacy and data protection rules, and the AI Act adds a separate product-safety layer on top, with its own obligations and its own enforcers.
For any site that feeds tracking data into recommendation, targeting or scoring models, that means two jobs running in parallel rather than one replacing the other. The two regimes touch at specific points, and knowing where they meet is the difference between a clean data flow and a double liability.
Does the EU AI Act replace cookie consent rules?
No. The AI Act sits alongside the existing cookie framework rather than absorbing it. Reading or writing a non-essential cookie still needs prior consent under Article 5(3) of the ePrivacy Directive. Processing the personal data those cookies generate still needs a lawful basis under the GDPR. The AI Act then governs the AI system that takes that data as an input. Three gates, three legal tests, and passing one does not excuse the others.
The first gate is collection. Before a non-essential tracker sets or reads a cookie, the visitor must give informed, prior consent, and refusing must be as easy as accepting. The second gate is processing: once you hold the data, the GDPR governs your purpose, your lawful basis, your transparency and your retention. The third gate is the system itself. Regulation (EU) 2024/1689, in force since 1 August 2024, classifies AI systems by risk and attaches obligations to each tier. A model trained on cookie-derived behaviour falls under that classification however clean the consent flow was.
Where the AI Act does deal with data directly, it is mostly about quality, not collection. Article 10 sets data and data-governance requirements for the training, validation and test sets of high-risk systems. It does not create a new consent rule for cookies.
Three gates between a cookie and an AI decision
How cookie data ends up inside an AI system
A tracking cookie assigns a stable identifier, and every page view, click and dwell time attaches to it. Over a session, and across sessions, that stream becomes a behavioural profile. The profile is then turned into features that feed recommendation engines, propensity models and automated ad bidding. Cookie data is rarely used raw. It is the raw material for models, which is exactly where the AI Act starts to apply.
Concrete examples make the pipeline clearer. An analytics cookie such as _ga supplies pseudonymous session behaviour that trains audience and churn models. A marketing cookie such as _gcl_au or _fbp feeds conversion modelling and automated bidding through real-time bidding pipelines. First-party profile cookies drive on-site personalisation and dynamic recommendations for signed-in users.
| Cookie or identifier | What it captures | Typical AI use |
|---|---|---|
_ga (Google Analytics) | Pseudonymous visitor ID, session behaviour | Audience modelling, churn and propensity scoring |
_gcl_au (Google Ads) | Ad-click attribution | Conversion modelling, automated bidding |
_fbp (Meta Pixel) | Browsing and conversion events | Lookalike targeting, ad optimisation |
| First-party profile cookie | Signed-in behaviour and preferences | On-site recommendation engines, dynamic personalisation |
The same data can also feed analytics dashboards that never touch a model. The regulatory weight only shifts once the data trains or drives an automated system that acts on people.
Which cookie-fed AI systems the AI Act regulates
Most of them fall into the lowest tiers. The AI Act sorts systems into unacceptable, high, limited and minimal risk. Ordinary on-site personalisation, audience modelling and ad optimisation are generally minimal or limited risk rather than high risk. Annex III lists employment, credit scoring, education, biometrics, law enforcement and critical infrastructure, not general website advertising, so most adtech is not on the high-risk list.
Where a system interacts directly with a person, a lighter duty applies. An AI chatbot running on session cookies, or a feature that generates synthetic content, carries transparency obligations under Article 50, which has applied since 2 August 2026: the person must be told they are dealing with an AI system or with generated media.
| Risk tier | Cookie-fed example | Core obligation |
|---|---|---|
| Unacceptable (prohibited) | Personalisation that manipulates a decision and causes harm | Banned outright under Article 5 |
| High risk | Rare in adtech: web signals feeding credit or hiring scores | Conformity assessment, logging, human oversight |
| Limited risk | AI chatbot on session cookies, AI-generated content | Transparency and disclosure (Article 50) |
| Minimal risk | Recommendation engine, audience modelling, ad optimisation | No specific AI Act duty, the GDPR still applies |
Timing matters here too. As of October 2026, the Digital Omnibus on AI, published in the Official Journal on 24 July 2026 as Regulation (EU) 2026/1744, has moved the obligations for stand-alone high-risk (Annex III) systems from 2 August 2026 to 2 December 2027, and those for high-risk systems embedded in products under Annex I to 2 August 2028. Even where a cookie-fed system would qualify as high risk, the compliance clock has moved.
Where personalisation crosses Article 5's red line
Only when it stops being persuasion and becomes manipulation that causes harm. Article 5(1)(a) prohibits AI that uses subliminal or purposefully manipulative techniques to materially distort behaviour and cause significant harm. Article 5(1)(b) prohibits exploiting vulnerabilities tied to age, disability or socio-economic situation. The Commission's guidelines of 4 February 2025 state that personalising ads on stated preferences is not inherently manipulative, and that GDPR compliance reduces the manipulation risk.
The edge cases are where cookie data matters. Optimising a model purely towards a conversion objective can, as a side effect, drift into manipulative tactics, and that side effect is the real question. Using cookie-derived signals of financial distress or age to push exploitative offers edges towards the Article 5(1)(b) prohibition.
Consent banners built with dark patterns already breach the GDPR. If such a banner were AI-optimised to suppress refusal below the visitor's awareness, it could attract Article 5 scrutiny on top. The threshold for the ban is high, since it requires significant harm, so most personalisation stays outside it. What changed is that the outer limit is now written into statute rather than left to consumer-protection principles.
What legal basis applies when cookie data trains an AI model?
Two questions, answered separately. Reading the cookie needs consent under ePrivacy if it is non-essential. Processing the resulting personal data to build or run an AI model needs a GDPR lawful basis. In Opinion 28/2024, adopted on 17 December 2024, the EDPB confirmed that legitimate interest can support AI model development and deployment, subject to a three-step balancing test, while treating model anonymity as a high, case-by-case threshold.
The trap is repurposing. If you collected analytics cookies on a consent basis to measure traffic, feeding that data into an ad-targeting model is a new purpose the original consent never covered. Legitimate interest does not rescue data that was gathered through a consent-gated tracker for a narrower stated purpose, because purpose limitation under Article 5(1)(b) of the GDPR still binds you.
An "anonymised" training set is not automatically outside the GDPR either. The EDPB expects documented evidence that re-identifying any individual is insignificant before a model counts as anonymous. Where a model drives decisions about people, the separate rules on automated decision-making and profiling can also apply.
How enforcement splits between the AI Act and cookie rules
Different regulators police each gate. Cookie and ePrivacy breaches fall to the national authorities each country names, usually the data protection authority. On 1 September 2025 the CNIL fined SHEIN EUR 150 million for placing advertising cookies before visitors made a choice and for continuing after a refusal. On the same day it fined Google EUR 325 million for showing ads between Gmail emails without consent and for invalid consent to advertising cookies when people created a Google account. Both decisions landed at the collection gate, under the French rules that transpose the ePrivacy Directive, and neither involved the AI Act.
The AI Act is enforced separately, by national market-surveillance authorities with the European AI Office, with fines reaching EUR 35 million or 7% of global turnover for prohibited practices. No AI Act penalty has yet targeted a cookie-fed advertising system, and enforcement so far concentrates on the consent gate.
A single practice can still trigger both regimes. An unlawful consent flow feeding a manipulative model would face a data protection authority on the cookie side and a market-surveillance authority on the system side, with two penalty ceilings in play. The wider relationship between GDPR and the AI Act extends well beyond cookies, but cookie data is one of the clearest places the two sets of rules meet on the same website.
The cookie rules themselves may move too. A separate Digital Omnibus proposal from November 2025 would bring cookie consent into the GDPR. As of October 2026 it is still a proposal.
Five steps for website owners
Start where the enforcement is. Keep the consent gate clean, because that is what regulators are fining today, then map which cookie data flows into any AI system and document a lawful basis for that specific use. The AI Act layer matters, but a leaking consent banner is the more immediate financial risk.
- Run a cookie scan and record which trackers set non-essential cookies and where that data goes.
- Confirm the consent flow blocks non-essential cookies until the visitor agrees, and that "Reject all" is honoured everywhere.
- Document the GDPR lawful basis for each downstream AI use, kept separate from the collection consent.
- Check that no personalisation or bidding model optimises towards manipulative or vulnerability-exploiting objectives.
- Track the AI Act timeline: Annex III high-risk obligations apply from 2 December 2027, but the Article 5 prohibitions and the Article 50 transparency duties already apply.
A cookie scan gives you the map for step one. Kukie.io detects first-party and third-party cookies and records what each one sets, which is the starting point for tracing the data into any model downstream.
Questions about the AI Act and cookie data
Does the EU AI Act require a separate cookie banner?
No. Cookie consent is governed by the ePrivacy Directive and the GDPR, not the AI Act. Your banner obligations are unchanged. The AI Act only adds duties for the AI system that uses the data those cookies collect.
Can Google Analytics data be used to train an AI model?
Only if the original consent and stated purpose covered that use. Analytics cookies collected to measure traffic cannot be silently repurposed for ad targeting, because GDPR purpose limitation and the ePrivacy consent both tie you to the purpose you disclosed.
Are AI recommendation engines banned under the AI Act?
No. Recommendation and personalisation engines are generally minimal risk and carry no specific AI Act obligation. They are prohibited only if they use subliminal or manipulative techniques that distort decisions and cause significant harm under Article 5.
Who enforces the AI Act compared with cookie rules?
National authorities, usually the data protection authority, enforce cookie and ePrivacy rules, as the CNIL did against SHEIN and Google in 2025. The AI Act is enforced by national market-surveillance authorities with the European AI Office.
Does anonymising cookie data remove GDPR and AI Act obligations?
Not automatically. The EDPB treats AI-model anonymity as a high, case-by-case threshold and expects evidence that re-identifying any individual is insignificant. If that bar is not met, the data stays personal and the obligations still apply.
When do the AI Act's high-risk rules take effect?
Under Regulation (EU) 2026/1744, the Digital Omnibus on AI published on 24 July 2026, obligations for stand-alone high-risk (Annex III) systems apply from 2 December 2027 and those for Annex I systems from 2 August 2028. Prohibited practices under Article 5 have applied since February 2025, general-purpose AI rules since August 2025 and the Article 50 transparency duties since 2 August 2026.
Map the cookie data behind your AI models
If you feed cookie data into recommendation, targeting or scoring models, the first step is knowing exactly what you collect. The Cookie scanner loads your site like a visitor and lists every first-party and third-party cookie it finds, with its category.
