Privacy × AI A weekly EU privacy and AI brief, with primary sources and practical context
Home Archive Add weekly reminder

Ireland's DPC sets out what it expects before an AI model is trained on personal data

Issue 5 · 30 September 2026 · Covers 23 to 30 September 2026 · About 8 min read
The weekly brief in video
1

Executive Summary

2

In Focus

AI and data protection Ireland / EU Supervisory report 25 September 2026

Ireland's DPC sets out what it expects before an AI model is trained on personal data

On 25 September 2026 the Irish Data Protection Commission published "Responsible Artificial Intelligence Innovation: Insights from the Data Protection Commission's Supervision of AI (2021-2025)". It is the DPC's own account of how it has supervised the creation, training and launch of AI products. The companies it names include Airbnb, Apple, DeepSeek, Google, LinkedIn, Meta, Microsoft, OpenAI, Pinterest, TikTok and X.

The DPC says its supervision unit for large technology companies engaged with controllers on approximately 180 AI products and services over the five years. Of those engagements, 143 concerned generative AI. AI accounted for 1 in 35 of the unit's engagements in 2021, and in 2025 for around 1 in 4 of those with large technology companies.

Most engagements ended in recommendations, with a median of three each. The DPC reports that 95% of these AI products launched without being delayed by its involvement, and that its interventions delayed or cancelled nine launches over the whole period.

The substance is in what the DPC says it looks for. On legal basis, the report starts from the EDPB's Opinion 28/2024, which recognised legitimate interests under Article 6(1)(f) of the GDPR as a possible basis for training AI models. It then describes how the DPC tests a controller's legitimate interests assessment against the three step test. The DPC examines the transparency given and whether it matches what people would reasonably expect, how easy it is to object, necessity, less intrusive alternatives, mitigations and the balancing itself.

Four positions in the report go further than restating the law, and each is the DPC's own assessment:

  1. The right to object should be available before training starts. The DPC's view is that there may be no effective way to exercise the Article 21 right if people are not given the chance to object before processing begins, and it says that, where applicable, an objection to AI training should be treated as an unconditional opt-out that is honoured automatically.
  2. Private messages face a particularly high threshold. The DPC considers it unlikely that a controller can show that people expect their private messages to be used for AI training, and assesses the threshold for relying on legitimate interests for that purpose as very difficult to achieve. It notes that the ePrivacy Directive separately protects the confidentiality of communications, and says it formed this assessment in the first half of 2026, outside the period the report otherwise covers.
  3. Training a new model will likely need a DPIA. The DPC considers that training a new AI model on personal data necessarily involves a new technology and likely large volumes of data, and therefore will likely require a data protection impact assessment.
  4. Transparency comes first. The DPC has emphasised to controllers that no amount of transparency after processing relieves them of the obligation to inform people before processing begins. It considers that the information should explain the risks, not only the fact of processing. It has also recommended that the information reach people who are not users of the service but whose data is processed.

On children, the report records that in most engagements on the development of AI models, particularly generative models, controllers chose not to train on children's data at all. Where a controller made that choice, the DPC recommended that it also exclude accounts where it could not be sure whether the user was a child. For users turning 18, the DPC considers that controllers must give a specific notice period, transparency and the chance to object before their data is used.

The report also details two cases in which the DPC intervened before or during processing.

In 2024 the DPC asked Meta to pause planned training on user data before it began, on the ground that Meta's notices and objection forms undermined its reliance on legitimate interests. Meta paused and re-engaged in February 2025 with a revised legitimate interests assessment and DPIA. The DPC concluded that the risks to individuals would be significantly reduced.

In August 2024 the DPC applied to the Irish High Court after X declined to stop training its Grok model on user data. X undertook to stop processing the user data collected while the mitigations were not in place and to delete the datasets containing that data. Both episodes were public at the time, and the report is the DPC's own account of them.

The report arrives while the EU legislator is deciding whether to answer the same question in the GDPR itself. The Commission's Digital Omnibus proposal of 19 November 2025 would insert an Article 88c under which processing for the development and operation of an AI system could rely on legitimate interests, subject to the balancing test and safeguards, while preserving laws that expressly require consent. The proposal is still under negotiation and none of it is law.

Why it matters

As lead supervisory authority for many of the largest technology companies operating in the EU, the DPC here gives a detailed account of how it has supervised AI training in the EEA. The report shows which arguments on legitimate interests the DPC has accepted and which it considers very unlikely to succeed. It is not a decision, binds no other authority and has not been tested in court.

Practical impact

Check the objection mechanism wherever training relies on legitimate interests. On the DPC's approach, it should be live, easy to use and reach every user on every device before any data enters a training set, and it should cover the full scope of the processing. Exclude private messages from training data unless the controller can show that both the legitimate interests threshold the DPC describes and the ePrivacy rules on confidentiality of communications are met.

Copies a clean, cited sentence for use in a note.
3

Key Developments

Video surveillance and wearables Germany Endorsement and draft guidance 23 September 2026

German authorities endorse Hamburg's smart glasses assessment and publish draft guidance on gym cameras

The Conference of the Independent Federal and State Data Protection Authorities, the DSK, met on 23 September 2026 and, in results published the next day, set out its views on smart glasses and on video surveillance in fitness studios. Both reach well beyond the devices and premises they name.

The first is smart glasses. The DSK endorsed the Hamburg Commissioner's examination report of 10 September 2026 on the Ray-Ban Meta AI Glasses, which issue 3 covered. In the DSK's assessment, the wearer is responsible for the recordings, and the GDPR applies to the wearer once images of other people are processed outside a purely personal or household context. The DSK considers that people must be informed before they are recorded, and that a warning LED is not enough. It also considers that using recordings of others to train AI generally lacks a legal basis, and that facial recognition with AI in public spaces would not be permitted. What had been one Land authority's assessment now carries the endorsement of the German supervisory authorities' joint conference. It remains an assessment, not a decision against any company.

The second is draft guidance on video surveillance in fitness studios, dated 1 September 2026 and published after the conference as a supplement to the DSK's general guidance on video surveillance by private bodies. It responds to rising complaints, and in particular to studios that operate partly or entirely without staff. Under the draft:

  1. Changing rooms, showers and saunas may never be monitored.
  2. Relaxation, eating and waiting areas may not be monitored, because people do not expect it there.
  3. Training floors are generally off limits. Narrow exceptions include a camera limited to where small equipment is stored, on a documented concrete risk; live monitoring where there is an elevated health risk and staff able to intervene at once; and, in exceptional cases, targeted surveillance after repeated serious offences.
  4. At the entrance, a camera for access control is justified only where electronic access control, staff at reception or barriers are not enough. Recording to investigate offences requires documented past incidents and must be limited in time, and recording alone is not considered able to prevent offences.
  5. Outside opening hours, when nobody is entitled to be present, blanket surveillance is generally permissible.
  6. Permanent workstations and the monitoring of staff behaviour or performance are excluded.

The draft narrows the legal bases operators most often rely on. Walking into a signposted area is not consent, and consent written into a membership contract is generally not freely given. Consent is conceivable only exceptionally, for a camera trained on a single machine that the member switches on for their own session and that captures no one else.

Surveillance is generally not necessary to perform the contract, unless the service itself is video based, such as video training analysis. The ordinary risk of a sports injury does not engage vital interests.

For other uses, that leaves legitimate interests. The draft says that a supposed deterrent effect does not justify cameras on its own, and it names emergency buttons, fall-alert wristbands and staff rounds as the answer to safety concerns. It also says that a sign on the wall is irrelevant to whether people should objectively expect to be filmed.

The DSK will open a public consultation on the draft. As at 30 September 2026, no deadline had been published in the DSK's materials.

Follow-up: the EDPB statement on smart glasses of 13 October 2026, in issue 7.

Why it mattersSmart glasses now have a common supervisory position, endorsed by the DSK, that organisations deploying or permitting them in Germany should take into account. The gym guidance is written for one sector, but its reasoning on consent, contract and deterrence is likely to be relevant to any private operator with cameras in spaces where people change, rest or exercise, and it follows the same line the Garante took on a pool changing area in issue 4.
Practical impactWrite a rule on recording third parties and on AI features for any organisation that issues or allows smart glasses at work. List every camera by area against the six categories above at gyms, pools, clubs and similar premises. Remove any cameras that cover changing, washing or rest areas, and start keeping the incident log the draft expects before relying on a camera at the entrance. Watch for the consultation and respond while the text is open.
Copies a clean, cited sentence for use in a note.
4

Enforcement Watch

Direct marketing and data brokers France Closure of injunction 17 September 2026

France: the CNIL accepts how Solocal now checks consent collected by its data suppliers

By deliberation SAN-2026-012 of 17 September 2026, published on 24 September, the CNIL's restricted committee closed the injunction it had imposed on Solocal Marketing Services in May 2025, and decided not to collect the penalty of 10 000 euro per day of delay that went with it. The original deliberation, SAN-2025-001 of 15 May 2025, fined the company 900 000 euro, among other things for prospecting by SMS and email without valid consent and for passing data to partners without a legal basis. It also ordered the company to stop prospecting without valid consent within nine months.

The company acquires prospect data mainly from data brokers and from competition and product test sites. To comply, it put in place checks of its partners' data collection forms, based on automated analysis, with human review where needed. The checks include, among other things, whether the form has a tick box requiring a positive action to consent, whether the required information is present, and whether the layout could undermine the free and unambiguous nature of consent. The restricted committee considered that, if scrupulously applied, those measures allow the company to ensure that consent is validly collected. It also stated that the company remains responsible for the validity of consents collected by its partners on its behalf. The committee added that the company could still be held liable if its checks fail to identify non-compliant forms.

Why it mattersThis closure describes the measures the CNIL accepted. It confirms that a buyer of marketing data cannot pass the consent question to its supplier: verifying the supplier's forms is part of the buyer's own compliance, and a verification system reduces the risk without transferring it.
Practical impactCheck the collection form used for each source of purchased or partner data before the first campaign message goes out. Keep a record of the check and repeat it when a partner changes its forms. A contractual warranty from the broker is not a substitute for these checks.
Copies a clean, cited sentence for use in a note.

Each story links to its supporting sources. Editorial analysis is clearly marked separately from factual reporting.

Sources and references

How the brief is prepared and checked

This newsletter is prepared with the help of AI and goes through several stages of review. Please check the original source before relying on the information presented here.