Ireland's Data Protection Commission published a 75-page report on September 25, 2026 setting out how it examined roughly 180 artificial intelligence products and services from companies including Meta, Google, Microsoft, OpenAI, TikTok and X between 2021 and 2025, most of them before they reached European users. Nine launches were delayed or cancelled because of the regulator's interventions. Transparency problems appeared in 72% of the recommendations it issued, and the report closes with a position that reaches beyond its five-year window: training AI on private messages, the DPC says, faces a legal threshold that will be "very difficult to achieve."
In Short
Ireland's privacy regulator has published five years of figures on how it reviewed about 180 AI products from Meta, Google, OpenAI, X and others, mostly before launch. It matters because the DPC is the lead EU regulator for many of the large platforms that sell advertising in Europe, so its views on AI training, opt-outs and transparency travel with those platforms across the bloc. Nine launches were delayed or cancelled, 59% of AI products launched in 2024 and 2025 received recommendations, and the DPC now says a company is unlikely to be able to justify training AI on private messages under legitimate interests.
What the report is
The document is titled "Responsible Artificial Intelligence Innovation: Insights from the Data Protection Commission's Supervision of AI (2021-2025)" and carries a September 2026 date on its cover. It is not a decision, a fine or a guideline. It records the work of one team, the Technology Multinational Supervision Unit (TMSU), which sits inside the DPC's Supervision Function and whose remit, according to the report, "is to engage with the large multinational tech companies based in Ireland." Its engagement statistics all come from that unit's internal records.
That distinction matters. Supervision, according to the DPC, "operates completely separately from other regulatory functions in the DPC, such as those involved in complaints or inquiries." It does not look backwards at alleged infringements. It engages "on a voluntary basis, primarily before a product or service launches," and the report describes the result as a "quasi-regulatory sandbox."
The companies named include several of the largest sellers of digital advertising. According to the DPC, the Supervision Function engaged on "the creation, training, and deployment of AI by companies, including Airbnb, Apple, Deepseek, Google, LinkedIn, Meta, Microsoft, OpenAI, Pinterest, TikTok, X (formerly Twitter), and others." The technologies covered include large language models, age assurance, facial recognition, recommender systems, personalisation and agents.
Most case studies in the report do not name the company involved. Meta and X, under its former name Twitter, are the exceptions, and Google appears in the list of open inquiries.
The numbers
The headline figure is growth. According to the report, "from 2021 to 2025, the DPC Supervision Function has seen a tenfold increase in AI-related engagements with large technology controllers." A footnote records that preliminary figures given to the Oireachtas Joint Committee on Artificial Intelligence on February 24, 2026 had described "over a six-fold increase." The final number uses the full calendar year of 2025.
Generative AI drove most of it. The DPC counted 25 generative AI engagements in 2023, 36 in 2024 and 82 in 2025, for 143 in total, against 37 non-generative engagements over the whole period. Generative AI therefore accounted for 79% of all AI engagements, even though, according to the report, generative engagements only began in 2023.
| Metric | Figure |
|---|---|
| AI products and services engaged on, 2021-2025 | Approximately 180 |
| Generative AI engagements | 143 (25 in 2023, 36 in 2024, 82 in 2025) |
| Non-generative AI engagements | 37 |
| Generative share of all AI engagements | 79% |
| Growth in AI engagements, 2021 to 2025 | Tenfold (preliminary figure, Feb 24, 2026: over six-fold) |
| AI share of Supervision engagements | 1 in 35 (2021); around 1 in 4 with large technology controllers (2025) |
| AI products launched in the EU that received recommendations, 2024-2025 combined | 59% |
| Recommendations per AI engagement | Median 3; highest 18; lowest 0 |
| Recommendations involving transparency | 72% |
| Legal basis as a share of data protection issues raised | 7% |
| Right to object as a share of data protection issues raised | 7% |
| Launches delayed or cancelled by DPC intervention | 1 (2022), 1 (2023), 4 (2024), 3 (2025); 9 in total (PPC Land calculation) |
| AI engagements that launched without delay due to DPC engagement | 95% |
| Minimum lead time the DPC says it needs | 8 weeks |
| Controllers engaged on agentic AI launches | 5 |
The report breaks the engagements down by technology type, with some engagements counted under more than one type.
| AI type | Engagements, 2021-2025 | Share of generative AI engagements | Share of non-generative AI engagements |
|---|---|---|---|
| Model/model creation | 46 | 23% | 19% |
| Chatbot | 22 | 13% | - |
| Uncategorised generative | 20 | 12% | - |
| Personalisation | 17 | 6% | 19% |
| Assistant | 13 | 6% | 3% |
| Video | 12 | 7% | - |
| Audio | 11 | 6% | 5% |
| Facial recognition | 9 | - | 24% |
| Image | 9 | 5% | 3% |
| Multimodal | 8 | 5% | - |
| Search | 7 | 4% | 3% |
| Agentic | 7 | 4% | - |
| Age assurance | 7 | - | 19% |
| Advertising | 7 | 4% | - |
| Persona/character | 6 | 3% | - |
| Social | 3 | 2% | - |
| Bias detection | 2 | - | 5% |
Advertising appears as its own category in seven engagements, and personalisation in 17. The report does not say which products those were.
One detail in the denominators deserves a note. The "1 in 4" figure for 2025 refers to engagements "with large technology controllers," while the totals plotted in the report's second chart reflect, according to its note, "all engagements with controllers and other stakeholders as reported in the DPC Annual Report." The ratio and the chart do not share a base, and the report does not print the yearly AI counts behind its first chart.
The report also sets this activity against wider adoption data. Citing Eurostat, it says AI use among EU enterprises rose from 13.5% to 20% in 2025, and that "52.7% of EU firms which considered using AI technologies and chose not to, cited data protection and privacy concerns as a reason why."
How a Supervision engagement works
According to the DPC, a Supervision engagement follows six stages:
- "An open and honest conversation regarding a controller's intended processing or the capabilities of launches, products, services, etc. that will process personal data."
- Questions from the DPC and answers from the controller.
- Engagement with peer EU regulators "to identify issues of concern." For high-profile matters, the unit holds what the report calls "flash meetings" early on.
- Recommendations, where necessary, "that indicate areas for improvement or point out likely issues for compliance."
- Follow-up by the controller on those recommendations.
- "Ongoing monitoring by the DPC post-launch."
For AI, the review may cover a Data Protection Impact Assessment, a Legitimate Interests Assessment and other risk documents, and may extend to monitoring how mitigations work in practice, "such as the operation of an objection form for compliance with GDPR Article 21." The process is voluntary and, according to the report, distinct from "the more formal prior consultation under Article 36 of the GDPR."
There is a limit written into the model. "The DPC will never approve or state that a process, product, or service is compliant with the GDPR through Supervision," the report says. The recommendations are "non-binding in nature," although where "redline issues" are found, the unit "can indicate that the DPC may consider other regulatory powers if the controller decides not to implement the recommendations."
Timing is the other variable. Most engagements, the DPC says, happened "shortly before the anticipated launch," and it considers "a minimum 8-week lead time is necessary to enable the DPC to engage substantively." Controllers that engaged "8 weeks or earlier, with good documentation have been predominantly able to launch without delay due to DPC engagement."
The report also states a position on impact assessments. The DPC "considers that the training of a new AI model will likely require a Data Protection Impact Assessment," because training a new model with personal data "will necessarily involve a new technology and likely large quantities of data" - two criteria drawn from the Article 29 Working Party's 2017 guidelines, WP248 rev.01.
What the recommendations said
The report quotes several recommendations verbatim, without naming the recipients. Among them:
- "The DPC recommends that clear retention periods or criteria to determine such periods, be developed and provided to users with regard to the transcription of audio data."
- "The DPC recommends that additional transparency notices should be published prior to the creation of the datasets that will be used for training and testing of the models."
- "If the controller intends to continue to process any personal data of users who object to the processing, the controller must demonstrate compelling legitimate grounds prior to launching in the EU/EEA."
- "The DPC recommends a critical re-evaluation and adjustment of the consent flow, to ensure that it meets the requirements of 'freely given' consent under Article 7(4) of the GDPR."
- "The DPC recommends that the controller develop and apply measures to remove, pseudonymise, anonymise, or otherwise protect full names of private individuals from the dataset prior to any AI training."
- "The DPC recommends that the controller take immediate steps to avoid collecting or processing personal data from pages, groups and elsewhere on the platform that would have a likely prevalence of special category data, or data capable of revealing special category data."
The analysis passed to controllers can be blunt. According to the report, it may include "Stating the DPC's opinion that a controller's legitimate interests balancing test has not been met" or "Informing a controller that the compatibility assessment of processing purposes has not been met."
The five issues cited most often in AI recommendations were transparency, technical and organisational measures, Article 25 (data protection by design and default), data minimisation and legal basis. On PPC Land's reading of the report's unlabelled bar chart, transparency was cited roughly four times as often as the next issue. The full ranking continues with objection, children-focused issues, fairness, retention, data subject rights, Article 9, purpose and compatibility, Article 35 and DPIAs, non-user data subjects, notice, Article 22 and the Article 15 right of access.
When engagement turns into enforcement
The DPC is careful to present Supervision as cooperative. Yet the report also catalogues where cooperation ran out.
The sharpest example involves X. According to the report, the DPC learned in July 2024 "that key mitigation measures were not in place with regards to Twitter training its Grok AI model using user personal data." The DPC asked Twitter to stop. "Twitter refused," the report says. On August 8, 2024, the regulator applied to the Irish High Court under section 134 of the Data Protection Act 2018, a power that lets the Commission seek an order "suspending, restricting or prohibiting the processing of personal data" where there is "an urgent need to act." Following a hearing, Twitter "undertook to cease processing user data that had been collected during the period of time when the mitigations were not in place, and to delete the datasets that included such data." Twitter's European entity was rebranded X Internet Unlimited Company in April 2025, according to the report.
Three formal inquiries under Article 58 of the GDPR were still open "at the time of writing":
| Company | Subject | Opened |
|---|---|---|
| Whether an Article 35 assessment was required before processing EU/EEA personal data for the PaLM 2 foundation model | September 12, 2024 | |
| X Internet Unlimited Company | Training of generative AI models, in particular the Grok large language model | April 11, 2025 |
| X Internet Unlimited Company | Processing of EU/EEA personal data, including children's, through Grok functionality within the X platform | February 17, 2026 |
PPC Land reported the April 2025 Grok training inquiry when it opened, and Grok's image generation has since drawn separate action elsewhere, including a European Commission DSA investigation in January 2026 and an Amsterdam court order in March 2026 against non-consensual nude imagery.
Post-launch monitoring produced its own example. In January 2025, according to the report, the Supervision Function found that the opt-out from AI training for "a particular chatbot was not available to non-subscribers using mobile devices, although access to the chatbot was available to all users." The cause was a bug. "Approximately 7.7 million mobile users in the EEA could potentially have been impacted," the report says. A fix was in place by February 2025, and the controller "committed to not training AI on any data collected during the period when the opt out was unavailable." The chatbot is not named.
Since 2023, the DPC has also required periodic reports from controllers building and deploying large language models. These rely on live data to evaluate the performance of the models, the performance of safeguards, "reactions of actual individuals or the wider public," and other real-world privacy issues. Each report has required the controller to re-evaluate its DPIA and LIA against what actually happened. In one case cited, the controller identified additional data points to exclude from training, fixed output errors reported by users and added work "to identify and block hallucinations."
Meta: the pause and the restart
Meta's case study is the most detailed in the report. In May 2024, Meta announced its intention to train AI on user data. Almost immediately, according to the DPC, "difficulties with users attempting to exercise their right to object and confusion around the transparency notices became apparent." After reviewing Meta's 2024 LIA, the DPC concluded that "Meta's implementation of transparency notices combined with multiple objection form issues (including difficulties in accessing and completing an objection) fundamentally undermined Meta's reliance of the legitimate interests lawful basis." It asked Meta to pause before training began, and Meta complied.
Meta came back in February 2025 with a revised LIA and DPIA. This time the notice and objection rollout "did not encounter the same issues that occurred in 2024." After consulting other EU and EEA authorities, the DPC issued further recommendations, which Meta implemented. Measures Meta committed to included "filtering of data and de-identification prior to training." The DPC concluded "there were no grounds justifying further injunctive regulatory action at that time" - while stating explicitly that it was "not determining that Meta had complied with the GPDR [sic]."
That restart is the one PPC Land readers saw from the outside. Meta said in April 2025 that it would use public posts and comments from adults to train generative AI under legitimate interests, with an objection form open until May 27, 2025. Privacy group noyb sent a cease and desist letter on May 14, 2025 demanding opt-in consent, and on May 23, 2025 the Cologne Higher Regional Court declined an emergency injunction, finding Meta had a legitimate interest. A noyb-commissioned survey later found that only 7% of German Meta users wanted their data used for AI. The DPC's report adds the regulator's side of that sequence, but it does not engage with those disputes.
Transparency in 72% of recommendations
"Transparency issues have been present in 72% of all recommendations made by the DPC Supervision Function," according to the report. Some controllers were reluctant to tell users about AI processing at all, occasionally for reasons they considered legitimate, such as fraud-detection systems that bad actors could evade if they knew how they worked. The DPC's answer was flat: "no amount of post-processing transparency will relieve a controller of its obligations to provide transparency prior to processing under Articles 5, 12, 13, and 14 of the GDPR."
The case studies show how far the DPC pushes this. Two controllers planned AI agents that capture screenshots of a virtual browser; the DPC recommended that they "tailor warnings specifically to the use of agents." In 2023, it asked one chatbot developer to publish notices explaining the privacy risks of generative AI, not just the fact of processing, and "since then, the DPC has made similar recommendations to 5 other controllers developing LLMs." In another case, a controller had addressed its AI notice only "to verified users, in a section of their website that was unlikely to be accessed by non-verified users." The DPC also flagged non-users - people whose data is processed without their using the service - as an area "consistently lacking."
Delivery channels matter too. Where users do not log in regularly, the report says controllers "should consider issuing transparency information external to their platform, such as through the use of email." On blog posts, the DPC is sceptical: "it is unclear how many average users of a service actually see or would regularly access such transparency information."
In three separate cases where controllers planned to train AI on user data without adequately informing users, the DPC recommended both a notice and a notification period. All three accepted.
Legal basis and the 2014 cut-off
Legal basis accounted for only 7% of issues raised, but the report describes those recommendations as "often expansive and of critical importance." Since the EDPB's Opinion 28/2024 in December 2024 confirmed legitimate interests under Article 6(1)(f) as a possible basis for AI training, the DPC has applied that guidance when assessing LLMs and other AI systems, and it analyses every claim to rely on Article 6(1)(f). According to the report, the factors it weighs include transparency and the reasonable expectations of users, "Ability and ease of an individual to exercise their rights, and any friction created by the implementation of data subject rights request forms," necessity, compatibility with original purposes, "whether there are any less intrusive ways to meet the goals of the processing," mitigations and the controller's balancing test.
Historical data receives particular scrutiny. One controller wanted to train generative AI on "all public user posts, comments and articles collected from the inception of the platform," arguing that this fell within the original purpose to "enable and improve" its services. The DPC asked for every historical privacy policy. Before 2014, it found, purposes were "confined largely to the improvement of cookies or failed to have any reference to improvement of service." The DPC recommended the controller not use any personal data collected before 2014 "at the very earliest," and the controller cut its training window to 2014 onwards. The platform is not named.
The underlying reasoning is spelled out. "Public discourse concerning AI, in particular regarding generative AI, would suggest that many individuals have only in recent years become aware of AI in any substantive sense," the report says, so "it is likely that few would expect their personal data to be processed for or by AI."
Private messages: "very difficult to achieve"
The most consequential position in the report sits outside its formal scope. During the first half of 2026, according to the DPC, it was asked to assess "the use of private messages for AI training, such as private communications between users via a messaging service." Its conclusion, word for word:
"Given that there is an inherent expectation of privacy to private communications, the DPC considers it is unlikely that a controller will be able to demonstrate that data subjects expect the usage of their private messages or other private data for AI training."
Awareness alone does not change that. The DPC "considers it insufficient to demonstrate that individuals are merely aware of the possibility of their private messages being used for AI training." A controller would have to show that the expectation of privacy "does not exist" or that it had informed people, "including third parties," to the point that they would expect it.
The necessity test tightens further. Any LIA would need to explain "why other data sources are inappropriate and why the private messages are the only suitable data," with synthetic data and publicly available sources cited as alternatives. The report then states: "a controller will have to meet a very considerably higher threshold when justifying the use of private messages than if the controller were to process publicly available information. It is the DPC's assessment that this threshold will be very difficult to achieve in the context of the use of private messaging data for AI training purposes."
The GDPR is not the only obstacle. The DPC points to the ePrivacy Directive's guarantee of confidentiality of communications and says "making private messages available for AI training would significantly erode the trust and confidence of users." Controllers "will have to demonstrate that any usage of private messages for AI training does not breach the ePrivacy Directive, as well as the GDPR." The report does not say which controller raised the question.
Objections without friction
Article 21's right to object featured in 7% of issues raised, though the report notes that one objection form often covered several products from the same company. With current technology, the DPC says, it "can be very difficult to remove any training data" from a trained model, which means "there may be no effective way for users to exercise their Article 21 rights if not provided the opportunity to opt out or object prior to the processing beginning."
The language here is some of the firmest in the report: "The DPC considers that any chilling effect on the exercise of data subjects rights is unacceptable and any friction to access an objection form or method should be minimised to the greatest extent possible." It singles out one phrase as misleading: "if your objection is honoured..." when the controller intends an unconditional opt-out.
Recommendations the DPC has made on objection include, according to the report:
- "Being clear on the scope, extent, and duration of an objection."
- "Ensuring that all users, regardless of what type of device is used, have received the transparency notices regarding this processing and the ability to opt out prior to their personal data being used for training."
- "Where applicable, making clear that any objection to AI training is treated as an unconditional and automatically honoured opt-out."
- "Ensuring users who have made objections are aware of the status of their request."
The DPC also considers it "insufficient to provide an objection form at the time the processing that AI training begins." And because the GDPR does not distinguish between data collected directly and third-party data, "controllers must ensure that individuals have the ability to object, even if the controller did not receive the personal data directly from the individual."
Minimisation, children and automated decisions
On data minimisation, the DPC concedes that "it is currently unclear from the state of the art what constitutes 'enough' data for the creation of LLMs." Measures it encourages include pseudonymisation or anonymisation, de-duplication, shorter collection windows, avoiding under-18s' data, avoiding sources likely to contain special category data, and avoiding "sources of personal data where the user may not be aware of the processing, such as dormant accounts." Where a controller's own risk assessment flags a data point for removal from one dataset, the DPC has recommended removing it from all.
One case shows minimisation at the deployment stage. Two controllers built voice AI that kept recording after the user's last interaction, "in one case for up to one hour," because that was "the maximum capacity of the AI." There was no wake word. The DPC recommended shorter recording windows and an audio reminder that the AI was still listening. "Simply because it is technologically possible to process personal data for a period of time does not mean that doing so complies with the principles of data minimisation," the report says. It also rejects the argument, made by "some organisations," that users are responsible for limiting what they share, calling it "clearly against the CJEU jurisprudence" in cases such as C-73/07 Satamedia.
On children, the DPC recommended to six controllers, covering seven products, greater transparency on data protection risks and "ongoing in-app transparency to support children in learning about tools to check the veracity of generative AI outputs." Most controllers chose not to train on children's data at all; where a controller made that choice, the DPC recommended it also exclude any account where it was "unsure if the user is a child or not." Three controllers planned to start training on users' data once they turned 18. The DPC recommended a dedicated notice period and objection opportunity for those users. Children's data has surfaced in court too: a German appeals court acknowledged in August 2025 that minors' data reaches Meta's training through adult users' posts.
On automated decision-making, one controller used AI to police disruptive behaviour in voice chats, with the power to ban users. The controller argued Article 22 was satisfied because another user triggered the system and appeals involved a human. The DPC disagreed: "triggering an AI system by a human or an appeals process with human intervention does not necessarily ensure compliance." The controller added human review before any ban.
Harmonisation, bias and the Omnibus
The report traces the EDPB's AI Opinion back to a request from the DPC. In summer 2023, it says, "it became clear to the DPC that there were a number of differing opinions, interpretations, and understandings when it came to processing personal data for the training of AI." The DPC shared insights through EDPB subgroups and the ChatGPT Taskforce, now the Generative AI Enforcement Taskforce, and requested an Article 64(2) opinion. The result was Opinion 28/2024, which "clarified when an AI model could be considered anonymous and that Article 6(1)(f) was a possible legal basis for training AI." Regulators have not fully converged since; a study PPC Land covered in March 2026 found that data protection authorities still diverge on how to apply legitimate interest, and the EDPB's July 2026 web scraping guidelines set their own legitimate interest test for scraped training data.
Bias detection is where the DPC held back. Testing whether a dataset is representative can require processing special category data, which Article 9(1) of the GDPR generally prohibits. Article 10(5) of the AI Act allows it exceptionally for bias detection and correction, but, according to the report, "Until the Digital Omnibus on AI was published in late 2025, it was unclear if this was intended to be an exception under 9(2)(g) of the GDPR or a new standalone exception." The DPC "refrained from taking a position" and says its stance "may change in the future." The Omnibus on AI reached a provisional agreement on May 7, 2026, with formal adoption still pending at the time.
On special category data in training sets more generally, the DPC took a position before any guidance existed, advocating de-identification, pseudonymisation or anonymisation. "When the DPC made such recommendations, no regulatory guidance required such measures," the report says, adding that the approach "is now being followed in the draft Digital Omnibus on AI, which has been welcomed by the EDPB."
The legislative picture is less settled than that sentence implies. In February 2026, the EDPB and EDPS rejected parts of the Commission's Digital Omnibus and questioned its proposed AI training provision. And four days before the DPC published, PPC Land reported that a leaked Irish Council Presidency compromise dated September 3, 2026 drops the unconditional objection right from the GDPR's AI training clause. The Presidency is run by the Irish government, not the DPC, but the contrast is plain: the regulator's own recommendations describe an "unconditional and automatically honoured opt-out" as a safeguard it has asked controllers to provide.
Agents and what comes next
Agentic AI is the area the DPC flags for the coming years. It has engaged with five separate controllers on agentic launches in the EU and has issued recommendations, but "the common themes across controllers were not yet clear at time of writing beyond a continuing lack of transparency and clarity on how the AI processes personal data."
The report also concedes a limit to its own remit. Image and video generators capable of producing "non-consensual intimate or sexualised images" may be better addressed by harassment, defamation or intimate image laws than by data protection law. Even so, the DPC says it "has and will continues to act" where personal data processing is in question, citing its February 2026 investigation into X. The report does not address supervision under the AI Act.
Why this matters for the marketing community
For advertisers and publishers, the DPC is not a distant regulator. As lead supervisory authority for companies whose European headquarters sit in Ireland, it oversees many of the platforms through which digital ad budgets flow in Europe. Seven of the engagements counted in the report concerned advertising AI, 17 involved personalisation, and the named companies include Meta, Google, Microsoft, LinkedIn, Pinterest, TikTok and X.
The report's positions therefore set expectations for products marketers already use or will be offered. The insistence on objection forms that work on every device, notices sent beyond the platform when users rarely log in, and a separate notice for users who turn 18 all describe how training on user data is expected to happen in Europe. LinkedIn's move to train generative AI on member data from November 3, 2025, under legitimate interest with an opt-out and with under-18 members excluded, matches several elements of that template.
The 95% figure can be read two ways. On the DPC's framing, almost every AI product cleared its review without delay. Yet 59% of AI launches in 2024 and 2025 left with recommendations, and the median engagement produced three. None of those recommendations are binding, and the DPC "will never approve" a product. How much changed in each case? That remains visible only to the regulator and the companies involved.
The report also arrives with a commission whose composition drew attention. Niamh Sweeney, who previously held public policy roles at Facebook and WhatsApp, became the DPC's third commissioner from October 13, 2025, an appointment noyb criticised. The report names no commissioner and carries no quoted statement.
What the report leaves out is short to list. It prints no yearly AI engagement counts or Supervision totals as numbers, only as unlabelled charts. It does not name the controllers in most case studies, including the chatbot whose opt-out bug touched about 7.7 million mobile users, the platform held to post-2014 data, or the company that raised private messages. It also gives two dates for EDPB Opinion 28/2024: adopted December 18, 2024 in a footnote, and issued December 19, 2024 in its glossary.
Key facts
- Report title: Responsible Artificial Intelligence Innovation - Insights from the Data Protection Commission's Supervision of AI (2021-2025)
- Publisher: Data Protection Commission (An Coimisiún um Chosaint Sonraí), Ireland
- Cover date: September 2026; the DPC's publication page shows September 24, 2026 and its news release "Data Protection Commission publishes AI Insights Report" shows September 25, 2026
- Length: 75 pages, with Appendix A (key takeaways) and Appendix B (statistics reference)
- Unit covered: Technology Multinational Supervision Unit (TMSU), Supervision Function
- Period covered: January 2021 to December 2025, with a section on private messages assessed in the first half of 2026
- Companies named: Airbnb, Apple, Deepseek, Google, LinkedIn, Meta, Microsoft, OpenAI, Pinterest, TikTok, X
- Open inquiries listed: Google PaLM 2 (opened September 12, 2024); X Internet Unlimited Company, Grok training (April 11, 2025); X Internet Unlimited Company, Grok on the X platform including children's data (February 17, 2026)
- High Court application: Section 134, Data Protection Act 2018, filed August 8, 2024 (Twitter, Grok training)
- Legal references: GDPR Articles 5, 6(1)(f), 6(4), 7(4), 9, 12-14, 21, 22, 25, 31, 35, 36, 58; AI Act Article 10(5); EDPB Opinion 28/2024; WP29 Guidelines WP248 rev.01 (October 4, 2017); CJEU C-73/07 Satamedia; ePrivacy Directive 2002/58/EC; EDPB Helsinki Statement (July 3, 2025)
- EU proposals referenced: Digital Omnibus, COM(2025) 837 final, and Digital Omnibus on AI, COM(2025) 836 final, both Brussels, November 19, 2025
- Source: https://www.dataprotection.ie/sites/default/files/uploads/2026-09/DPC-AI-Insights-Report-AC.pdf (publication page: https://dataprotection.ie/en/dpc-guidance/publications/dpc-AI-insights-report)
Timeline
- 2021 - AI accounts for 1 in every 35 Supervision engagements
- 2022 - One AI launch delayed or cancelled due to DPC intervention
- 2023 - DPC starts requiring monitoring reports from controllers creating and deploying LLMs; 25 generative AI engagements; one launch delayed or cancelled
- Summer 2023 - DPC begins sharing AI training insights through EDPB subgroups and the ChatGPT Taskforce
- May 2024 - Meta announces plans to train AI on user data; the DPC later asks it to pause, and Meta complies
- July 2024 - DPC finds key mitigations missing for Twitter's Grok training; Twitter refuses to stop
- August 8, 2024 - DPC applies to the Irish High Court under section 134 of the Data Protection Act 2018
- September 12, 2024 - DPC opens an inquiry into Google's PaLM 2
- December 2024 - EDPB issues Opinion 28/2024 on AI models
- 2024 - 36 generative AI engagements; four launches delayed or cancelled
- January 2025 - Monitoring finds a chatbot's AI-training opt-out missing for mobile non-subscribers, potentially affecting about 7.7 million EEA users; fixed by February 2025
- February 2025 - Meta returns with a revised LIA and DPIA
- April 2025 - Twitter International Unlimited Company is rebranded X Internet Unlimited Company
- April 11, 2025 - DPC opens an inquiry into Grok LLM training
- April 2025 - Meta says it will train AI on adults' public posts, with objections open until May 27, 2025
- May 14, 2025 - noyb sends Meta a cease and desist letter over AI training
- May 23, 2025 - Cologne Higher Regional Court declines to block Meta's AI training
- June 2025 - noyb survey finds 7% of German Meta users want their data used for AI
- August 12, 2025 - German appeals court acknowledges children's data reaches Meta's AI training via adult posts
- September 17, 2025 - Irish government appoints Niamh Sweeney as third DPC commissioner
- September 18, 2025 - LinkedIn sets November 3, 2025 for training AI on member data
- November 19, 2025 - European Commission publishes the Digital Omnibus and the Digital Omnibus on AI
- 2025 - 82 generative AI engagements; three launches delayed or cancelled; AI reaches around 1 in 4 Supervision engagements with large technology controllers
- January 26, 2026 - European Commission opens a DSA investigation into X over Grok
- February 10, 2026 - EDPB and EDPS adopt a joint opinion rejecting parts of the Digital Omnibus
- February 17, 2026 - DPC opens an investigation into Grok functionality on the X platform
- February 24, 2026 - Preliminary figures to the Oireachtas Joint Committee on AI put AI engagement growth at over six-fold
- March 2026 - Study finds data protection authorities still diverge on legitimate interest for AI training
- March 26, 2026 - Amsterdam court orders X and xAI to stop Grok generating non-consensual nude images
- First half of 2026 - DPC assesses the use of private messages for AI training
- May 7, 2026 - Provisional agreement reached on the Digital Omnibus on AI
- July 7, 2026 - EDPB adopts web scraping guidelines for generative AI
- September 3, 2026 - Irish Council Presidency compromise drops the unconditional opt-out from the GDPR AI clause
- September 25, 2026 - DPC publishes its AI Insights Report covering 2021-2025
Related PPC Land coverage
- Irish DPC launches Grok LLM training inquiry - The April 2025 inquiry into X's use of EU users' public posts to train Grok, one of three AI inquiries the report lists as open.
- Meta to use public posts to train AI models, users can opt out - Meta's April 2025 plan to train generative AI on adults' public content under legitimate interests, with an objection deadline of May 27, 2025.
- Meta faces legal battle as noyb sends cease and desist over AI training - noyb's May 2025 demand that Meta seek opt-in consent before training on EU data.
- German court allows Meta's AI training with public data - The Cologne Higher Regional Court's refusal of an emergency injunction on May 23, 2025.
- Only 7% want Meta to use their data for AI, finds survey - A Gallup survey for noyb of 1,000 German Meta users on AI training.
- German court confirms Meta AI training includes children's data despite protections - The Schleswig-Holstein court's August 2025 ruling on minors' data in adult posts.
- LinkedIn expands AI training to include user data starting November 3 - LinkedIn's opt-out based plan to train generative AI on member data in the EEA, Switzerland and the UK.
- Former Meta executive becomes data protection commissioner - Niamh Sweeney's appointment to the DPC and the criticism it drew.
- GDPR's AI training legal battle: regulators converge but still clash - A study of 19 data protection authority guidelines on legitimate interest for AI training.
- Europe's privacy watchdogs reject Commission's plan to narrow GDPR protections - The EDPB-EDPS joint opinion on the Digital Omnibus and its AI training provision.
- EU AI Act gets its first real haircut - high-risk deadlines pushed to 2027 - The provisional deal on the Digital Omnibus on AI.
- EDPB blocks AI firms from using consent as an excuse to scrape - EDPB Guidelines 03/2026 on web scraping for generative AI training.
- EU Council draft drops unconditional opt-out from GDPR AI clause - The leaked Irish Presidency compromise on the GDPR's AI training article.
- Europe confronts X over Grok risks: child safety fears trigger DSA probe - The European Commission's January 2026 DSA investigation into Grok on X.
- Dutch court bans Grok from generating non-consensual nude images - The Amsterdam preliminary relief ruling against xAI, X Corp and X Internet Unlimited Company.
Summary
Who: Ireland's Data Protection Commission, through its Technology Multinational Supervision Unit, and the technology companies it supervised, including Airbnb, Apple, Deepseek, Google, LinkedIn, Meta, Microsoft, OpenAI, Pinterest, TikTok and X.
What: A 75-page report on five years of AI supervision covering about 180 AI products and services. It records a tenfold rise in AI engagements, nine launches delayed or cancelled, recommendations for 59% of AI products launched in 2024 and 2025, and transparency issues in 72% of recommendations. It also states that training AI on private messages faces a threshold the DPC calls "very difficult to achieve."
When: Published September 25, 2026, covering January 2021 to December 2025, with a section on private messages assessed in the first half of 2026.
Where: Ireland, with effects across the EU and EEA, because the DPC is lead supervisory authority for many large technology companies with European headquarters in Ireland.
Why: The DPC says early, voluntary engagement lets it shape AI products before launch. For marketers, the report shows the conditions under which the platforms they buy advertising on can train AI on user data in Europe: notices, notice periods, objection forms that work on every device, and a much higher bar for private messages.
Discussion