Ireland’s cross-border GDPR cases take a median 6.2 years to decide. Forecasters give strong AGI better than even odds of arriving first. The EU’s AI Act is built to be enforced the same way. data/code
TLDR:
The median cross-border case takes 6.2 years (counting still-open investigations).
66% of the 2018-2020 cross-border cases were still open at 4.5 years.
If AI Act cases take as long, strong AGI will probably arrive before they close, so oversight has to happen before a model ships.
Under the GDPR, cross-border enforcement is led from the country where a company has its EU headquarters, and most of big tech (OpenAI and Anthropic included) has theirs in Dublin. That makes Ireland’s Data Protection Commission (DPC) the primary privacy regulator for the tech industry in Europe. Since data protection law is one of the main ways tech companies are regulated, the DPC’s record is approximately the best evidence there is on enforcement speed. The question matters beyond privacy, since the EU’s AI Act and most of the new American state AI laws are set up to be enforced similarly, with investigations, fines, and appeals.[1] To assess this, I compiled the 69 investigations with published or announced final decisions under the Data Protection Act 2018, from August 2019 through September 2026, and measured each one from formal commencement to final decision. Then I compared the durations to forecasts of when artificial general intelligence will arrive. They’re roughly the same length.
Figure 1 (below) compares the durations of cross-border cases with AGI horizons, measured from September 25, 2026. Every decided case in the figure took longer than the 1.6 years remaining until the forecasters’ 25% date for strong AGI (described below).
Figure 1. Blue bars are every cross-border inquiry that ended in a fine of €1 million or more, commencement to final decision, labeled with the fine. Gray bars are still-open inquiries, shown to date. The red lines mark the Metaculus forecast for when the chance that strong AGI has arrived reaches 25% and 50%.
Figure 2. The blue line is the share of 2018-2020 cross-border cases decided within X years, counting the 14 still-open or discontinued cases as censored. It flattens after the last decision because 43% of cases are still undecided. The red curves give the forecast probability that weakly general AI (light red) or strong AGI (dark red) arrives within X years of September 25, 2026; the curves start near 26% and 10% because forecasters put those chances on the milestones having already been reached.
If an investigation opened today took as long as the median cross-border case (6.2 years), forecasters would give weakly general AI an 84% chance of existing before it concluded (59% for the strong form). Google’s real-time bidding inquiry is already at 7.3 years, where AGI probabilities reach 86% and 64%.
Structurally, the GDPR and AI Act differ in that the European Commission’s AI Office enforces the AI Act, so its cases won’t require back-and-forth between national regulators the way cross-border GDPR cases do. But this doesn’t seem to be driving the investigation length; the procedural histories of fourteen of the fifteen cross-border cases with fines indicate that more than four-fifths of the time passed before any other EU regulator saw a draft decision. Instead, most of the length is accounted for by investigating, drafting, and allowing the company to respond. The AI Act requires the same steps before the Commission can issue a fine. Its procedural rules, adopted in July 2026, give the Commission five years from a violation to impose a fine (up to ten years if investigative steps keep restarting the clock).
But investigating after the fact is only one type of enforcement. The table below compares it with three alternatives (emergency orders, clear-cut duties such as reporting requirements, and approval before release) on how fast each worked, and on whether an intervention at that speed arrives before the technology transforms (ordered from weakest for AI safety to strongest).
In August 2024, X gave a High Court undertaking to stop using the public posts it had collected from Europeans that summer to train Grok. The DPC opened a formal inquiry eight months later.
Potentially useful for containment, but it may be difficult to unwind a model’s training or deployment after the fact.
Reporting duties can be enforced quickly but are narrow in scope. Other tools benefit from the information they provide.
Prevention (approval before release)
Not required by the GDPR. But the DPC says its voluntary pre-launch supervision delayed or stopped nine of about 180 AI products between 2021 and 2025.
Strongest on paper for irreversible harms. But untested and requires global oversight.
In data protection, emergency powers have sometimes reversed or changed behavior. When the DPC took X to the High Court in August 2024, the company agreed to stop training Grok on posts it had collected that summer. Italy's data protection agency blocked both ChatGPT (for a month) and DeepSeek over privacy policy issues. In the same vein, Article 93 of the AI Act lets the Commission require mitigation or withdraw a model, prior to an investigation concluding an infringement occurred. But these emergency powers require decisive action from regulators, knowledge that events are happening, and, in the AI context, it may not be straightforward to put the genie back in the bottle (e.g. unwind a model’s training or deployment) after the fact.
Simpler rules may be easier to enforce quickly. Investigating whether a public body had appointed a data protection officer took 81 days. Whether a hospital’s security measures were “appropriate” took more than six years. The EU’s AI Act contains both kinds of rules. A developer has two weeks to notify the Commission once a model’s training run crosses a compute threshold (simple). Developers of the most capable models must also report “serious incidents” without undue delay (complex). Within weeks of its fining powers switching on in August, a Commission spokesperson confirmed that OpenAI had notified the Hugging Face intrusion to the AI Office but had not filed a report on a later incident that outside researchers uncovered (the open question is whether it was sufficiently “serious” to require a disclosure). Where possible, reporting rules should be simple and categorical to sidestep debates about eligibility, providing more information to enable decisive action.
Approval before a model is released (like a new drug), the strongest tool in the table, has no formal track record in Europe’s digital regulation. The AI Act does have a two-week notice for crossing a compute threshold, and developers who signed its code of practice also file a safety report with the AI Office before launch. But companies aren’t required to wait for a response before launching. This is different from, for example, EU merger control, which prohibits a deal from closing until the Commission clears it (or approves by default by not ruling within a few months).
If AGI forecasts are anywhere close, enforcement that depends on investigations will keep arriving model generations too late. Regulation aimed at catastrophic risk has to operate at training and deployment, before something has gone wrong. Prevention has worked before. The DPC says its pre-launch reviews (made on a “voluntary basis”[2]) delayed or stopped nine AI launches between 2021 and 2025. A pause after the AI Act’s two-week notice, with a fixed review period like merger control’s, would give the AI Office a set window to review a model before it ships. Doing this well requires a muscular agency, and the AI Office may be under-resourced for its mandate. As with data protection – where researchers, journalists, and civil society provided information about or raised the salience of potential infringements – the AI Act may benefit from third parties sounding the fire alarm to enable oversight.
Appendix
Data and method
Durations are measured from formal inquiry commencement to final decision, excluding pre-inquiry complaint handling and appeals (thus understating the full timeline). Figure 2’s blue curve and the 6.2-year median are Kaplan-Meier estimates over the 30 cross-border cases that were formally begun by the end of 2020 (statutory inquiries plus three complaint decisions), 16 of them decided, 13 still open on the basis date and one discontinued in 2022; without the three complaint decisions the median is 6.6 years, so 6.2 years is a floor. Censored cases with unpublished start dates are given the latest-possible starts (e.g. December 31 of the year when they first appeared in an Annual Report), so the estimate is conservative. Eight open cases with no DPC status report since its 2020 to 2023 Annual Reports are assumed still open.[3]Eight Law Enforcement Directive decisions are excluded. Start dates are exact to the day for 63 of the 69 concluded inquiries; for the other six (all domestic cases), no published document gives the start date, so I inferred it from what was available (a twin inquiry into the same incident, correspondence dates in the decision, the DPC’s Annual Reports, or the month encoded in the DPC’s case number). The share of time before other regulators saw a draft comes from the procedural histories in fourteen of the fifteen fined cross-border cases; the fifteenth, Google’s location-data decision, isn’t published yet. Durations, open-inquiry ages, and the forecast curves are all measured as of September 25, 2026. Sources are the DPC decisions register and PDFs (plus two decisions announced but not yet published), DPC Annual Reports and press releases, company filings, and Metaculus questions 5121 and 3479 as captured on September 25, 2026. The Annual Reports mention at least nine further final decisions from 2023 to 2025 (mostly no-infringement outcomes in complaint cases) that were never posted to the decisions register and so cannot be included.
Inspired by
BlueDot’s AI alignment project idea: “Data protection law has been one of the main ways tech companies have been regulated in the past... examine how well this has worked, and in particular what enforcement has worked, and where there are gaps in enforcement that are being exploited...then take these learnings and suggest ways AI regulations could be enforced, evaluating multiple enforcement strategies against how effectively they will reduce catastrophic risks.” On reflection, anchoring on data protection might have been a worse starting point than the Digital Services Act or Digital Markets Act, which are digital platform regulations that (like the AI Act) are enforced by the European Commission. Enforcement under these laws has proceeded somewhat faster.[4] Still, even the Commission's fastest decisions took 13 months, and most DSA cases are still open. Frontier labs ship new models faster than that.
American state enforcement seems to take about as long or only somewhat less time (e.g., the 40-state investigation of Google’s location tracking opened after an August 2018 AP story and settled in November 2022, 4.3 years later). The DPC's investigation of the same Google location setting resulted in a September 2026 €403 million decision.
The DPC describes this process as follows: “The DPC’s Supervision Function operates completely separately from other regulatory functions in the DPC, such as those involved in complaints or inquiries. Rather than looking at infringements of the GDPR that are alleged to have occurred in the past, the Supervision Function engages with controllers on a voluntary basis, primarily before a product or service launches” (page 9).
If instead I stop tracking each of these eight cases after its last mention in an Annual Report (i.e. treat the closing date as unknown after last mention), the median falls to 5.4 years, and the chance that strong AGI arrives before a median-length case closes falls from 59% to 56%. If I code each case as decided at the time of its last Annual Report mention, the median is 4.4 years and the chance is 49%. One of the eight (La Quadrature du Net’s Apple ad identifier complaint) may be an Apple complaint that the DPC’s 2023 Annual Report says was decided in November 2023. Giving it a November 2023 conclusion date leaves the median at 6.2 years.
The Commission’s platform cases have been faster than the DPC’s, though its record is short, making the censoring from “still open cases” more influential. It has opened 15 formal proceedings under the Digital Services Act since December 2023 and fined three companies after 1.6 to 2.3 years. Ten of the fifteen cases haven’t reached a decision yet (the oldest has been open 2.4 years). Under the Digital Markets Act, the first fines occurred 13 months after the cases opened, and Google’s €890 million in fines came after 28 months, meaning none of the decisions met the Act’s 12-month timeline target. Two of the seven non-compliance cases still remain open (Apple’s contract-terms case has been open 27 months). Four proceedings specifying how a gatekeeper must comply finished within their binding six-month deadline.
Ireland’s cross-border GDPR cases take a median 6.2 years to decide. Forecasters give strong AGI better than even odds of arriving first. The EU’s AI Act is built to be enforced the same way. data/code
TLDR:
Under the GDPR, cross-border enforcement is led from the country where a company has its EU headquarters, and most of big tech (OpenAI and Anthropic included) has theirs in Dublin. That makes Ireland’s Data Protection Commission (DPC) the primary privacy regulator for the tech industry in Europe. Since data protection law is one of the main ways tech companies are regulated, the DPC’s record is approximately the best evidence there is on enforcement speed. The question matters beyond privacy, since the EU’s AI Act and most of the new American state AI laws are set up to be enforced similarly, with investigations, fines, and appeals.[1] To assess this, I compiled the 69 investigations with published or announced final decisions under the Data Protection Act 2018, from August 2019 through September 2026, and measured each one from formal commencement to final decision. Then I compared the durations to forecasts of when artificial general intelligence will arrive. They’re roughly the same length.
How long GDPR enforcement actually takes
Of 61 concluded GDPR cases, the median investigation took 1.9 years from opening to final decision. But more than half of those cases are against Irish public bodies and domestic firms. The 15 cases that ended in fines of €1 million or more, all against large platforms, took a median of 4.4 years. LinkedIn’s behavioral advertising case took 6.2 years from inquiry to decision. Ireland’s health service reported ransomware at a hospital laboratory in November 2018; the final decision came in June 2026 (7.6 years later).
And these numbers understate things, because many of the slowest cases remain unresolved. Of the 27 open cross-border investigations at the end of 2020, 13 were still open as of September 25, 2026, and one more had been discontinued. Treating the unfinished cases as censored observations, the median for that cohort comes out to at least 6.2 years. Google’s real-time bidding inquiry has been open 7.3 years, and Tinder’s has been open 6.6 years (the DPC sent the company a draft decision in July 2026, more than six years in). Google’s location-data inquiry ended with a €403 million fine announced on September 21, 2026, 6.6 years after it opened. Even a final decision doesn’t conclude the process, since fines only become collectable after a court confirms them or appeals end. By the end of 2025 the DPC had imposed €4.04 billion in fines, but roughly €20 million of that had actually been collected.
Figure 1 (below) compares the durations of cross-border cases with AGI horizons, measured from September 25, 2026. Every decided case in the figure took longer than the 1.6 years remaining until the forecasters’ 25% date for strong AGI (described below).
Figure 1. Blue bars are every cross-border inquiry that ended in a fine of €1 million or more, commencement to final decision, labeled with the fine. Gray bars are still-open inquiries, shown to date. The red lines mark the Metaculus forecast for when the chance that strong AGI has arrived reaches 25% and 50%.
What the AGI forecasts say
Metaculus asks when someone will demonstrate a strong form of AGI (a single system that passes an adversarial Turing test, assembles a model car, and aces exam/coding benchmarks). As of September 25, its 1,843 forecasters put the median arrival at March 2031 (4.5 years out), with a 25% chance by April 2028. Metaculus has another question about weaker AGI with a median in September 2027 (a year away).
Figure 2. The blue line is the share of 2018-2020 cross-border cases decided within X years, counting the 14 still-open or discontinued cases as censored. It flattens after the last decision because 43% of cases are still undecided. The red curves give the forecast probability that weakly general AI (light red) or strong AGI (dark red) arrives within X years of September 25, 2026; the curves start near 26% and 10% because forecasters put those chances on the milestones having already been reached.
If an investigation opened today took as long as the median cross-border case (6.2 years), forecasters would give weakly general AI an 84% chance of existing before it concluded (59% for the strong form). Google’s real-time bidding inquiry is already at 7.3 years, where AGI probabilities reach 86% and 64%.
The first AI-training inquiries suggest that AI-related GDPR enforcement is not much faster. Since September 2024 the DPC has been examining whether Google did the required risk assessment before training PaLM 2, the model behind Bard, on Europeans’ personal data, and since April 2025 whether X lawfully used its European users’ public posts to train Grok. Neither case has been decided yet. Google retired Bard months before its inquiry even opened.
What this means for AI regulation
Structurally, the GDPR and AI Act differ in that the European Commission’s AI Office enforces the AI Act, so its cases won’t require back-and-forth between national regulators the way cross-border GDPR cases do. But this doesn’t seem to be driving the investigation length; the procedural histories of fourteen of the fifteen cross-border cases with fines indicate that more than four-fifths of the time passed before any other EU regulator saw a draft decision. Instead, most of the length is accounted for by investigating, drafting, and allowing the company to respond. The AI Act requires the same steps before the Commission can issue a fine. Its procedural rules, adopted in July 2026, give the Commission five years from a violation to impose a fine (up to ten years if investigative steps keep restarting the clock).
But investigating after the fact is only one type of enforcement. The table below compares it with three alternatives (emergency orders, clear-cut duties such as reporting requirements, and approval before release) on how fast each worked, and on whether an intervention at that speed arrives before the technology transforms (ordered from weakest for AI safety to strongest).
Strategy
What the record shows
Against catastrophic risk
Investigate, then fine
Cross-border cases opened by 2020 take a median of at least 6.2 years, contested practices often continue while a case runs, and the biggest fines wait on appeals.
Weak. Better than even odds that strong AGI arrives before such a case closes.
Emergency pauses and orders
In August 2024, X gave a High Court undertaking to stop using the public posts it had collected from Europeans that summer to train Grok. The DPC opened a formal inquiry eight months later.
Potentially useful for containment, but it may be difficult to unwind a model’s training or deployment after the fact.
Issue clear-cut reporting duties
Clear-cut requirements are enforceable on a fast time scale. A Data Protection Officer appointment case took 81 days.
Reporting duties can be enforced quickly but are narrow in scope. Other tools benefit from the information they provide.
Prevention (approval before release)
Not required by the GDPR. But the DPC says its voluntary pre-launch supervision delayed or stopped nine of about 180 AI products between 2021 and 2025.
Strongest on paper for irreversible harms. But untested and requires global oversight.
In data protection, emergency powers have sometimes reversed or changed behavior. When the DPC took X to the High Court in August 2024, the company agreed to stop training Grok on posts it had collected that summer. Italy's data protection agency blocked both ChatGPT (for a month) and DeepSeek over privacy policy issues. In the same vein, Article 93 of the AI Act lets the Commission require mitigation or withdraw a model, prior to an investigation concluding an infringement occurred. But these emergency powers require decisive action from regulators, knowledge that events are happening, and, in the AI context, it may not be straightforward to put the genie back in the bottle (e.g. unwind a model’s training or deployment) after the fact.
Simpler rules may be easier to enforce quickly. Investigating whether a public body had appointed a data protection officer took 81 days. Whether a hospital’s security measures were “appropriate” took more than six years. The EU’s AI Act contains both kinds of rules. A developer has two weeks to notify the Commission once a model’s training run crosses a compute threshold (simple). Developers of the most capable models must also report “serious incidents” without undue delay (complex). Within weeks of its fining powers switching on in August, a Commission spokesperson confirmed that OpenAI had notified the Hugging Face intrusion to the AI Office but had not filed a report on a later incident that outside researchers uncovered (the open question is whether it was sufficiently “serious” to require a disclosure). Where possible, reporting rules should be simple and categorical to sidestep debates about eligibility, providing more information to enable decisive action.
Approval before a model is released (like a new drug), the strongest tool in the table, has no formal track record in Europe’s digital regulation. The AI Act does have a two-week notice for crossing a compute threshold, and developers who signed its code of practice also file a safety report with the AI Office before launch. But companies aren’t required to wait for a response before launching. This is different from, for example, EU merger control, which prohibits a deal from closing until the Commission clears it (or approves by default by not ruling within a few months).
If AGI forecasts are anywhere close, enforcement that depends on investigations will keep arriving model generations too late. Regulation aimed at catastrophic risk has to operate at training and deployment, before something has gone wrong. Prevention has worked before. The DPC says its pre-launch reviews (made on a “voluntary basis”[2]) delayed or stopped nine AI launches between 2021 and 2025. A pause after the AI Act’s two-week notice, with a fixed review period like merger control’s, would give the AI Office a set window to review a model before it ships. Doing this well requires a muscular agency, and the AI Office may be under-resourced for its mandate. As with data protection – where researchers, journalists, and civil society provided information about or raised the salience of potential infringements – the AI Act may benefit from third parties sounding the fire alarm to enable oversight.
Appendix
Data and method
Durations are measured from formal inquiry commencement to final decision, excluding pre-inquiry complaint handling and appeals (thus understating the full timeline). Figure 2’s blue curve and the 6.2-year median are Kaplan-Meier estimates over the 30 cross-border cases that were formally begun by the end of 2020 (statutory inquiries plus three complaint decisions), 16 of them decided, 13 still open on the basis date and one discontinued in 2022; without the three complaint decisions the median is 6.6 years, so 6.2 years is a floor. Censored cases with unpublished start dates are given the latest-possible starts (e.g. December 31 of the year when they first appeared in an Annual Report), so the estimate is conservative. Eight open cases with no DPC status report since its 2020 to 2023 Annual Reports are assumed still open.[3] Eight Law Enforcement Directive decisions are excluded. Start dates are exact to the day for 63 of the 69 concluded inquiries; for the other six (all domestic cases), no published document gives the start date, so I inferred it from what was available (a twin inquiry into the same incident, correspondence dates in the decision, the DPC’s Annual Reports, or the month encoded in the DPC’s case number). The share of time before other regulators saw a draft comes from the procedural histories in fourteen of the fifteen fined cross-border cases; the fifteenth, Google’s location-data decision, isn’t published yet. Durations, open-inquiry ages, and the forecast curves are all measured as of September 25, 2026. Sources are the DPC decisions register and PDFs (plus two decisions announced but not yet published), DPC Annual Reports and press releases, company filings, and Metaculus questions 5121 and 3479 as captured on September 25, 2026. The Annual Reports mention at least nine further final decisions from 2023 to 2025 (mostly no-infringement outcomes in complaint cases) that were never posted to the decisions register and so cannot be included.
Inspired by
BlueDot’s AI alignment project idea: “Data protection law has been one of the main ways tech companies have been regulated in the past... examine how well this has worked, and in particular what enforcement has worked, and where there are gaps in enforcement that are being exploited...then take these learnings and suggest ways AI regulations could be enforced, evaluating multiple enforcement strategies against how effectively they will reduce catastrophic risks.” On reflection, anchoring on data protection might have been a worse starting point than the Digital Services Act or Digital Markets Act, which are digital platform regulations that (like the AI Act) are enforced by the European Commission. Enforcement under these laws has proceeded somewhat faster.[4] Still, even the Commission's fastest decisions took 13 months, and most DSA cases are still open. Frontier labs ship new models faster than that.
American state enforcement seems to take about as long or only somewhat less time (e.g., the 40-state investigation of Google’s location tracking opened after an August 2018 AP story and settled in November 2022, 4.3 years later). The DPC's investigation of the same Google location setting resulted in a September 2026 €403 million decision.
The DPC describes this process as follows: “The DPC’s Supervision Function operates completely separately from other regulatory functions in the DPC, such as those involved in complaints or inquiries. Rather than looking at infringements of the GDPR that are alleged to have occurred in the past, the Supervision Function engages with controllers on a voluntary basis, primarily before a product or service launches” (page 9).
If instead I stop tracking each of these eight cases after its last mention in an Annual Report (i.e. treat the closing date as unknown after last mention), the median falls to 5.4 years, and the chance that strong AGI arrives before a median-length case closes falls from 59% to 56%. If I code each case as decided at the time of its last Annual Report mention, the median is 4.4 years and the chance is 49%. One of the eight (La Quadrature du Net’s Apple ad identifier complaint) may be an Apple complaint that the DPC’s 2023 Annual Report says was decided in November 2023. Giving it a November 2023 conclusion date leaves the median at 6.2 years.
The Commission’s platform cases have been faster than the DPC’s, though its record is short, making the censoring from “still open cases” more influential. It has opened 15 formal proceedings under the Digital Services Act since December 2023 and fined three companies after 1.6 to 2.3 years. Ten of the fifteen cases haven’t reached a decision yet (the oldest has been open 2.4 years). Under the Digital Markets Act, the first fines occurred 13 months after the cases opened, and Google’s €890 million in fines came after 28 months, meaning none of the decisions met the Act’s 12-month timeline target. Two of the seven non-compliance cases still remain open (Apple’s contract-terms case has been open 27 months). Four proceedings specifying how a gatekeeper must comply finished within their binding six-month deadline.