DataPacts

Canadian Privacy Commissioners' OpenAI Investigation (PIPEDA Findings #2026-002)

Checked

At a glance

PIPEDA Findings #2026-002 is the report, published on May 6, 2026, of the joint investigation by the Office of the Privacy Commissioner of Canada (OPC) and the privacy offices of Quebec, British Columbia and Alberta into how OpenAI OpCo, LLC collected, used and disclosed personal information to develop and train ChatGPT.

Independent reference. Every buyer is covered on the same terms, every figure carries its source and check date, and no buyer pays to be listed or placed higher.

Request an independent introduction

Tell us what data your company holds. We reply with the buyers that fit and what their terms leave out. Your details reach a buyer only after you agree to a named introduction.

IndependentYour details stay with us until you agreeNot selling is a valid outcome

The four offices found that the way OpenAI trained its GPT-3.5 and GPT-4 models did not comply with their laws. As of November 17, 2026 the federal matter stands "well-founded and conditionally resolved" while OpenAI implements its commitments.

The report contains findings and recommendations, and no fine. This page is part of the independent reference on AI training data licensing, it gives no legal advice, and DataPacts is not affiliated with any party.

Every statement comes from the OPC's own pages, and how figures are sourced and dated is explained on the methodology page.

PIPEDA Findings #2026-002 at a Glance: Offices, Laws, Dates and Status

The investigation ran from April 2023 to May 6, 2026 under four private-sector privacy laws, and each office reached its own conclusion under the law it enforces. The table lists the record as of 2026-11-17, each row tied to the report or to an OPC page.

Field Value Source
Report Joint Investigation of OpenAI OpCo, LLC, PIPEDA Findings #2026-002 report header
Offices Office of the Privacy Commissioner of Canada (OPC); Commission d'accès à l'information du Québec (CAI); Office of the Information and Privacy Commissioner for British Columbia (OIPC-BC); Office of the Information and Privacy Commissioner of Alberta (OIPC-AB) report header
Laws PIPEDA (S.C. 2000, c. 5); Quebec's Act Respecting the Protection of Personal Information in the Private Sector; PIPA-BC; PIPA-AB para. 1
Respondent OpenAI OpCo, LLC (registered in Delaware in 2018; headquartered in San Francisco), a subsidiary of OpenAI, Inc. paras. 1 to 3
Opened April 2023 (OPC, on a complaint); joint investigations from May 2023 paras. 14, 15; OPC announcements 2023-04-04, 2023-05-25
Scope ChatGPT on GPT-3.5 and GPT-4; later models and image or video services not assessed para. 16
Published 2026-05-06 report; OPC news release
Outcome OPC: well-founded and conditionally resolved; OIPC-BC and OIPC-AB: consent well-founded and unresolved; CAI: consent and retention (Issues 2 and 6) well-founded and unresolved paras. 553 to 557
Penalty none in the report report
Status commitments under implementation with quarterly reports; no completion published by the offices as of 2026-11-17 para. 550; OPC pages read 2026-11-17

What the Four Offices Investigated: ChatGPT on GPT-3.5 and GPT-4

The offices examined ChatGPT and the GPT-3.5 and GPT-4 models that powered it when the investigation began, and did not assess later models or OpenAI's image and video services (para. 16). They examined seven issues: purpose, consent, openness, accuracy, access and correction, retention and accountability.

The method combined OpenAI's written representations, interviews and OPC laboratory testing between November 2023 and May 2024 (para. 19). OpenAI did not grant access to its systems, so its own acknowledgment that it processes personal information served as evidence (para. 50).

The offices rejected OpenAI's arguments that the British Columbia office lacked jurisdiction (paras. 34 to 40) and that it had no real and substantial connection to Canada before ChatGPT's release on November 30, 2022 (paras. 41, 42).

The Canadian consent rules and the business-transaction exception are set out on the page on Canada's rules for licensing data for AI training.

Where OpenAI Said Its Training Data Came From

OpenAI told the offices that it collects training data from four sources, each of which may contain personal information (para. 51).

  • Publicly accessible Internet sources, which OpenAI said are "the vast majority" of its training datasets, gathered from third parties such as Common Crawl or by its own crawler, GPTBot.
  • Data licensed from media outlets, a large stock image vendor and other specialized sources.
  • User interactions with ChatGPT.
  • Conversations written by human AI trainers.

OpenAI also stated that licensed datasets were less than 1 % of its pre-training datasets at the time of GPT-3.5 and GPT-4, a figure the offices could not verify (para. 159).

The total size was not confirmed beyond OpenAI's representation of "trillions of words" (para. 127). The web archive OpenAI named as a source is described on the page on Common Crawl, its contents and legal status.

OpenAI Investigation Timeline, 2023 to 2026

Eight dated steps take the matter from the complaint of April 2023 to the restatement of the outcome in the Commissioner's speeches of October 2026. The table gives each step with its source, as of 2026-11-17.

Date Step Source
2022-11-30 ChatGPT released; OpenAI later argued it had no Canadian connection before this date, which the offices rejected paras. 41, 42
2023-04-04 OPC announces its investigation on a complaint OPC announcement 2023-04-04
2023-05-25 OPC, CAI, OIPC-BC and OIPC-AB announce the joint investigation OPC announcement 2023-05-25; para. 15
2023-11 to 2024-05 OPC laboratory testing of ChatGPT (GPT-3.5 and GPT-4) para. 19
undated Preliminary Report of Investigation issued with recommendations para. 20
2025-09-02 OpenAI's response to the Preliminary Report report, footnote 133
2026-05-06 report of findings and OPC news release; OpenAI's Canadian blog post due at the same time report; para. 550
2026-10-08, 2026-10-15, 2026-10-21 the Privacy Commissioner of Canada restates the outcome in three speeches, without an implementation update OPC speeches

Findings on Scraped Web Data and Licensed Datasets

The offices found that OpenAI collected significant amounts of personal information to train GPT-3.5 and GPT-4, that the collection from publicly accessible websites and licensed datasets was overbroad, and that OpenAI had no valid consent for it.

On purpose (Issue 1), the offices found the collection not necessary and proportional, contravening subsection 5(3) of PIPEDA and the provincial equivalents (para. 105). The collection from user interactions was proportional, so that aspect was not well-founded (para. 106).

On consent (Issue 2A), OpenAI had no implied consent for scraped and licensed data (para. 119). The use was outside reasonable expectations (para. 153), and OpenAI had to offer a choice (para. 155).

The OPC, OIPC-AB and OIPC-BC found no valid consent for collection through GPTBot or Common Crawl, contravening section 6.1 and Principle 4.3 of Schedule 1 of PIPEDA (paras. 166, 167).

OpenAI did not claim the publicly available exception, and the offices noted that social media and professional profiles collected from websites are not "publicly available" under the regulations (paras. 160, 162).

The OPC still closed both issues as conditionally resolved, because a new filtering tool and other measures significantly reduce the risk for future models (paras. 115, 175, 176).

OpenAI states that, fine-tuned on a small subset of an open-source benchmark, the tool reached 98 to 99 % recall with a 3 to 6 % false positive rate (para. 111).

The offices held that data licensed from a third party is generally collected for a purpose unrelated to the one it was first collected for, so OpenAI had to ensure the people concerned had been given a choice before it obtained the data (para. 157).

OpenAI answered that its Data Partnerships do not seek sensitive or personal data, that partners had to warrant compliance with data protection laws, and that its contracts now add assurances on publication, notices and consents (para. 158). These are OpenAI's statements, not findings.

The offices took note of the share they could not verify and expect "reasonable efforts", such as written agreements with the third parties and monitoring, so that the data was obtained lawfully with appropriate consent or a valid exception (para. 159).

How a data holder documents that consent is covered on the page on customer consent before licensing data for AI.

OpenAI also argued that datasets from news outlets such as the Guardian Media Group or News Corp may count as publicly available information (para. 164). The offices maintained that such information would generally not qualify (para. 165). The report states no deal values.

Each Office's Conclusion, Issue by Issue

The OPC closed every issue except one as well-founded and conditionally resolved under PIPEDA, while British Columbia and Alberta left consent unresolved and Quebec's commission left consent and retention unresolved. The table gives the report's wording as of 2026-05-06.

Issue OPC (PIPEDA) OIPC-BC and OIPC-AB CAI (Quebec) Paragraphs
1 Appropriate purpose (scraped and licensed data) well-founded, conditionally resolved contravention found jointly with the OPC (para. 105) well-founded, conditionally resolved 105, 115, 557
1 Appropriate purpose (user interactions) not well-founded joint finding not separately stated 106
2 Consent (2A scraped and licensed data; 2B user chats; 2C disclosure) well-founded, conditionally resolved 2A and 2C well-founded and unresolved well-founded and unresolved 176, 185, 311, 387, 391, 557
3 Openness well-founded, conditionally resolved no separate conclusion stated no finding 414, 557
4 Accuracy well-founded, conditionally resolved no separate conclusion stated no finding 455, 557
5 Access, correction, deletion well-founded, conditionally resolved no separate conclusion stated well-founded, conditionally resolved 474, 483, 497, 509, 557
6 Retention well-founded, conditionally resolved no separate conclusion stated well-founded and unresolved 529, 557
7 Accountability well-founded, conditionally resolved no separate conclusion stated well-founded, conditionally resolved 546, 557

The British Columbia and Alberta laws do not allow implicit or deemed consent to rest on shifting expectations and risk mitigation (para. 183).

Their offices find that OpenAI's models "are based on scraped data for which OpenAI has not obtained, and cannot obtain, consent" under PIPA-BC and PIPA-AB, yet they joined the recommendations and the monitoring (paras. 555, 556). The CAI made no finding on openness and accuracy (para. 557).

Recommendations and OpenAI's Commitments

OpenAI disagreed with the findings and said it complied with the Acts in most respects, but it committed to a set of measures that the OPC accepted as addressing the intent of the recommendations (paras. 549, 552, 553).

The recommendations of the Preliminary Report fall in two windows (para. 548). Within 3 months of the final report, OpenAI was to share a plan to limit training data to what is necessary, including ceasing collection from sources with significant personal information "including but not limited to social media and discussion forums", and a consent plan covering licensors' consent. Within 6 months it was to provide transparency, accuracy, export, retention and accountability measures.

The Preliminary Report also recommended that OpenAI untrain its models of personal information collected without consent, or show why that is impossible, and never use such information to train new models. Contract terms on deleting or unlearning data are covered on the page on deletion and machine unlearning clauses.

The commitments differ from the recommended measures, and the OPC assessed them against the intent of each recommendation (para. 552).

The table compares them as of 2026-05-06. No commitment matches the untraining or the social-media cessation recommendation as worded, and OpenAI told the offices it retired GPT-3.5 and GPT-4 (para. 551). The offices have not published implementation status as of 2026-11-17.

Topic Recommended (Preliminary Report) OpenAI committed to Window Paragraphs
Limiting collection plan to limit training data to what is necessary, including ceasing collection from sources with significant personal information such as social media and discussion forums none as worded; points to its filtering tool and retired models 3 months 548, 551
Consent and licensed sources consent plan; licensors' consent must permit OpenAI's own use, otherwise direct consent none as worded 3 months 548
Data already collected untrain models or show why impossible; never use for new models none as worded 3 months 548
Transparency plain-language information on categories and sources of training data Canadian blog post at the time of the report; expanded "How ChatGPT and our foundation models are developed" article; notice before the first prompt in signed-out ChatGPT recommended 6 months; committed at publication and 3 months 548, 550
Access accessible export format and a way to challenge it more accessible exports and information on how to challenge them 6 months 548, 550
Retention formal retention and deletion policy including training datasets report on controls for datasets kept as historical benchmarks; continued re-evaluation 6 months 548, 550
Other accuracy and accountability measures test refusal of names and birth dates of minor family members of public figures; quarterly reports until all commitments are met 6 months; quarterly 548, 550

Quebec's Commission d'accès à l'information added recommendations on consent and retention under Quebec's Private Sector Act (para. 558).

  • Inform people at initial collection that published data may be scraped and used for AI training, including via third-party scrapers.
  • Show a notice before first use of the free web version.
  • Do not use chats for training by default unless users activate it (section 9.1).
  • Inform users about retention, or destroy or anonymize the information once the purposes are fulfilled.

The CAI will monitor and may consider further verification, recommendations or orders (para. 559). The Quebec rules are set out on the page on Quebec Law 25 and AI training data.

What the Findings Mean for Companies Whose Data Contains Personal Information

The findings concern OpenAI as the collector of training data, but three passages speak directly to a company that licenses datasets containing personal information to an AI developer: the requirement of a choice before a third party's data is collected for an unrelated purpose (para. 157), the expectation of written agreements and monitoring (para. 159), and the recommendation that licensors obtain consent permitting the developer's own use (para. 548).

The report holds scraped and licensed data to the same consent test. The differences between the two routes are set out on the page on licensed vs scraped training data.

The OPC accepted implied consent for future training where the risk is "significantly and meaningfully mitigated" (para. 175), while British Columbia and Alberta did not accept it for the same practice (para. 184). The report binds no seller and sets no rule for a contract. This page is information, not legal advice.

PIPEDA Findings #2026-002 is the tracked regulator report in which four offices examined consent for training data across scraped, licensed and user-generated sources, while other regulators and courts have reached different results. Every tracked proceeding with its status is on the case tracker of AI copyright lawsuits over training data.

The table compares three related actions on personal data and AI, ordered by date, as of 2026-11-17 for this row set with the check date of each status in its cell.

Proceeding Body and number Question Status Source
ANPD v Meta ANPD, Voto n. 11/2024/DIR-MW/CD Meta's use of user data for AI training preventive measure 2024-07-02; later lifting not checked (as of 2026-11-09) gov.br/anpd
Italian Garante v. OpenAI Garante provvedimento n. 755 of 2024-11-02; Tribunale di Roma n. 4153/2026 ChatGPT processing, including training on personal data opposition upheld, published 2026-03-18 (checked 2026-11-10) garanteprivacy.it
Meta AI training on EU user data Oberlandesgericht Köln, 15 UKl 2/25 interim injunction against Meta's training on EU user data interim application rejected 2025-05-23; main proceedings not found (as of 2026-11-09) nrwe.justiz.nrw.de

OpenAI's own licensing program and its litigation list belong on the page on OpenAI's licensing program and litigation list, not here.

More About PIPEDA Findings #2026-002

Decided Versus Still Open After the Findings

The findings are final reports, not court rulings, and what remains open is implementation.

  • Decided: contraventions for GPT-3.5 and GPT-4 (paras. 105, 167); OPC well-founded and conditionally resolved (para. 553); British Columbia and Alberta consent unresolved (para. 556); Quebec consent and retention unresolved (para. 557).
  • Open: whether the 3- and 6-month commitments were met (no publication found as of 2026-11-17); the outcome of the quarterly reports; any further CAI action (para. 559); whether OpenAI challenged the findings in court (no record found).

Frequently Asked Questions About the Canadian OpenAI Investigation

Did Canada fine OpenAI?

No, the report of May 6, 2026 contains findings, recommendations and OpenAI's commitments, but no penalty. The OPC continues to work with OpenAI toward final resolution (para. 554).

What does "well-founded and conditionally resolved" mean?

The OPC found the complaint justified and accepted OpenAI's commitments as resolving it, subject to their implementation. The OPC monitors the implementation (paras. 553, 554).

Did all four offices reach the same result?

No, British Columbia and Alberta found consent well-founded and unresolved, and Quebec found consent and retention unresolved. The OPC found every issue conditionally resolved except the user-interaction purpose aspect (paras. 556, 557).

Does the finding apply to current ChatGPT models?

The contraventions concern GPT-3.5 and GPT-4, which OpenAI says it has retired. The offices note that the findings may still be relevant to similar models (paras. 16, 551).

Where can I read PIPEDA Findings #2026-002?

On priv.gc.ca, dated May 6, 2026, with the OPC news release of the same day.