After Bill C-11 and Bill C-27, Bill C-36 is Ottawa's third attempt in six years to replace or amend the Personal Information Protection and Electronic Documents Act (PIPEDA). Introduced on June 15, 2026, the bill would repeal PIPEDA's privacy provisions and enact the Protecting Privacy and Consumer Data Act (PPCDA). It was tabled eleven days after the federal AI strategy promised modern privacy legislation as a foundation for public trust in artificial intelligence. The bill has yet to be debated at second reading in the House of Commons, so its text remains open to amendment. Unlike its stalled predecessor, Bill C-27, which sought to pair privacy reform with a standalone Artificial Intelligence and Data Act (AIDA), Bill C-36 abandons a separate AI statute. Ottawa is instead betting that a modernized, comprehensive privacy regime can do the heavy lifting of building public trust in AI.
The PPCDA is comprehensive: it establishes a streamlined enforcement framework and introduces substantial administrative monetary penalties. Two of its provisions, however, appear destined to face serious difficulty in today's AI industry: those governing the disposal of personal information and those governing automated decision-making. Applied to modern generative AI, neural networks and algorithmic decision pipelines, both provisions reveal a fundamental structural flaw. Bill C-36 appears to have been drafted with a traditional database mindset, one in which personal data sits neatly in indexed rows and columns, waiting to be retrieved, modified or deleted. Rather than pioneering a forward-looking regulatory framework, Ottawa risks walking directly into the technical and conceptual traps that the European Union has spent the past several years attempting to untangle under the General Data Protection Regulation (GDPR). If Canada is to avoid enacting provisions that are obsolete on the day they receive royal assent, should the bill reach that stage at all, Parliament must give serious consideration to whether these provisions can work in practice and must examine the hard lessons learned across the Atlantic.
I.Bill C-36 Gives Canadians a Right to Delete. AI Models Are Not Built to Forget.
Section 54 of the bill allows an individual to require an organization to dispose of their personal information in three circumstances: where the information was collected, used or disclosed in contravention of the Act; where the individual has withdrawn consent; or where the information is no longer necessary for the product or service the individual requested. Disposal under the bill means the permanent and irreversible deletion or anonymization of the information.
In a conventional cloud database, where customer data is stored in rows, complying with a deletion request is straightforward. Machine learning is different. Training data is not “stored” inside a model like a file in a folder, ready to be retrieved. Instead, the correlations, syntactic patterns and factual fragments within that data are mathematically diffused across billions of numerical parameters, or weights, and the influence of any one person's data is spread across all of them. Once an individual's personal data has been baked into a model, it cannot simply be plucked out, any more than a single ingredient can be removed from a baked cake. (See e.g., ScienceDirect) The problem is compounded by scale: large organizations may receive enormous volumes of disposal requests every day, each of which would require them to adjust their systems.
Europe has confronted the same problem, and its experience is instructive. The GDPR's right to erasure in Article 17 is equally silent on trained models, and complaints have accumulated without resolution. In April 2024, the privacy advocacy group noyb filed a complaint with Austria's data protection authority after ChatGPT repeatedly gave a public figure’s wrong birth date, and OpenAI refused the person’s requests for rectification and erasure, on the basis that correction was not possible and that it could only block prompts containing his name, which would suppress all information about him and not just the date of birth. Although the complaint centred on inaccurate information, OpenAI's response exposes the core difficulty: the only way to stop the model repeating one fact about a person was to stop it saying anything about him at all. In a similar situation, in March 2025, noyb filed a second complaint in Norway after ChatGPT described a man as a convicted child murderer. Noyb argued that the false information could still remain in the model and that its erasure cannot be guaranteed unless the entire model is retrained. Noyb’s point of argument was that if an AI architecture fundamentally makes compliance with GDPR erasure and accuracy rights impossible, the model’s training itself constitutes an unlawful processing activity. This case brought the structural impossibility of “machine unlearning” to the center of European enforcement.
Nor have European regulators and courts agreed on what the right requires once personal data is inside a model. These positions from different EU members were slow to emerge, and they remain in conflict to this day. At the strict end, in May 2024, the European Data Protection Board's (EDPB) ChatGPT taskforce stated that technical impossibility cannot excuse non-compliance, and the EDPB's Opinion 28/2024 holds that a model trained on personal data cannot be presumed to be anonymous. France's data protection authority, the CNIL, takes a more pragmatic line, as it expects developers to retrain where they still hold the training data, but permits requests to be batched and accepts robust output filters as an alternative, recognizing that retraining may be disproportionate, as where a public figure asks to be erased from a large language model.
Meta tried to head off the problem before training rather than after; when it resumed training on the public posts of adult EU users, it first gave users an opportunity to object. In May 2025, the Irish Data Protection Commission, Meta’s lead GDPR regulator, cleared the way for the training to proceed after Meta adopted safeguards the Commission had recommended, and required Meta to report on their effectiveness by October 2025. Germany's Higher Regional Court of Cologne relied in part on Meta’s safeguards when it declined to block the training in May 2025, which caused uproar among privacy advocates.
At the permissive end, the Hamburg data protection authority rejects the premise altogether: in its assessment, language models do not store personal data, so the right to erasure attaches to a system's inputs and outputs rather than to the model itself.
i.The Bill's Answers Fall Short
Rather than confronting these problems, Bill C-36 offers exits. Organizations are never required to dispose of de-identified information (s. 54(3)), and they may use such information without consent for their internal research, analysis and development (s. 21). The wording seems broad enough to cover model training. Yet by the bill's own definition, de-identified information still carries a risk of re-identification, and the bill treats it as personal information (s. 2(2)). The exemption therefore cannot rest on the idea that the data is no longer about anyone. An organization could strip names from its training data, train its model, and never face a disposal request, even though the people behind that data could still be identified. Rather than solving the problem of deleting personal information from AI models, the bill exempts it. The bill’s logic is that anonymized data can be kept because it no longer threatens anyone; de-identified data, by the bill's own admission, still does. If that risk remains, what protection is the bill offering the individuals whose data is kept?
The bill also allows an organization to refuse a disposal request under paragraph 54(2)(f) where “the disposal of the information would have an undue adverse effect on the organization that outweighs any potential adverse effect on the individual resulting from the retention of the information.” This pathway raises difficulties of its own. Every refusal on that ground must be reported to the Commission (s. 54(4)(b)). A company facing thousands of requests relating to a single model must therefore either retrain repeatedly or file thousands of refusal notices. The CNIL at least permits requests to be batched and the response deadline to be extended where volume requires it. Bill C-36 provides for neither.
Consider a large organization that has trained a large language model on personal data and then receives a high volume of disposal requests. Those requests are materially indistinguishable from one another: removing any one individual's data from the model would most likely require the same effort as removing any other's. Either every request falls within the paragraph 54(2)(f) exception, or none does. Only two outcomes are therefore realistic, and neither is desirable:
the organization is not required to delete any consumer data, because every request falls within the paragraph 54(2)(f) exception, which defeats the purpose of section 54 as a whole; or
the organization is compelled to devote enormous labour and resources to retraining its model every day.
Nor does the bill offer any solution for more extreme scenarios, in which a significant proportion of an organization's customers request the disposal of their data. Suppose a company has trained its AI model on data from one million subscribers. A report then reveals serious cybersecurity weaknesses at the company that could adversely affect consumers' data, and 90 per cent of subscribers ask the company to dispose of their personal information. What is the organization to do with the model it has already trained? Removing all of that data is almost certainly impossible without dismantling the AI system entirely. Parliament should resolve these problems before the bill becomes law.
II.Automated Decisions: A Human in the Loop, on Paper
The second major blind spot in Bill C-36 concerns its regulation of automated decision systems (ADS): algorithms increasingly relied upon to screen job applicants, assess mortgage applications, price insurance premiums and flag suspicious financial activity.
Under sections 62 and 63 of the PPCDA, organizations must publicly disclose their use of automated decision systems. Where an ADS makes a prediction, recommendation or decision that could have “a legal or similarly significant effect” on an individual, that individual is entitled to a plain-language explanation of how the system arrived at the outcome. Section 63(6) also creates a new right to make written representations to an employee of the organization who is able to review the decision.
Automated decision-making raises an obvious question: how much human involvement is needed to take a decision outside the rules? Under Article 22 of the GDPR and under Quebec's Law 25, the strict automated-decision protections apply almost exclusively to decisions based solely, or exclusively, on automated processing. That approach invites rubber-stamping: place a person at the end of the pipeline and argue that the rules no longer apply. The Court of Justice of the European Union’s Schufa case is a clear example of this. In 2021, a German resident was refused a loan after Schufa, a private credit agency, supplied the bank with a poor credit score, and the agency refused to explain how the score was calculated, citing trade secrets. In its December 2023 ruling, the Court of Justice of the European Union held that the score itself was an automated decision under Article 22 where a lender “draws strongly” on a probability value or score to establish, implement or terminate a contractual relationship with the data subject, stretching the provision to reach an outcome that a human formally decided.
To its credit, Bill C-36 attempts to close this notorious loophole in the GDPR. It defines an automated decision system as any technology that “assists or replaces the judgment of human decision-makers,” from rules-based systems to neural networks. Its obligations attach to predictions and recommendations as well as to decisions. A human sign-off at the end of the process does not take a system outside the Act.
Yet by adding the single word “assists,” Bill C-36 swings the regulatory pendulum to the opposite extreme without resolving the core governance problem:
The threshold dilemma. Because the definition captures any technology that “assists” human judgment through rules-based systems, regression analysis or machine learning, the bill does not establish where ordinary productivity software ends and an automated decision system begins. If a recruiter uses an AI tool to summarize a résumé, or a loan officer uses basic statistical software to calculate a debt-to-income ratio, has the technology “assisted” the decision? By failing to define the degree of algorithmic influence required to trigger section 63, the bill risks either burdening businesses with compliance obligations for trivial tools or diluting the rules altogether.
The rubber-stamp problem persists. Although keeping a human in the loop no longer exempts an organization from providing an explanation, Bill C-36 places complete faith in human intervention under section 63(6) as the ultimate consumer safeguard. Yet the text says nothing about what constitutes meaningful human oversight.
The weakness lies in what follows once a system is covered. An individual receives an explanation on request and an opportunity to make written representations to an employee able to review the output (ss. 63(4)–(6)). Nothing in the bill prohibits fully automated decisions, and nothing gives the individual a right to object or to insist on a human decision in the first place.
Nor does the bill specify what the review must involve. It does not require that the reviewer have the authority to change the outcome, understand how the system works, or reach a conclusion independently of it. It does not require a reconsideration or a reasoned response. A reviewer who glances at the same score and confirms it would appear to satisfy the provision. The rubber stamp has simply moved from the front of the process to the back.
Two further gaps compound the problem. First, the rights apply only where the output could have a legal or similarly significant effect on the individual, a threshold the bill does not define. Second, section 63 is absent from the list of provisions whose contravention can attract an administrative monetary penalty (s. 113(1)). An organization that ignores a request for an explanation risks a complaint and a compliance order, but no penalty for the contravention itself.
III.What Parliament Should Fix
None of this requires rewriting the bill. Four amendments at committee would close most of the gaps:
Clarify whether the disposal right extends to models trained on personal information, and what compliance requires. The CNIL's framework offers a ready template: retraining where the training data is still held, batching of requests, and robust output filters as a fallback.
Decide what happens when a model trained on de-identified data produces output that identifies a person: section 75 forbids organizations from using such data to identify individuals, but the bill does not say whether a model doing so counts, or what remedy the individual has.
Establish a minimum standard for human review, requiring a reviewer who has the authority to change the outcome, is not bound by the system's output, and must give reasons.
Add section 63 to the provisions whose contravention can attract an administrative monetary penalty.
The bill does empower Cabinet to make regulations respecting disposal (s. 139(1)(e)), so some of these questions could be settled later. Europe's experience, however, shows the cost of delay: years of conflicting regulatory positions while complaints go unanswered. Whether Canadians' right to deletion extends to the systems built from their data is a policy choice, not a technical detail. It belongs in the statute, and committee is the place to put it there.