In March 2013, the European Commission fined Microsoft €561 million for failing to provide a browser-choice screen required by binding commitments. The Commission found negligence. Microsoft’s explanation involved a technical error, failures of internal communication "due to human error" and inadequate monitoring. An obligation accepted at the top had failed in its implementation. Commission decision.
This is a useful starting point for considering whether artificial intelligence (AI) companies should substantially regulate themselves. Even an organisation with considerable technical resources can fail to deliver a clear undertaking. Good intentions, where they exist, still need an institutional means of surviving contact with the business.
The more difficult question is who should decide whether the undertaking was sufficient in the first place.
By self-regulation, I mean leaving companies substantial discretion to determine the relevant safeguards, assess their own compliance and control what outsiders can inspect. In the above case, Microsoft knew in advance that a breach of their earlier technical commitments towards the European Commission could result in a fine up to 10% of their total annual turnover. Nonetheless, they still failed in installing the necessary internal safeguards, and they failed to detect the breach for a "significant" period of 14 months - until it was reported to them by the European Commission (system error, anyone?). Internal standards operating within an enforceable legal framework are a necessary proposition. The issue is whether the people selling a technology should also have the final word on the risks others must accept from it.
What the enforcement record can tell us
The past fifteen years provide substantial material for this discussion, although a collection of large fines can easily mislead. Competition proceedings, privacy settlements and consumer refunds address different legal obligations. Their amounts cannot sensibly be added together and presented as a single measure of corporate wrongdoing.
My research was by no means comprehensive, but I looked into 30 separate cases. They arose under existing regulation. They are not a controlled experiment proving that self-regulation always fails. An enforcement record selects for problems, says little about misconduct successfully prevented, and reveals weaknesses in public supervision as well as corporate governance.
Nevertheless, the detail matters, so let me share with you a few examples.
Facebook’s 2019 settlement with the US Federal Trade Commission (FTC) resolved allegations that it had violated an earlier privacy order. It included a $5 billion civil penalty, without a general admission of the allegations. Its governance provisions required an independent board privacy committee and external assessment. The remedy strengthened a regime that had already required a privacy programme and independent assessments. It therefore raises questions about the quality and enforcement of oversight, as well as internal policy. Settlement order.
Microsoft and Facebook illustrate different problems: an established failure to implement binding commitments, and allegations resolved through stronger, enforceable privacy governance. Neither requires us to assume that every employee was indifferent to compliance. Organisations can contain conscientious people while distributing incentives and responsibility in ways that leave important obligations unattended.
For AI, this suggests a practical test: when commercial and safety judgements conflict, who has the information and authority to require a different decision? A safety committee’s existence tells us very little until we understand its powers.
The user’s choice is the aim of the product's service - or is it the product?
Consumer protection brings the issue closer to everyday experience. A company determines which option appears first in a user experience (UX), what happens by default and how much UX effort it takes the user to leave or cancel a service or a product. Those design choices can materially affect what customers choose.
Amazon’s September 2025 Prime settlement required a $1 billion civil penalty and a separate $1.5 billion consumer redress fund, alongside changes to Amazon Prime enrolment and cancellation. It followed a partial court ruling against Amazon; the company otherwise neither admitted nor denied the allegations, subject to the order’s exceptions. The distinction between a penalty and money intended for consumers, including the respective sums, is material. But so is the order’s insistence on obtaining consumers’ express informed consent before charging them and a simple cancellation option. A seamless A-to-Z experience shouldn't feel like crossing the Amazon River; but why did it require the regulator's intervention to make it flow better? Court order.
Epic Games’ separate Fortnite settlements, finalised in 2023, addressed allegations concerning children’s privacy, default communications settings affecting children and teenagers, and unwanted purchases. The outcomes included a $275 million privacy penalty and $245 million for consumer redress in the billing matter. Epic pointed to changes it had already made and argued that the application of older laws had evolved. Their epic statement was: "Statutes written decades ago don't specify how gaming ecosystems should operate. The laws have not changed, but their application has evolved and long-standing industry practices are no longer enough." That explanation deserves consideration; it also demonstrates how practices familiar within an industry can become difficult to defend outside it. Privacy order, billing order, Epic’s response.
For AI, the immediate issue is how a user’s apparent choice is produced, or is itself processed as the product. If an AI assistant recommends a purchase, selects a service or acts on a user’s behalf, its defaults and commercial relationships deserve scrutiny. A reassuring conversation with a friendly AI agent cannot, by itself, establish informed consent. We should examine what the system actually enables the user to understand and control, as opposed to having human choices controlled by AI agents.
Market power does not disappear inside an artificially intelligent assistant
In September 2024, the Court of Justice upheld the approximately €2.4 billion Google Shopping fine. The case concerned Google’s treatment of its own comparison-shopping service relative to rivals. The abusive exploitation of a dominant position is deemed to hinder competition and is likely to cause harm to consumers. The Court did not, however, establish that every instance of favouring one’s own service is unlawful. Court summary.
The question for AI is whether an assistant could become an important route through which customers reach other businesses, while favouring the provider’s own services. This is a scenario requiring examination in light of the future deployment of AI capabilities across existing digital platforms, not a finding against any particular AI product. Technical sophistication would not remove the underlying conflict of interest.
Competition can discipline providers, but its effectiveness depends partly on whether users can understand the difference between competing offers and move between them. A user who cannot detect a biased AI recommendation cannot readily reward the provider making a better one. Reputation has similar limits when the conduct that matters is difficult to observe.
This is where the case for relying predominantly on voluntary restraint becomes weakest. The party best placed to identify a problem may also benefit from leaving it unresolved.
The serious case for self-regulation
The strongest argument for self-regulation rests on knowledge and speed. Providers can identify technical problems and change their systems without waiting for a legislative process. Customers can demand safeguards, while reputational damage and legal liability give companies reasons to respond. The argument is strongest where informed buyers can compare performance and exercise a credible choice.
Voluntary governance can also have substance. The US National Institute of Standards and Technology’s 2023 AI Risk Management Framework provides a structured approach to identifying, assessing and managing risks, while deliberately allowing flexible adoption. Developed by a public body with private and public input, it illustrates a practical alternative to prescribing every technical safeguard in legislation. However, the existence or adoption of this framework does not establish its effectiveness. NIST framework.
Legislators have, for a long time, found it necessary yet difficult to balance the benefits of prescribing solid regulations against the corresponding risk of detailed regulations. Arguably, compliance costs may protect established firms from smaller challengers. If authorities depend heavily on incumbents for technical advice, rules intended to constrain market power could strengthen it. Industry expertise deserves a substantial role, but so does scepticism about whose interests a proposed rule would serve.
Furthermore, the enforcement of regulations can sometimes be as difficult as enacting them. In September 2024, the European General Court annulled the European Commission’s Google AdSense decision of 20 March 2019 in its entirety, finding that the Commission had failed to establish the infringements at issue. The European Commission appealed and the case remains pending. Judgment, appeal notice.
Even the successful Google Shopping case (imposing a fine of approximately €2.4 billion) took seven years from the European Commission’s 2017 decision to the final appeal ruling. That interval does not prove the remedy was ineffective, and does not establish seven years without corrective action. In fact, the Commission's decision of 27 June 2017 required a remedy within 90 days of notification, and Google announced changes on 27 September 2017. However, it may challenge the assumption that public enforcement will necessarily provide a timely and final correction. Court chronology.
Regulation must be flexible enough to allow regulators to respond when the facts change, which is not always easy to achieve. However, the case of Google's commitments over its Privacy Sandbox proved that regulators can adapt to changing circumstances. The UK's Competition and Markets Authority (CMA) had previously accepted binding commitments over Google’s Privacy Sandbox proposals in 2022, without making an infringement finding. In October 2025, following consultation, the CMA released those commitments after concluding that it had reasonable grounds to believe its original competition concerns no longer arose. Google had abandoned plans to remove third-party cookies from Chrome and the subsequent proposal for a standalone choice prompt; the CMA also took account of Google's decision to retire several Privacy Sandbox technologies. The CMA credited the commitments with ensuring stakeholder engagement to guard against Google favouring its own advertising services. An obligation can serve a purpose and still cease to be justified when the conduct it addresses changes. Commitments decision, release decision.
The difficulty for self-regulation is that flexibility does not resolve the underlying conflict over whose interests count. A buyer of a technology product may accept its risks even though some of the consequences will be borne by employees, competitors or others who had no say in its adoption. An employer might buy an AI recruitment tool because it reduces costs. The employer accepts the commercial risk of a poor purchase, but applicants may bear a different risk: being unfairly screened out. They did not choose the tool, negotiate its safeguards or agree to that risk. Both a supplier and its customer can be satisfied with a product while its use harms someone outside their agreement. The difficulty for public regulation is that legal authority does not supply technical competence. A defensible system must address both weaknesses.
What should follow for AI
I recognise that the general analogy made above between non-AI and AI technology products has limits. A subscription interface, a search market and a general-purpose AI model involve different technologies and potential harms. These cases cannot establish the scale of future AI risks, or settle whether a particular model should require permission before release. They simply illuminate familiar problems of information, incentives and control, albeit at the risk of introducing smart solutions in the wrong frameworks.
Nevertheless, I would draw three practical conclusions.
First, scrutiny should reflect the likelihood and severity of harm. An application rephrasing a routine, non-sensitive email needs different scrutiny from one used to make decisions affecting employment or access to essential services. Where serious harm is plausible, competent independent scrutiny needs access to the evidence required to test the provider’s claims, with appropriate protection for confidential information. If the provider controls the scope of every test and which failures are disclosed, the assessor’s nominal independence offers limited reassurance.
Second, responsibility must follow actual control. The company developing a model, the business adapting it and the organisation using it may each make consequential decisions. Duties should identify those decisions and their owners. Otherwise, responsibility risks disappearing into contracts while the affected person is sent from one company to another.
Third, success should be measured through implementation and outcomes. A fine’s size cannot tell us whether the conduct stopped, whether an injured person obtained redress or whether a market became more open. Equally, describing penalties as merely a cost of doing business requires evidence about incentives and behaviour. It should not become a substitute for doing that work.
Industry standards, internal testing and professional judgement remain essential. They become more credible when an outsider can challenge the assumptions, establish whether obligations have been met and require corrective action. Public authorities, in turn, need sufficient expertise to make that challenge useful and sufficient discipline to justify its cost.
My conclusion is that technical expertise gives AI companies an indispensable role in designing safeguards. It does not confer authority to decide, alone, which risks everyone else must bear. The appropriate intensity of supervision will remain contested, and should change with the evidence. Its independence should be a starting condition.