How Big Tech Learned to Make Governments Afraid of Regulating AI
If you wanted to design a system for maximizing the amount of time that a powerful new technology could operate without meaningful oversight, you could do worse than the current approach to AI regulation. Legislation that takes years to negotiate gets amended into toothlessness before passage. Regulatory agencies tasked with oversight lack both the technical expertise to understand what they are overseeing and the jurisdictional authority to act on what they discover. International coordination on standards collapses into lowest-common-denominator agreements that commit no one to anything meaningful. And over all of it, the companies being regulated sit at the tables where the rules are written, shaping the language of their own constraints.
This is not paranoia. It is a documented pattern, refined over three decades of technology industry regulatory engagement, and it is being applied to AI with particular effectiveness.
To understand the mechanics, start with what might be called the complexity defense. Every consequential technology generates a class of expertise specific to itself, and regulatory agencies depend on the industry they regulate to supply that expertise. This is not corruption, at least not initially — it is an unavoidable feature of technical regulation. The Food and Drug Administration must rely on pharmaceutical company scientists to understand the drugs being submitted for approval. Financial regulators must rely on bank quants to understand the products being reviewed. And AI regulators must rely on AI company researchers to understand the systems being evaluated.
This dependence is manageable when the regulated industry has a genuine interest in safety — when, for example, drug companies would face massive liability for unsafe products regardless of regulatory failure. It becomes problematic when the industry’s interest in regulatory delay outweighs its interest in safe products. In the AI case, the asymmetry is stark: the costs of AI harms are largely externalized onto users and society, while the benefits of rapid, unregulated deployment accrue to shareholders. This creates an incentive structure in which industry participants genuinely believe the things they tell regulators — “we need more time to develop best practices,” “mandatory standards will freeze the technology at an early stage,” “innovation should lead regulation” — while those things happen to serve the industry’s immediate financial interest.
The framing of “innovation versus safety” is the second pillar of the regulatory playbook, and it is remarkably durable despite being analytically indefensible. The claim is that regulating AI will slow innovation, and slowing innovation has costs — delayed medical breakthroughs, foregone productivity gains, competitive disadvantage relative to less regulated jurisdictions, primarily China. The costs of regulation are always vivid and near-term: jobs at risk, products delayed, investment deterred. The benefits of regulation are diffuse and long-term: harms prevented that might not have happened, risks mitigated that might not have materialized.
This asymmetry in the legibility of costs and benefits is not unique to AI — it applies to almost every safety regulation ever proposed. The pharmaceutical industry used exactly the same arguments against the 1962 Kefauver-Harris amendments, which required drugs to be proved effective before approval, and which were passed only because the thalidomide disaster made the costs of under-regulation suddenly visible and impossible to dismiss. The financial industry used the same arguments against Dodd-Frank regulation after the 2008 crisis, and largely succeeded in weakening it during implementation.
The innovation-versus-safety framing is also empirically questionable. The countries with the strictest automotive safety regulations do not have the least innovative car industries. The countries with the strictest pharmaceutical regulations produce a disproportionate share of pharmaceutical innovation. Safety regulation has historically accelerated innovation in important ways by creating clear targets for products to meet, by generating consumer confidence that increases market size, and by eliminating competition from low-quality products that would otherwise crowd out investment in high-quality ones.
The EU AI Act is the most instructive recent case study in how the regulatory playbook operates. Europe began serious work on AI regulation earlier than any other major jurisdiction, publishing an initial draft in April 2021 that was substantive, technically informed, and — crucially — contained provisions that would have meaningfully constrained the behavior of frontier AI developers. By the time the Act was finalized and passed in 2024, the provisions that mattered most had been substantially weakened. The high-risk categorization that would have triggered serious oversight requirements was narrowed. The obligations on general-purpose AI models — the large language models at the heart of the AI industry — were reduced from strict requirements to transparency and compliance obligations that were far easier to satisfy in form without much changing in substance.
This did not happen because European regulators are incompetent or captured. It happened because the lobbying apparatus deployed by American tech companies in Brussels is one of the most sophisticated in the world, because member state governments with domestic tech industry interests intervened to protect those interests, and because the sheer pace of AI development meant that the regulatory categories being defined in 2021 no longer described the technology being deployed in 2024. The Act that passed was better than nothing, but it was substantially less than what its initial proponents had hoped for.
The revolving door between industry and government accelerates this dynamic in ways that are difficult to address directly. The technical experts who can staff AI regulatory positions are the same people who can earn ten times as much working for AI companies. The civil servants who develop deep expertise in AI policy are the people most attractive to industry as lobbyists and public affairs executives. This creates a natural flow of talent from regulatory agencies to regulated companies that brings with it informal relationships, shared assumptions, and a tendency to see regulatory problems through an industry-friendly lens.
There is a historical comparison that does not get made often enough. The tobacco industry’s response to evidence of health harms in the 1950s and 1960s was not to deny that evidence immediately existed, but to systematically manufacture doubt about it. The strategy, documented in detail in internal memos that only became public through litigation, was to fund alternative research, question methodology, emphasize uncertainty, and call for more study before action. The goal was not to permanently defeat the science — that was impossible — but to delay regulatory action long enough to extract maximum value from an unregulated market. The industry achieved roughly thirty additional years of largely unregulated operation after the initial scientific evidence became clear.
The AI industry’s situation is different in important ways — the harms are less directly lethal, the technology has more genuine benefits, and the industry’s relationship with evidence is more complex. But the structural incentive is the same: delay is profitable. Each additional year of operation without meaningful oversight is a year of revenue, market consolidation, and infrastructure build-out that will be very difficult to undo once regulation arrives.
The way out of this dynamic is not technologically simple, but it is politically intelligible. Regulatory agencies need stable, well-funded technical capacity that is not dependent on industry for expertise — which means paying civil servant AI researchers competitive salaries, or something closer to competitive salaries than current government pay scales permit. The innovation-versus-safety framing needs to be challenged every time it appears, with the historical evidence that safety regulation and innovation are not in structural conflict. And the legislative timelines need to be dramatically compressed, because in a technology that develops on an eighteen-month product cycle, a five-year regulatory process is effectively no regulation at all.
None of this is easy. But it is worth being clear about what the alternative is: continuing to allow the companies developing the most powerful technology in human history to be the primary authors of the rules governing how they develop it.