AI Safety Lessons from Internet & Pharma Regulation
Internet and pharmaceutical regulation offer concrete, hard-won lessons for AI safety and alignment: both fields show that top-down rules alone fail without technical standards, that liability drives real safety improvements, and that pre-market review works best for high-risk products but stifles innovation when applied too broadly. This article explains what AI governance can borrow from the FDA’s drug approval process and from two decades of internet platform regulation, and where those analogies break down.
Why These Two Domains Matter for AI
AI safety and alignment aim to ensure advanced AI systems do what humans intend and avoid causing harm. Policymakers and researchers often look for regulatory models to borrow from, and two of the most frequently cited are the pharmaceutical industry and the internet. These are useful because they sit at opposite ends of the regulatory spectrum: drugs are heavily regulated before they reach the market, while internet platforms were largely unregulated for years and are now being governed through piecemeal laws. Neither is a perfect template for AI, but each offers evidence about what works and what fails.
What Pharmaceutical Regulation Gets Right
The FDA’s drug approval process is built on a simple idea: you must prove a product is safe and effective before people can use it. That pre-market review is the single most important lesson for AI. It forces developers to generate evidence, not just claims. For AI, this suggests that high-risk uses—like medical diagnosis, autonomous driving, or hiring decisions—should require demonstrated safety and alignment before deployment, not after.
Another lesson is the use of standardized testing. Drug trials use randomized controlled trials with clear endpoints. AI alignment would benefit from similarly standardized benchmarks and evaluation protocols, so that a claim like “our model is aligned” means the same thing across different companies and systems. The FDA also requires post-market surveillance: even after approval, drugs are monitored for rare or long-term side effects. AI systems should have analogous monitoring for drift, emergent behavior, or misuse.
What Internet Regulation Teaches Us
The internet grew with minimal regulation, and the consequences are instructive. Section 230 of the US Communications Decency Act gave platforms immunity for user content, which allowed services like social media to scale but also created incentives to ignore harmful content. The lesson for AI is that liability shields can reduce safety investment. If an AI developer is not responsible for harms caused by their system, they have less reason to spend on alignment research.
Europe’s GDPR introduced a different approach: ex-post enforcement with heavy fines. It forced companies to care about data protection because violations were costly. For AI, a similar regime—where developers face meaningful penalties for demonstrable harms—could create market pressure for better alignment. However, GDPR also showed that compliance can become bureaucratic and focused on paperwork rather than actual safety, a risk for AI regulation that relies too heavily on documentation.
How to Choose Between Pre-Market and Post-Market
The key question is when to intervene. Pre-market review (like drugs) is appropriate when the risk is high and the technology is well-understood enough to define clear tests. For AI, that applies to narrow, high-stakes uses: a model that controls a power grid or makes loan decisions. Post-market regulation (like the internet) is better for broad, rapidly evolving technologies where we cannot predict all harms in advance. Most AI systems fall into this category, so a hybrid is likely: pre-market certification for a small set of high-risk applications, and post-market monitoring and liability for everything else.
A practical middle ground is a “risk-based” approach, similar to the EU’s proposed AI Act. It classifies AI into risk tiers and applies stricter rules to higher tiers. This avoids the cost of full pre-market review for a chatbot while still requiring it for a medical device. The challenge is that risk classification itself can be gamed or become outdated as AI capabilities grow.
Common Problems with Regulatory Analogies
One common mistake is assuming AI is a single product like a drug. AI is more like a general-purpose technology—like electricity or the internet itself—that is used in countless ways. Regulating all AI as if it were a drug would be impossibly slow and would kill beneficial uses. Conversely, treating AI like the internet and leaving it unregulated repeats the mistake of letting harmful systems scale without accountability.
Another problem is the “alignment tax”—the cost in performance or speed that comes from making a model safer. Drug regulation also has costs, but they are accepted because the alternative is unsafe medicines. For AI, the cost of alignment is often invisible until a system is deployed, and developers may resist it. Regulatory pressure can make that tax acceptable, but only if the rules are clear and consistent.
Costs and Trade-offs
Regulation is not free. The FDA’s drug approval process costs hundreds of millions of dollars and takes years, which is why small companies often struggle. For AI, pre-market review could similarly favor large incumbents who can afford the compliance burden. Internet regulation, on the other hand, has been cheaper for companies but has led to a patchwork of laws that are hard to comply with and enforce. The ideal AI regime should minimize both the cost of compliance and the cost of harm, which argues for targeted, not blanket, regulation.
Alternatives to Traditional Regulation
There are non-regulatory tools that can complement or even substitute for government rules. Technical standards, like those from the IEEE or ISO, can create common benchmarks without legal mandates. Industry self-governance, such as voluntary commitments to pause dangerous AI research, can work but is fragile because it depends on good faith. Insurance is another lever: if insurers require safety audits before covering AI products, they can drive alignment without a new agency. Finally, procurement rules—where governments only buy AI that meets safety standards—can shift the market, since the public sector is a large buyer.
Frequently Asked Questions
Is AI regulation like drug regulation actually feasible?
Partially. For narrow, high-risk uses, pre-market approval is feasible and already happening in some sectors (e.g., medical AI devices must get FDA clearance). But for general-purpose AI, it is not feasible to test every possible use in advance, so a drug-style model would be too slow and too limited.
Why not just apply internet-era rules to AI?
Because the internet’s laissez-faire approach led to widespread harms like misinformation and privacy violations that are now hard to reverse. AI can cause physical harm (e.g., in autonomous vehicles) or large-scale economic harm (e.g., automated fraud), which argues for more proactive oversight than what the internet received.
What is the “alignment” problem exactly?
Alignment means ensuring an AI system’s goals and behaviors match human intentions. It is hard because we cannot specify every desired behavior in advance, and a system might find unintended ways to satisfy a goal. It differs from safety, which is about preventing accidents, though the two overlap.
Can AI companies self-regulate effectively?
History suggests self-regulation works only when there is a credible threat of government action or market punishment. Tech companies have made voluntary safety commitments, but without enforcement, there is little incentive to follow through, especially when competitors do not.
What role do international agreements play?
AI is global, so a single country’s rules can be bypassed by moving development elsewhere. International coordination, like the OECD’s AI principles, is useful but non-binding. The most effective approach may be “Brussels effect,” where the EU’s strict rules become de facto global standards because companies comply to access its market.
How can we measure if AI regulation is working?
Use the same metrics as drug regulation: reduced incidence of known harms, successful recalls of unsafe systems, and evidence that safety testing is actually catching problems. For alignment specifically, we need public benchmarks that track how often AI systems behave in ways humans would not approve of, which is still an open research area.
Related
Method: the standard formula for this calculation, rounded as each result states. See how the calculators work or report an error.