Microsoft President Warns: AI Regulation Is Looser Than Milk — Who Fills the Safety Gap?

Microsoft President Brad Smith recently made a sharp comparison in public: artificial intelligence is subject to less regulation than milk. At first glance this sounds like rhetoric, but on closer thought it points to a concrete industrial reality — the dairy industry has mandatory inspection, traceability, and recall systems, while AI models' capability boundaries, risk assessments, and deployment conditions still lack legal constraints of comparable magnitude.

Milk Has Pasteurization — What Does AI Have?

Smith's core argument is not complicated. From pasture to shelf, milk undergoes microbial testing, temperature control, and expiration labeling, and if a problem arises it can be traced to a specific batch and forcibly pulled from shelves. By contrast, when a large model with hundreds of billions of parameters is released, the outside world typically only sees a technical report and limited evaluation data.

This asymmetry produces two consequences. First, risks are invisible: spoiled milk can be identified by the naked eye, but AI's biases, hallucinations, and abuse pathways are hidden in weights and training data. Second, responsibility is untraceable: when milk has a problem, the producer can be found, but when AI causes harm, responsibility is often blurred among developers, deployers, and users.

Smith's warning is not meant to stifle innovation. His company, Microsoft, is both one of OpenAI's largest investors and a driver of the Copilot product line. The fact that someone selling AI is calling for rules on AI is itself worth pondering — but it also shows that leading industry players are beginning to realize that scaled deployment without a safety baseline will ultimately backfire on the entire market.

Pasteurization is a low-temperature sterilization process invented by the 19th-century French scientist Louis Pasteur. By heating a liquid to 60–85°C and holding it there for a period, it kills pathogenic microorganisms without significantly damaging the original flavor and nutrients. It was first used for wine and beer, then extended to milk, greatly reducing milk's disease risk and becoming a foundational standard for the global dairy industry. The rollout of pasteurization went through a long process from resistance to mandate: early dairy farmers worried about rising costs, consumers were skeptical of milk whose taste had been changed by heating, and only after frequent public health incidents did countries gradually legislate it into mandatory inspection procedures. This history is often cited in regulatory discussions because it shows that for a safety standard to become common industry practice, it often requires the combined effect of technological maturity, public consensus, and institutional mandate. For AI, there is currently no widely accepted "baseline safety process" comparable to pasteurization; practices such as model evaluation, red teaming, and watermark traceability are not yet unified, and this is precisely the institutional gap Smith's analogy seeks to highlight.

Regulation Is Not a Brake — It's a Guardrail

The debate around AI regulation is often simplified into a binary opposition between "regulate" and "don't regulate." Smith's framework is closer to "how to regulate." His milk analogy is actually implying that mature regulation does not destroy an industry; rather, it builds trust infrastructure.

The history of the dairy industry can serve as a reference. Pasteurization also met resistance in its early days, but it eventually became the standard because it gave consumers the confidence to buy and allowed compliant companies to earn a premium. What the AI field currently lacks is precisely this kind of "dare to use" trust. When enterprise customers deploy AI customer service, medical assistance, or financial risk control, their biggest concern is often not effectiveness, but who is responsible if something goes wrong and where the compliance boundaries lie.

From this perspective, the absence of regulation is not freedom but uncertainty. Uncertainty drives up transaction costs, and ultimately the entire ecosystem pays the bill.

Who Opposes, Who Watches

On the same day Smith spoke out, this discussion on Hacker News had only one comment and was not particularly hot, but the topic itself is not marginal. AI regulation worldwide is in a phase of fragmented advancement: the EU's AI Act uses risk tiers, the United States leans toward industry self-regulation and executive orders, and China has issued administrative measures around generative AI services.

The disagreement centers on two points. First, speed: the innovation camp worries that premature legislation will lock in technological paths, while the safety camp believes the cost of a wait-and-see approach is rising exponentially. Second, level: should we regulate the model itself or the application scenario? Milk regulation regulates the product; if AI regulation only targets models, it may miss the amplifying effects of the deployment stage.

As a company spanning both the model layer and the application layer, Microsoft's position is naturally delicate. Smith chose to anchor on "milk," an everyday consumer product, which is highly effective for communication but also evades a question: milk's risks are relatively certain, while AI's risks are still evolving. Applying an old regulatory template to new technology may not be enough.

The EU AI Act was formally adopted in 2024 and is the world's first regional framework to systematically legislate on AI. It adopts a "risk tiering" approach, dividing AI applications into four levels: unacceptable risk, high risk, limited risk, and minimal risk. Scenarios with unacceptable risk (such as social credit scoring and manipulative AI) are prohibited; high-risk scenarios (such as hiring, credit, healthcare, and critical infrastructure) must meet compliance obligations including data governance, risk management, technical documentation, and human oversight; limited-risk scenarios mainly bear transparency obligations, for example chatbots must inform users that they are interacting with AI. The regulation targets both model providers and deployers, and sets additional provisions for general-purpose AI models (GPAI). This model of "defining risk by scenario and defining obligations by risk" is similar to the logic of setting inspection standards by product category in dairy regulation, but the dynamic nature of AI risk and its ability to spread across scenarios make tiered design face challenges such as blurred boundaries and lagging updates in practice.

Safety Is Not a Cost Item — It's Competitiveness

Setting aside the metaphor, the essence of Smith's warning is this: the AI industry is currently enjoying a regulatory honeymoon, and honeymoons do not last forever. Once a major safety incident occurs — whether social unrest triggered by deepfakes or systemic failure of automated systems — public sentiment could force rules far stricter than those of today.

Proactively establishing safety standards is more advantageous than passively accepting punitive regulation. This has precedents in aviation, pharmaceuticals, and finance. When leading industry players participate in rule-making, they can both reduce the compliance shock and turn safety capabilities into competitive barriers.

For developers and enterprise users, the signal is equally clear: before regulation lands, internal governance and risk assessment should not wait. Model cards, red teaming, usage policies, and human fallback mechanisms — actions that seem "unnecessary" today may become tomorrow's market entry threshold.

Smith's milk analogy may not be precise enough, but it successfully puts a question on the table: when AI capabilities have already penetrated critical decisions, are we still willing to accept a governance framework looser than dairy regulation? The answer will determine how far AI can go in the next decade.

Red teaming originated in the military and cybersecurity fields. It refers to organizing an independent adversarial team to simulate the behavior of real attackers or abusers and proactively seek out system vulnerabilities, biases, jailbreak paths, and harmful outputs. In AI development, red teaming is usually conducted before model release, covering dimensions such as prompt injection, data leakage, discriminatory generation, and dangerous advice. The test results are fed back to the development team to harden the model and update usage policies. Complementing it, a model card is a standardized document that records a model's training data sources, applicable scenarios, known limitations, evaluation performance, and ethical considerations, helping deployers judge whether the model suits their business. These two practices are currently mostly adopted voluntarily by companies and have not yet become legal requirements, but they constitute core tools of internal AI governance. The "safety baseline" Smith calls for relies, at the operational level, to a large extent on mechanisms like these moving from voluntary to mandatory, thereby forming auditable and traceable industry practices.

One-line Summary

Microsoft President Smith uses the milk regulation analogy to call for AI to establish a traceable, accountable safety baseline.

Paragraph Summary

Microsoft President Brad Smith argues that AI is subject to less regulation than milk: the dairy industry has mandatory inspection, traceability, and recall systems, while large models' biases, hallucinations, and abuse risks are hard to trace. This regulatory gap makes risks invisible and responsibility untraceable, raising the uncertainty and transaction costs for enterprises deploying AI. Global AI regulation is advancing in fragments: the EU legislates by risk tiers, the US leans toward industry self-regulation, and China has issued measures for generative AI. The disagreement centers on legislative speed and regulatory level. Smith warns that once a major safety incident occurs, public sentiment could force stricter rules, and leading companies that proactively participate in setting safety standards can turn compliance pressure into competitive barriers.