The Mineral Oil Lesson for Artificial Intelligence: Safety Is a Context, Not a Label

What does a jar of white mineral oil have to do with artificial intelligence, patient empowerment, and the future of medicine?

At first glance, almost nothing. One belongs to the quiet, technical world of food regulation. The other belongs to the noisy frontier of genomics, digital health, and large language models. Yet they illuminate the same underlying problem: how do we decide that something is safe enough to enter ordinary life?

The answer is rarely found in the object alone. It depends on context, dosage, users, incentives, evidence, and what happens after deployment. A substance can be acceptable in food only under specified conditions. An algorithm can be accurate in a trial yet become unreliable when it meets different populations, new behaviors, and changing institutions.

The central lesson is easy to miss because one example looks futuristic and the other looks mundane:

Safety is not a permanent property of a thing. It is a relationship between a thing, its conditions of use, and the people living with its consequences.

That idea should change how we think about digital health. The goal is not simply to produce more data, faster diagnoses, or more personalized interventions. The goal is to build systems whose benefits remain intelligible, measurable, and accountable as they move from controlled settings into the untidy world of human life.

The quiet intelligence of a narrow rule

A regulation allowing white mineral oil to be used in food does not offer a sweeping declaration that mineral oil is universally safe. It establishes a bounded permission: this material may be used, in food, under specified conditions. The apparent modesty of the rule is its strength.

Such a rule recognizes several facts at once. A substance cannot be judged apart from its route of exposure. The amount matters. The purpose matters. The quality of the material matters. The population consuming it matters. Safety is therefore not a simple label attached to a chemical. It is a conclusion produced by narrowing the circumstances until evidence can meaningfully support a claim.

This is a powerful contrast with the way technological products are often discussed. We ask whether an artificial intelligence system is safe, whether a wearable device is accurate, or whether a digital therapeutic works. These questions sound sensible, but they are incomplete. Safe for whom? Accurate under what conditions? Effective compared with which alternative? What happens when people misunderstand it, rely on it too much, cannot access it, or use it in a setting the designers never imagined?

The humble regulatory rule contains a mental model that the digital world urgently needs: permission should be contextual, conditional, and revisable.

Consider a hypothetical symptom checking system. In a clinical study, it performs well for English speaking adults with reliable internet access who describe their symptoms clearly and follow instructions. After launch, it is used by people with limited health literacy, intermittent connectivity, different patterns of disease, and cultural understandings of pain. Its accuracy may decline, not because the underlying model suddenly changed, but because the environment changed.

The system was never a free floating intelligence. It was a component in a larger social arrangement. Its performance depended on language, trust, access, clinician availability, and the choices people made after receiving its advice.

This is why a regulation for a familiar substance can be conceptually relevant to a frontier technology. Both require us to replace the question “Is it safe?” with a more disciplined sequence:

1. Safe for which use?
2. Safe at what intensity or scale?
3. Safe for which populations?
4. Safe compared with what alternative?
5. How will we know if the answer changes?

The last question is the one most often neglected.

From product approval to living systems

Modern medicine has traditionally relied on a useful fiction: that a medical product can be evaluated as a relatively stable object. A drug has a composition. A device has a design. A clinical trial has a protocol. Regulators can then examine evidence and decide whether expected benefits outweigh expected risks.

Digital systems disrupt this model because they do not remain entirely stable after release. Algorithms are influenced by new data. Users adapt to recommendations. Clinicians change their workflows. Institutions alter their incentives. Online narratives reshape what people believe and how they describe their symptoms. A system can therefore alter the very environment in which it is being evaluated.

This creates a feedback loop:

The system influences behavior, behavior changes the data, the data changes the system, and the revised system influences behavior again.

Imagine a tool designed to identify patients at high risk of hospitalization. If clinicians begin prioritizing the patients it flags, those patients may receive earlier treatment and avoid hospitalization. The tool could then appear less accurate because the outcome it predicted was prevented. Alternatively, if the tool is trusted too much, clinicians may neglect patients it does not flag, creating a different pattern of harm.

The algorithm cannot be judged only by asking whether its initial predictions were correct. We must also ask what its predictions caused people and institutions to do.

This is where patient empowerment becomes more demanding than simply giving patients access to more information. If people are asked to contribute continuous data through phones, watches, surveys, or home devices, they are not merely data sources. They are participants in a changing system. Their experiences determine whether a measurement captures what matters, whether an intervention fits into daily life, and whether a statistically significant result represents a meaningful improvement.

A patient who reports fewer symptoms but can no longer work has not necessarily experienced a successful outcome. A treatment that extends life while producing intolerable fatigue may be judged differently by different people. A digital tool that reduces clinic visits may improve access for one patient and deepen isolation for another.

The relevant evidence must therefore include both biological signals and lived consequences. This is not a sentimental addition to scientific rigor. It is a condition of scientific validity. If the measurement excludes the outcomes people actually care about, the measurement is precise but wrong in the most important sense.

A system cannot be patient centered if patients are invited to provide data but not to define what counts as a good outcome.

The same principle appears in ordinary regulation. A narrowly defined permission does not merely protect consumers from a substance. It also makes the claim testable. By specifying the conditions, regulators create a boundary within which evidence can be interpreted. Patient centered development performs a similar function by identifying the outcomes, burdens, and tradeoffs that should organize evaluation.

The real risk is not only bad technology

Public discussion often treats digital health risk as a problem of technical error. The algorithm might hallucinate. The device might misread a signal. The model might reproduce bias. These are real dangers, but they are only one layer of the problem.

A more consequential risk is misplaced confidence. People may assume that a polished interface represents reliable knowledge. Clinicians may treat a probability as a diagnosis. Policymakers may mistake data volume for representativeness. Patients may accept recommendations without understanding their limitations, especially when information is delivered in fluent, personalized language.

Large language models make this danger unusually vivid. Their ability to answer questions in language suited to a person’s literacy and numeracy could make health information far more accessible. Yet the same fluency can conceal uncertainty. A confident answer can travel through families, social networks, and online communities faster than a careful correction.

The problem is not solved by telling users to be more skeptical. Skepticism is unevenly distributed, and people do not make decisions in an informational vacuum. Someone facing pain, financial pressure, or limited access to care may reasonably accept an answer that feels immediate and comprehensible. The burden must therefore fall partly on system design and governance.

A trustworthy system should make its boundaries visible. It should distinguish evidence from inference, uncertainty from confidence, and general information from individualized advice. It should create clear escalation paths when symptoms are urgent or information is insufficient. It should be tested across the populations and circumstances in which it will actually be used.

This is another place where narrow regulatory thinking is useful. A label that says a product is acceptable under defined conditions does not insult the user’s intelligence. It gives the user a map. Digital systems need equivalent maps: what the system knows, what it does not know, who was represented in its testing, how often it is monitored, and what to do when its recommendation conflicts with lived experience.

Transparency, however, should not mean dumping technical documentation on the public. The important question is not whether a person can inspect the entire model. It is whether the person can understand the system’s practical limits well enough to make a safer decision.

A new framework: the four boundaries of responsible innovation

To move from general optimism or fear toward useful judgment, every health technology should be examined through four boundaries.

1. The exposure boundary

What exactly is being introduced into people’s lives, and at what intensity?

For a substance, this may mean concentration, route, and frequency. For a digital system, it includes notification volume, decision authority, data collection, and how often users are encouraged to consult it. A tool that offers occasional reminders is not equivalent to one that continuously interprets a person’s body and behavior.

Excessive exposure can create harms even when each individual interaction seems harmless. Constant monitoring may increase anxiety. Repeated alerts may produce fatigue. Endless personalization may make people feel that ordinary variation is a medical problem.

2. The context boundary

Where and by whom will the technology be used?

A clinical trial can provide a controlled environment, but deployment occurs in homes, workplaces, pharmacies, schools, and communities. The relevant context includes language, income, disability, geography, cultural expectations, and the availability of human support.

The same tool can be beneficial in a well staffed clinic and dangerous in a setting where it becomes a substitute for professional care. Context is not background information. It is part of the intervention.

3. The authority boundary

What decisions is the system allowed to influence, and who remains accountable?

A model that helps organize questions for a doctor has a different risk profile from one that recommends medication changes. A patient portal that displays laboratory results is different from a system that interprets them without clinician involvement.

As systems become more capable, their authority often expands informally before it expands legally. People begin to rely on them because they are convenient. Responsible design must make authority explicit before convenience turns into dependence.

4. The learning boundary

How will the system be monitored and corrected after release?

Accuracy at launch is not enough. Developers and regulators need mechanisms for detecting drift, subgroup failures, unexpected uses, and changes in behavior. Patients need channels to report harms that may not appear in standard performance metrics.

This suggests a lifecycle model of trust. Approval should be understood not as a finish line but as the beginning of an evidence relationship. The system earns continued permission by demonstrating that it remains useful and safe under real conditions.

What organizations can do now

The practical implications are less glamorous than announcing a revolutionary platform, which is precisely why they matter.

First, define the intended use narrowly enough to evaluate. “Improves health” is not an adequate claim. Specify the population, setting, decision, outcome, and acceptable tradeoffs.

Second, measure outcomes that patients recognize as meaningful. Include function, burden, quality of life, access, and the ability to carry out ordinary activities, not only laboratory values or engagement statistics.

Third, test the system in the environments where it will operate. If a product is intended for broad public use, evidence from a narrow and unusually supported population is insufficient.

Fourth, build monitoring into the product rather than treating it as a compliance exercise. Track performance by subgroup, watch for changing patterns, and create a process for updating or withdrawing the system when conditions change.

Fifth, make limits actionable. Tell users what the system cannot determine, what warning signs require human care, and how to challenge or correct an output.

Finally, treat regulation as an enabling technology. Good rules do not merely slow innovation. They define the conditions under which innovation can become trustworthy enough to scale. A clear boundary can create more durable freedom than an ambiguous promise of safety.

Key Takeaways

• Judge technologies in context, not in isolation. Ask who uses them, under what conditions, at what intensity, and compared with which alternative.
• Let patients define meaningful success. A technically impressive outcome is not necessarily a valuable one if it ignores function, burden, dignity, or access.
• Separate accuracy from authority. A system may offer useful information without being qualified to make the final decision.
• Monitor after deployment. Algorithms, users, and institutions change together, so safety requires continuous measurement and correction.
• Prefer conditional trust to blind confidence. The most credible claim is not “this is safe,” but “this is appropriate for this use, within these boundaries, with this oversight.”

The future of health technology will not be decided only by the sophistication of models or the quantity of data. It will be decided by whether society can turn complexity into understandable conditions of trust.

That is why a mundane rule about white mineral oil deserves more attention than it usually receives. It represents a form of institutional wisdom: do not grant a material an unlimited blessing. Specify where it belongs, how it may be used, what evidence supports it, and what would require the judgment to be revisited.

Digital health needs the same humility. The most advanced system is still entering a human environment that can magnify its benefits, distort its measurements, and distribute its harms unevenly. We should not ask technology to become infallible before allowing it to help. We should demand something more realistic and more rigorous: bounded permission, visible limits, patient defined value, and the capacity to learn when reality proves us wrong.