What the opioid crisis can—and cannot—teach us about artificial intelligence
The same capacity can relieve or harm. The difference lies in dose, delivery and power.
The analogy appears neat at first.
Assistive AI is morphine: controlled, purposeful and administered to relieve a genuine burden. Generative AI is heroin: intoxicating, excessive and liable to replace human agency with dependence.
It is also the wrong distinction.
Morphine can be indispensable medicine, but it can also produce dependence, overdose and profound harm. Heroin is manufactured from morphine; under the name diamorphine, it even retains limited medical uses in some jurisdictions. The molecules do not occupy morally opposite worlds. Their effects depend upon potency, dose, method of delivery, clinical purpose, supervision, patient vulnerability and the system through which they are supplied.
AI presents a similar difficulty. “Assistive” and “generative” are not stable moral categories. A generative model can help a doctor prepare clinical notes, allow a disabled person to communicate more easily or give a student individualised explanations. A supposedly assistive recommendation system can promote extremism, encourage compulsive use or quietly narrow the information a person encounters.
The same system may assist one user, replace another worker, mislead a third and place a fourth under surveillance.
The relevant question is not simply what kind of AI a system is. It is how the system is delivered, who controls it, what behaviour it encourages and who bears the consequences when it fails.
This is where the opioid comparison becomes more useful.
Opioids are powerful because they interact with systems already present in the body. They do not introduce the capacity for relief or pleasure from nothing; they act upon mechanisms that evolved for pain, reward and survival. Their medical value and their danger arise from the same underlying potency.
AI also acquires much of its power by entering existing human systems. It works through language, judgement, memory, social trust and our willingness to delegate. A model that summarises a difficult document may save time. A model that supplies every interpretation may gradually weaken the reader’s habit of forming one. A system that helps someone begin a task may be liberating; one that makes beginning without it feel impossible has changed the nature of the relationship.
This should not be confused with clinical addiction. Using an AI assistant is not pharmacologically equivalent to opioid dependence, and careless comparisons can trivialise a public-health crisis that has killed and injured enormous numbers of people. The resemblance is structural rather than biological. In both cases, a useful intervention can become a dependency when it is made continuously available, aggressively promoted and embedded within institutions before its long-term effects are understood.
The word dose is helpful here.
For AI, dose might mean frequency, reach or depth of delegation. Asking a system to correct punctuation is a small intervention. Allowing it to determine what news a person sees, whether a job applicant receives an interview or which patient receives further investigation is something else entirely.
The distinction is not merely quantitative. A higher dose can change the role of the technology. Occasional assistance may leave human skill intact; constant substitution may cause that skill to atrophy. A person who uses AI to challenge an argument remains engaged in reasoning. A person who routinely asks it what to believe has delegated more than labour.
Institutions can develop the same dependency. A workplace introduces AI to reduce administrative burden, then reorganises staffing around the assumption that the system will always be available. Junior positions disappear. Experienced workers lose opportunities to practise the tasks on which their judgement was built. Eventually, the organisation may no longer possess the human capacity to operate without the tool, correct it or recognise when it is wrong.
What began as assistance becomes infrastructure. Infrastructure acquires power because withdrawing it is no longer easy.
The opioid crisis also demonstrates that harm cannot be understood solely as the consequence of irresponsible individual use. Its history includes prescribing practices, misleading assurances, commercial promotion, regulatory failure, inadequate treatment and the later saturation of illegal drug supplies with highly potent synthetic opioids. The United States has not simply “stabilised” opioid use through careful control; overdose remains a severe public-health problem, even as the substances driving it have changed over time. The CDC describes successive waves involving prescription opioids, heroin and illegally manufactured synthetic opioids.
The lesson is not that powerful things should be prohibited. It is that risk expands when institutions profit from distributing a powerful intervention while the costs are carried elsewhere.
That lesson should make us wary of the way AI is being introduced. Technology companies are rewarded for adoption, scale and market capture. The people exposed to errors may be patients, students, workers or citizens who never chose the system and have no practical ability to refuse it. Benefits can be immediate and easily advertised; harms may be delayed, dispersed and difficult to attribute.
A company can point to the time its product saves while ignoring the expertise it displaces. It can advertise accuracy in a controlled evaluation while leaving users to discover how the system behaves in unfamiliar conditions. It can call a product a “copilot” even when the human operator lacks the information, authority or time required to overrule it.
The language of assistance can conceal a transfer of responsibility.
The comparison with medicine suggests that AI should be governed according to its use, exposure and potential harm rather than by a single distinction between “assistive” and “generative.” We do not regulate every drug identically simply because each alters the body. Nor should a tool that generates birthday invitations be treated like one that influences diagnoses, employment, education or criminal justice.
Higher-risk systems require stronger evidence. They should be tested against the populations and conditions in which they will actually operate. Their failures should be recorded, investigated and disclosed. People affected by consequential decisions need a meaningful route to explanation, correction and redress.
Human oversight must also mean more than placing a person beside the machine so that responsibility has somewhere to land. The person must understand the system’s limitations, have enough time to evaluate its output and possess genuine authority to reject it. Otherwise, “human in the loop” becomes a ceremonial phrase attached to an automated decision.
This is especially important in healthcare. AI can help analyse information, reduce clerical work and extend scarce expertise. It can also produce confident falsehoods, reproduce biases or encourage people to rely on systems that were never tested for their circumstances. The World Health Organization’s guidance on generative models in health therefore calls for defined tasks, independent auditing, stakeholder involvement and continued assessment after deployment—not simply optimism accompanied by a disclaimer. Its recommendations recognise both the medical promise and the potential for serious harm.
Post-release monitoring matters because no laboratory can anticipate every real-world use. A system changes when people incorporate it into routines, trust it, resist it or learn to exploit it. Safety is not a property demonstrated once before launch. It is a condition that must be maintained.
The opioid analogy has limits that should remain visible. AI does not act directly upon the body in the way a drug does. Its dangers are not reducible to addiction, and its benefits are not equivalent to pain relief. AI is a sociotechnical system: its effects emerge through institutions, interfaces, data, labour practices and unequal distributions of power.
But that difference strengthens the central lesson. The character of a technology cannot be inferred from its intended purpose or the language used to market it. A medical product can cause social catastrophe when commercial incentives overwhelm caution. An intelligent system can produce widespread harm while every participant insists that it was merely helping.
We should therefore resist both technological panic and technological innocence. AI is neither digital heroin nor an inherently benevolent assistant. It is a family of capabilities whose effects depend upon where they are placed and what we permit them to replace.
Responsible use does not mean drawing one bright line and declaring everything on one side safe. It means asking harder questions of every deployment.
What is the smallest effective dose? What evidence supports the claimed benefit? Can the user refuse it? Does the system preserve human competence or quietly make that competence uneconomical? Who monitors long-term effects? Who profits from greater dependence? Who is accountable when the harm appears somewhere other than where the revenue is collected?
The fine line is not between morphine and heroin. It is not between assistive and generative AI.
It runs through every application: between relief and reliance, augmentation and substitution, consent and imposition, public benefit and private extraction.
Crossing it is rarely a single dramatic act. More often, it happens through a series of convenient decisions that no one finds important enough to resist.
