The Real AI Risk Is Not Intelligence. It Is Unaccountable Power

What if the most dangerous thing about artificial intelligence is not that machines may become smarter than us, but that they may become very good at doing what institutions already do: hiding responsibility behind a system?

That possibility links two conversations that are usually kept apart. One concerns superintelligent machines, whose decisions may eventually move faster than human comprehension. The other concerns familiar human institutions: immigration enforcement, political parties, entertainment franchises, elite social networks, and public figures who make harmful choices while everyone nearby looks away.

The common problem is not intelligence. It is power without meaningful accountability.

A machine that decides humans are a threat would be terrifying. So would a bureaucracy that treats people as categories, a media system that quietly defines who is desirable, or an elite network that turns vulnerable people into currency. In each case, harm becomes easier when no single person feels fully responsible, when procedure substitutes for judgment, and when those most affected are excluded from the conversation.

The central question, then, is not simply whether AI will behave well. It is this: Can we build systems that remain answerable to the people who bear their consequences?

The danger begins when responsibility becomes diffuse

A powerful technology is often described as a tool. Tools, in this framing, are neutral. A hammer can build a house or cause injury. Artificial intelligence can improve education and healthcare, or it can be used for malicious purposes. The observation is true, but incomplete.

The deepest danger of a tool is not only what an individual chooses to do with it. It is what happens when the tool becomes embedded in a system where responsibility is distributed so widely that nobody can be held to account.

Consider a hypothetical automated benefits system. It denies thousands of people access to healthcare. The programmer says the model followed its training data. The agency says the vendor designed the system. The vendor says the client selected the parameters. The administrator says the result was generated by the model. The citizen receives a denial, but no person can explain the decision or reverse it.

This is not an evil machine. It is something more ordinary and potentially more dangerous: a chain of human decisions that has been made to look like an act of nature.

The same pattern appears in institutions that use masks, uniforms, official language, or legal procedure to distance power from its effects. A person is detained by an officer whose identity is concealed. A political leader proposes expanding the authority of a force while defenders debate cameras, uniforms, and technical safeguards. The debate over visible accountability matters, but it can also become a way of avoiding a prior question: should this institution possess the power it is being given at all?

This distinction is crucial. Governance is not the same as decoration. Body cameras, disclosure forms, revised branding, and new oversight offices may be useful. But they do not automatically alter the incentives, mission, or power structure that produced the harm.

An institution can be made more legible without being made more just.

Cosmetic reform is the politics of appearances

There is a recurring institutional reflex when a system is criticized. It changes the surface while protecting the core.

A dating franchise may announce a return to traditional love. That phrase sounds timeless, even wholesome. But what does it mean in practice? If traditional means commitment and shared values, many people may welcome it. If it means fixed gender roles, a provider husband, a passive wife, and a narrow image of desirability, the phrase becomes a coded restoration of an older hierarchy.

The important insight is that tradition is often less a description of the past than a marketing strategy for selecting which parts of the past should return.

The same pattern appears in technology. Companies announce responsible AI principles, ethical guidelines, and safety commitments. These can be valuable. Yet if they leave untouched the concentration of decision making, the opacity of models, the economic pressure to deploy quickly, and the absence of recourse for those harmed, then responsibility becomes a public relations layer placed over an unchanged machine.

A system may say it values fairness while continuing to define fairness without consulting the people most likely to be misclassified. It may say it supports human oversight while giving humans only a ceremonial role, asking them to approve decisions they cannot understand and are discouraged from challenging.

This is why the question “Does the system have safeguards?” is too weak. We should ask four harder questions:

1. Who designed the system?
2. Who benefits when it expands?
3. Who bears the cost when it fails?
4. Who has the authority to stop it?

These questions expose the difference between accountability and theater. A system is accountable only when the people affected can identify a responsible decision maker, understand the grounds for action, contest the outcome, and obtain repair when harm occurs.

Without those conditions, reform becomes a costume.

The people closest to harm are not a focus group

One of the most revealing themes in these debates is the insistence that affected communities must be engaged directly, not merely represented by a press release.

When a public figure makes a comment that harms Black audiences, issuing a statement to a general audience is not the same as having a difficult conversation with Black media and community members. The difference is not symbolic. It determines whether the person is willing to encounter the knowledge that their own social position may have prevented them from acquiring.

This is a general rule for institutional intelligence: the people who experience a system from below often understand its failures more accurately than the people who administer it from above.

A policymaker sees a budget. A family sees a separation. An executive sees a risk metric. An employee sees retaliation. A producer sees audience demographics. A participant sees that the supposedly universal ideal of romance was designed around someone else.

The exclusion of affected voices produces a specific kind of blindness. It allows institutions to mistake their own defaults for reality.

That blindness is visible in cultural production. A rock song can sound energetic, rebellious, and commercially effective while having shed the musical influences that gave rock much of its original force. The result is not necessarily bad music. It is music that performs rebellion through a standardized product. Its posture says resistance, while its structure says safety.

This is a useful analogy for AI. A system can appear innovative while reproducing old assumptions. It can use sophisticated language, impressive computation, and novel interfaces while quietly encoding the same social hierarchy that existed before automation.

The question is not whether a system looks modern. The question is whether it has learned anything from the people whom previous systems ignored.

The “Epstein class” is a model of social permission

The most unsettling form of complicity is not always direct participation. It is the willingness to remain near obvious wrongdoing because proximity offers access, status, money, or excitement.

That is the deeper significance of the idea of an enabling elite. The problem is not limited to identifying individual predators. It includes the network that recognizes something is wrong, benefits from the arrangement, and decides that the cost of intervention is too high.

This network is held together by what might be called social permission. One person normalizes the behavior. Another provides access. A third offers credibility. A fourth stays silent. A fifth says the facts are complicated. Eventually, the system no longer requires everyone to be cruel. It only requires enough people to be comfortable.

This matters for AI because advanced systems will not arrive in a moral vacuum. They will be built, purchased, deployed, and defended by institutions. If those institutions already reward silence, obscure responsibility, and treat vulnerable people as disposable, AI may increase the scale and speed of those tendencies.

Imagine an automated surveillance system used to identify “suspicious” people. The model is trained on historical enforcement data. Those data reflect unequal policing. The system then identifies the same communities as high risk, which justifies additional surveillance. Officials describe the output as objective because it came from a model. The model appears accurate because the institution keeps generating the same evidence it used to train the model.

No single moment needs to contain explicit malice. The system can reproduce injustice through a series of reasonable looking decisions.

That is how social permission becomes computational permission. The machine does not need to hate anyone. It only needs to inherit a world in which some people are treated as more disposable than others.

The future danger of AI may not be that machines become alien. It may be that they become perfect participants in our existing excuses.

A better framework: speed, scale, opacity, and recourse

To evaluate any powerful system, we need more than a binary distinction between beneficial and harmful uses. A practical framework is to examine four properties: speed, scale, opacity, and recourse.

Speed asks how quickly the system can act compared with the time available for human judgment. If a decision can be made in milliseconds but challenged only after months, the system is functionally unaccountable even if an appeal exists on paper.

Scale asks how many people can be affected by one error. A mistaken human decision may harm one person. A flawed model can reproduce the same mistake across millions of cases before anyone notices.

Opacity asks whether those affected can understand why the system acted. Complexity is not an excuse for secrecy. If a decision cannot be explained in terms a person can use to challenge it, it should not control that person’s life.

Recourse asks whether someone with real authority can stop, reverse, and repair the decision. An apology without correction is not recourse. A review process that cannot change the result is not recourse. A hotline that routes people back into the same automated loop is not recourse.

The risk becomes especially high when all four properties converge. A fast, scalable, opaque system with weak recourse is not merely efficient. It is a mechanism for distributing harm while concentrating power.

This framework also clarifies why some reforms are insufficient. Adding a camera may improve visibility, but it does not necessarily provide recourse. Publishing principles may improve transparency, but it does not necessarily reduce scale. Creating a review committee may add oversight, but if the committee lacks authority, it is only a witness.

The goal should be to design friction where friction protects people. Not every process should be optimized for speed. A pause before a deportation, a human conversation before a benefits termination, or a challenge from an affected community before deployment may look inefficient from the perspective of the institution. From the perspective of the person exposed to irreversible harm, it is civilization.

Key Takeaways

1. Do not confuse visibility with accountability. Ask who can reverse a decision, not merely whether the decision is recorded.

2. Treat “traditional,” “neutral,” and “efficient” as contested terms. Every supposedly universal standard contains assumptions about whose lives count as normal.

3. Put affected people inside the decision process early. Do not wait for a scandal before inviting the people most exposed to harm into the room.

4. Audit incentives, not only outputs. If an institution benefits from expansion, it will often interpret evidence in ways that justify expansion.

5. Demand a named human owner for high stakes decisions. If nobody can explain, contest, and repair an outcome, the system is not ready to govern anyone.

The hardest challenge posed by artificial intelligence is therefore not forecasting the moment when a machine becomes superintelligent. It is recognizing how often humans already surrender judgment to systems they claim not to control.

We fear a future in which a machine decides that humanity is a threat. We should also fear a present in which institutions decide that certain humans are acceptable losses, then hide that decision behind tradition, procedure, data, or public relations.

The remedy is not blind faith in human judgment. Humans are biased, vain, tribal, and capable of extraordinary cruelty. The remedy is to build systems in which judgment remains contestable, power remains visible, and the people most affected are impossible to ignore.

A truly intelligent society would not merely create faster decision makers. It would create better reasons to stop, listen, and change course.

That may be the most important rule of the road: the more powerful the system, the more answerable it must remain to the people beneath it.