Humanwashing: The AI Governance Problem Nobody Wants to Name

There is a term that appeared in the academic literature this year that deserves to travel much further than conference proceedings.

Humanwashing.

Researchers at Swansea University published a paper at the 2026 Hybrid Human-Artificial Intelligence conference arguing that the phrase “human in the loop” is increasingly used the way “green” was used by corporations in the 1990s — not to describe a substantive practice, but to make a system appear safer than it is. The analogy is precise and the implication is serious: organizations are deploying the language of human oversight as a reputational and regulatory shield, without examining whether the oversight they’re claiming actually functions.

This is a stronger claim than saying human-in-the-loop is imperfect. It says that in many cases, naming a human reviewer is the organizational act — that the behavioral reality behind the name is treated as a detail. And it predicts something specific: that as regulatory requirements for AI oversight become more exacting, the gap between what organizations claim and what they have built will become increasingly difficult to hide.

Behavioral science has been accumulating the evidence for this gap for years. The question is whether the organizations doing the humanwashing know they’re doing it.

There is a charitable version of how we arrived here. Most organizations that placed a human reviewer in an AI workflow genuinely believed they were doing something meaningful. The reviewer is present. The button exists. The audit trail is real. Nobody set out to deceive regulators or the public. They set out to be responsible and reached for the most available tool — a person in the process.

The behavioral problem is that this is where most of them stopped. Once the person was placed in the loop, the question of whether that person could do the work the loop required was treated as either obvious or someone else’s problem. Obvious, because humans are competent agents who will notice what is in front of them. Someone else’s because the reviewer is an operational matter, not a governance one.

Both assumptions are wrong, and behavioral science can say precisely why.

The reviewer who is processing thirty cases an hour is not failing to try. They are responding rationally to an environment that has made careful review impossible. The reviewer who approved a concerning output and said nothing is not timid. They are responding to an incentive structure in which overriding requires written justification, supervisor notification, and a reasonable expectation of being questioned — while approving requires nothing. The reviewer who cannot describe how the AI system they oversee tends to fail is not negligent. They were trained on how to use the tool and never taught anything else.

None of these failures are individual failures. They are the predictable output of institutional environments that organizations designed, in most cases without asking whether those environments were capable of producing the oversight they claimed to be providing.

That is the behavioral definition of humanwashing: not deliberate deception, but an organizational practice of naming oversight without designing for it.

The metric most commonly offered as proof that human oversight is working is the override rate. Reviewers are overriding rarely, which means either the AI is performing well, or the reviewers are satisfied with its outputs, or both. Either way, the loop appears to be functioning.

Behavioral science finds this interpretation almost entirely empty.

A low override rate is consistent with a highly accurate AI system. It is equally consistent with a reviewer who has learned that overriding is socially and procedurally costly, and has adjusted their behavior accordingly. It is consistent with a reviewer who is processing cases too quickly to evaluate them carefully. It is consistent with a reviewer who does not know what an error would look like. It is consistent with a reviewer who noticed something concerning, weighed the effort of acting against the effort of approving, and approved.

All of these situations produce the same metric. The number distinguishes none of them. Using a low override rate as evidence of effective oversight is not a measurement of the oversight process. It is a measurement of how often the AI and the reviewer arrive at the same output — which is a very different thing.

This matters because humanwashing survives on exactly this kind of metric. An organization that reports a 2% override rate and calls it evidence of functioning oversight has made a claim that the data cannot support. It has named the number as evidence of something the number does not measure. That is humanwashing in its most operational form.

Here is where the argument becomes genuinely unpopular.

A significant portion of the AI governance industry is built around two interventions: training reviewers to use AI tools better, and providing AI explanations so reviewers can understand why the system produced a given output. Both are well-intentioned. Both are, in many implementations, likely to make humanwashing worse rather than better.

Training that covers how to use the tool but not how it fails produces reviewers who are more fluent with the system without being better at catching its errors. Fluency and error detection are different skills. A reviewer who can navigate the interface efficiently and understand the output categories has not necessarily developed any capacity to recognize when an output in front of them is specifically wrong. The training has increased the appearance of competence without necessarily increasing the substance of it.

Explanations are more complex and more disturbing. A growing body of experimental evidence suggests that AI explanations can increase acceptance of AI advice regardless of whether that advice is correct. People treat the explanation as a signal to trust and not as material to evaluate. The mechanism is well-understood: an explanation that presents a plausible rationale for an output creates an impression of transparency that substitutes for the hard cognitive work of actually evaluating whether the output is right. The explanation is doing the opposite of what it claims. In systems where explanations are provided specifically to support meaningful oversight, this effect — reviewed in the research literature — means the explanation may be undermining the oversight it purports to enable.

Neither of these claims is absolute. Training done well, on failure modes and known error patterns, does improve detection. Explanations designed to support genuine evaluation rather than to create impressions of transparency can help. But the versions of training and explainability that most organizations are implementing are not those versions. They are the versions that look like oversight from the outside.

The most troubling dimension of humanwashing is not regulatory. It is distributional.

When a human-in-the-loop process fails — when the reviewer approves an output that is wrong and the wrong output produces a harmful outcome — the accountability structure in most organizations places the reviewer closest to the outcome. They were there. They approved. The audit trail says so.

Research on what has been called the moral crumple zone describes this dynamic precisely: in automated systems, responsibility tends to concentrate on the human nearest the failure, regardless of whether that human had any real control over what happened. The reviewer who was processing cases at a rate that made careful review impossible, who had never been taught what the system’s errors look like, who knew that overriding would be questioned and approving would not — that reviewer becomes the person who failed to catch it.

The organization that built that environment, that set that workload, that designed those incentives, that never examined whether the conditions for effective oversight were present — that organization is protected by the very claim of human oversight it never meaningfully provided.

This is why humanwashing is not only a governance problem. It is an ethics problem. It distributes blame away from the institutional actors who are responsible for the conditions and onto the individuals who were placed in positions that those conditions made untenable.

If an external assessor came to your organization and asked to examine your human oversight arrangement — not whether a reviewer exists, but whether the reviewer can actually do the work — what would they find?

Would they find reviewers who have time for careful engagement with each case? Would they find training that covers not just how to use the system but specifically how it fails and what its errors look like? Would they find a process where overriding is as easy as approving, where concerns can be raised without social cost, where escalation paths are clear and used? Would they find any mechanism by which reviewers learn what happened to the cases they passed through?

Or would they find a button, an audit trail, and a low override rate?

Those are two very different things. The growing regulatory pressure to demonstrate that human oversight is effective rather than merely present will eventually require organizations to answer that question on the record. The organizations that have been treating human-in-the-loop as a claim will find themselves needing to defend it as a practice.

Humanwashing has a shelf life. The behavioral conditions that make oversight real are not a secret. What they require from organizations — workload, training on failure modes, genuinely frictionless override, feedback loops that reach reviewers — is known. The gap between claiming oversight and designing for it is a choice. It is time to start treating it as one.