This is today’s edition of The Download, our weekday newsletter that provides a daily dose of what’s going on in the world of technology. The problem with thinking you’re part Neanderthal There’s a theory tha
The Oversight That Isn’t There
The doctrine of “human in the loop” has become the ethical fig leaf of our age. In defence circles, it is the phrase that makes autonomous weapons systems politically acceptable. In enterprise boardrooms, it is the phrase that makes AI deployment feel responsible. In both cases, the humans nominally in the loop are operating at the wrong timescale to exercise anything resembling genuine judgement.
Consider what “human oversight” actually means when a hypersonic missile closes at Mach 8. The engagement window is measured in seconds. The sensor fusion, threat classification, and targeting calculations happen in milliseconds. The human operator is not making a decision — they are performing a ritual that allows an institution to say a decision was made. This is not a failure of implementation. It is a structural feature that organisations have strong incentives to preserve, because the alternative — admitting that the human is decorative — triggers obligations they would rather not face.
The uncomfortable parallel for executives is direct. When a machine learning model flags a credit application, a fraud alert, or a hiring candidate, and a human reviews it in thirty seconds, that human is not exercising oversight. They are providing institutional cover. The system is already autonomous; you have simply inserted a person who can absorb the blame.
When Speed Becomes the Architecture
The deeper problem is that speed is not incidental to these systems — it is their core value proposition. An air defence system that requires a thirty-second human deliberation is worse than useless; it is dangerous. An algorithmic trading system that pauses for human review has already missed the opportunity. A fraud-detection model that holds transactions for human approval has already annoyed the customer and slowed the business.
This creates a structural trap. Organisations adopt AI specifically because it operates faster than human cognition. Then they insert humans into the loop to satisfy governance requirements. Then they discover that the human bottleneck defeats the purpose. So they tune the system to only escalate the most ambiguous cases — which means the human is reviewing precisely the decisions where they have least confidence and least data. The US military’s Iron Dome partnership with Israel illustrates this exactly: the system can intercept incoming rockets autonomously, and does so routinely, but technically requires operator authorisation. In practice, the operator confirms after the fact, or the engagement window closes before they can act. The loop exists on paper. The autonomy exists in reality.
The enterprise version of this plays out in content moderation at scale. Meta reportedly reviews approximately three million pieces of content daily through human moderators — but the AI pre-filters what reaches them, and the pace means each decision receives seconds of attention. Research from the University of Michigan found that moderators under time pressure reverse AI recommendations at rates barely above chance. The human adds latency. It does not add judgement.
The Accountability Gap Nobody Wants to Close
Here is where the institutional incentive becomes most visible. Genuine human oversight is expensive. It requires slow systems, trained reviewers, documented reasoning, and meaningful authority to override. Ceremonial oversight is cheap. It requires a dashboard, a button, and a policy that says “human reviewed.”
The defence sector has spent a decade debating autonomous weapons under international humanitarian law, precisely because genuine accountability for machine decisions is legally and politically intractable. If an autonomous system kills civilians, who is responsible? The operator who pressed confirm without understanding the targeting calculus? The engineer who wrote the classification model? The procurement official who approved the system? The answer nobody wants to give is: nobody, in any meaningful sense. The “human in the loop” doctrine exists not to enable accountability but to defer this question indefinitely.
Boards face an identical structure when deploying AI in high-stakes decisions. A bank that uses a model to deny loans can point to a human reviewer and claim compliance with fair lending requirements. But if the reviewer is approving four hundred decisions per shift with no tools to interrogate the model’s reasoning, the accountability chain is fiction. The regulatory frameworks — the EU AI Act’s “human oversight” requirements, the FCA’s model risk guidance — are written in ways that can be satisfied by ceremony rather than substance. Most organisations will satisfy them by ceremony, because substance is costly and ceremony is sufficient to avoid enforcement.
The executives who will face genuine exposure are those who discover this after something goes wrong. At that point, the question “was there a human in the loop?” becomes “what did that human actually have the capacity to do?” These are very different questions, and only the second one matters in a courtroom or a parliamentary inquiry.
Decision framework for genuine versus ceremonial oversight:
If your AI system makes decisions faster than a human can meaningfully evaluate a single case — and you have inserted human review anyway — acknowledge that you have an autonomous system with a compliance wrapper, and govern it accordingly: invest in model audits, explainability infrastructure, and statistical monitoring rather than individual case review.
If your human reviewers are overriding the model at rates below 5% consistently, treat that as a signal that oversight has become ceremonial; either invest in making override genuinely viable or remove the pretence and accept accountability for the system’s outputs directly.
If the consequence of a wrong decision is reversible (a content recommendation, a marketing offer, a search ranking), light-touch human governance is proportionate — periodic audits, aggregate outcome monitoring, and clear escalation paths are sufficient.
If the consequence is irreversible — a loan denial, a benefits decision, a targeting classification — the human in the loop must have the time, tools, and authority to actually reverse the machine. If they do not, you do not have oversight. You have liability camouflage. The distinction will matter when the decision you cannot undo becomes the decision you cannot explain.