A human will check it at the end. That sentence has a remarkable capacity to reassure.
Yet having a person at the end is not the same as having another judgment at the end. Someone reads the AI's explanation, checks that it makes sense, and approves it. A person certainly participated. Participation can be recorded in a log.
What is harder to record is what that person prevented us from missing.
Splitting the average
Sander Lanbo Shi and colleagues' preprint, “Towards Human-AI Complementarity for AI Safety,” tries to make this less visible dimension measurable. A qualification belongs at the beginning: the study has serious limitations in its comparisons and ethical procedures. Its findings cannot be adopted directly as operating policy. We read it here for what it tried to measure, as well as what it found.[1]
The study collected human judgments on 3,957 questions across eight safety-related datasets. It compared no AI assistance, assistance that supplied candidate answers, and assistance that broke tasks into smaller questions. Its principal human evaluation used majority votes from three people.
Rather than merely comparing overall accuracy, it separated questions the AI answered correctly from those it got wrong.
On questions Sonnet answered incorrectly, the human group's accuracy was 36.7% in the unassisted condition, compared with 20.6% when candidate answers were supplied. But this was not a causal effect established by random assignment. Differences between participants and implementation across conditions cannot be ruled out. We cannot say that showing assistance caused the human errors.
Even so, the framing matters. Improvement in people's average score and their ability to correct AI's mistakes are different things. Without measuring them separately, we cannot tell them apart.
The distance between good advice and good oversight
What follows is a reflection prompted by that approach to measurement.
When AI is right, help that makes its answer easier to understand is welcome. It may make us faster, reduce uncertainty, and increase correct answers. But oversight also has a role when a system that usually works well goes off course.
Simply appointing someone to disagree with AI every time is not the answer. Overturning a correct judgment is a failure too. What we need is not a greater volume of objections, but a route to different evidence.
For instance, do we bundle checking an explanation's consistency and searching for evidence it leaves out into the single word “review”? The first can happen inside the world presented to us. The second takes time to look beyond it.
When a polished option is in front of us, we may want to scrutinize the choices rather than fill a blank page. That is not necessarily laziness. The screen itself is proposing a particular job. The more imposing the approval button becomes, the smaller every action other than approval can appear.
This research cannot justify rejecting a particular interface design. But we can decide what we want to evaluate: speed, overall accuracy, or the ability to rescue a particular kind of failure. Calling all of them important and merging them into one average can make one disappear.
The reader needs another pair of eyes too
The paper's constraints cannot be pushed into a corner of the footnotes. The authors state that they did not obtain ethics review and did not establish an explicit informed-consent procedure. The evaluation platform's code is public, but the human evaluation data and the code combining human and AI answers are not. Independent replication faces obstacles.
An interesting question and a sufficiently reliable answer are also different things.
This is not a warning designed to make research less enjoyable. It is a distinction that helps a question remain useful. If another experiment changes these numbers, the need to examine which AI errors people corrected does not disappear. Conversely, the importance of that question does not strengthen this study's comparison method.
We can treat the finding as a hypothesis while awaiting research with randomization and stronger independent verification. Or we can use safe evaluation tasks of our own to test whether another judgment actually does useful work. Instead of treating the presence of a human as a completed safeguard, we can observe what the human changed. That sequence avoids selling assurances beyond the limits of the research.
One more distinction matters: someone being right is not the same as their judgment being adopted. Different answers may exist, yet the potential for complementarity never becomes a result if we do not know which one to choose. A mechanism for collecting opinions and a mechanism for comparing their evidence are separately necessary.
If titles or confidence become our shortcut, the work of judgment shifts into a different job: deciding whom to trust. Some such shift may be inevitable, but it is better made visible. An organizational chart that shows only the final approver can leave hidden what determined the answer along the way.
In this sense, good oversight is more complicated than adding a person. Which information comes first? Where can someone stop? What can be checked when an objection is raised? Oversight is an activity made up of many conditions. A human name on the result does not, by itself, tell us that activity occurred.
A record saying “reviewed by a human” may be necessary for locating responsibility. But there is still another question about the quality of judgment.
Which mistake failed to get through because that person was there?