Human–AI collaboration
The unit of evaluation is the workflow
A good model assessment is only one part of a sound decision process. The handoffs, review and human contribution need examining too.
A model can produce a sound assessment and still be part of a poor decision process. Its answer may arrive after the important judgment has already been made. A reviewer may see the conclusion without the evidence needed to challenge it. An uncertainty that was visible during analysis may disappear when the result is summarised.
These are questions about the arrangement of the work. They are central to hybrid intelligence because the behaviour of the combined process depends on what each participant contributes, what each can inspect and how one contribution changes the next.
For analytical work, I think the useful unit of evaluation is the workflow that produces the judgment. A model evaluation is an essential part of that inquiry. It leaves open what happens when the output becomes an input to someone else's reasoning.
Follow the judgment through the process
Consider a team reviewing a proposal. An AI extracts relevant evidence, identifies possible difficulties and drafts an assessment. An expert checks the draft before making a recommendation. This description sounds straightforward, but it leaves several consequential choices unspecified.
Does the expert form an initial view before seeing the AI's assessment? Can they inspect the source material directly? Are omitted issues visible anywhere? What happens when the expert disputes the model's interpretation? Does the revised assessment preserve the reason for the change, or simply replace the earlier answer?
Each choice creates a different process. Showing a draft first might save effort, while also making that draft the starting point for the review. Requiring an independent assessment may provide a useful comparison, but consume much of the time the system was intended to save. These are trade-offs to investigate in the intended setting.
The handoff deserves particular attention. It is the point at which one participant's uncertainty, assumptions and evidence must become intelligible to another. A concise conclusion may be convenient to read while being difficult to examine. A long explanation may reveal the structure of an argument while imposing too much review effort. The evaluation needs to consider both.
Specify the human contribution
The phrase “human oversight” can conceal very different activities. A person might check whether cited evidence exists, judge whether a causal mechanism fits the case, choose the organisation's tolerance for uncertainty, or authorise an action. These activities call for different information and different expertise.
A review stage should therefore have an explicit purpose. If it is intended to catch unsupported factual claims, reviewers need access to the relevant sources. If it is intended to challenge the framing of a problem, a checklist of extracted facts will be insufficient. If the reviewer controls a decision threshold, they need to understand what the score represents and how it behaves.
This also makes the limits of review easier to investigate. A reviewer may correct a conspicuous error while missing a material omission. Counting edits alone would tell us little about that difference. Useful evidence would include the kinds of errors that reach the reviewer, those they identify, those they introduce and those that remain in the final assessment.
Compare processes on the same question
An evaluation can begin with a small set of clearly specified alternatives: the current expert process, an AI assessment under defined conditions, and a proposed combined workflow. The point of comparison is the same underlying analytical task, with equivalent access to the information available at the time.
Decision quality requires more than a single accuracy figure. For document-based analysis, some claims can be checked directly against evidence. Other conclusions depend on interpretation or a stated decision policy. Those should be assessed separately, so that disagreement about judgment does not hide a factual mistake, and factual consistency is not mistaken for a sound overall conclusion.
Effort belongs in the comparison too. A workflow that improves the first draft but doubles the time spent verifying it may have a different value from the one initially expected. Conversely, a slower process may be justified if it reliably exposes a consequential issue that was previously missed. The relevant trade-off depends on the task and the cost of its errors.
I would also examine how performance varies between cases. An average can obscure whether the workflow helps with routine assessments while failing on the ambiguous cases for which expert involvement matters most.
What my research contributes
In our study of layer-appropriate augmentation, we examined how structured reasoning and output calibration interact in an LLM-assisted assessment pipeline. The findings were conditional on the model, task and expert reference. They support treating the arrangement of the method as something to evaluate explicitly.
That study does not establish how every aspect of a live expert–AI interaction will behave. It evaluates assessments against an expert reference; it does not measure the effects of repeated use on a working team's attention, review habits or decisions. Those are further empirical questions. The distinction matters when moving from a promising analytical method to an operational workflow.
My interest is in making that transition inspectable. We should be able to explain what the AI contributes, what the expert contributes, where either can correct the other and what evidence supports the resulting judgment. Those explanations provide a basis for testing the process again when its model, task or human review arrangements change.