Moving research meetings from data updates to cognitive windows in the age of artificial intelligence

Synthesis: When a generative system can produce the analysis, the polished product stops being evidence that a trainee understands the reasoning behind it. Prunuske and Adams-Phillips propose that laboratory research meetings act as "cognitive windows" that make reasoning visible across three dimensions: metacognitive calibration, conceptual explanation, and collaborative sensemaking. Through an illustrative case, a confident AI summary that never saw the incubator's temperature logs and a technician who stayed silent about it, they argue that what needs restructuring is the meeting, not the deliverable.

Perspective piece: the authors propose a framework for research meetings and illustrate it with a hypothetical case. No data were collected or analyzed, so claims here are arguments, not findings.

Key Findings

1. The AI product cannot stand in for the reasoning. Generative systems produce language probabilistically rather than from grounded understanding, so they can emit explanations that are fluent, plausible, and unsupported by the evidence in hand. That shifts the central apprenticeship problem from finding information to judging the validity of AI-generated reasoning, and it means mentors have to ask not only what a trainee produced but whether the trainee can justify it. The authors add a second reason the product is unreliable as evidence: people, faculty included, do not reliably separate AI-generated from human-authored work when the output is polished.
2. Three windows, each with its own question. The metacognitive window asks whether the mentee recognizes what they know, what they assume, and what they still need to learn. The conceptual window asks whether the mentee is reasoning from a coherent mental model supported by evidence. The collaborative window asks whether understanding is being built collectively in a climate that permits uncertainty, critique, and revision. The authors present the three as complementary and deliberately concise rather than exhaustive, chosen so a mentor can actually use them in a meeting.
3. Cognitive apprenticeship is the lens, and AI thins it. Modeling, coaching, Scaffolding, articulation, reflection, and exploration are the methods by which expert thinking becomes visible and a learner moves toward independent practice. As generative AI mediates analysis and explanation, the authors argue that articulation, reflection, and exploration in particular may appear less often in a trainee's independent work, which is what makes the live meeting the place where they have to be deliberately re-enacted.
4. The prompts are the intervention. The paper supplies diagnostic questions per window rather than advice to "ask more". For metacognition: what do you feel most confident about here and what remains uncertain, what assumptions might be shaping your conclusion, how did you judge whether the AI output was reliable for this context. For conceptual reasoning: what mechanism explains this result, can you draw the system as a model or workflow, how would your interpretation change if a key assumption were wrong. For collaborative reasoning: what alternative explanations do others see, where does the group actually disagree, whose expertise is missing from the discussion.
5. Uneven development across windows is diagnostically useful, not a failure. A trainee can reason well conceptually while calibrating poorly, which produces overconfidence in a flawed interpretation. Another can participate actively in discussion without conceptual depth. Because the windows separate these, a mentor can aim questions and Feedback at the specific dimension that is weak instead of treating "understanding" as one quantity.
6. In the case, authority substitutes for scrutiny. Sam's AI summary attributes inconsistent assay results to reagent lot variability. The explanation is plausible and incomplete: the tool never saw the incubator's time-stamped temperature logs, which sat in a separate monitoring system. Because the summary looked confident and objective, nobody asked what information it lacked, and the perceived objectivity made it harder for Lena, a technician with less positional authority, to raise the incubator hypothesis she had. A grant deadline and a request for quick updates closed the moment.
7. Restructured, the same moment becomes assessable. The authors replay the meeting with one question per window: what data went into the analysis and what was missing, sketch a model that incorporates both biomarkers and predicts the effect of external changes, and ask the room directly whether others see alternative explanations, which is collaborative sensemaking in practice. The last prompt does double duty, because Sam's answer tells the mentor whether Sam can explain his own results, and it surfaces the control experiment that would otherwise have gone unmentioned.
8. The case doubles as a faculty development artefact. A facilitator guide turns the scenario into a 45 to 60 minute session: individual reading, small-group discussion, whole-group debrief, and an optional application step in which participants commit to one or two changes and one specific question for their own meetings. Its guidance is to keep discussion off AI accuracy and on how reasoning gets evaluated, and to surface the tension between an efficient meeting and effective mentoring.

What this means for practice

• Ask for the reasoning chain before the result. Have the trainee state which data the analysis used, which data existed but were not available to the tool, and what they would verify next. A single question about the tool's inputs exposes whether the summary was understood or accepted.
• Demand a second representation. A sketch of the mechanism, a workflow diagram, or an annotation of the figure reveals conceptual coherence in a way that prose repeating the AI output cannot, and it makes gaps available for revision.
• Name the epistemic status of AI output, and set that norm in advance. Calling a generated analysis a hypothesis to be tested rather than an authority to be deferred to is what gives a lower-authority team member permission to challenge it, and that person is often the one holding the missing explanation. Decide in advance how AI-assisted analyses will be reported and interrogated, so the norm exists before anyone is under time pressure.
• Budget agenda time for reasoning. The failure in the case is partly structural: a deadline-compressed round of quick updates leaves no room for the questions that would have caught the gap. If reasoning is worth assessing, it needs protected minutes.
• Use the case and facilitator guide as a mentoring training session. It requires no data collection and targets mentors rather than trainees, which makes it usable in departmental or center-level professional development.

Limitations

• This is a perspective piece with no data. It does not report whether the framework changes what mentors notice, whether trainees' scientific judgment improves, or whether either holds up under real meeting loads.
• The three windows were derived from patterns the authors observed in mentoring interactions and then aligned with learning theory. They were not validated as a measurement instrument, and the authors describe them as concise rather than exhaustive.
• The case is an illustrative composite rather than an empirical case study with collected data, so the analysis demonstrates what the framework would reveal, not what it did reveal.
• The claim that AI mediation reduces articulation, reflection, and exploration in trainees' independent work is argued from cognitive apprenticeship and from adjacent research on illusions of competence, offloading, and AI-text detection. It is not measured in this paper.
• Scope is one-to-one and small-group mentoring in a laboratory. The paper does not address large labs, remote or multi-institution teams, or coursework settings where a mentor sees a trainee rarely.
• No outcome evidence accompanies the facilitator guide: nothing here reports changed mentor behavior or trainee learning after the session.

Citation

Prunuske, A. J., & Adams-Phillips, L. C. (2026). Moving research meetings from data updates to cognitive windows in the age of artificial intelligence. CBE-Life Sciences Education, 25(4), pe1.