Begin with the purpose

A faster answer, a more accurate decision and a deeper understanding are different aims. Before describing a partnership as successful, identify the task and the outcome that matter. This is Chitta’s proposed approach to examining collaboration.

Choose the comparison carefully

In the published 2024 review, the combined system performed better than humans alone on average, but worse than the better standalone performer. These are different comparisons: helping a person does not necessarily mean exceeding the stronger of the person or the AI.

The included experiments were published from January 2020 through June 2023. Results varied considerably across tasks and settings. The authors reported potential publication bias for the comparison with humans alone. The selected tasks and participant groups may not represent practical deployments. This is a task-performance synthesis, not a finding about lasting learning or every current AI system.

Performance is not the whole learning question

A school mathematics field experiment distinguishes assisted practice from later unaided performance. Its findings make tutor design and the assessment setting important to the discussion. The study concerns a particular population, subject and implementation. Its guarded tutor largely mitigated the negative unaided-exam effect, but did not show a positive exam effect relative to the control group. These outcomes do not establish a universal claim about AI and human learning.

Make the contribution inspectable

When presenting a case, record who did what, which system was used, when it was used, what was compared and how success was assessed. Confidence in the wording of an answer is not a substitute for checking its sources.

Reflect before relying

Choose one claim in an assisted answer. What source could check it? Which part remains uncertain? This prompt invites examination; Chitta is not collecting private conversations in this preview.

Follow the inquiry

Define success before comparing

These are proposed questions for reading a result. Then ask a separate question: what can the person do after assistance ends? An improved joint output does not itself measure independent understanding.