There may be another important direction for future AI development: evaluate not only the quality of a model’s immediate answer, but the quality of the human work and thinking that follows it.
Early signals of this shift are already appearing in education. Recent AI-learning initiatives are beginning to evaluate whether an AI intervention leads to a better subsequent learner action, revision, or outcome, rather than measuring the AI response in isolation.
See, for example, the Digital Promise K-12 AI Infrastructure Program and its work on AI-enabled formative assessment and evaluation of AI coaching:
We think this idea can be extended much further.
Instead of evaluating only:
PROMPT → MODEL RESPONSE
future AI systems could also study:
HUMAN QUESTION
→ MODEL INTERVENTION
→ HUMAN NEXT MOVE
→ NEXT QUESTION
→ IDEA / DECISION DEVELOPMENT
→ PROJECT CHANGE
→ LONGER-TERM OUTCOME
The important question becomes:
What became possible because this intervention happened?
A model response may be correct and useful in the moment but have little downstream value.
Another response may introduce one conceptual distinction that the user later modifies, combines with other knowledge, tests, generalises, and eventually turns into a new method, product, research direction, or professional decision.
That difference is invisible if evaluation stops at the end of the model response.
A small real example: from prosody to a new design route
We have already observed this kind of trajectory in our own work.
An AI conversation initially introduced two ideas: prosody and higher-order mechanics.
Neither idea contained the eventual solution.
The human follow-up was to ask a different kind of question: if research already shows that prosody carries meaning in live speech, why does so much of that knowledge remain largely inside research rather than becoming usable learning design?
That question opened a new path.
The work developed roughly like this:
PROSODY
+
HIGHER-ORDER MECHANICS
→ prosody as a meaning-bearing signal
→ language as a broader system of signals
→ signals as evidence for interpretation
→ relationships between evidence: aligned, partial, ambiguous, conflicting
→ medium ≠ function
→ audio is one possible realization of prosodic work, not the construct itself
→ prosody tasks that can begin without audio
→ a broader learning goal:
help learners hear the music in live speech.
The final idea was not present in the original model response.
It emerged through repeated human questioning, reinterpretation, combination with research, methodological design, and subsequent development.
This is exactly the kind of downstream value that conventional response-level evaluation does not capture.
The AI intervention mattered.
But so did the quality of the human questions that followed it.
Preserve question lineage, not only answer history
This suggests preserving question lineage, not only answer history.
The quality of the original human question determines the search space, the evidence that becomes relevant, the relationships that can be noticed, and therefore the foundation on which subsequent reasoning is built.
A professional AI workspace could therefore preserve:
QUESTION
→ EVIDENCE RETRIEVED
→ MODEL INTERVENTION
→ HUMAN RESPONSE
→ FOLLOW-UP QUESTION
→ DECISION
→ EXPERIMENT / IMPLEMENTATION
→ DOWNSTREAM DEVELOPMENT
Over time, this could reveal which kinds of questions and model interventions repeatedly generate productive work.
It could also reveal something more subtle: some interventions may appear modest at the time but later become highly generative, while impressive-looking answers may produce almost no durable downstream value.
Model capability must remain an explicit variable
Model capability must also remain an explicit variable.
The same question may produce very different downstream trajectories depending on:
QUESTION QUALITY
× MODEL CAPABILITY
× REASONING EFFORT
× AVAILABLE CONTEXT
× HUMAN EXPERTISE
This matters both for evaluation and for future model training.
A weaker outcome should not automatically be attributed to a weak question if the model was unable to explore its potential.
Nor should a strong downstream result automatically be attributed to the model when much of the intellectual development came from the human.
Longitudinal interaction data could eventually support a different kind of training and evaluation dataset:
QUESTION
→ MODEL INTERVENTION
→ HUMAN REACTION
→ NEXT QUESTION
→ REVISION
→ EXPERIMENT
→ PROJECT OUTCOME
→ LATER REFLECTION
This would allow AI companies to investigate something that conventional benchmarks largely miss:
the generative impact of a model on human thinking and work over time.
The next generation of AI should not be evaluated only by the quality of its answers, but also by the quality of the work and thinking that follows them.