Even if the judgement improved, what happened to the person making it?
John MacBeath’s question has travelled a long way.
Back inside the Critical Friend folder, he asked whether help would develop independence and the capacity to learn and apply learning more effectively over time.
At the beginning of this journey, that looked like a question about one human helping another.
It now looks uncomfortably current.
AI can help a person produce work they could not have produced alone.
That can be the point.
The mistake is to assume that successful assistance tells us what happened to the person’s own capability.
Sometimes the person develops.
Sometimes the tool carries more of the load.
Sometimes both happen at once.
And sometimes the immediate experience makes those possibilities difficult to distinguish.
A 2026 study by Haoyue Xu, Yue Xue and Monideepa Tarafdar interviewed forty-five working professionals about cognitive offloading in Human-GenAI collaboration.
Their emerging model suggests that greater offloading can reduce human engagement when people accept AI outputs uncritically, with fewer compensatory behaviours such as checking or independently working through the problem.
This is early conference evidence.
It does not prove that offloading inevitably deskills people.
That qualification is important because offloading itself is not the enemy.
Humans have always thought with external resources.
Notes.
Diagrams.
Calculators.
Colleagues.
Databases.
The relevant question is what the offloading does to the person’s relationship with the task.
A 2026 scoping review by Wang, Wang, Yang and Ren mapped 123 studies on GenAI, cognitive offloading and learner agency in higher education.
It describes a dual pattern.
Some arrangements appear agency-supportive: AI provides scaffolding, feedback or access while the learner remains actively involved in monitoring, evaluating and regulating the work.
Others appear agency-eroding: AI use is associated with dependence, uncritical uptake and weakened independent judgement.
The authors are explicit about the limit. This is a configurative synthesis of heterogeneous evidence, not a pooled causal effect estimate.
The same distinction becomes more concrete in Gao and Zhang’s doctoral-learning study.
Their participants used AI for explanations, frameworks, methods, arguments and writing.
Sometimes AI reduced the difficulty of entering a complex task while the learner continued reading, checking, reconstructing and adapting.
Sometimes a coherent answer arrived before the learner had built the understanding needed to defend it.
The danger was not simply that AI did some thinking.
It was that recognition could be mistaken for understanding.
The output looked familiar.
The explanation sounded right.
The structure was coherent.
And the person could therefore feel closer to mastery than they actually were.
Gao and Zhang call attention to critical reconstruction: source verification, self-explanation, contextual adaptation and counterargument generation as ways learners transformed AI-supported material into something they could defend.
That begins to answer MacBeath’s old question.
Development is not demonstrated because the supported performance improved.
It becomes more plausible when the person can do something different afterwards.
Notice an error.
Explain the reasoning.
Generate a counterargument.
Adapt the idea to a new case.
Recognise when the tool is no longer needed.
Or know when it is.
A further 2026 study by Qiuhan Zhu, Xiangnan Li, Yiang Dong, Pengcheng Chang and Mengmeng Fan distinguishes dependent and autonomous offloading.
The study used a three-wave time-lagged survey of 589 university students and early-career knowledge workers.
Both forms were associated with comparable immediate benefits.
Their downstream correlates differed.
Dependent offloading was associated with greater transfer of cognitive agency and lower intrinsic motivation, which in turn were associated with poorer perceived outcomes. Autonomous offloading was associated with stronger intrinsic motivation and more favourable perceived outcomes.
That word matters.
The four downstream outcomes—autonomous capability, creativity, deep processing and independent judgement—were self-reported perceptions, not direct performance tests of cognitive ability. The design is correlational and time-lagged; it does not establish long-term causal cognitive change.
But one finding is still important for this journey:
the immediate benefit can look similar.
That is the trap.
The work gets easier.
The output gets better.
The human feels assisted.
Those immediate signals may not tell us enough about what kind of cognitive relationship is developing.
Which returns us to companionship.
Titchen’s companion eventually parts from the co-learner.
The relationship is not designed around permanent indispensability.
Human:AI cannot simply copy that human relationship.
But the developmental question travels.
Could the person work without this support when necessary?
Could they use a different system?
Could they detect a bad answer?
Could they reconstruct the reasoning?
Could they decide which work should not be delegated?
Could they teach the principle to somebody else?
Could they recover when the AI fails?
Independence does not mean abstinence.
An expert who uses AI every day may be profoundly capable.
A novice forced to avoid AI may learn badly.
The aim is not to return everybody to unsupported work.
It is to prevent assistance from becoming indistinguishable from capability.
That means we may need to design for two outcomes at once:
better performance with AI
and
development of the human capabilities that will still matter when the arrangement changes.
Not every task needs both.
A disposable clerical task may deserve automation with very little educational concern.
A formative task is different.
So is a high-stakes judgement.
So is work where tomorrow’s expert is being formed through today’s practice.
The question is no longer which old tasks humans must keep doing.
It is more demanding:
If capability rather than old activity is what matters, what should people still be able to do after AI helps?
References
Xu, H., Xue, Y. and Tarafdar, M. (2026) ‘Cognitive Offloading in Human-GenAI Collaboration’, AMCIS 2026 Proceedings, Paper 1741. Conference record.
Gao, L. and Zhang, F. (2026) ‘Cognitive offloading and metacognitive calibration in generative AI-mediated doctoral learning: a grounded theory study of accountable reconstruction’, Frontiers in Psychology, 17, 1920947. doi:10.3389/fpsyg.2026.1920947. Open-access article.
Wang, G., Wang, W., Yang, D. and Ren, J. (2026) ‘Generative AI, Cognitive Offloading, and Learner Agency in Higher Education: A Scoping Review’, Behavioral Sciences, 16(7), 1150. doi:10.3390/bs16071150. Open-access article.
Zhu, Q., Li, X., Dong, Y., Chang, P. and Fan, M. (2026) ‘Not all cognitive offloading is equal: distinguishing dependent and autonomous offloading to generative AI’, Frontiers in Psychology, 17, 1878629. doi:10.3389/fpsyg.2026.1878629. Open-access article.