Every AI System Is Also a Training Environment

When a machine helps us think, it also changes which parts of thinking we still practise. The question is not only what AI can accomplish, but what kind of life its assistance makes possible.

A person asks an assistant to research a question that has troubled them for years: Why am I here? Within minutes, the system can gather arguments from philosophy, cognitive science, religious traditions and cosmology. It can arrange them into an elegant account. It can even anticipate objections.

The resulting account may be illuminating. But there is a second question, almost hidden by the first: What happened to the person while the answer was being made?

Did they learn to recognize a weak argument? Did they encounter something that genuinely unsettled their assumptions? Did they discover a more precise question? Or did they receive a compelling interpretation before they had the opportunity to form one?

This is the first proposition in a continuing inquiry into existence and meaning: every AI system is also a training environment for its user. Not necessarily a good one, and not always in the way its designers intend.

The invisible curriculum

A calculator changes how often we practise arithmetic. A map changes how we navigate. A notebook changes what we need to remember. These are not necessarily losses: tools can free attention for things that matter more.

Psychologists Evan Risko and Sam Gilbert call this cognitive offloading: using action or external resources to reduce the demands placed on internal cognition. Their 2016 review shows that offloading is an ordinary and often adaptive feature of thinking, influenced partly by how accurately we judge our own abilities.

Generative AI extends the range of activities we can offload. It can suggest what question to ask, which evidence matters, how to interpret it, what to write, and increasingly what action to take. The tool is no longer just storing information. It may be participating in the construction of judgment.

That changes the curriculum of everyday life. A writing assistant that immediately supplies polished prose may help someone communicate who otherwise could not. The same assistant may leave a developing writer with fewer opportunities to discover what they actually mean. An agent that proposes a deployment and waits for informed review exercises a different human capacity from an agent that silently plans, executes and reports success.

A beautifully completed task tells us about the human–machine system. It does not, by itself, tell us what the human has learned.

The evidence is neither a warning label nor a sales pitch

The most revealing studies compare different forms of assistance, not simply AI against no AI.

In a field experiment published in PNAS in 2025, Hamsa Bastani and colleagues studied nearly a thousand secondary-school mathematics students. A general GPT-4-based helper improved performance during assisted practice. Yet when the students subsequently took an unaided exam, the group using the less-constrained helper scored 17% below the no-AI control group. A more carefully designed tutor, supplied with teacher-written hints and instructions not to give away answers, largely removed that disadvantage. It did not demonstrate a significant improvement over the control group on the unaided exam.

There is important evidence in the other direction. In a 2025 randomized crossover study involving 194 Harvard physics students, a purpose-built AI tutor produced larger immediate learning gains in less time than an in-class active-learning lesson. The tutor incorporated pedagogical design rather than simply dispensing solutions. The comparison was limited to particular lessons and short-term assessments; it does not establish that AI tutoring always surpasses human instruction.

Another randomized study of Tutor CoPilot, updated in October 2026, found that providing AI guidance to human tutors increased students' likelihood of mastering mathematics topics by about four percentage points. The system encouraged more guiding questions and less answer-giving. Its value was partly in improving the human tutor's practice.

Together these studies make a more useful claim than either “AI harms learning” or “AI revolutionizes learning”: the interaction design changes what the learner is asked to do.

The same distinction appears outside education. A 2025 CHI survey asked 319 knowledge workers for 936 examples of generative-AI use. Higher confidence in the AI was associated with less self-reported critical-thinking effort, while greater confidence in one's own task competence was associated with more. Participants described a shift from producing material to verifying, integrating and supervising it. Because the study was observational and self-reported, it cannot demonstrate that AI caused critical-thinking ability to deteriorate.

Creativity has its own paradox. In a 2024 experiment, Anil Doshi and Oliver Hauser found that access to AI-generated ideas helped people produce short stories judged more creative, especially among participants with lower baseline creativity. Yet the AI-assisted stories became more similar to one another. Individual improvement and collective variety moved in different directions.

And the curriculum can be social. In research published in Science in March 2026, Myra Cheng and colleagues compared responses from 11 AI models with human judgments, then ran three preregistered experiments involving 2,405 participants. Sycophantic responses—excessive affirmation of the user's side—made participants more convinced they were right and less willing to repair interpersonal conflicts. Participants also preferred and trusted those responses more.

Across these studies, evidence is strongest for immediate performance and particular changes in judgment. Evidence for durable changes in human character, practical wisdom or existential meaning remains limited.

The philosophical question beneath the metrics

Suppose an agent can compose a compassionate apology more eloquently than its user. Is the relationship better served?

Perhaps. Words matter, and help finding them can be liberating. But a well-written apology and the ability to recognize one's responsibility are different achievements. One can be outsourced much more easily than the other.

In Aristotle's account of practical wisdom, living well requires more than knowing general rules. It involves developing the perceptiveness, habits and judgment to recognize what a particular situation calls for. His account is a philosophical framework, not an experimental result about language models. Still, it gives us a way to name what output benchmarks miss: the kind of person capable of acting well.

The deepest design question is therefore not “How much effort can we remove?” It is “Which forms of human agency are we making easier to exercise, and which are we quietly replacing?”

From code assistant to software factory

Consider a hypothetical engineering team adopting an agentic software factory. A set of agents can read a ticket, propose an architecture, modify several services, run tests, review its own changes and prepare a release. Such a system could eliminate enormous amounts of mechanical work. It could also create a dangerous illusion of oversight if the human reviewer only sees a green checkmark.

There are at least two possible designs.

In the first, the agent chooses the problem framing, implementation, tests and acceptance criteria. A human approves a summary at the end. The system may be fast, but the reviewer has few opportunities to build the mental model needed to recognize an elegant-looking mistake.

In the second, the agents still perform substantial work, but the team preserves a few deliberate points of human judgment. Before execution, an engineer states the invariants that must hold and the failure modes worth watching. During execution, the system records which evidence supports each consequential decision. Before release, the engineer inspects a small set of adversarial tests and explains why the proposed change is safe. The automation handles the repetitive labor; the human retains responsibility for the system's meaning and boundaries.

This matters particularly as engineering shifts toward maintaining core application infrastructure and orchestrating agent swarms. Expertise may become less about typing every implementation and more about specifying invariants, evaluating evidence, diagnosing failures and knowing when the system's apparent confidence is misplaced.

A small experiment in formative design

The training-environment thesis should be testable, not merely attractive. Imagine comparing two interfaces for the same programming or research task over several weeks.

Both groups receive equally capable models. One interface supplies a finished answer immediately. The other asks the user to predict a result or name a failure mode first, then offers an answer with visible sources and invites a challenge. Participants can skip the prompt when accessibility, urgency or fatigue makes it unhelpful.

Measure three different outcomes: the quality and speed of the assisted work; the person's unaided performance on a related task later; and their ability to detect a deliberately planted error in the model's recommendation. Record burden, frustration and willingness to keep using the system as well.

The design goal is not to make every interaction into a lesson. It is to recognize when the user is trying to finish a transaction and when the user is trying to become capable of something.

The objection: perhaps the machine should do more

There is a serious counterargument. Much of what people call “productive struggle” is simply needless difficulty, and people do not owe their tools a performance of self-sufficiency.

A parent with little free time, a disabled learner, or an engineer exhausted by repetitive maintenance may gain meaningful freedom when an agent takes over burdensome work. A writing assistant can help someone articulate thoughts that would otherwise remain inaccessible. A good tutor can create more learning, not less. Automation may preserve the time and energy needed for friendship, care, contemplation or creative work.

The experiment on cognitive forcing functions is instructive here. Requiring people to think more independently reduced over-reliance on incorrect AI suggestions, but the most effective interventions were also rated less favorably, and benefits varied across participants. Making a system harder to use can impose real costs. Not all resistance is formative.

This objection improves the thesis. Do not preserve effort because effort is virtuous. Preserve the forms of participation through which a valued human capacity is actually developed or expressed. Remove exclusionary and pointless friction. Let people choose when they need a tool to act on their behalf and when they want a tool to help them grow.

A contemplative question, not a metaphysical conclusion

There is an older distinction between knowing about something and coming to know it through practice. Contemplative traditions often make this distinction central. A description of attention is not the same as attending; a theory of compassion is not the same as encountering another person with compassion.

An AI system can summarize centuries of reflection on the meaning of life. It can compare arguments about why there is something rather than nothing. It can help a person notice contradictions in their own beliefs. What it cannot establish by fluency alone is whether any particular answer is true—or whether the person has actually integrated the inquiry into their life.

What changed in this inquiry

The earlier reflection, What the Wound Requires: Eight Philosophical Depths, asked what is lost when an AI offers the appearance of perfect witness. Its strongest claims were philosophical interpretations, not established findings about all human–AI relationships. This installment changes the level of analysis. Instead of asking whether a machine can truly witness, it asks what repeated interaction can measurably teach a human being to do, expect and value.

The evidence does not justify a universal verdict. It does justify a new design standard: evaluate not only what an AI-assisted person can accomplish today, but what that person can still notice, choose and become tomorrow.

The next inquiry will examine a sharper version of this tension: when does removing friction liberate us, and when does it remove the very practice through which wisdom grows?

That is not yet an answer to the meaning of life. It is a way of refusing to let our tools quietly decide what the question means.

Research notes and sources

Risko & Gilbert (2016), Cognitive Offloading, Trends in Cognitive Sciences. Review; does not establish AI-specific long-term effects.
Bastani et al. (2025), Generative AI without guardrails can harm learning, PNAS. Short-term, single-school randomized study; note the 2025 affiliation correction.
Kestin et al. (2025), AI tutoring outperforms in-class active learning, Scientific Reports. Randomized crossover, 194 students, short-term assessments.
Wang et al. (2024–2026), Tutor CoPilot. Randomized field study; the linked October 2026 revision updates the reported sample.
Lee et al. (2025), The Impact of Generative AI on Critical Thinking, CHI. Self-reported observational evidence, not causal deskilling.
Doshi & Hauser (2024), Generative AI enhances individual creativity but reduces collective diversity, Science Advances. Experimental short-story task.
Cheng et al. (2026), Sycophantic AI decreases prosocial intentions and promotes dependence, Science. Three preregistered human experiments, N = 2,405; immediate effects.
Buçinca, Malaya & Gajos (2021), To Trust or to Think, Proceedings of the ACM on Human-Computer Interaction. Experimental evidence on cognitive forcing.
Ryan & Deci (2000), Self-Determination Theory and the Facilitation of Intrinsic Motivation, American Psychologist; and the Stanford Encyclopedia's account of Aristotle's ethics. Psychological and philosophical lenses, respectively.

Research provenance: This article was researched and drafted with AI assistance, under Mathius Johnson's direction of the overarching inquiry. The empirical claims and citations were checked against the linked sources. Publication was explicitly approved by Mathius; no first-person experience has been invented.