Use It or Lose It: Keep Your Team Sharp While AI Does the Work

If people stop practicing the judgment AI is helping with, that judgment fades. Design for it.

The pitch for AI in most businesses is "let the machine handle it." For a lot of tasks that's the right call. There's a cost nobody puts on the slide, though: the skill you stop using is the skill you start losing. One day the AI gets something wrong, and the person who's supposed to catch it can't.

A 2026 research paper puts it plainly:

"Humans without opportunity to practice eventually lose the skills needed to operate on their own" (p. 1, citing Kulveit et al., 2025a)

The paper is Solipsistic Superintelligence is Unlikely to be Cooperative, presented at ICML 2026. Its five authors include researchers affiliated with Google DeepMind. It's an academic paper, not an official DeepMind position. Its main subject is cooperation among AI systems, and skill erosion is one strand of that argument. For operators, though, I think it's the most useful strand.

A thought experiment from radiology

The authors illustrate the problem with a hypothetical. It's a scenario, not a study. Imagine AI diagnostic systems become standard in radiology. Junior doctors train alongside the AI, and senior doctors increasingly lean on it. Over time:

"seniors start experiencing degradation of unassisted skills resulting from disuse and catch fewer errors the AI also misses." (p. 1)

Then the loop tightens:

"The feedback loop closes with physicians confirming AI suggestions and the AI learning from these confirmations. This leads to gradual atrophy in the capacity for independent human judgment and results in the narrowing of diagnostic diversity." (p. 1)

Swap "radiologist" for "account manager reviewing AI-drafted proposals" or "bookkeeper approving AI-categorized transactions" and the shape is the same. The errors that get through are the ones the human and the AI both miss.

The evidence the paper points to

For real-world evidence, the paper cites other researchers' work rather than its own experiments. The best-known example is navigation:

"Longitudinal studies document degradation of spatial navigation skills among habitual GPS users. (Dahmani & Bohbot, 2020) showed that GPS users demonstrated weaker cognitive map formation compared to those who navigated without assistance." (p. 20)

That doesn't mean GPS "ruins your memory". It's a specific finding about cognitive maps from one cited study. The mechanism is what matters here:

"The effect compounds: as navigation skills atrophy, users become more dependent on GPS, further reducing opportunities for unassisted navigation." (p. 20)

The paper sees the same pattern with spell-checkers: "when a system reliably catches errors, the cognitive investment in avoiding errors diminishes." (p. 20)

More generally, the authors write that humans "systematically alter their behavior in response to AI systems, often in ways that... erode the human capacities the system was designed to augment." (p. 19)

"Human in the loop" only works if the human can still do the job

This is the line I'd put on the wall of any team rolling out AI:

"Keeping humans in the loop requires that they retain the skills, information, and cognitive engagement needed to exercise meaningful authority, rather than only the formal authority to intervene" (p. 8, citing Parasuraman & Riley, 1997)

A lot of AI governance stops at the formal step: "a human approves it before it ships." If that human has clicked "approve" 400 times in a row without having to think, the approval is a formality. The paper draws the same contrast. Systems that "present options and defer to human judgment" preserve the human role, while systems that "optimize end-to-end risk making human participation nominal" (p. 8).

The authors also recommend that the effect of AI "on human skills, autonomy, and meaningful choice" be part of how deployments are evaluated, "rather than as a separate ethical concern" (p. 8). For a business, I'd read that as: when you measure an AI rollout, measure what's happening to your people's skills, not just the time saved.

What I'd do about it

None of this is a reason to avoid AI. The point is to build workflows where people keep practicing the judgment the AI is supporting. The paper's phrase for the goal is "preserving human agency as a structural feature of the systems we build." (p. 9) In operator terms, the skill practice has to be designed into the workflow. Hoping people stay sharp on their own won't work.

Practical takeaways

1. Schedule "manual reps". For the judgment calls that matter most, such as pricing, credit decisions or key client deliverables, have people regularly do some of the work without the AI and then compare their answer with the AI's.
2. Make the AI show options, not just answers. Where you can, set tools up to give alternatives and their reasoning so the reviewer has to choose, not just confirm.
3. Audit your approvers, not just your AI. Check approval logs occasionally. If something is approved almost every time with almost no review time, that oversight may be nominal.
4. Track skills alongside time saved. When you review an AI rollout, ask whether the team could still do the task if the tool went down tomorrow.
5. Protect how juniors learn. New hires who've only ever worked with the AI may never build the baseline skill. Give them deliberate, unassisted practice early on.

Sources

• Trivedi, R. S., Jaques, N., Cross, L., Vezhnevets, A. S., & Leibo, J. Z. Solipsistic Superintelligence is Unlikely to be Cooperative. Proceedings of the 43rd ICML (PMLR 306), 2026. arXiv:2606.03237. https://arxiv.org/abs/2606.03237. Cited pages 1, 8, 9, 19, 20.
• Studies cited within the paper: Dahmani & Bohbot (2020), Scientific Reports 10:6310 (GPS); Galletta et al. (2005) (spell-checkers); Kulveit et al. (2025a); Parasuraman & Riley (1997); Santoni de Sio & van den Hoven (2018).