Sociotechnical systems in software and AI: The deskilling question

Are we quietly losing expertise as AI tools take over, or is this the familiar rhythm of tool-assisted work? A sociotechnical view helps distinguish harmless automation from dangerous capacity-deskilling and points to practical steps teams can take to preserve human judgement and craft.

Introduction — why deskilling feels different this time

When a new tool arrives, a familiar script plays out: efficiency increases, some tasks disappear, and people worry about lost jobs and eroded craft. The current wave of AI feels larger and faster, and that accelerates concerns about deskilling. But not all deskilling is the same. Recent research urges us to separate harmless task obsolescence from what scholars call capacity-deskilling — a loss of underlying human abilities we later find hard to recover.

Two kinds of deskilling

Think of deskilling in two buckets. The first is activity-deskilling: tools take over repetitive work and free humans for higher-value work. That's been happening for centuries — compilers, spreadsheets and automated tests all replaced some grunt work but created new kinds of expertise.

The second, and more worrying, is capacity-deskilling: when our environment makes it harder to cultivate or maintain core human capacities such as judgement, pattern recognition, or diagnostic skill. A recent normative account argues that while activity-deskilling can be benign or even positive, capacity-deskilling warrants deliberate concern because it erodes personhood and long-term resilience (Springer Nature).

Evidence that tools change the way we think and perform

Empirical studies show the pattern is not just hypothetical. For instance, a multinational medical study found gastroenterologists detected fewer adenomas after AI assistance was removed — a drop from 28% to 22% in detection rate. The implication is stark: the tool improved momentary performance but long-term dependence weakened human ability when the tool wasn't available (Erigo summarising The Lancet Gastroenterology & Hepatology).

Other work from academic labs points to subtler effects: language models can shift framings and prioritisation in scientific communication, so repeated use changes how people select and interpret evidence. These shifts are quiet, cumulative and often invisible until you test for them (Erigo; MIT studies referenced there).

Why sociotechnical thinking matters

A socio-technical lens says we should design not only the algorithm, but the environment around it. That means policy, training, workflow, monitoring and culture all matter. For AI to be responsible in practice we need holistic evaluation — technical accuracy is necessary but not sufficient (NIST PDF). Engineering systems approaches, cross-disciplinary curricula and socio-technical governance are all part of the response.

What this looks like in software engineering

Software teams already live in the land of tool-assisted work. Compilers and IDEs changed what junior engineers learn; modern code assistants accelerate that trend. Some practitioners argue the change is evolutionary — like the shift after compilers — and that developers who adapt will thrive (LinkedIn commentary).

That adaptation is real, but the sociotechnical point is: how you adopt matters. A team that uses AI as an unforgiving autopilot will lose diagnostic muscle; a team that uses it as a tutor, checker and laboratory will preserve and extend expertise. Practical behaviours like asking ‘‘How would I test this without the tool?’’ or prompts that encourage DRY/KISS thinking nudge the balance toward skill retention (practices noted in practitioner discussions online).

What the labour-economics view adds

Economists and think tanks remind us of one-way fungibility: experts displaced into routine roles can still do non-expert work, but the reverse is hard. If expertise becomes stranded in ever-more-automated workflows, the pipeline of future experts dries up. That makes the long-term effects of automation uncertain and risky (Stanford HAI; Autor's framework).

Practical guidance for practitioners and managers

If we accept that some forms of deskilling are dangerous, what do we do? Here are grounded, pragmatic steps teams can take to keep capacity alive while reaping the advantages of AI.

Measure skill retention, not just throughput. Run periodic blind assessments where practitioners complete tasks without AI assistance to detect degradation early — the medical study shows why this is useful (Erigo).
Keep humans ‘in the loop’ as learners, not just monitors. Design workflows so the human makes the final decision and explains it back. Encourage human-generated candidate solutions before showing AI suggestions, which reduces automation bias.
Rotate tasks and roles. Ensure engineers and analysts spend scheduled time doing core diagnostic and exploratory work without assistance to keep mental models sharp.
Build explainability into daily practice. Require short rationales for accepting or rejecting AI outputs. That practice both reduces automation bias and documents tacit knowledge.
Design sandboxes for ‘skill practice’. Create training environments where people deliberately practise without AI, and others where they collaborate with AI — both scenarios provide contrastive learning.
Foster a socio-technical governance loop. Use multidisciplinary reviews — engineers, domain experts, ethicists, operations — to assess whether systems cultivate or erode human capacities (NIST socio-technical approach).
Encourage meta-skills: prompt design, verification and tool hygiene. Practitioners who learn to craft good prompts, validate outputs and refactor AI-generated artefacts are less likely to become passive consumers of automation (practitioner accounts such as LinkedIn threads).
Plan for fallbacks and maintain expertise pathways. Ensure career and training pathways exist so expertise can be passed on. Consider apprenticeship, pair programming and documentation rituals to keep tacit knowledge alive (Stanford HAI concerns about one-way fungibility).

Organisational culture and incentives

Tools reflect what organisations reward. If metrics favour short-term throughput over learning and quality, teams will optimise for speed and erode harder-to-measure capacities. Counter that by embedding learning objectives into performance assessment and investing in cross-disciplinary training (NIST; socio-technical governance).

Conclusion — deliberate stewardship beats nostalgia

Is AI deskilling us in ways we will regret? The evidence says: sometimes. There are documented cases where assistance improved immediate performance but reduced human capacity when the tool was absent. But history also shows tools can amplify human potential when we build the right socio-technical environment around them.

The sensible middle path isn’t technophobia or blind adoption. It is deliberate stewardship: design systems that treat humans as capable partners, measure and preserve core skills, and make room for learning. Do that, and this wave of tool-assisted work becomes another chapter in skill evolution rather than a wholesale loss of expertise.