IEEE Spectrum breed · Technologie

AI Efficiency Could Cost Us the Next Generation of Experts

The increasing adoption of AI in the workplace, particularly generative AI, is leading to a decline in junior employment, potentially hindering the development of future experts. This trend mirrors historical challenges with automation in safety-critical fields like aviation and nuclear power, where deliberate manual steps were retained to maintain human proficiency.

A decade ago, a nuclear plant control system was designed with deliberate manual steps, even though automation could have handled them. This was to ensure operators remained proficient, as automation only fails when confused or in trouble, leaving inexperienced humans to manage critical situations.

Studies show that after companies adopt generative AI, junior employment falls while senior employment grows. This suggests AI automates tasks that previously served as training grounds for junior roles. While some attribute this to remote work hindering mentorship, both scenarios point to a broken apprenticeship channel.

Expertise is not downloaded; it's earned through hands-on experience, including failures and debugging. Over-reliance on AI spares newcomers these formative experiences, resulting in a cohort that can supervise models but lacks the intuition to identify when an AI is critically wrong.

Aviation provides a cautionary tale. The automation paradox means increasingly capable automation reduces human practice, leaving operators only with the most difficult situations. The loss of Air France flight 447 was attributed to pilot competence failure, not hardware issues, as they couldn't recover from a stall after relying on autopilot for thousands of hours.

The aviation industry responded by reintroducing deliberate manual practice, recognizing skill decay as a hazard. Some airlines even encourage hand-flying during initial climb and descent, trading minor fuel efficiency for maintaining essential flying skills. This highlights that a perfectly optimized system producing incompetent operators is not optimized at all.

The concept of a "manual gate" in AI-augmented work involves intentionally incorporating human control points in workflows. These gates are not for speed or safety alone, but specifically to exercise and preserve skills that would otherwise atrophy. Organizations must decide which competencies are critical to maintain in humans for potential future failures.

For software teams using AI, a manual gate could be implemented around debugging. Junior engineers would be tasked with reproducing failures, tracing root causes, and writing automated tests with AI assistance turned off. Only after diagnosis would AI assist with fixes, allowing for comparison and surfacing disagreements before critical issues arise.

This approach reframes junior roles not as overhead to be eliminated but as essential training for future senior staff. Protecting these formative tasks, even if less efficient in the short term, is crucial for long-term capability. Dismantling this training apparatus mortgages future expertise.

Implementing deliberate manual gates is less efficient in the short term, a difficult sell in a market focused on quarterly results. Executives carrying "unnecessary" human roles replaced by AI face board scrutiny. This strategy is more feasible for founders with control, private companies, or institutions with long-term horizons, or those subject to regulatory requirements for skill demonstration.

Deliberate inefficiency, when viewed as insurance rather than waste, is key. In safety-critical engineering, this practice has always been intentional. As AI takes over expertise-forging tasks, the smart move is not to resist automation but to intentionally retain human control to ensure someone can fly when the automation inevitably fails.

AI-samenvatting op basis van de bron.