Does AI augmentation protect workers from skill erosion?

The abstract of "Unaccountable Delegation, Fading Skills" says its analysis "highlights four findings," and the excerpt carries only the first: "augmentation is not inherently safe because overreliance on agents can gradually erode workers' skills and oversight." The finding sits against a labeling scheme. The authors applied a structured prompt to 2,078 O*NET job tasks and produced 8,356 risk scenarios, each labeled by severity and by deployment mode, automation or augmentation. The claim is that the second label does not work as a safety label. A scenario where the agent supports a worker instead of replacing them can still produce risk, and the title's "fading skills" names that risk.

The mechanism, as far as the excerpt states it, is a chain with two links. Overreliance on the agent erodes skills, and the same erosion reaches oversight. The word "gradually" matters: the harm is an accumulation, not an event, so no single agent output is the failure. On the vault's reading, the two links close a loop. Augmentation is usually justified by the human staying in a position to check the agent, and a worker whose skills have faded is less able to do the checking. The excerpt does not spell this loop out; it names skills and oversight together and leaves the connection to the reader.

This qualifies rather than contradicts Does AI risk increase with the autonomy we give it?. That note ties risk to ceded autonomy and recommends a governed spectrum. The workplace paper adds that the low-autonomy end of the spectrum has a risk of its own, one that changes form from a single failure with wide consequences to a slow loss of human capacity. The excerpt does not compare the size of augmentation risk with automation risk, so monotonicity is not challenged. The two-mode label is also coarser than the Does machine agency exist on a spectrum rather than binary? taxonomy, though the excerpt does not say how modes were assigned. A loosely convergent concern appears in Which AI risks are already harming individual users today?, where an expert survey scored autonomy erosion as already occurring. That survey used a different method and a different setting, so the agreement is thematic, not evidentiary.

What the excerpt does not establish. It reports no measurement of actual skill change in any worker, no share of augmentation scenarios that involve erosion, and no timescale for "gradually." The validation it describes, 45 workers across 10 job roles plus an independent LLM judge, tested whether the scenarios were plausible and matched their job tasks. It did not test whether erosion happens. The other three findings are cut off. At the strength the evidence allows, this is a plausibility claim about workplace scenarios: augmentation deserves its own risk assessment instead of being treated as the safe fallback.