IBM released a global workforce study Sept. 21 finding that 60% of surveyed employees worry artificial intelligence is eroding their skills, while 46% of organizations do not involve their chief human resources officer when AI strategy is defined. The numbers frame a practical management problem: companies are deploying systems that change jobs faster than they are deciding how employees will preserve judgment, challenge outputs and own the work.
Analysis. The IBM Institute for Business Value and Oxford Economics surveyed 1,500 HR leaders or equivalent executives and 8,800 full-time employees between April and June 2026. Their responses describe perceptions and organizational practices, not a controlled measure of skill loss. The distinction is essential. Workers’ concerns are real evidence about their experience, but the survey alone cannot prove that AI caused a decline in critical thinking.
The accountability gap sits inside the workflow
IBM’s release says 80% of HR leaders believe AI creates invisible work, including checking recommendations, correcting mistakes, supplying context and handling exceptions. Forty-two percent of employees say AI increases their work or that the additional work goes unrecognized. Another 43% say blame falls on them when something goes wrong with AI. Taken together, those answers suggest that the apparent productivity of an automated step may depend on human labor that never appears in the metric.
The study also reports that 36% of HR leaders see unclear accountability as a complication in AI deployment. Only 28% say they have a joint roadmap with IT supported by a shared operating cadence. Those are governance findings, not evidence that any particular organization’s deployment is unsafe. They point to a design question every employer can answer: Who has authority to approve, contest or override an AI recommendation, and how is that work recognized?
IBM argues that organizations should distinguish human-led, AI-assisted and AI-executed activities. That classification is useful only if it changes the operating procedure. A label in a policy document does not tell an employee which evidence to inspect, when to escalate an uncertain output or whether rejecting a model’s suggestion will be treated as sound judgment rather than resistance to adoption.
Other evidence supports the concern, with limits
Microsoft’s 2026 Work Trend Index, a separate survey of AI users in 10 markets, found 50% named quality control of AI output and 46% named critical thinking as human skills that become more important as AI takes on work. Eighty-six percent said they treat AI output as a starting point, not a final answer. That does not validate IBM’s percentages; the populations and questions differ. It does reinforce the broader observation that the worker’s role is moving toward evaluation and responsibility.
A 2025 study by Microsoft Research and Carnegie Mellon University surveyed 319 knowledge workers about 936 examples of AI-assisted work. It found that greater confidence in the AI was associated with less reported critical-thinking effort, while greater confidence in one’s own ability was associated with more. The authors described a shift toward verification, integration and stewardship. Association and self-report do not establish that AI permanently damages cognition. They do show why training people to accept fluent output without checking it would be a poor substitute for expertise.
The OECD’s 2026 review of AI and skills adds a wider labor-market perspective. It says most workers do not need advanced model-building skills; they need digital fluency, the ability to analyze information and human capabilities such as problem-solving. The OECD also finds that workers who receive training are more likely to report positive outcomes from AI use. Training matters, but so do transparency and accountability in the workplace.
What leaders should measure next
The strongest implication of IBM’s findings is not to ban AI or to add another generic course. It is to measure the work around the tool: time spent validating outputs, the frequency and cost of corrections, whether employees can override recommendations, and whether less experienced staff still get opportunities to learn the underlying task. These are proposed management measures inferred from the research, not results IBM says it tested.
HR belongs in that discussion because job design, evaluation and development are its remit. Yet IBM’s own survey says 72% of organizations make limited or no use of AI inside the HR function. The company sells AI products and HR consulting, so its recommendations carry a commercial interest and should be weighed alongside independent evidence. The survey’s clearest value is narrower: it exposes a gap between the speed of technical adoption and the care given to human capability.
For executives, the decision is less about whether to ask employees to use AI and more about what the organization wants them to remain able to do without it. That answer should determine where human review is mandatory, where practice is protected and where the productivity gain is truly coming from.