A 2026 study of 2,257 employees found that psychological safety determines whether people start using AI at work, and has almost no bearing on what happens after that.

When HR and learning teams invest in psychological safety to support AI adoption, the research backs that choice for the first step. A study published in February 2026 (arXiv:2602.23279) examined AI tool adoption across a global consulting firm and found that psychological safety reliably predicted whether employees began using AI tools, but had no significant relationship to how often they continued to use them afterward. For program designers in HR and leadership development, that distinction reshapes where effort should go and when.

What the Research Actually Found

The conventional wisdom in AI implementation programs treats psychological safety as a broad enabler of adoption and continued use. The 2026 arXiv study tests that assumption directly and finds it is only partially correct. A sample of 2,257 employees in a global consulting firm shows that psychological safety’s influence is real but bounded. It operates at the point of first engagement and falls away after that.

Using logistic and linear regression, the researchers tested whether psychological safety predicted both whether employees adopted AI tools and how intensively they used them afterward. On the adoption question, the relationship was significant and consistent. On the sustained-use question, measured by frequency and duration, the relationship disappeared entirely. No subgroup tested showed a different pattern, whether by experience level, seniority, or region.

The research team concluded that psychological safety functions as an antecedent to the decision to try, but not to subsequent usage intensity. That split has a structural explanation. Getting people to start and keeping them going are driven by different conditions entirely.

Why Adoption and Continued Use Are Different Problems

Before someone uses a new tool for the first time, the primary barrier is interpersonal risk. People hesitate because of fear of looking incompetent, concern that visible mistakes will count against them, and the sense that experimenting falls outside expected role boundaries. Psychological safety reduces that risk directly. When the interpersonal cost of trying and failing is low, people are willing to experiment.

After the first attempt, the barrier shifts. Continued use depends on whether the tool produces enough value to displace existing habits and whether workflows have been restructured to accommodate it. It also depends on whether the person has developed enough skill to get genuine benefit and whether the work environment reinforces use over time. These are capability, habit, and systems conditions, and they respond to a different set of interventions.

Programs that treat the two stages as a single problem tend to invest in conditions that produce a strong first wave of adoption. A plateau follows, and the program has no mechanism to address it, because the design was built for a different barrier. The adoption gap and the sustained-use gap each require a specific answer.

The Implications for HR Program Design

For HR teams building AI adoption programs, the research offers a sequencing principle. Psychological safety investments return the most value at the start, when the primary barrier is interpersonal risk. Leader modeling of experimentation, explicit permission to fail, and visible organizational endorsement of learning over performance all work at this stage, clearing the first threshold.

Once adoption begins, the program design question shifts to what sustains use. Three conditions consistently underpin continued use that HR programs underweight. The first is extended skill development, because a single onboarding module does not build the fluency needed for AI to feel better than established methods. The second is workflow integration, since individual adoption rarely holds without restructured processes that create regular opportunities for use.

A third condition is manager behavior during the post-adoption phase, which the next section addresses in detail. The practical implication for program design is to measure adoption and sustained use as separate outcomes. Each outcome points to different program elements and different management behaviors to develop.

The Implications for Leadership Program Design

Managers play a different role at each stage of AI adoption, and leadership programs built around psychological safety behaviors alone will produce managers who are effective at the start and limited after that. The initiation phase and the sustained-use phase each demand a distinct set of managerial behaviors. That distinction is what separates programs that generate early enthusiasm from programs that sustain change over time.

During the initiation phase, the manager’s primary function is climate. This means modeling experimentation visibly, normalizing mistakes openly, and ensuring that early failures do not quietly affect how team members are assessed. These behaviors are well established in the psychological safety literature, and leadership programs generally develop them.

During the sustained-use phase, the manager’s function shifts toward enabling. This means adjusting workloads so that ongoing AI practice is realistic and giving specific feedback on quality of use, not just whether tools are being used. Redesigning team workflows is the third element, embedding AI use into how work gets done rather than treating it as an individual add-on.

The manager’s enabling behaviors are learnable but rarely included in leadership development programs focused on psychological safety. The research explains why that gap matters. Those sustained-use conditions require deliberate investment in the later stage of any program.

Designing for the Full Arc of AI Integration

The gap the research identifies, between what psychological safety predicts and what sustained use requires, creates a specific design challenge for organizations building AI adoption programs. Within the Human Performance Intelligence framework, this maps onto two separate challenge areas, each requiring different program designs. The interpersonal safety conditions (CH-AI-05) govern the first stage, while the workload, capability, and integration conditions (CH-WF-04) govern what follows.

HR and leadership programs built around a single intervention lever will consistently encounter a predictable ceiling, strong early engagement followed by a plateau the program logic cannot address. The more productive design principle is to treat AI integration as a staged process with distinct human performance requirements at each stage. That means building programs that shift their focus deliberately as adoption progresses.

The research asks program designers to be precise about which conditions belong at each stage of adoption. Psychological safety interventions belong at the front of the program, while sustained-use conditions, including skill development, workflow redesign, and manager enabling, belong in what follows. Building both into the same design is what produces AI integration that holds over time.