Employees May Speak More Honestly to AI Than to Their Managers

Employees in scenario-based experiments were more willing to voice concerns to an algorithmic leader than to a human one, according to a 2026 series of studies with participants in China — a finding that should make HR leaders reexamine what their AI adoption dashboards actually measure. The concern raised to a chatbot never reaches the org chart, which means the management system, not the technology, is where dissent disappears.

The study, published in Scientific Reports, found that perceived fairness and psychological safety explained the difference. The effect did not hold in emotionally charged situations, but for analytical and cognitive tasks it was consistent. One explanation is structural: an algorithm cannot embarrass an employee at the next staff meeting, damage their reputation with the executive team, or remember a challenge from six months earlier.

For HR, the implication is a measurement blind spot. Two employees can look identical on a usage dashboard — one accelerating work with AI, the other using it as the only place they feel safe saying what they think. Usage numbers will not distinguish them.

What does the research say about psychological safety itself?

A study covering five workplace settings — from hospital nurses to knowledge workers, plus four years of sales data across 229 retail stores — found the relationship between psychological safety and performance is not simply linear. At high psychological safety without matching accountability, performance on routine work could decline; collective accountability buffered that drop. The problem was not that people felt too safe. It was safety operating without the discipline teams also need.

A newer conceptual paper reinforces the point: psychological safety lowers the interpersonal risk of speaking up, but safety alone does not create the preparation and shared norms that keep teams sharp. Aviation's sterile cockpit rules offer the model — challenge during critical moments stays welcome, while decision rights and operating discipline tighten. The practical translation for people processes: keep candor maximal during pilots and strategy debates, then increase accountability once work enters production. Psychological safety does not go down; accountability goes up.

Does a healthy culture make people trust AI?

Less than leaders assume. One recent study found inclusive leadership did improve psychological safety, but psychological safety did not significantly predict trust in AI. What mattered more was an individual's tolerance for uncertainty — relevant because AI adoption is an uncertainty event, with capabilities shifting every few months and outputs that can be confidently right and impressively wrong.

A separate study of employees in South Korea, published in Humanities and Social Sciences Communications, found AI adoption was associated with lower psychological safety, which in turn correlated with greater employee depression. Ethical leadership — transparency and fairness about what is changing — softened but did not eliminate the relationship. The studies draw on different countries, samples and methodologies, so neither is universal proof, but together they separate two problems HR often conflates: whether employees can question what is happening around them, and whether they can work productively with a probabilistic tool. The second requires AI literacy and judgment, not another culture initiative.

The canonical warning remains the 1986 Space Shuttle Challenger disaster: engineers raised O-ring concerns before launch, managers under schedule pressure overruled them, and seven astronauts died 73 seconds after liftoff. Organizations still miss what they already know when people are afraid to speak.

What should HR change now?

The column, written by advisor and researcher Khalid Kark, who brings 25 years of enterprise technology experience, lays out three moves:

• Let accountability flex by phase. Encourage aggressive challenge during pilots and architecture reviews, where disagreement is cheap. Tighten decision rights at execution without closing channels for safety, ethical or operational concerns.
• Measure whether dissent changes decisions. Ask which initiative was redesigned, which AI pilot was narrowed or killed, and which executive decision shifted because someone challenged the prevailing view. If no one can name one, another engagement survey adds little.
• Build a bridge from AI conversations to human decisions. Not by reading employees' private chats — that surveillance would destroy candor — but through anonymized escalation channels, prompt-to-retrospective workflows, or a structured path from an AI conversation into an investment review.

The design test is simple: can an employee move a concern discovered with AI into the organizational decision process in one deliberate action? As generative and agentic AI embeds in daily work, the concern may show up there first.