When Employees Let AI Say What They Can't: The Rise of "Safety by Proxy" at Work

Marcus had spent six months building the case for a regional expansion. The data supported it, leadership had signed off, and he'd laid out the plan to his team with confidence. Two days later, an anonymous email arrived. No note, no name, just a forwarded document. Someone had asked an AI tool to build the strongest case for the expansion, and the strongest case against it.

The counterarguments weren't shocking. Marcus had seen versions of them in his own data: a crowded market, an overestimated customer base, partnerships that hadn't actually been secured yet. What was new wasn't the content. It was the delivery: typed, circulated, unowned. When Marcus asked who'd sent it, nobody answered.

Researchers Jayshree Seth and Amy C. Edmondson call this pattern "safety by proxy." It happens when someone uses AI's perceived neutrality to voice a concern they don't feel safe raising in person. Think of it as an anonymous survey nobody commissioned and nobody can easily wave off. It tells leaders what people actually think, and it tells them something else too: that speaking honestly, in that room, felt too risky.

This isn't about people using AI to draft emails or brainstorm ideas, which happens constantly for practical reasons that have nothing to do with fear or hierarchy. Safety by proxy is narrower. It's what happens when a real, substantive concern gets funneled through a machine because the human channel felt closed. Suggestion boxes have existed for decades, but AI outputs currently carry more perceived weight and legitimacy than a typed suggestion ever did. That difference matters.

Decades of workplace research, including Edmondson's own foundational work on team psychological safety, have shown that whether people speak up depends almost entirely on whether they believe they can do so without personal cost. Teams that get this right, where people feel safe taking interpersonal risks for the sake of shared goals, tend to learn faster and adapt better under pressure. The traditional fix has always rested on leaders: model humility, invite pushback, respond without defensiveness. What that framework never anticipated was a workaround that lets people bypass the leader entirely.

Four Ways AI Is Reshaping Who Speaks and Who Stays Quiet

Seth and Edmondson describe four distinct patterns emerging where AI's perceived neutrality, the belief that a machine is less biased than a human boss, intersects with the willingness to speak up.

The first is the workaround itself. People who believe honesty carries real personal risk use AI to surface a concern without attaching their name to it. A 2026 peer-reviewed study of board directors found this exact dynamic in action. One director explained it bluntly: if a director voices criticism, they might face repercussions, but if AI says it, "it's out there, without being tied to a particular director." For leaders, this is about as direct a signal as it gets that something important is going unsaid.

The second is what the researchers call the equalizer effect. Psychological safety is rarely spread evenly across an organization. It shifts depending on someone's status, identity, or how new they are to a team. AI can level that terrain a little. A team member from an underrepresented group can route a concern anonymously. A non-native speaker can ask AI to sharpen language they'd otherwise struggle to phrase under pressure. In these moments, AI functions less like a mouthpiece and more like a coach, helping people say what they already believe but couldn't quite articulate.

The third pattern is testing. In organizations working to build a more open culture but not quite there yet, people sometimes use AI as a low-stakes rehearsal space, checking whether an argument holds up before bringing it to a live audience. Whether that habit accelerates real psychological safety or simply substitutes for it isn't yet clear. The evidence doesn't tell us which way it cuts.

The fourth is the mirror, and it's the one with the clearest upside for leaders. Power suppresses honest input, one of the oldest findings in management research. The higher someone's authority, the more people shape what they say around what they think that person wants to hear. A leader who asks AI to generate the strongest possible objections to their own decision gets access to input their own authority would otherwise choke off. The machine doesn't modulate its answer based on who's asking.

At 3M, this mirror pattern shows up in the company's stage-gate process for new product development, a culture long built on the belief that good ideas can come from anywhere. Established scientists use AI to stress-test ideas outside their expertise before raising them in cross-functional meetings. Newer employees use it to rehearse arguments before presenting to senior colleagues. Cross-functional teams hold dedicated AI sessions before major decision points, asking the system to surface the toughest technical, market, and regulatory questions a reviewer might raise. Seth herself practices this before presenting new AI use-cases to teams that may feel unready for change: she asks the AI to play skeptic first, so she arrives with cleaner examples and a more honest account of the limitations.

The workaround pattern, though, was crystallized for Seth in a single line she heard at a recent symposium on generative AI in R&D: "Someone can just AI it." The era in which a manager could control which concerns reached the table, she noted, has quietly ended.

Why It Matters

None of this is cause for alarm. Read correctly, each pattern is an opening rather than a threat. Research shows 83% of business leaders believe psychological safety directly affects whether AI initiatives succeed, according to a joint report from Infosys and MIT Technology Review Insights. Building that safety isn't a nice-to-have culture project anymore. It's a strategic necessity.

When an anonymous AI-generated document lands on your desk, the instinct is to judge its content. The better instinct is to ask what its existence reveals. A concern that arrives unattributed tells you someone felt they couldn't raise it directly, and that gap is itself a measurement of your culture. Responding with curiosity reads the signal correctly. Responding with defensiveness proves the point the workaround was trying to make.

It's worth remembering that AI's neutrality can be misused too, dressing up self-interested resistance as principled concern. Judge the substance, ask who benefits, and treat the anonymous channel as a reason to look closer, not as a verdict either way. And anonymity has limits: most AI platforms log who ran a query. The room may feel anonymous. The system rarely is.

The better sequence, then, isn't announce and react. It's think with AI first, asking what the strongest objections to a decision might be, then genuinely listen to your people, then announce. AI can surface the general shape of risk. Only human conversation tells you which of those risks are actually alive in your specific team, and why. Skip that last step, running the AI stress-test and ignoring what it turns up, and you're not building safety. You're just performing it, and teams notice the difference fast.