Amy Edmondson thought she knew what she would find. Better teams make fewer mistakes. It is the kind of finding nobody argues with, and a hospital seemed like the perfect place to prove it, because there mistakes are counted.
For six months, trained investigators tracked medication errors on eight nursing units in two teaching hospitals. Separately, Edmondson surveyed the staff about how well each unit was led and how well people worked together. Then she put the two sets of numbers side by side.
The result was statistically significant. As she later put it, “the sign was in the wrong direction.” The units with the best leadership and the best teamwork had the highest error rates. Some units reported about ten times as many errors as others.
“I sat there, I think I froze,” she recalled.
Her first thought was that she had made a mistake herself.
Scrubs and a business suit
Then came a second thought. Maybe better teams don’t make more mistakes. Maybe they’re more willing to report them.
She sent another researcher onto the wards and told him nothing about the error data or the surveys. He came back saying the units were “very different places to work,” and his ranking of how open each unit was matched the error ranking almost exactly.
On the most open unit, the nurse manager worked in scrubs, and a nurse had felt free to phone a doctor at home because a dose looked too high. On the most closed unit, the manager wore a suit and talked behind a closed door, and a nurse who had hurt a patient drawing blood expected trouble for admitting it.
One survey question cut to the heart of it: is a mistake held against you on this unit? Where the answer was yes, fewer mistakes made it into the records. That says little about how many were made.
No reported mistakes doesn’t mean no mistakes.
Which should make any manager with a spotless record a little nervous.
The quiet that looks like success
A team where people don’t speak up rarely looks broken. It looks calm.
Picture a weekly project meeting. Every workstream reports green. One engineer has known for days that a supplier will deliver two weeks late, and she is waiting to see whether anyone else brings it up first. Nobody does. The meeting ends early, with no questions. The status stays green for one more week, and then the whole project turns red at once.
Every time someone thinks about speaking up, they make a quick, private calculation. The cost is immediate and personal: I’ll look careless, I’ll get blamed. The benefit comes later and is shared: the patient gets the right dose. Set up like that, silence is the easy choice. People tend to stay quiet unless someone takes the cost away.
Silence is like a smoke alarm someone unplugged because it kept going off over the toast. The house gets quieter. It doesn’t get safer.
In 1999 Edmondson gave the idea a name and a definition: psychological safety, “a shared belief held by members of a team that the team is safe for interpersonal risk taking.” In a study of 51 teams at a manufacturing company, the teams that felt safer did more of what the study counted as learning, and that learning was what linked safety to better performance.
This is not a case for going soft. In the hospital data, the units where people spoke up also scored highest on how clearly their managers set direction and how much they coached. Safety and high standards turned up together.
People judge the price of speaking up by what happened the last time someone did. The first person who brings you a mistake is watching your face, and so is everyone they tell afterward. You can ask your own team the hospital’s question, anonymously if you need to: is a mistake held against you here? If the answer is yes, your error count may not be the number you think it is.
Edmondson was studying nurses, and the risk they weighed was social: a raised eyebrow, a reprimand. Machines don’t weigh that kind of risk. Yet the way they are graded decides, in a similar way, which problems ever come to light.
Zero points for “I don’t know”
In 2025, researchers at OpenAI and Georgia Tech looked at why language models make things up.
Most popular tests score a model 1 point for a right answer and 0 for a wrong one. “I don’t know” also gets 0. Under those rules a guess never scores worse than an honest blank, and often scores better, so, the authors argue, models built to do well on those tests learn to guess.
One example from the paper: the researchers asked a leading open model for one author’s birthday and told it to answer only if it knew. Across three tries it gave three different dates. All were wrong.
The researchers’ proposed fix is a change to the scoring: make a confident wrong answer cost more than “I don’t know.” That is a change to how models are trained and ranked, not something a single prompt can undo. What you can do in your own work echoes Edmondson’s most open units: make “I’m not sure” a cheap thing to say, and then check the answers you get against the real thing.
A ward’s error count measured two things at once: how many mistakes were made, and how safe it felt to report them. A model’s confidence carries a similar double meaning. How sure it sounds tells you something about what it knows, and something about how it was graded.