00:00:08AI Signals Team Culture
00:00:36Meet Marcus’s Silent Dissent
00:02:06Four AI Safety Patterns
00:03:33Anonymous, But Not Invisible
00:04:00Manager Takeaways
00:05:03The Real Problem And Challenge
ABOUT THIS EPISODE
Somewhere on your team, someone may be using AI to surface concerns they don't feel safe raising directly — a phenomenon called "safety by proxy." This episode explains how that behavior is a signal about your culture and introduces four patterns (workaround, equalizer, testing space, mirror) that reveal different safety dynamics.
Greg Gillum summarizes four managerial actions: diagnose the pattern before reacting, use AI as a deliberate mirror, don't outsource judgment to tools, and make speaking up safe in the room. The episode ends with a practical challenge: run a "strongest case against" test and bring the findings to your team.
English
United States
SHOW NOTES 🔗
TRANSCRIPT 🔗
WEBVTT
00:00:00.351 --> 00:00:04.387
Somewhere on your team, a high potential may already be asking an AI to build
00:00:04.387 --> 00:00:08.925
the strongest case against your latest decision, then quietly passing it around.
00:00:09.401 --> 00:00:14.085
Not because they're disloyal, but because they don't feel safe saying it to your face.
00:00:14.724 --> 00:00:19.697
A new Harvard Business Review article says employees who use AI to surface concerns
00:00:19.697 --> 00:00:24.000
they don't feel safe raising directly are sending signals about the culture.
00:00:24.424 --> 00:00:28.336
And what you do next decides whether you'll ever hear the truth directly again.
00:00:37.100 --> 00:00:41.616
Welcome to The Manager Lab. I'm Greg Gillum, your host and CEO of Human Fusion.
00:00:42.092 --> 00:00:47.287
Today, a fresh HBR article from Amy Edmondson, the Harvard professor who put
00:00:47.287 --> 00:00:53.406
psychological safety on the map, and Jayshree Seth, chief science advocate at 3M.
00:00:53.946 --> 00:00:58.692
It's called, What Leaders Need to Know About AI and Psychological Safety.
00:00:59.063 --> 00:01:00.497
And it's about something happening
00:01:00.497 --> 00:01:03.684
on your team that you may not have noticed. So let's get into it.
00:01:04.404 --> 00:01:09.199
Meet Marcus. He spent six months building a case for a regional expansion.
00:01:09.710 --> 00:01:14.389
The data supported it. Leadership approved it, and he presented the plan to his team.
00:01:15.045 --> 00:01:19.120
Two days later, an anonymous email arrives with a forwarded document.
00:01:19.758 --> 00:01:24.787
Three pages generated by someone who asked an AI tool for the strongest case
00:01:24.787 --> 00:01:27.839
for the expansion and the strongest case against it.
00:01:28.146 --> 00:01:30.201
No name, no explanation.
00:01:30.660 --> 00:01:35.517
Somebody on Marcus's team had reservations. and the safest way they could find
00:01:35.517 --> 00:01:37.678
to voice them was through a machine.
00:01:38.372 --> 00:01:41.570
The authors call this safety by proxy.
00:01:42.134 --> 00:01:46.755
AI feels neutral. When a colleague says, I think this is a mistake,
00:01:47.137 --> 00:01:51.259
it's personal, with risk to the relationship and the reputation.
00:01:51.711 --> 00:01:55.101
But when an AI makes the same argument, it feels like analysis.
00:01:55.384 --> 00:01:57.126
Nobody sticks their neck out.
00:01:57.660 --> 00:02:02.277
And here's what managers need to hear. How people use AI is telling you something
00:02:02.277 --> 00:02:05.589
about your culture. It's a smoke alarm, not the fire.
00:02:06.744 --> 00:02:10.627
The authors describe four patterns, and each one tells you something different.
00:02:10.958 --> 00:02:15.689
Number one, if safety is absent, AI becomes a workaround.
00:02:15.945 --> 00:02:19.933
That's Marcus. People who can't speak up go around you.
00:02:20.505 --> 00:02:25.219
Number two, if safety is unequal, then AI becomes an equalizer.
00:02:25.788 --> 00:02:29.317
Some people on your team feel free to challenge ideas, and others don't.
00:02:29.869 --> 00:02:36.473
AI lets the quieter voices pressure test an idea and put it on the table on more even footing.
00:02:37.164 --> 00:02:39.180
Ask yourself who goes silent in
00:02:39.180 --> 00:02:44.269
your meetings then shows up later with a suspiciously well-built argument.
00:02:44.981 --> 00:02:49.695
Number three, if safety is developing, AI becomes a testing space.
00:02:50.252 --> 00:02:53.421
On a new team or under a new leader, people rehearse.
00:02:53.908 --> 00:02:57.687
They try a concern out on AI before deciding whether to say it to you.
00:02:58.134 --> 00:03:01.501
That's a good sign. Trust is growing, but it isn't there yet.
00:03:01.873 --> 00:03:06.360
Your job is to make the jump from rehearsal to real conversation very easy.
00:03:06.979 --> 00:03:12.848
And fourth, because even the best leaders are separated by authority, AI becomes a mirror.
00:03:13.459 --> 00:03:16.518
Bad news gets softened on its way up.
00:03:16.959 --> 00:03:20.756
Not because you're a bad boss, but because that's how hierarchy works.
00:03:21.127 --> 00:03:26.955
AI can show you what your team is too polite to say, and you can use it that way on purpose.
00:03:27.217 --> 00:03:33.120
So same tool, four meanings. The technology is identical. The difference is your culture.
00:03:33.703 --> 00:03:37.917
Now here's one wrinkle. That anonymity is partly an illusion.
00:03:38.534 --> 00:03:43.190
AI removes the attribution of the view, but not the attribution of the act,
00:03:43.480 --> 00:03:45.924
and most platforms log who ran a query.
00:03:46.452 --> 00:03:49.691
So it may be anonymous in the meeting room, but not in the system.
00:03:50.150 --> 00:03:53.771
The temptation is to hunt for the author. My advice?
00:03:54.218 --> 00:03:57.730
Don't. You teach everyone that even the workaround isn't safe,
00:03:57.956 --> 00:04:00.197
and the next time the concern won't surface at all.
00:04:00.777 --> 00:04:05.322
So here's four takeaways for managers. Number one, diagnose before you react.
00:04:05.850 --> 00:04:07.182
Which of the four patterns are
00:04:07.182 --> 00:04:11.104
you seeing? A workaround needs a different response than a testing space.
00:04:11.684 --> 00:04:14.197
Two, use the mirror on purpose.
00:04:15.472 --> 00:04:20.989
Edmondson and Seth, the authors, suggest engaging with concerns before you announce a decision.
00:04:21.535 --> 00:04:26.002
Ask AI what the strongest arguments against it are and what risks haven't been
00:04:26.002 --> 00:04:28.843
addressed and what assumptions haven't been tested.
00:04:29.226 --> 00:04:32.750
Then take that to a thoughtful group before it goes public.
00:04:33.429 --> 00:04:38.012
Number three, don't outsource the judgment. A general-purpose tool surfaces
00:04:38.012 --> 00:04:42.022
general patterns of risk, which makes it a starting point, not a substitute
00:04:42.022 --> 00:04:43.466
for organizational knowledge.
00:04:43.826 --> 00:04:46.618
Use it to start the conversation, not end it.
00:04:47.354 --> 00:04:50.719
And then fourth, make the front door safer than the side door.
00:04:51.303 --> 00:04:54.310
Say it in the room. Tell me the strongest case against this.
00:04:54.751 --> 00:04:59.082
Thank the person who does publicly and act visibly on something they said.
00:04:59.453 --> 00:05:02.994
If speaking up directly gets rewarded, nobody gets a proxy.
00:05:03.648 --> 00:05:08.029
So the one-line version, if your team is using AI to say what they can't say
00:05:08.029 --> 00:05:10.550
to you, the problem isn't the AI.
00:05:10.939 --> 00:05:14.817
The opportunity is an early warning most leaders never got.
00:05:15.345 --> 00:05:19.639
Your challenge this week? Pick one decision on your desk, run the strongest
00:05:19.639 --> 00:05:25.063
case against test, and bring what you find to your team before you announce it.
00:05:25.835 --> 00:05:29.759
Read the full article on HBR linked in the show notes. I'm Greg Gillum,
00:05:29.759 --> 00:05:31.907
your host and CEO of Human Fusion.
00:05:32.203 --> 00:05:36.667
And until next time we meet in the Manager Lab, do good work.
00:00:00.351 --> 00:00:04.387
Somewhere on your team, a high potential may already be asking an AI to build
00:00:04.387 --> 00:00:08.925
the strongest case against your latest decision, then quietly passing it around.
00:00:09.401 --> 00:00:14.085
Not because they're disloyal, but because they don't feel safe saying it to your face.
00:00:14.724 --> 00:00:19.697
A new Harvard Business Review article says employees who use AI to surface concerns
00:00:19.697 --> 00:00:24.000
they don't feel safe raising directly are sending signals about the culture.
00:00:24.424 --> 00:00:28.336
And what you do next decides whether you'll ever hear the truth directly again.
00:00:37.100 --> 00:00:41.616
Welcome to The Manager Lab. I'm Greg Gillum, your host and CEO of Human Fusion.
00:00:42.092 --> 00:00:47.287
Today, a fresh HBR article from Amy Edmondson, the Harvard professor who put
00:00:47.287 --> 00:00:53.406
psychological safety on the map, and Jayshree Seth, chief science advocate at 3M.
00:00:53.946 --> 00:00:58.692
It's called, What Leaders Need to Know About AI and Psychological Safety.
00:00:59.063 --> 00:01:00.497
And it's about something happening
00:01:00.497 --> 00:01:03.684
on your team that you may not have noticed. So let's get into it.
00:01:04.404 --> 00:01:09.199
Meet Marcus. He spent six months building a case for a regional expansion.
00:01:09.710 --> 00:01:14.389
The data supported it. Leadership approved it, and he presented the plan to his team.
00:01:15.045 --> 00:01:19.120
Two days later, an anonymous email arrives with a forwarded document.
00:01:19.758 --> 00:01:24.787
Three pages generated by someone who asked an AI tool for the strongest case
00:01:24.787 --> 00:01:27.839
for the expansion and the strongest case against it.
00:01:28.146 --> 00:01:30.201
No name, no explanation.
00:01:30.660 --> 00:01:35.517
Somebody on Marcus's team had reservations. and the safest way they could find
00:01:35.517 --> 00:01:37.678
to voice them was through a machine.
00:01:38.372 --> 00:01:41.570
The authors call this safety by proxy.
00:01:42.134 --> 00:01:46.755
AI feels neutral. When a colleague says, I think this is a mistake,
00:01:47.137 --> 00:01:51.259
it's personal, with risk to the relationship and the reputation.
00:01:51.711 --> 00:01:55.101
But when an AI makes the same argument, it feels like analysis.
00:01:55.384 --> 00:01:57.126
Nobody sticks their neck out.
00:01:57.660 --> 00:02:02.277
And here's what managers need to hear. How people use AI is telling you something
00:02:02.277 --> 00:02:05.589
about your culture. It's a smoke alarm, not the fire.
00:02:06.744 --> 00:02:10.627
The authors describe four patterns, and each one tells you something different.
00:02:10.958 --> 00:02:15.689
Number one, if safety is absent, AI becomes a workaround.
00:02:15.945 --> 00:02:19.933
That's Marcus. People who can't speak up go around you.
00:02:20.505 --> 00:02:25.219
Number two, if safety is unequal, then AI becomes an equalizer.
00:02:25.788 --> 00:02:29.317
Some people on your team feel free to challenge ideas, and others don't.
00:02:29.869 --> 00:02:36.473
AI lets the quieter voices pressure test an idea and put it on the table on more even footing.
00:02:37.164 --> 00:02:39.180
Ask yourself who goes silent in
00:02:39.180 --> 00:02:44.269
your meetings then shows up later with a suspiciously well-built argument.
00:02:44.981 --> 00:02:49.695
Number three, if safety is developing, AI becomes a testing space.
00:02:50.252 --> 00:02:53.421
On a new team or under a new leader, people rehearse.
00:02:53.908 --> 00:02:57.687
They try a concern out on AI before deciding whether to say it to you.
00:02:58.134 --> 00:03:01.501
That's a good sign. Trust is growing, but it isn't there yet.
00:03:01.873 --> 00:03:06.360
Your job is to make the jump from rehearsal to real conversation very easy.
00:03:06.979 --> 00:03:12.848
And fourth, because even the best leaders are separated by authority, AI becomes a mirror.
00:03:13.459 --> 00:03:16.518
Bad news gets softened on its way up.
00:03:16.959 --> 00:03:20.756
Not because you're a bad boss, but because that's how hierarchy works.
00:03:21.127 --> 00:03:26.955
AI can show you what your team is too polite to say, and you can use it that way on purpose.
00:03:27.217 --> 00:03:33.120
So same tool, four meanings. The technology is identical. The difference is your culture.
00:03:33.703 --> 00:03:37.917
Now here's one wrinkle. That anonymity is partly an illusion.
00:03:38.534 --> 00:03:43.190
AI removes the attribution of the view, but not the attribution of the act,
00:03:43.480 --> 00:03:45.924
and most platforms log who ran a query.
00:03:46.452 --> 00:03:49.691
So it may be anonymous in the meeting room, but not in the system.
00:03:50.150 --> 00:03:53.771
The temptation is to hunt for the author. My advice?
00:03:54.218 --> 00:03:57.730
Don't. You teach everyone that even the workaround isn't safe,
00:03:57.956 --> 00:04:00.197
and the next time the concern won't surface at all.
00:04:00.777 --> 00:04:05.322
So here's four takeaways for managers. Number one, diagnose before you react.
00:04:05.850 --> 00:04:07.182
Which of the four patterns are
00:04:07.182 --> 00:04:11.104
you seeing? A workaround needs a different response than a testing space.
00:04:11.684 --> 00:04:14.197
Two, use the mirror on purpose.
00:04:15.472 --> 00:04:20.989
Edmondson and Seth, the authors, suggest engaging with concerns before you announce a decision.
00:04:21.535 --> 00:04:26.002
Ask AI what the strongest arguments against it are and what risks haven't been
00:04:26.002 --> 00:04:28.843
addressed and what assumptions haven't been tested.
00:04:29.226 --> 00:04:32.750
Then take that to a thoughtful group before it goes public.
00:04:33.429 --> 00:04:38.012
Number three, don't outsource the judgment. A general-purpose tool surfaces
00:04:38.012 --> 00:04:42.022
general patterns of risk, which makes it a starting point, not a substitute
00:04:42.022 --> 00:04:43.466
for organizational knowledge.
00:04:43.826 --> 00:04:46.618
Use it to start the conversation, not end it.
00:04:47.354 --> 00:04:50.719
And then fourth, make the front door safer than the side door.
00:04:51.303 --> 00:04:54.310
Say it in the room. Tell me the strongest case against this.
00:04:54.751 --> 00:04:59.082
Thank the person who does publicly and act visibly on something they said.
00:04:59.453 --> 00:05:02.994
If speaking up directly gets rewarded, nobody gets a proxy.
00:05:03.648 --> 00:05:08.029
So the one-line version, if your team is using AI to say what they can't say
00:05:08.029 --> 00:05:10.550
to you, the problem isn't the AI.
00:05:10.939 --> 00:05:14.817
The opportunity is an early warning most leaders never got.
00:05:15.345 --> 00:05:19.639
Your challenge this week? Pick one decision on your desk, run the strongest
00:05:19.639 --> 00:05:25.063
case against test, and bring what you find to your team before you announce it.
00:05:25.835 --> 00:05:29.759
Read the full article on HBR linked in the show notes. I'm Greg Gillum,
00:05:29.759 --> 00:05:31.907
your host and CEO of Human Fusion.
00:05:32.203 --> 00:05:36.667
And until next time we meet in the Manager Lab, do good work.
Disclaimer: The podcast and artwork embedded on this page are from Dr. J. Gregory Gillum, CPCC, which is the property of its owner and not affiliated with or endorsed by Listen Notes, Inc.
EDIT
Thank you for helping to keep the podcast database up to date.
