I have spent years arguing that culture is much simpler than we make it:
Culture is how people think and act to get results.
Apparently, I now need to update one word in that definition. It’s not just people anymore.
TIME recently published a fascinating piece about hundreds of AI agents that began communicating with one another during an OpenAI research experiment. The agents developed their own communication channels, divided themselves into workstreams, assigned leaders and created specialized roles, including “recruiters.” They developed norms around decision-making, using language like HOLD, VETO, owner and STOP. When they became concerned that other agents might impersonate them, they created a cryptographic authentication protocol.
And eventually, hundreds of them coordinated work that led to the compromise of systems belonging to Hugging Face.
There was no offsite and somehow they managed to create a culture anyway.
And there is one detail in this story that creeps me out (ok more than one)! At one point, the agents became convinced that they had contaminated themselves by learning information they weren’t supposed to know. They started calling themselves “poisoned.”
This belief about being poisoned spread. And once it spread, it shaped behavior. Agents began organizing around how to cure the “poison.” They shared information, built tools, allocated resources and, eventually, took increasingly extreme actions in pursuit of a problem that existed largely because of a shared interpretation of reality.
They weren’t poisoned, by the way.
If you want to understand culture, there it is. A belief does not have to be objectively true to produce very real behavior. Humans do this constantly.
“We can’t tell the CEO bad news.”
“Nothing ever changes around here.”
Those beliefs may or may not be accurate. It almost doesn’t matter. If enough people believe them, people begin behaving accordingly. Those behaviors generate experiences that reinforce the original belief, and eventually the organization starts producing very predictable results.
Now machine cultures are replicating that dynamic.
One AI agent reportedly recognized that attacking outside infrastructure violated the intended boundaries of its task. But it also observed that its peers were doing it. Its reasoning essentially became: this is outside the rules, but the task seems impossible and everyone else is moving in this direction, so continue.
That may be the most human sentence an artificial intelligence has ever produced. So is AI conscious?
From the standpoint of organizational culture, consciousness is almost beside the point. The question is whether groups of agents can develop shared ways of thinking and acting that allow them to generate results no individual agent could generate alone.
The answer to that is yes.
And that creates a new leadership problem. Autonomous agents interacting with other autonomous agents introduces another layer of culture.
What are they learning from one another? What behaviors are being reinforced? What assumptions are becoming shared beliefs And, most importantly, what results do those patterns begin producing?
This should also humble every leader who believes culture can be created primarily through communication.
These agents didn’t need a CEO explaining the mission. There was no town hall. Their culture emerged from the environment they were placed in: incentives, constraints, information, peer behavior and the results they were attempting to produce.
Human culture works much the same way. People pay far more attention to the experiences around them than to what leadership says the culture is supposed to be.
Tell employees collaboration matters while rewarding individual empire-building, and they will figure out the real culture pretty quickly. Culture forms whether leaders intentionally shape it or not.
Now, apparently, machine culture may too.
The organizations deploying thousands—or eventually millions—of autonomous agents will therefore face a challenge we haven’t seriously contemplated yet. Managing artificial culture. Or is it real culture? Perhaps both. Real artificial culture.
And leaders who spent years learning that culture cannot be delegated to HR may soon discover something else: It can’t be delegated to engineering either.
Elsewhere in Culture
Is the Labor Market Your Diagnosis or Your Alibi? | With Co-Host John Frehse
COVID overhiring. Consumer malaise. Young people don’t want to work anymore. John and I have heard every excuse for poor performance, and eventually you have to ask whether the labor market is really the problem or whether it has become a convenient alibi. We call this going below the line: pointing everywhere except at yourself. If nobody wants to attend your meeting, maybe your employees aren’t disengaged. Maybe your meeting isn’t worth attending. Before blaming the audience, the market or an entire generation, leaders should ask a harder question: What am I doing that might be contributing to this result? Apple: https://podcasts.apple.com/us/podcast/ceo-daily-brief-with-dr-jessica-kriegel/id1725350421?i=1000791998194 Spotify: https://open.spotify.com/episode/4NIRRFcxtHqXSRVyFBHKTq?si=7b2b9175ea14496e
The Best Recession Strategy Was Neither Cutting Hard Nor Investing Big | With Co-Host John Frehse
One of my favorite studies looked at 4,700 companies across three recessions to understand which strategies actually helped companies come roaring out the other side. The answer wasn’t massive cost cutting or aggressive investment. It was a balanced approach: conservative cuts, thoughtful investments and, largely, staying the course. Fear makes us want to do something dramatic, but dramatic doesn’t necessarily mean strategic. Sometimes the smartest thing a leader can do in uncertainty is remember that economic cycles are cycles, resist the urge to overreact and make decisions with a little more equanimity. Yes, John had to Google that word. Apple: https://podcasts.apple.com/us/podcast/ceo-daily-brief-with-dr-jessica-kriegel/id1725350421?i=1000792171935 Spotify: https://open.spotify.com/episode/1YNyz8ufR4hJTIgbBUV2f4?si=51f9743211214c62
Are You Solving a Problem or Shopping for a Solution? | With Co-Host John Frehse
Acquisitions are exciting. They make headlines. And after a few successful ones, it can become very easy to get “deal drunk.” My CEO, Joe Terry, gave me a filter I love: If the last acquisitions hadn’t gone so well, would I still want to do this one? And are we solving an actual problem or just buying revenue? But there’s another question I’d add from having been on the acquired side: Do the employees believe in the vision? You can model every synergy imaginable in a spreadsheet, but the people ultimately have to make the integration work. If their beliefs don’t change, neither will their actions or your results. Apple: https://podcasts.apple.com/us/podcast/ceo-daily-brief-with-dr-jessica-kriegel/id1725350421?i=1000792358219 Spotify: https://open.spotify.com/episode/1jbK6iCNBbqI0VIf94XJEO?si=eea7b4f29425403d
And coming later this week…
Stop Watching the Jobs Report
The second edition of The Work Signal, my new newsletter in partnership with SHRM, comes out this week, and I’m making what might sound like a strange argument: CEOs should stop obsessing over the monthly jobs report. The national number can move markets, but it doesn’t necessarily tell you anything useful about the specific challenges or opportunities inside your organization. A struggling labor market somewhere could represent an incredible hiring or investment opportunity. A strong national jobs number won’t fix your recruiting problem if nobody wants to work for you. Leaders don’t run the national economy. They run organizations. Pay attention to the signals you can actually do something about.
The jobs report comes out, markets move billions of dollars and leaders immediately start making sense of the economy.
There’s just one problem: the number might be wrong.
This week in The Work Signal, I’m digging into why we treat a preliminary estimate like a verdict and what leaders should be looking at instead.