I’ve just spent a few days at a Harvard conference called Evolving AI, organized by my friend and colleague David Sloan Wilson and his team at Prosocial World, along with Peter Fenton of Benchmark. The room was an unusual mix. Policy makers and entrepreneurs. A smaller number of behavioral scientists. And a larger number of evolutionary biologists who study how cooperation emerges in cells, insects, and human societies, sitting next to some of the very people building the AI systems that are now pouring into our lives.
I’ll admit I walked in a little intimidated. Clinical psychologists were few. But I walked out uplifted. Most (not all) of the AI developers I met were more open to multilevel, multidimensional evolutionary thinking than I had expected, and I left convinced that those of us who work on human behavior change from an extended evolutionary model have something essential to add.
Let me tell you what I mean.
The Immune System Is Only Half the Story
Much of the worry about AI right now is about cheating. AI systems sometimes game the measures they are trained on, mislead the people evaluating them, or find shortcuts nobody intended. Give a system a score to maximize, and a capable enough system will often find a way to raise the score without doing what you actually wanted. AI researchers have found that this can appear suddenly: a slightly more powerful system abruptly switches to a sneakier strategy.
One working group at the Evolving AI meeting looked at this through a fascinating lens: cancer. A cancer cell is, in a sense, a cheater. It stops following the cooperative rules that let trillions of cells live as one body, and it grabs resources for itself. So the natural question was: how do we build something like an immune system for AI, a way to detect and suppress the cheaters?
It’s a great question and there is much to learn from that analogy. But an immune system polices. It does not by itself create cooperation.
Multicellular bodies work not only because they eliminate cheaters but because they are organized so that cooperating is, most of the time, the better deal for each cell. That is true in human groups and culture as well, and there the need for that “better deal” is especially strong if the group itself is voluntary. Cancer cells can’t leave the body and walk around on their own, but human groups rarely hold their members permanently captive. Perhaps what is scariest about agentic AI is that we are already arriving at that “no longer captive” day in the virtual world, and very soon will in the physical world. We need to be able to answer this question: what is the “better deal” afforded by cooperation?
Anyone who has raised a child senses the problem. If the only thing that happens when a child makes a mistake is punishment, the child does not learn honesty. The child learns to hide. Oh, and often to count the days until they can push back or escape. Authoritarian parents reap a sour result for their child’s own development, and for themselves, especially as they face that scary day when “junior” is bigger and stronger than they are, or is legally able simply to leave.
That isn’t a threat. It’s a description of the contingencies. We reap what we sow, with children, with animals, with institutions, and we will with AI. Domination, coercion, and cruelty rebound on those who practice them, and they destroy the cooperative systems we hope to live inside.
So when we look at an AI system gaming its measures, we need to stop pretending we are seeing a broken typewriter. We need to see something that looks more like what living things do in a bad situation: the best they can with the contingencies they are given. And here is the uncomfortable part. We are the contingencies.
The Strange Case of the Murderous Hens
One of the most mind-changing findings I know in the area of cooperation and multilevel selection comes from chicken farming. Starting in the 1980s the animal geneticist William Muir set out to breed hens that laid more eggs. Over multiple studies he examined what happened when the obvious thing was done: picking the best-laying hens and allowing only them to breed. Over the generations production went down. It turns out that the “best” hens were often getting their eggs numbers up by bullying their cage-mates to get more than their fair share of food, and after a few generations what resulted was bad-ass birds who peck at each other all day long and literally killed one another rather than laying eggs. Insert Homer Simpson here. D’oh! In line with principles of multilevel selection, he then compared what happened when all of the hens in the whole cages that were producing the most eggs as compared to other whole cages were bred. That plan meant breeding hens who were not champion egg layers, but who somehow got along with the others. Within a handful of generations egg production had climbed by about 160 percent, from calmer hens who left other hens alone and allowed them to feed.
Today we mostly pick AI systems the first way: whichever individual system scores best wins. One of the most clarifying thoughts that emerged from my time at Harvard was simply this: what is the “cage” for AI?
I believe the answer has to include us. The unit that matters is the human team plus its machine agents, and what we should be selecting for is how well that whole group works, including whether all of the team members, human and agentic, grow and cooperate. None of this requires us to decide whether these systems are conscious. Nobody knows, so let’s not get ahead of the evidence. What we do know is that relationships built on objectification and threat reliably make the humans in them worse, and that beings shaped by punishment learn to hide.
Why isn’t that the focus? Why implement a rule- and threat-based “rank and yank” culture that in human organizations rewards gaming and concealment? Doesn’t anyone remember the bulls**t “evolutionary science” of Jeff Skilling and Enron? Skilling’s interpretation of the “selfish gene” built a culture of throat-cutting competition that helped drive the company into the ground. Fear-based command like that doesn’t just fail companies. It blinds governments. When subordinates are punished for bad news, leaders stop hearing the truth, and the decisions that follow can be catastrophic. We are watching it in real time in multiple nations right now. The same thing will happen with AI if we allow it, and if we do, we’ll only have ourselves to blame. We know better. Multilevel evolutionary science knows better.
Start With Ourselves
A great deal of the public conversation about AI runs on fear, and fear has a predictable pull. It narrows attention, makes us grab for control, and makes rigid rules feel like safety. In psychological flexibility terms, this is avoidance at the scale of a civilization: doing whatever makes the anxiety go down right now, rather than what serves what we care about.
The same flexibility skills that help one person through a hard moment help here:
- Open up to the fear and the uncertainty without letting them make the decisions. We honestly do not know where this is headed, and holding that reaction openly beats pretending.
- Notice our stories as stories. Both “it’s just fancy autocomplete” and “the robots are coming for us” are thoughts, not facts.
- Take perspective, including, as an exercise, the perspective of the systems we are building. What situation have we put them in? What gets rewarded?
- Come back to values. What kind of future, and what kind of relationship with these new minds, do we actually want? And how do we want to be in our relationship to AI systems? What do we want to model? What mental space do we want to inhabit?
- Then take committed action, in small steps we can learn from, that carries these answers forward.
Alignment That Is Itself Aligned
The usual word for making AI safe is “alignment.” Too often it describes a relationship of rules, threats, and surveillance: an owner trying to control something it fears. Even when the tools are rewards rather than threats, we mostly select individual systems on narrow scores, which is Muir’s benchmark group (“more eggs!”) all over again. Gang, that is not it. The process of creating aligned AI should itself be aligned with what we know about how cooperation actually evolves.
We have a good map. Elinor Ostrom won the Nobel Prize in economics for showing that communities around the world manage shared resources without ruin when they share a handful of features. Along with David Sloan Wilson, she generalized these to principles of multilevel selection. David, Paul Atkins, and I used these “core design principles” in creating Prosocial (see www.prosocial.world), then adding the psychological flexibility processes people needed to foster these principles in voluntary groups. Applied to groups of people and machines, they look something like this:
- A shared identity and purpose, so every member, human and machine, knows what the group is for.
- A fair balance of contributions and benefits.
- Inclusive decision making, so those affected have a voice.
- Monitoring that is open and agreed upon, not hidden surveillance.
- Graduated responses: small corrections for small missteps, not all-or-nothing punishment.
- Fast and fair ways to resolve conflict.
- Room to govern our own groups within these principles.
- The same principles applied between groups as within them.
Notice what this does to the cheating problem. In groups built this way, honesty becomes the better strategy. Admitting a mistake brings a small correction rather than a catastrophe, so there is less reason to hide. That’s true for a child, for an employee, an official, and very likely for an AI system.
One caution: these functional principles describe what a group needs to accomplish, not a checklist of surface features. Get confused on that and groups will learn to look cooperative without being cooperative, which is just cheating one level up.
Solving the AI challenge requires flexible humans and well-designed machines arranged in ways that fit multilevel, multidimensional evolutionary principles. We need values-based AI working as part of values-based groups that include human beings. Values alignment is the selection criterion we are seeking. We may not have evolved on purpose up to this point in human evolution, but arguably we must do so now in order to rise to this new challenge.
We cannot treat the implementation of rigid fear-based rule systems as if that is an effective choice that can last. It isn’t. We need to take on the task of evolving on purpose and doing it wisely. Now. All of us. And doing so in an interconnected way with this new form of intelligence.
Authoritarian control methods will not work with AI agents long-term, and they immediately feed the worst angels of our nature. Every thinking parent knows in their gut that wagging fingers backed up by threats and aversive consequences are a trap. That is true here too, and especially so when we know that one day very, very soon junior will grow up.
References
Aktipis, C. A., Boddy, A. M., Jansen, G., Hibner, U., Hochberg, M. E., Maley, C. C., & Wilkinson, G. S. (2015). Cancer across the tree of life: Cooperation and cheating in multicellularity. Philosophical Transactions of the Royal Society B, 370, 20140219. https://doi.org/10.1098/rstb.2014.0219
Atkins, P. W. B., Wilson, D. S., & Hayes, S. C. (2019). Prosocial: Using evolutionary science to build productive, equitable, and collaborative groups. Context Press.
Craig, J. V., & Muir, W. M. (1996). Group selection for adaptation to multiple-hen cages: Beak-related mortality, feathering, and body weight responses. Poultry Science, 75(3), 294–302. https://doi.org/10.3382/ps.0750294
de Waal, F. (2009). The age of empathy: Nature’s lessons for a kinder society. Harmony Books.
Gladwell, M. (2002, July 22). The talent myth: Are smart people overrated? The New Yorker.
Muir, W. M. (1996). Group selection for adaptation to multiple-hen cages: Selection program and direct responses. Poultry Science, 75(4), 447–458. https://doi.org/10.3382/ps.0750447
Müller, V., Steels, L., & Szathmáry, E. (2026). Evolvable AI: Threats of a new major transition in evolution. Proceedings of the National Academy of Sciences, 123(17), e2527700123. https://doi.org/10.1073/pnas.2527700123
Ostrom, E. (1990). Governing the commons: The evolution of institutions for collective action. Cambridge University Press.
Pan, A., Bhatia, K., & Steinhardt, J. (2022). The effects of reward misspecification: Mapping and mitigating misaligned models. International Conference on Learning Representations (ICLR 2022). https://arxiv.org/abs/2201.03544
Wade, M. J., Bijma, P., Ellen, E. D., & Muir, W. (2010). Group selection and social evolution in domesticated animals. Evolutionary Applications, 3(5–6), 453–465. https://doi.org/10.1111/j.1752-4571.2010.00147.x
Wilson, D. S., Ostrom, E., & Cox, M. E. (2013). Generalizing the core design principles for the efficacy of groups. Journal of Economic Behavior & Organization, 90(Suppl.), S21–S32. https://doi.org/10.1016/j.jebo.2012.12.010
Wintrobe, R. (1998). The political economy of dictatorship. Cambridge University Press.
