Hide table of contents

I started working in cooperative AI almost a year ago, and as an emerging field I found it quite confusing at times since there is very little introductory material aimed at beginners. My hope with this post is that by summing up my own confusions and how I understand them now I might help to speed up the process for others who want to get a grasp on what cooperative AI is.

I work at Cooperative AI Foundation (CAIF) and there will be a lot more polished and official material coming from there, so this is just a quick personal write-up to get something out in the meantime. We’re working on a cooperative AI curriculum that should be published within the next couple of months, and we’re also organising a summer school in June for people new to the area (application deadline April 26th).

Contradicting definitions

When I started to learn about cooperative AI I came across a lot of different definitions of the concept. While drafting this post I dug up my old interview preparation doc for my current job, where I had listed different descriptions of cooperative AI that I had found while reading up:

  • “the objective of this research would be to study the many aspects of the problems of cooperation and to innovate in AI to contribute to solving these problems”
  • “AI research trying to help individuals, humans and machines, to find ways to improve their joint welfare”
  • “AI research which can help contribute to solving problems of cooperation”
  • “building machine agents with the capabilities needed for cooperation“
  • “building tools to foster cooperation in populations of (machine and/or human) agents”
  • “conducting AI research for insight relevant to problems of cooperation”

To me this did not paint a very clear picture and I was pretty frustrated to be unable to find a concise answer to the most basic question: What is cooperative AI and what is it not?

At this point I still don’t have a clear, final definition, but I am less frustrated by it because I no longer think that this is just a failure of understanding or failure of communication - the situation is simply that the field is so new that there is no single definition that people working in the field agree on, and it is still an ongoing discussion where the boundaries should be drawn.

That said, my current favourite explanation of what cooperative AI is is that while AI alignment deals with the question of how to make one powerful AI system behave in a way that is aligned with (good) human values, cooperative AI is about making things go well with powerful AI systems in a messy world where there might be many different AI systems, lots of different humans and human groups and different sets of (sometimes contradictory) values.

Another recurring framing is that cooperative AI is about improving the cooperative intelligence of advanced AI, which leads to the question of what cooperative intelligence is. Here also there are many different versions in circulation, but the following one is the one I find most useful so far:

Cooperative intelligence is an agent's ability to achieve their goals in ways that also promote social welfare, in a wide range of environments and with a wide range of other agents.

Is this really different from alignment?

The second major issue I had was to figure out how cooperative AI really differed from AI alignment. The description of “cooperative intelligence” seemed like it could be understood as just a certain framing of alignment - “achieve the goals in a way that is also good for everyone”.

As I have been learning more about cooperative AI, it seems to me like the term “cooperative intelligence” is best understood in the context of social dilemmas (collective action problems). The most famous model is the prisoner’s dilemma, in which it would be best for two agents collectively to cooperate, but where the rational decision for each individually is to defect, leading to a collectively worse outcome (e.g. arms race dynamics). Another famous model of a social dilemma is the tragedy of the commons which can be used to model overexploitation of shared resources (e.g. climate change).

The point is that in social dilemmas it is rational for each agent to defect, even if everyone would be better off if they were all cooperating. If we imagine a world with multiple powerful AI systems, social dilemmas are not solved by default by alignment. If two different groups have their own aligned AI systems, an interaction between these systems where they make rational and aligned choices can still lead to poor outcomes for both groups even if a better solution were possible. I understand “cooperative intelligence” as a concept that aims to capture the kind of capability that would be required in such a situation to achieve a collectively better outcome.

This is not only relevant for a situation with superhuman artificial intelligence. Even for more narrow or limited systems these kinds of dynamics could occur, leading to significantly harmful outcomes even if each system in isolation was safe.

The dual-use aspect of cooperative intelligence

In this paper which outlines the agenda for cooperative AI, the elements of cooperative intelligence are described as follows:

  • Understanding: The ability to take into account the consequences of actions, to predict another’s behaviour, and the implications of another’s beliefs and preferences.
  • Communication: The ability to explicitly and credibly share information with others relevant to understanding behaviour, intentions and preferences.
  • Commitment: The ability to make credible promises when needed for cooperation.
  • Norms and institutions: Social infrastructure — such as shared beliefs or rules — that reinforces understanding, communication and commitment.

When I first read this I had just gone through Bluedot’s AI Safety Fundamentals course and I suspect I’m not the only one reacting to this list of capabilities (at least the first three points) with the thought that this sounds dangerous. Yes, these are capabilities that are useful for cooperation, but they are also the very same capabilities that are needed for deception, manipulation and extortion.

I still think this is true - these are potentially dangerous capabilities - but I also think these capabilities are being developed outside of cooperative AI. I think the right framing of cooperative AI is less as a field pushing capabilities forward in each of these areas, and more as an initiative to study these emerging capabilities and to work on achieving differential development in this area. What we need is to understand how we can promote beneficial cooperation while suppressing harmful tendencies and dynamics. There is clearly cause to be extremely cautious with this, and I think it’s important to recognize the components of cooperative intelligence as dual-use capabilities.

My current model of how cooperative AI fits into the AI safety landscape

As cooperative AI is an emerging field the boundaries towards and overlaps with other fields are not yet well established. Below I share my current model of how I understand cooperative AI to fit in the wider landscape - note though that this is very much subject to change, and I would very much appreciate counter-suggestions and challenges in comments! 


 

My current model of how cooperative AI fits into the AI safety landscape


 

  • I here use “beneficial AI” to denote all work on AI that aims to make the world better.
  • I see “AI safety” as a subfield of beneficial AI that aims to ensure that powerful AI systems do not cause catastrophic harm
  • I see “AI alignment” as a subfield of AI safety that aims to ensure that powerful AI systems do not pursue goals that are bad for humanity
  • I think of “multipolar safety” as another subfield of AI safety, separate from AI alignment
  • I am somewhat uncertain about what falls under “AI ethics” - my tentative framing is that it is about ensuring that AI systems are fair and inclusive, and that there is some overlap between this and AI safety

What Cooperative AI is not (in my current model):

  • Most work on AI ethics is not cooperative AI: e.g. detecting and removing bias in AI systems
  • Most work on beneficial AI is not cooperative AI: e.g. using AI tools to improve areas such as healthcare, agriculture, education
  • While work on alignment is often relevant for cooperative AI, AI alignment of one system to one set of values is not cooperative AI
  • AI safety includes work that is neither alignment nor cooperative AI, for example work that deals with harm caused by malicious use of AI systems

What Cooperative AI is (in my current model):

  • Multipolar safety is a subset of cooperative AI: Everything that is to do with preventing serious harm arising from a dynamic with multiple very powerful (potentially superhuman) AI systems in the world
  • Cooperative AI includes aspects of AI safety that is not necessarily related to the most powerful systems (and therefore outside of “multipolar safety”), as catastrophically harmful dynamics could occur even with more limited systems
  • Cooperative AI also includes aspects of beneficial AI outside of AI safety that is about realising the positive potential of advanced AI to provide solutions for large-scale cooperation among humans. To count as cooperative AI in my mind, such work should be about improving the general capabilities of AI systems for solving cooperation problems (though a specific use case might be used for testing and demonstration).

What I am still very uncertain about:

  • What (if any) is the overlap of cooperative AI, AI ethics, and AI safety? Perhaps preventing catastrophic harm that is somehow tied to failures of fairness or inclusion?
  • What (if any) is the overlap of AI ethics and cooperative AI, outside of AI safety? Perhaps some work on preference aggregation and AI enabled improvements to fair governance systems?


 

49

0
0

Reactions

0
0

More posts like this

Comments7
Sorted by Click to highlight new comments since:

No need to reply to my musings below, but this post prompted me to think about what different distinctions I see under “making things go well with powerful AI systems in a messy world.”

That said, my current favourite explanation of what cooperative AI is is that while AI alignment deals with the question of how to make one powerful AI system behave in a way that is aligned with (good) human values, cooperative AI is about making things go well with powerful AI systems in a messy world where there might be many different AI systems, lots of different humans and human groups and different sets of (sometimes contradictory) values.

First of all, I like this framing! Since quite a lot of factors feed into making things go well in such a messy world, I also like highlighting “cooperative intelligence” as a subset of factors you maybe want to zoom in on with the specific research direction of Cooperative AI.

Another recurring framing is that cooperative AI is about improving the cooperative intelligence of advanced AI, which leads to the question of what cooperative intelligence is. Here also there are many different versions in circulation, but the following one is the one I find most useful so far:

Cooperative intelligence is an agent's ability to achieve their goals in ways that also promote social welfare, in a wide range of environments and with a wide range of other agents.

As you point out, a lot of what goes under “cooperative intelligence” sounds dual-use. For differential development to have a positive impact, we of course want to select aspects of it that robustly reduce risks of conflict (and escalation thereof). CLR’s research agenda lists rational crisis bargaining and surrogate goals/safe pareto improvements. Those seem like promising candidates to me! I wonder at what level to best to intervene with a goal of installing these skills and highlighting these strategies. Would it make sense to put together a “peaceful bargaining curriculum” for deliberate practice/training? (If so, should we add assumptions like availability of safe commitment devices to any of the training episodes?) Is it enough to just describe the strategies in a “bargaining manual?” Do they also intersect with an AI's “values” and therefore have to be considered early on in training (e.g., when it comes to surrogate goals/safe pareto improvements)? (I feel very uncertain about these questions.)

I can think of more traits that can fit into, “What specific traits would I want to see in AIs, assuming they don’t all share the same values/goals?,” but many of the things I’m thinking of are “AI psychologies”/“AI character traits.” They arguably lie closer to “values” than (pure) “capabilities/intelligence,” so I’m not sure to what degree they aren’t already covered by alignment research. (But maybe Cooperative AI could be a call for alignment research to pay special attention to desiderata that matter in messy multi-agent scenarios.)

To elaborate on the connection to values, I think of “agent psychologies” as something that is in between (or “has components of both”) capabilities and values. On one side, there are “pure capabilities,” such as the ability to guess what other agents want, what they’re thinking, what their constraints are. Then, there are “pure values,” such as caring terminally about human well-being and/or the well-being (or goal achievement) of other AI agents. Somewhere in between, there are agent psychologies/character traits that arose because they were adaptive (in people it was during evolution, in AIs it would be during training) for a specific niche. These are “capabilities” in the sense that they allow the agent to excel at some skills beneficial in its niche. For instance, consider the cluster of skills around “being good at building trust” (in an environment composed of specific other agents). It’s a capability of sorts, but it’s also something that’s embodied, and it comes with tradeoffs. For comparison, in role-playing games, you often have only a limited number of character points to allocate to different character dimensions. Likewise, the AI that’s best-optimized for building trust probably cannot also be the one best at lying. (We can also speculate about training with interpretability tools and whether it has an effect on an agent's honesty or propensity to self-deceive, etc.) 

To give some example character traits that would contribute towards peaceful outcomes in messy multi-agent settings:

(I’m mostly thinking about human examples, but for many of these, I don’t see why they wouldn’t also be helpful in AIs as well with “AI versions” of these traits.)

Traits that predispose agents to steer away from unnecessary conflicts/escalation: 

  • Having an aversion to violence, suffering, other “typical costs of conflict.” 
  • ‘Liking’ to see others succeed alongside you (without necessarily caring directly about their goal achievement).
  • a general inclination to be friendly/welcoming/cosmopolitan. Lack of spiteful or (needlessly) belligerent instincts.

Agents with these traits will have a comparatively stronger interest in re-framing real-world situations with PD-characteristics into different, more positive-sum terms.

Traits around “being a good coalition partner” or “being good at building peaceful coalitions” (these have considerable overlap with the bullet points above):

  • Integrity, solid communication, honesty, charitable, not naive (i.e., is aware of deceptive or reckless agent phenotypes, is willing to dish out altruistic punishment if necessary), self-aware/low propensity to self-deceive, able to accurately see other’s perspective, etc.

“Good social intuitions” about other agents in one’s environment: 

  • In humans, there are also intuition-based skills like “being good at noticing when someone is lying” or “being good at noticing when someone is trustworthy.” Maybe there could be AI equivalents of these skills. That said, presumably AIs would learn these skills if they’re being trained in multi-agent environments that also contain deceptive and reckless AIs, which opens up the question: Is it a good idea to introduce such potentially dangerous agents solely for training purposes? (The answer might well be yes, but it obviously depends on the ways this can backfire.)

Lastly, there might be trust/cooperation-relevant procedures or technological interventions that become possible with future AIs, but cannot be done with humans:

  • Inspecting source codes. 
  • Putting AIs into sandbox settings to see/test what they would do in specific scenarios.
  • Interpretability, provided it makes sufficient advances. (In theory, neuroscience could make similar advances, but my guess is that mind-reading technology will arrive earlier in ML, if it arrives at all.) 

To sum up, here are a couple of questions I'd focus on if I were working in this area: 

  • To what degree (if any) does Cooperative AI want to focus on things that we can think of as “AI character traits?” If this should be a focus, how much conceptual overlap is there with alignment work in theory, and how much actual overlap is there with alignment work in practice as others are doing it at the moment? 
  • For things that go under the heading of “learnable skills related to cooperative intelligence,” how much of it can we be confident is more likely good than bad? And what’s the best way to teach these skills to AI systems (or make them salient)?
  • How good or bad would it be if AI training regimes are the way they are with current LLMs (solo-competition, AI is scored by human evaluators) vs whether training is multi-agent or “league-based” (AIs competing with close copies, training more analogous to human evolution). If AI developers do go into multi-agent training despite its risks (such as the possibility for spiteful instincts to evolve), what are important things to get right?
  • Does it make sense to deliberately think about features of bargaining among AIs that will be different from bargaining among humans, and zoom in on studying those (or practicing with those)?

A lot of the things I pointed out are probably outside the scope of "Cooperative AI" the way you think about it, but I wasn't sure where to draw the boundary, and I thought it could be helpful to collect my thoughts about this entire cluster of things in once place/comment.

Thanks, I found this post helpful, especially the diagram.

What (if any) is the overlap of cooperative AI […] and AI safety?

One thing I’ve thought about a little is the possiblility of there being a tension wherein making AIs more cooperative in certain ways might raise the chance that advanced collusion between AIs breaks an alignment scheme that would otherwise work.[1]

  1. ^

    I’ve not written anything up on this and likely never will; I figure here is as good a place as any to leave a quick comment pointing to the potential problem, appreciating that it’s but a small piece in the overall landscape and probably not the problem of highest priority.

Thanks - yes I agree, and study of collusion is often included into the scope of cooperative AI (e.g. methods for detecting and preventing collusion between AI models is among the priority areas of our current grant call at Cooperative AI Foundation).

Oh, interesting, thanks for the link—I didn’t realize this was already an area of research. (I brought up my collusion idea with a couple of CLR researchers before and it seemed new to them, which I guess made me think that the idea wasn’t already being discussed.)

Executive summary: The author shares their understanding of cooperative AI as an emerging field focused on making things go well in a world with multiple powerful AI systems and diverse human values, distinct from but related to AI alignment.

Key points:

  1. Cooperative AI lacks a single agreed-upon definition, but broadly aims to promote cooperation and social welfare among multiple AI systems and humans with diverse values.
  2. Cooperative intelligence is an agent's ability to achieve goals in socially beneficial ways across varied environments and interactions, and is relevant for addressing social dilemmas between AI systems.
  3. The capabilities involved in cooperative intelligence (understanding, communication, commitment, norms) are dual-use and require caution.
  4. Cooperative AI overlaps with AI safety on multipolar scenarios and catastrophic risks, and with beneficial AI on using AI to foster large-scale human cooperation.
  5. Key uncertainties include the boundaries and overlaps between cooperative AI, AI ethics, and AI safety.

 

 

This comment was auto-generated by the EA Forum Team. Feel free to point out issues with this summary by replying to the comment, and contact us if you have feedback.

Thank you for this post. The short description of AI Ethics is interesting. I spent time thinking about this issue when researching private law and AI - I ended up effectively stumbling into a definition centred around safe design on one hand and fair output on the other, but that did not feel quite right. I prefer your broader takeaway of fairness and inclusivity. 

I really like the diagram! I found myself wondering where we would fit AI Law / AI Policy into that model. I think it is a very useful tool as an explainer. 

Thank you Shaun!

I found myself wondering where we would fit AI Law / AI Policy into that model.

I would think policy work might be spread out over the landscape? As an example, if we think of policy work aiming to establishing the use of certain evaluations of systems, such evaluations could target different kinds of risk/qualities that would map to different parts of the diagram?

Curated and popular this week
Relevant opportunities