Collective Intelligence in Human-AI Teams

Empirical Bayesian ToM improves Human-AI teams

First, an apology to my steadfast subscribers ; ) for the hiatus!
Over the past few months I have been working with multiple research groups from the SPAR program on topics of unadaptable foundation models, detecting misinformation online, and quantitative safety metrics for agents using an Active Inference framework. You’re invited to read more about them on my profile if you’re interested. I should probably even post about them here soon. Regardless, the program has officially concluded (though the projects are still in progress), so I can redirect my attention.

---

Here we have a work from 2023 by Westby and Riedl demonstrating that the addition of Bayesian agents to human communication networks improves the collective intelligence of the group, beyond what humans alone would achieve. This may not sound surprising at first when you read this as “AI in the mix makes humans smarter”, but that isn’t what’s going on here. As we will see, the AI aid is not from some knowledge base but in the way it enhances the theory of mind (ToM) projections of the human members. The authors further introduce metrics for measuring human ToM abilities.

ToM is the practice of inferring the beliefs, opinions, knowledge, etc., of the concealed mind of another agent/human. To achieve effective collaboration, this skill is essential. And we do it all the time. Any time you are predicting how someone might react, what they know, what their eyes are telling you, etc., that’s ToM and it’s non-trivial as to if and to what degree it is achievable in contemporary artificial agents. As the authors suggest (pg. 1),

As human civilization shifts further toward knowledge work (Autor 2014) where the most value is realized if members fully use and integrate their unique expertise, this ability is increasingly important.

Bayesian ToM Agent

Aspiring to human-AI collective intelligence (CI) goals in a team setting demands agents that adapt to ever-changing team dynamics, grasp the subtleties of human communication, and align their actions to support team goals. The authors construct a Bayesian agent that “can form ad-hoc mental models about its teammates based exclusively on observations drawn from human communication” (pg. 2).

Each human has a representative agent, and agents model themselves with an “Ego” model, and others with “Alter” models. The experimenters use teams of five to solve the Hidden Profile task in which players each have one of 5 private clues, and there is another shared public clue. The team then tries to ascertain the culprit with these. Communication channels are randomized and limited between players (Fig. 1a).

A player’s agent reads all their incoming and outgoing messages, updating their beliefs (Fig. 1b). The authors cite previous evidence that humans often miss opportunities to share their unique information under a false notion that others are already aware of it, “imputing their own knowledge on others”. (To be honest, I’m rather surprised to hear this, especially if they know their info is unique—which I suspect would be a more dull and predetermined experiment, so my guess is that players don’t know which of their info is private vs. public.)

And this is where the AI is anticipated to help: to notice when info is uniquely held when the human does not. By maintaining the ToM of their Alter models and their own Ego models, they can notice such discrepancies between them.

Results

Without going into all the details of the findings, I’ll outline a few interesting results:

1. Human Decision Biases. Humans do not pay enough attention to information that rules out alternatives. This is more pronounced with ambiguous information. There is a bias towards one’s own information over what others have. The agents were shown to, unsurprisingly, not carry this with them.

2. Predictive Power and Interventions. By observing the initial 25% of team communication, the Bayesian agents explained approximately 8% of the variation in the final team performance. This predictive capability allowed the agent to suggest timely interventions that improved overall performance by up to 4%.

3. ToM Team Performance. Teams that displayed better ToM abilities performed better. This effect is further enhanced for individuals with stronger ToM who also have high member connectivity. This result supports ToM as an important tool in team goals, and CI.

My Thoughts

What a cool demonstration. I think they could have explained the experiment more clearly. For instance, it wasn’t specified whether humans knew which info was private or not. A more clear example of this and just how the experiment would look like unfolding would’ve helped to visualize things.
However, I quite like this area of human-AI synergy. I particularly like the idea of AI helping counter biases that we hold—these are such pernicious things to squash since they fly so well under the radar.