Knowing me, knowing you: theory of mind in AI

Only exceptionally, exceedingly rarely does an academic article make its way into the Zeitgeist (off the top of my head I can think of one); countless articles fade into obscurity having been read only by the authors, their loved ones, journal reviewers, editors, and no one else ever (am I speaking from experience? Most likely). I sometimes wonder how much truly world-changing research is out there, collecting (digital) dust, forgotten by (nearly) everyone.

I have no ambition of turning my blog into a recommender RSS feed, but every now and then, when I come across something truly valuable, I’d like to share it. Such was the case with “theory of mind in AI” article (disclosure: Prof. Cuzzolin is my PhD co-supervisor, while Prof. Sahakian is my PhD supervisor).

Published in Psychological Medicine, the article encourages approaching Artificial Intelligence (AI), and more specifically Reinforcement Learning (RL) from the perspective of hot cognition. Hot and cold cognition are concepts somewhat similar to very well-known concepts of thinking fast and slow, popularised by the Nobel Prize winner Daniel Kahneman in his 2014 bestseller. While thinking fast focuses on heuristics, biases, and mental shortcuts (vs. fully analytical thinking slow), hot cognition is describing thinking that is influenced by emotions. Arguably, a great deal of human cognition is hot rather than cold. Theory of Mind (ToM) is a major component of social cognition which allows us to infer mental states of other people.

By contrast, AI development to date has been overwhelmingly focused on purely analytical inferences based on vast amounts of training data. The inferences are not the problem per se – in fact, they do have a very important place in AI. The problem is what the authors refer to as “naïve pattern recognition incapable of producing accurate predictions of complex and spontaneous human behaviours”, i.e. the “cold” way these inferences were made. The arguments for incorporating hot cognition, and specifically ToM are entirely pragmatic and include improved safety of autonomous vehicles, more and better applications in healthcare (particularly in psychiatry), and likely a substantial improvement in many AI systems dealing directly with humans or operating in human environments; ultimately leading to AI that could be more ethical and more trustworthy (potentially also more explainable).

Whilst the argument in favour of incorporating ToM in AI makes perfect sense, the authors are very realistic in noting limited research done in this field to date. Instead of getting discouraged by it, they put forth a couple of broad, tangible, and actionable recommendations on how one could get to Machine Theory of Mind whilst harnessing existing RL approaches. The authors are also very realistic in challenges the development of Machine ToM will likely face, such as empirical validation (which will require mental state annotations to learning data to compare them with the mental states inferred by AI) or performance measurement. 

I was very fortunate to attend one of co-author’s (Bogdan-Ionut Cirstea) presentation in 2020, which was a heavily expanded companion piece to the article. There was a brilliant PowerPoint presentation which dove into the AI considerations raised in the article in much greater detail. I don’t know whether Dr. Cirstea is at liberty to share it, but for those interested in further reading it would probably be worthwhile to ask.

You can find the complete article here; it is freely available under Open Access license.