One pillar of Dunbar’s account is that increasing social sophistication depends upon the ability to reason about mental states with increasingly high levels of recursion. Relevant evidence of such recursive mentalising comes from two sources: tests that explicitly require recursive mentalising and tests where recursive mentalising is one available interpretation of behaviour on strategic games. Neither source of evidence suggests that levels of recursive ability vary in the ways required to support Dunbar’s claims.

Explicit tests of adults’ recursive mentalising build on classic tests of children’s reasoning about 1^st order false beliefs and 2^nd order beliefs about false beliefs (e.g., Wimmer & Perner, Reference Wimmer and Perner1983; Perner & Wimmer, Reference Perner and Wimmer1985). Dunbar and colleagues (Oesch & Dunbar, Reference Oesch and Dunbar2017; see also Kinderman, Dunbar, & Bentall, Reference Kinderman, Dunbar and Bentall1998; O’Grady et al., Reference O’Grady, Kliesch, Smith and Scott-Phillips2015; Stiller & Dunbar, Reference Stiller and Dunbar2007) developed social narratives and asked participants about the validity of statements such as:

This question is designed to test 9 levels of recursive mentalising, and participants typically answer questions from level 1 to level 9. As described in the target article, most adults appear to manage 4 or 5 levels, with some appearing to manage 6-8, and an extensive programme of studies finds correlations between individual differences in levels of success and many socially relevant factors including cooperativeness (Paal & Bereczkei, Reference Paal and Bereczkei2007), social network size (Stiller & Dunbar, Reference Stiller and Dunbar2007), grey matter volume in regions of the “social brain” (Lewis et al., Reference Lewis, Rezaie, Brown, Roberts and Dunbar2011; Powell et al., Reference Powell, Lewis, Dunbar, García-Fiñana and Roberts2010), and (negative correlations) with psychopathic traits (Vonk et al., Reference Vonk, Zeigler-Hill, Ewing, Mercer and Noser2015).

A weakness in this otherwise impressive evidence base is its exclusive reliance on questions structured in the same way as the example above, because any problem with this method endangers the conclusions based upon it. Wilson et al. (Reference Wilson, Hruby, Perez-Zapata, van der Kleij and Apperly2023) identified a critical problem by demonstrating that these questions are prone to high levels of false positives. In the above example, the correct answer “false” would be given by someone who reasoned at 9 levels of recursion, but also by someone who reasoned at any level above level 1. Wilson et al. compared performance on examples of the original stimuli with new narratives and questions that greatly reduced the potential for false positives. On the original stimuli, adults performed at the high levels previously reported, but much more poorly on the new narratives, even when participants were strongly incentivised and given unlimited time to respond. With the normal incentives in place, above-chance performance was only observed for level-5 reasoning in one out of four cases, and even then only 9/33 participants gave the correct answer. Presumably, many adults are capable of higher levels of recursive mentalising given external support from pen, paper, or spreadsheet. However, without such support many showed little evidence of reasoning above level-2, as observed in many 7-year-old humans. This revised estimate of adult humans’ recursive reasoning performance leaves little room for variance that could explain differences between individuals or species, as required by Dunbar’s account.

Evidence from strategic behaviour is also relevant and is both cited and analysed in Dunbar’s target article. However, it is no more supportive of high levels of recursive mentalising. Such studies have participants play simple strategic games such as hide-and-seek over multiple trials against artificial opponents that implement strategies of varying sophistication. For example, a simple “level-0” opponent “hider” might adapt their hiding to avoid the location where a participant had looked more frequently on previous trials. A “level-1” participant seeker who took account of this strategy (and so biased their search away from their previous search pattern) would find the level-0 hider more frequently. A “level-2” hider would model such a dependency of the seeker’s behaviour on the hider’s past behaviour. They would again be beaten by a “level-3” participant seeker who took account of this strategy to bias their search behaviour.

Such studies are elegant demonstrations of strategic behaviour at varying levels of sophistication (e.g., Charpentier & O’Doherty, Reference Charpentier and O’Doherty2018; Devaine et al., Reference Devaine, Hollard and Daunizeau2014; Lee & Seo, Reference Lee and Seo2016; Yoshida, Dolan, & Friston, Reference Yoshida, Dolan and Friston2008; Yoshida et al., Reference Yoshida, Seymour, Friston and Dolan2010). However, these studies seldom provide evidence that human adults reason beyond “level 2,” a result that converges strikingly with evidence of strategic reasoning from behavioural economics (e.g., Camerer, Reference Camerer2003), and the findings from explicit recursive mentalising described above. Although evidence of higher levels of recursive sophistication has been reported (e.g., Yoshida et al., Reference Yoshida, Dolan and Friston2008), success reflects learning over many trials. This affords bootstrapping of performance whereby an increasingly sophisticated response strategy may develop over time even though the level of recursion on a given trial remains low (Wilson et al., Reference Wilson, Hruby, Perez-Zapata, van der Kleij and Apperly2023). Finally, no such study includes a control condition that allows recursive mentalising to be distinguished reliably from recursive reasoning about behaviour. Indeed, one of the first papers to adopt these methods states “we are not suggesting that players actually compute their optimal strategy explicitly or indeed are aware of any implicit inference on the opponent’s policy” (Yoshida et al., Reference Yoshida, Dolan and Friston2008, p. 12).

In sum, evidence from studies of strategic behaviour suggests that levels of recursion are generally low, that sophisticated behaviour is the product of learning not one-shot reasoning, and such studies do not clearly resolve the question about whether mentalising is involved at all.

Altogether, Dunbar’s hypothesis about cognitive solutions to the pressures of group living is left without a candidate social-cognitive process to underwrite it. This poses a challenge for the target article. It also exposes a critical weakness in current research on mentalising in adults, which provides clear grounds for believing that mentalising abilities vary over and above general cognitive abilities, but no satisfactory account of why (Yeung, Apperly & Devine, Reference Yeung, Apperly and Devine2024).