The distinction matters because it identifies where the AI debate often goes astray. Much popular discussion treats embodiment as the decisive barrier — machines cannot understand because they have no bodies. Collins's framework does not deny the importance of embodiment, but it locates the more consequential barrier in social participation rather than physical embodiment. A well-embodied AI without social membership would still lack collective tacit knowledge. A disembodied AI with hypothetical social membership would still possess something real.
For the AI transition described in Segal's You On AI, the somatic dimension matters in some domains and less in others. Surgery, craftwork, physical engineering — these domains require somatic competence that current AI systems cannot replicate. Much of software development requires somatic competence that is relatively simple (typing, reading code on screen), with the cognitive dimensions dominating. Collins's framework predicts that AI will integrate more easily into cognitively-dominated domains than into somatically-dominated ones — but that even in cognitively-dominated domains, the collective tacit knowledge barrier remains.
The concept was refined across Collins's work in the 2000s, consolidated in Tacit and Explicit Knowledge (2010). Collins's position on somatic tacit knowledge has been shaped by his engagement with Hubert Dreyfus, whose What Computers Still Can't Do emphasized embodiment as the central barrier to AI. Collins accepts the importance of embodiment but redistributes the analytic weight toward the social.
Body as learner. The body has its own intelligence, learned through repetition, operating below conscious access.
Engineerable in principle. Unlike collective tacit knowledge, somatic knowledge could in principle be given to a suitably instrumented machine.
Not the decisive barrier. Collins locates the structural barrier to AI in social participation rather than bodily embodiment.
Variable by domain. Some fields (surgery, craft) are heavily somatic; others (software, writing) are less so — AI integration will reflect this variation.