Underfold
Long-ish post.. connecting Anthropic J-space > Anthropomorphism > Seemingly Conscious AI > a new Hazard for Mental Health under HSE guidance?
Anthropic dropped a paper on J-space yesterday, having only read the post (linked below) and not the full paper, do take my comments with a huge grain of salt. Also, I’m not a neuroscientist, so theres that too.
From a technical point of view this is interesting for future audibility of outputs, to understand what Claude is thinking below the surface is key to governing AI outputs now and in the future. As we all know CoT can, as far as I’m aware, only shows what the model chose to write down, not how the answer was achieved; feel free to correct that statement.
The ‘issue’ I have with Anthropic is its use of language, conscious, unconscious, mind… all anthropomorphising technology. I don’t believe anthropomorphism is a ‘good’ way to translate technology to the masses. The problem is everyone is anthropomorphising AI, this can lead to psychiatric issues, we’re seeing the cases play out already.
This triggered a thought from the last AI Ethics Workshop I attended about ‘Seemingly Conscious AI Risks’ (link below), which, shockingly (sarcasm) are the same as todays AI Risks; maybe I used my own J-space 😉.
A paragraph from the paper: Current AI governance approaches address system capabilities, application domains, and organizational [sic] risk processes but largely do not include risks that stem from users perceiving systems as conscious, independent of what those systems can do.
At the same time, organisations have a general duty of care towards health, safety and welfare, including mental health.
So, heres the contradiction, the academic paper is saying that organisations aren’t considering the impact of seemingly conscious AI on their workforce, while the HSE does bring a general duty of care. There is a risk that can crystallise.
Let’s play forward Claude anthropomorphism, a potentially vulnerable employee & then psychiatric crisis. Does the organisation have a general duty of care when implementing AI, I’d say yes, and increasingly so, especially for known vulnerable employees. The direction of travel for HSE interventions for work-related stress and even HSE February 2026 outlook signalled increased enforcement on this subject and psychosocial risks. AI-interaction harm isn’t in anyone’s hazard taxonomy yet, as far as I know.
Here we are, AI described in anthropomorphic terms, academia showing organisations are not considering the psychological impacts of deploying AI, and the HSE starting to move in that direction.
How does an employer address the cross issue of someone with a diagnosed mental health condition or disorder where they are susceptible to manipulation (or psychosis, managed or unmanaged), who are advised that they shouldn’t use AI, or only in limited scopes, how does an organisation protect the employee, and itself?
Just some rambling thoughts.
References: