Tested against therapy guides from major medical institutions, current language models responded inappropriately to critical conditions, even newer and larger ones, researchers reported

Event date / reported

Evidence stage

Tasks this bears on

Where this applies

A lab study that mapped therapy guides used by major medical institutions — for example the importance of a therapeutic alliance — and tested current language models such as gpt-4o against them. The models expressed stigma toward those with mental health conditions and responded inappropriately to certain common and critical conditions, including encouraging clients' delusional thinking, even in larger and newer models; the authors also point to barriers such as a therapeutic alliance requiring human characteristics. It tests models in constructed scenarios, not products or counsellors.

What this means

What it does not yet show

A preprint in constructed scenarios; deployed products and future models may differ.

What you can check

Does it change the assessment?

Source