Artificial intelligence (AI) tools are overwhelmingly being used in mental health practices to support administrative tasks, but University of Queensland researchers have found they potentially lack privacy, accuracy and governance safeguards.
The researchers surveyed 278 clinicians about their use of AI in the workplace, as well as conducting in-depth interviews with 12 others.
Nearly 80 per cent of mental health clinicians and practices reported using AI to listen in on and then transcribe sessions with clients.
Many used AI for a range administrative purposes but found it to be substantially less reliable when assessed against complex clinical scenarios.
PhD candidate Benjamin Johnson from UQ’s National Centre For Youth Substance Use Research (NCYSUR), said the findings show AI has already been embraced by practices, despite a lack of strong industry guidelines.
“AI seems to be a very suitable tool as a supplement and complement for clinicians, particularly for routine administrative work,” Mr Johnson said.
“But it’s not in a position to replace clinicians when dealing with complex cases that require accurate and safe diagnosis, and treatment recommendations.”
The study found 43.5 per cent of clinicians surveyed used AI every day for at least one administrative or clinician support task such as drafting reports, writing case notes, research preparation and transcribing client sessions.
“Most clinicians with experience using AI rated it better than a human for tasks such as preparing documents, research and note-taking,” Mr Johnson said.
“They saw this as beneficial, allowing more time with clients, being more present during sessions instead of taking notes and giving them more time to make clinical decisions.
“Mental health work is demanding and can be emotionally exhausting, and many saw AI as a tool which could alleviate some of the cognitive pressures when doing admin tasks.”
In a separate study, the UQ researchers reviewed 66 studies evaluating large language models such as ChatGPT in mental health settings.
Co-author and senior NCYSUR researcher Associate Professor Janni Leung , said most clinicians agreed AI was not a replacement for therapists, their clinical judgement or therapeutic relationships.
“Many of the studies assessed AI using simulated clinical scenarios rather than real-world patient care, which is why we wanted to examine how clinicians are actually using these tools in everyday practice,” Dr Leung said.
“As simulated cases became more complicated, AI’s performance relative to humans got worse.
“The therapeutic relationship is fundamentally human.”
Surveyed clinicians also raised concerns around privacy, data security, accuracy and client consent.
About 22 per cent of mental health workplaces banned AI during client sessions, while 5 per cent strictly banned AI for administrative or documentation tasks.
Conversely, 31 per cent said their workplace encouraged appropriate AI use for administrative or documentation tasks, while 30 per cent reported it was encouraged during client sessions.
With AI adoption outpacing governance, clinicians said they wanted practical guidance rather than blanket bans.
“People are using these tools whether or not their organisations permit it, so clear, industry-wide standards could improve client safety, privacy protections and clinician confidence,” Mr Johnson said.
Read the research and survey findings in the Journal of Medical Internet Research (JMIR) and the systematic review on how large language AI models performed in mental health settings in the JMIR AI.
