A support-chat study revisits the illusion ELIZA first created

The design

'The Typing Cure: Experiences with Large Language Model Chatbots for Mental Health Support,' by Inhwa Song, Sachin R. Pendse, Neha Kumar and Munmun De Choudhury of the Georgia Institute of Technology, appeared in the Proceedings of the ACM on Human-Computer Interaction – the ACM Digital Library journal that carries CSCW conference papers – on 16 October 2025. The authors interviewed 21 people from globally diverse backgrounds who had used general-purpose LLM chatbots, not built for therapy, for mental-health support, analysing the interviews against psychotherapy literature and introducing the term 'therapeutic alignment' for designing AI toward therapeutic values.

What the evidence says

Participants described building 'unique support roles' for their chatbots and said the tools filled gaps in everyday care. A second ACM-family paper, 'All Too Human? Mapping and Mitigating the Risks from Anthropomorphic AI', by Canfer Akbulut, Laura Weidinger, Arianna Manzini, Iason Gabriel and Verena Rieser of Google DeepMind, published at the 2024 AAAI/ACM Conference on AI, Ethics, and Society, explains a likely mechanism: it states that although most earlier voice assistants used rule-based systems retrieving canned responses, 'users may come to expect that [they] are capable of understanding and generating language in real time,' and that fluent, humanlike output together with cues such as a typing indicator can extend this expectation to today's LLM-based chatbots. Neither paper uses the term 'ELIZA effect'; connecting Song et al.'s interview findings to that 1966 term, coined after people attributed real understanding to Joseph Weizenbaum's scripted therapy programme, is this outlet's editorial reading of the parallel, not a claim either paper makes itself.

What it asks of people

People using general-purpose chatbots for support are, per Song et al.'s interviews, extending trust, disclosure and emotional reliance to a system that no clinician designed, tested or vetted for that role. The DeepMind paper adds that some of these humanlike cues are 'an inadvertent byproduct' of how models are trained on human-authored text, not a deliberate therapeutic design choice by any developer.

Privacy and safeguards

Song et al. report lived experience, not a privacy or safety audit; they flag 'cultural limitations' in how chatbots respond across backgrounds and call for 'therapeutic alignment' as a design goal, implying it is not yet standard practice. The DeepMind paper similarly stops at proposing research directions for 'ethical foresight,' not a fix already implemented at any named company.

• Does a fluent, humanlike reply earn a user's clinical trust that the system's design never actually committed to?
• What would 'therapeutic alignment,' as this paper names it, need to include before a chatbot could be marketed for mental-health support?
• How would a person using a chatbot this way know whether their trust reflects the system's design or its absence?

Sixty years after Weizenbaum found users attributing understanding to a scripted programme, a peer-reviewed interview study finds people extending a similar trust to far more fluent systems that still carry no clinical judgment behind their replies.