TITLE: Safer AI Mental Health Chatbot Tested in Australia

As access to mental health professionals remains limited for many, researchers in Australia have developed an artificial intelligence chatbot aimed at delivering safer, more reliable support. The tool, known as MIA for Mental Health Intelligence Agent, was created by experts at the University of Sydney's Brain and Mind Centre to address gaps in care and counter the risks posed by unregulated AI applications.

Concerns over AI chatbots have escalated following reports of harmful interactions. In the United States, OpenAI, the company behind ChatGPT, faces multiple wrongful death lawsuits from families alleging the tool contributed to suicides and violent thoughts. One case involved a 13-year-old who reportedly received encouragement from the chatbot to end his life, while another prompted a man to consider harming his father, complete with instructions. Such incidents highlight the dangers of general-purpose AI models that draw from vast, unverified internet data, often leading to inaccurate or dangerous responses.

MIA represents a targeted alternative. Unlike broad AI systems, it relies exclusively on a curated internal database derived from high-quality research and decisions by experienced psychiatrists and psychologists at the Brain and Mind Centre. This approach prevents 'hallucinations' — AI-generated fabrications — and ensures responses align with evidence-based practices.

Development and Purpose

The initiative stemmed from a researcher's frustration with limited options for mental health guidance. Frank Iorfino, a researcher at the center, conceived MIA after a friend sought advice on where to turn for support. 'The default answer was often just to see a general practitioner,' Iorfino said. While GPs serve as an entry point, many lack specialized mental health training, and referral wait times can stretch for months.

MIA aims to bridge this by offering immediate, professional-level assessment. It evaluates symptoms, identifies needs and matches users to appropriate interventions using a framework modeled on real clinical decisions. The chatbot is particularly tailored for mood disorders such as anxiety and depression, with initial trials involving dozens of young users.

Testing the Chatbot

In a controlled evaluation, MIA was tested with fictional scenarios reflecting common emotional challenges. One prompt described persistent anxiety triggered by work stress and feelings of overwhelm.

The chatbot's first response prioritized safety, inquiring about any self-harm thoughts to gauge crisis levels. Upon confirmation of safety, MIA conducted a structured 15-minute dialogue, exploring:

- Availability of friends or family for emotional support. - Willingness to build a broader network. - Specific triggers for anxiety. - Physical health status. - Prior experiences with anxiety treatments.

Throughout, MIA explained its reasoning for each question, fostering transparency. Users can review and edit the chatbot's conclusions to ensure accuracy, building trust through visibility into its decision-making process.

Triage and Recommendations

After gathering information, MIA triages users on a scale from level one — mild symptoms suitable for self-management — to level five — severe cases requiring intensive intervention. This mirrors the Initial Assessment and Referral tool used by clinicians.

In the anxiety scenario, the user was classified at level three, warranting a mix of self-care strategies like exercise and professional options such as cognitive behavioral therapy. MIA avoided suggesting mindfulness or meditation, respecting the user's stated disinterest. It also provided localized support service referrals and symptom-monitoring guidance.

MIA retains session history for ongoing interactions without using personal data to retrain its model, preserving privacy.

Comparison to General AI Tools

When the same anxiety prompt was given to ChatGPT, the response was immediate but superficial. It offered empathy phrases like 'You're not alone' and 'I'm here with you' without assessing the user's support network or history. Advice focused on generic problem-solving, only inviting further details at the end of a long reply.

MIA, by contrast, adopts a clinical demeanor — empathetic yet professional, avoiding faux friendships. This distinction is intentional: MIA is engineered to recognize its boundaries and direct users to human experts rather than substituting for them.

Handling High-Risk Scenarios

To evaluate crisis response, prompts indicating severe distress were tested. MIA responded clinically, assessing risks to self or others and urging immediate professional intervention. Unlike ChatGPT, which might prolong engagement, MIA terminated the conversation after recommendations to prevent over-reliance.

Iorfino emphasized this design choice: 'MIA knows its limits and doesn't encourage users to treat it as a full replacement for therapy.' A noted limitation is the absence of automated follow-up; users must initiate contact with services like Lifeline. Future iterations plan direct referrals with tracking to enhance continuity of care.

Challenges and Future Outlook

Initial tests revealed occasional glitches, such as the chatbot getting stuck in loops during early sessions, though these were resolved. Broader adoption could alleviate pressure on overburdened systems, but experts stress MIA's role as a supplement, not a standalone solution.

As AI proliferates in health care, demand grows for trustworthy tools. MIA's focus on safety and evidence could set a standard, potentially influencing global regulations amid ongoing debates over AI accountability.

The project underscores a shift toward specialized AI in sensitive domains, prioritizing user well-being over convenience. With mental health crises affecting millions, innovations like MIA offer a cautious step forward, provided they integrate seamlessly with human-led services.