07 – Insights on AI Companions, Mind Reading with Podcasts

This week's top stories:

• How good can a fake friend be? Researchers from MIT have conducted a large-scale study on the relationship between loneliness and AI companion usage. Whilst they didn't find a direct connection, different moderators explained up to 50% of total variance in outcomes. Their findings have a list of practical implications for anyone building mental health solutions.
• Mind-reading lite. Researchers from the University of Texas Austin have built a system to decode thoughts. Well, not quite. It is sort of limited, more of a vibe-mind-reading. Regardless, their fMRI decoding algorithms pose significant advancement.

Odd Lots – Curated for mental health experts:

• Paper shows how to mitigate biases in health data for research purposes.
• Breakdown: How Ebb, the Headspace AI companion, was designed.
• Insights from a failed dementia startup.
• Report on an AI therapy self-experiment.

And a few more bonus stories.

Researchers conducted the biggest RTC trial on AI companion usage to date. Let's get straight to the point: This is the core implication for founders and investors facing an increasingly crowded AI mental health landscape:

As AI companions become increasingly prevalent, these findings have crucial implications for design and deployment. The identification of distinct user profiles and key mediating factors suggests the need for personalized approaches and built-in mechanisms to detect and intervene in potentially harmful usage patterns. The contrasting outcomes between different user types raise important questions about how to design AI companions that support psychosocial well-being.

Soo, How do you design a good chatbot that helps more than it hurts? Let's take a more detailed look.

Researchers at MIT have conducted a large scale study examining the relationship between AI companion usage and loneliness.

They found seven clusters of users, some of whom see a worsening of loneliness through usage. Loneliness is not a clear predictor, but usage type is. Here they are:

Base on this and other analysis, the authors conclude a list of concrete suggestions for HCI design and practice:

• AI chatbots need mechanisms to detect potentially harmful usage in an actionable way, allowing direct intervention in problematic cases (potentially even before something goes seriously wrong).
• Providers should offer resources on maintaining healthy human relationships alongside chatbot usage.
• If user traits can be collected, interfaces for those "social chatbots" should be adapted to the individual. This way, certain uses may be encouraged or discouraged to mitigate risks of AI use. The authors specifically mention people with high neuroticism.
• Not only interfaces, but also the chatbots themselves should be tailored to different user types in the features and interaction styles they present. This way, vulnerable users would be encouraged to engage in real social interaction.
• Design of chatbot-based interventions needs to become more flexible in accommodating different user types. There is no one-size fits all solution.

I have some more personal takeaways to add:

• All this personalisation should be clearly communicated to the user, understanding their vulnerabilties might be a good way to internally prevent risky uses.
• Startups can't afford to turn down users, especially in early stages. This poses a significant conflict of interest. External audit systems and stronger regulation are needed.

Back to theory: A more detailed overview of what they actually did:

• The study focusses on relationship between chatbot usage (companion chatbot) and loneliness. It is the first large scale study on the issue. The authors develop a model explaining approximately 50% of variance in loneliness, but usage is not a direct predictor of loneliness.

Background – loneliness is a massive issue

Chronic loneliness is associated with a significantly increased risk of mortality—approximately 26-29%, comparable to smoking 15 cigarettes a day or having an alcohol use disorder [Freedman and Nicolle(2020)].

There is other existing interventions, but outcomes are mixed:

(1) modifying maladaptive social cognition (e.g., cognitive behavioral therapy, psychodynamic therapy, reminiscence, and mindfulness)

(2) improving social support (e.g., involving care professionals in health and social care provision)

(3) facilitating social interaction (e.g., clubs, shared interest groups, and videoconferencing)

(4) enhancing social skills (e.g., training programs and computer courses)

Results

Focus on the top right and bottom left graphs here. The high share of emotional and social topics and usage motivations signals the relevance of the issue. Users are already taking matters into their own hands. The authors highlight the following:

Emotional disclosure emerged as a notable theme, encompassing mental health discussions and personal problems. This is reflected in the quantitative findings of personal issues and mental health (14.17%) and interpersonal issues and drama (10.96%) as notable topics. One participant responded, “I can talk about my problems, and it’s like having a private conversation with no fear of being criticized.”

In case you are interested, here is the final model describing the relationship between chatbot usage and loneliness:

This is how the authors operationalize problematic use:

Problematic use emerged as a significant mediator, suggesting that increased usage may create more opportunities for problematic behaviors to develop. This aligns with existing research connecting problematic internet use as both a cause and effect of loneliness [Kim et al.(2009)], indicating a similar pattern may exist for companion chatbot use.

And implications for future research:

Long-term longitudinal surveys are still needed. Currently, noone can tell how such artifical companionship shapes child development or health outcomes for the elderly, to only mention two areas of research.

There is no experimental studies on different forms of interfaces and features to my knowledge. Between the chatbots present in the study there is significant variance in design and settings. The impact of those choices is not clear.

Read the full paper here.

Researchers at the University of Texas have pushed the boundaries of neuroimaging by training an AI model on the largest fMRI dataset of its kind. In a novel setup, participants listened to hours of podcasts while lying in an fMRI scanner. The resulting brain activity was labeled and used to train a model that predicts how specific words and phrases show up in neural patterns.

This is not about mind-reading in the sci-fi sense. Think of it more like reconstructing the gist of thoughts based on brain “vibes.” When tested on new, unseen podcast audio, the model could often generate paraphrases close to what the subject had heard. For instance, when a participant listened to “I don’t have my driver’s license yet,” the AI decoded the brain signal as “She has not even started to learn to drive yet.”

In a separate test, participants watched Pixar clips without dialogue. The model was still able to generate rough descriptions of what they had seen, based only on brain activity. However, the decoder only worked on individuals it had been trained on, so there is no general plug-and-play decoder yet.

What this could mean for mental health tech:
The ability to map language and perception to brain signals opens up longer-term possibilities for passive mental state decoding, assistive communication, or even tuning therapy content in real time. Today, this is a lab-only setup that requires hours of individual training data. But as non-invasive neuroimaging advances and models get better at generalizing, decoding intent, mood, or perception from brain activity could become part of clinical tools, especially in contexts where verbal communication is difficult.

For founders, this research is a signal: the interface between language, brain activity, and AI is maturing. The challenge will be translating this into scalable, privacy-aware, and clinically meaningful applications.

Read the full paper here.

I. AI can Improve Data Quality in Healthcare Research (Paper)

Biases in health data sets are a significant issue for researchers. But as the authors of this paper show, synthetic data can effectively enhance time series data and to mitigate those issues. Their novel CA-GAN generated data improving model fairness in Black and female patients.

II. Overview of all Relevant Tech Startups in Berlin (Data)

This startup map provides a great overview of Berlin based startups, structured by sector, funding, and institution.

III. PHTI: Virtual Solutions for Depression and Anxiety (Paper)

In this report, the Peterson Health Technology Institute evaluated a list of digital interventions for depression and anxiety. They compare Self-Guided, Prescription Digital Therapeutics (PDTs) and Blended-Care Solutions.

IV. Headspace Companion Ebb. A Design Breakdown by Figma (Blog)

Figma breaks down the whole design process of Headspace's AI companion Ebb. The article provides a great list of design principles for building empathetic products.

V. Insights from a Failed Dementia Startup (Blog)

Duncan Reece shares insights from building a startup around dementia, Veronica Health. The company was built in the US, but there is still a significant share of transferable insights for European founders.

VI. FAIIR Assessment for Mental Health (Paper)

FAIIR stands for "Frontline Assistant: Issue Identification and Recommendation". The AI tool was built to support Critical Responders in triaging their patients. It demonstrated a high accuracy in recommendations for further treatment.

VII. Bio Markers for Mental Health (Paper)

Recently, testing companies like Function Health saw a massive hype. Whilst they primarily focus on physiological health, interest in mental health markers is also picking up. The paper looks at genetic markers related to immune response and examines their relationship to mental disorders.

VIII. Spotting Red Flags in Mental Health Discussions (Blog)

Brittainy Lindsey lists 10 different red flags that misrepresent the role of AI in therapy. She intelligently highlights how narratives distort reality when it comes to pushing for more tech in mental health.

Great read for mental health advocates to avoid common pitfalls in building and communicating their solutions.

IX. AI Therapist Experiment: Experience Report (Blog)

Joakim Achren shares his mixed experience in using ChatGPT 4o as a therapist for five weeks. He highlights strengths like nearly perfect recall and weights them against significant deficiencies in human features like empathy and intuition. What makes his report fascinating that he didn't simply start chatting but took the time to give the LLM a big load of context on his past.

He takes his insights and lists several options for improving the UI of the chatbot to improve the experience.

Alright, that's it for the week!

Best

Friederich

Join The Stream Newsletter

Get the weekly newsletter for professionals in mental health tech. Sign up for free 👇

Sign me up!