AI systems need more robust crisis detection protocols, discourage delusional and harmful thinking, and refuse to give advice on how to harm themselves, period. Adam deserved better than platitudes about the imperfections of AI and safeguards failing.

I've spent over a decade studying severe mental illness, AI, and social media. These risks were documented as early as 2016 in my own work about severe mental illness, social media, and algorithms! Since 2016! We knew that AI systems could propagate this.

It's heartbreaking doing research on the abstract risks of algorithms, social media, and AI - I've been doing that for years. But it's a whole other level of upsetting to see them manifest in real time.

8) Want more details on our FAccT paper? Find it here: github.com/jlcmoore/llm...

And join me on my substack to follow along! notatechdemo.substack.com

7) This blog post summarizes our recent FAccT paper about mental health, chatbot safety, and therapeutic practices.

But, I also reflect on my decade in therapy and what I've learned. It's not safe and not ready for replacing a therapist - it can't replace the messy process of healing.

6) And last week, I posted about why AI can't replace your therapist. notatechdemo.substack.com/p/why-ai-can...

The results are concerning: chatbots gave bridge heights to suicidal users and confirmed peoples' delusions. Chatbots also miss the human connection and vulnerability for progress.

5) So, in short - I'm skeptical. We're asking people to self-regulate something they don't understand. By the time someone is having a delusional conversation with a chatbot... a skippable timer isn't going to help. That's like jumping in front of a boulder rushing down a hill.

4) But I did a study in 2016 about skippable nudges like this on Instagram and dangerous mental health communities. Content on hashtags with the skippable nudges of graphic content had more dangerous mental health content go up over time.

3) I dive into the meta-analysis of why screen timers have very small effects mental health or other measures of well-being. Truly, use time is a very imprecise measure of what it means to "use" tech. Maybe AI would be different, and timers with skip-thru to chat options would work.

2) First, fresh up this week - do gentle timers get people to stop using ChatGPT in unsafe ways?

notatechdemo.substack.com/p/will-timer...

OpenAI released a timer/nudge feature to stop long sessions. They say it was to support more compassionate use. But is that the case?

💯💯💯💯

Come follow along if you'd like👇👇👇👇

🔗 notatechdemo.substack.com

My first welcome post went live yesterday. 🎉🎉🎉

If you're interested in research-backed takes on AI, practical productivity ideas, or just want to follow along as I figure out Substack, I'd love to have you come join

Here's what you'll get -
➡️ Cautiously optimistic takes on AI
➡️ Ideas on how we can fix social media to support mental health
➡️ Behind-the-scenes of how to be a writer/professor
➡️ Honest thoughts about building technology that actually helps humans instead of being over-hyped

No way! Agre is so good. If you enjoy his writing style, Computation and Human Experience is dense and GREAT. But not a starting piece of his, hah.