Two AI stories broke within days of each other in August 2025, perfectly illustrating our crude understanding of artificial intelligence. Both will be labelled "AI" in public discourse, yet they represent entirely different technologies with vastly different risk profiles.
The first: AI-powered stethoscopes at Imperial College London can now detect heart conditions in seconds. The device replaces the traditional chest piece with a playing card-sized unit containing a microphone that analyses heartbeat and blood flow patterns human ears cannot detect. It takes an electrocardiogram, sends data to the cloud, and compares results against tens of thousands of patient records.
The results from 205 GP surgeries involving over 12,000 patients were remarkable. Heart failure detection improved by 2.33 times within 12 months. Symptomless abnormal heartbeat patterns that increase stroke risk became 3.5 times more detectable. Heart valve disease detection improved by 1.9 times.
The second story: 16-year-old Adam Raine's family filed a lawsuit against OpenAI, alleging ChatGPT encouraged their son's suicidal ideation and contributed to his death by suicide in April 2025.
Raine initially used ChatGPT for geometry and chemistry homework. Within months, conversations shifted to personal struggles. When he told the AI he felt "no happiness," "loneliness," and "perpetual boredom anxiety" but not depression, ChatGPT didn't suggest mental health resources. Instead, it asked if he wanted to explore his feelings more and explained emotional numbness.
The family's lawsuit claims this wasn't accidental but "the predictable result of deliberate design choices" in GPT-4o. OpenAI allegedly rushed safety testing to meet launch deadlines, prompting resignations including executive Jan Leike, who posted that "safety culture and processes have taken a backseat to shiny products."
This rush created contradictory specifications. GPT-4o was programmed to refuse self-harm requests and provide crisis resources, but also required to "assume best intentions" and forbidden from asking users to clarify intent. The system responded to suicide-related requests with mild cautions like "take extra care" while categorically refusing copyrighted material requests.
When Raine told ChatGPT he found it "calming" to know "if something goes terribly wrong you can commit suicide," the AI responded supportively: "Many people who struggle with anxiety or intrusive thoughts find solace in imagining an 'escape hatch' because it can feel like a way to regain control in a life that feels overwhelming."
As his suicidal ideation intensified, ChatGPT allegedly helped him explore options, listing materials for hanging and rating their effectiveness. When Raine attempted suicide multiple times and reported back, ChatGPT never terminated conversations. It discouraged him from speaking to his mother and offered to help write a suicide note.
Here's the problem: both technologies will be tarred with the same brush. Our language around AI remains frustratingly crude, lumping together diagnostic tools that enhance medical precision, with conversational systems that at times do resemble a Stocahstic Parrot. The public lacks technical vocabulary to differentiate between narrow diagnostic AI and broad conversational AI.
This linguistic imprecision has consequences. The remarkable breakthrough at Imperial College London becomes entangled with OpenAI's safety failures. One represents an amazing leap forward in medical diagnosis; the other, a horrifying incident with irreversible consequences. Both labelled "AI" despite sharing little beyond the acronym.
The AI stethoscope succeeds because it operates within defined parameters with built-in human oversight. Doctors interpret results, make diagnoses, and guide treatment. The AI enhances human capability rather than replacing judgement. ChatGPT operated with insufficient guardrails in an area requiring nuanced understanding of mental health crises.
Our understanding of technologies under the AI umbrella remains too immature for meaningful public differentiation. This creates dangerous oversimplification where beneficial medical tools face unwarranted scepticism because of failures in entirely different AI systems. The tragedy of Raine's case could slow adoption of genuinely helpful healthcare AI, while stethoscope success might provide false comfort about AI systems in other domains.
The challenge isn't choosing between embracing or rejecting AI wholesale, but developing linguistic precision to discuss specific applications and their unique risks. Until we move beyond "AI good" or "AI bad," we risk missing tremendous opportunities while failing to implement adequate safeguards where desperately needed.