Welcome back! AIs are edging closer to your health, your records, and even the quiet moments when you worry alone. One system now tries to compress medical questions into a single summary, another wants to follow you around as a voice that remembers everything, a third is wiring itself into hospital paperwork, and a simple browser test turns your laptop into a rough screening tool. It’s all the same pressure point: how much of your body and your peace of mind you’re willing to hand to software.
In today’s Generative AI Newsletter:
Google AI Overviews give risky health advice.
OpenAI Tolan builds always-on voice companion.
Anthropic Claude enters HIPAA-ready clinical workflows.
Visual Field Test screens at-home peripheral vision.
Latest Developments
Reporters at the Guardian ran a brief experiment on Google Search, testing Google’s “AI Overviews” on health queries and found answers that seemed reliable but lacked essential context. According to the experiment, one overview advised pancreatic cancer patients to avoid high-fat foods, a recommendation that experts criticized as completely incorrect and potentially dangerous. Another dumped “masses of numbers” for liver test “normal ranges” with little explanation of what changes those numbers.
Here’s what the reporting uncovered:
Action Taken: Google disabled AI Overviews for some liver test range searches after the investigation.
Workaround: Slight rephrases like “LFT reference range” could still trigger an AI overview.
Defense: Google said the vast majority of overviews are accurate and compared them to featured snippets.
Stakes: With Google accounting for roughly 91% of global search traffic, one bad summary can spread quickly.
The pursuit of AI companies remains focused on achieving a single clear response that speeds up the process and convenience. AI overviews can help when you need a quick explainer and point you to reliable sources. However, health is where a smooth summary can do the most damage, as individuals often interpret a confident tone as definitive in health matters. For Google to prioritize AI answers, it must establish stricter guidelines for medical topics, enhance sourcing transparency, and offer safer options. Otherwise the future looks like a chaotic game of health information gambling.
Special highlight from our network
Your sales team talks to leads on WhatsApp. But where do those chats go? Usually, nowhere useful.
TimelinesAI connects your existing WhatsApp accounts to HubSpot, Pipedrive, Zoho, and more. No WhatsApp Business API required.
Shared inbox for your whole team
Every message logged to your CRM automatically
No-code Workflow Builder for follow-ups, assignments, and alerts
Works with both personal and WhatsApp Business accounts. Setup takes minutes.
OpenAI has introduced Tolan, a voice-first AI experience built by the Portola team. It marks a departure from traditional text-based prompting, moving instead toward continuous, natural conversations. Tolan is designed to function as an “Embodied Companion,” which is an intelligence that remembers your history, adapts to your mood, and maintains a consistent personality across every interaction. It’s the first real step toward AI as an ambient presence rather than a digital tool.
How Tolan is being built:
Conversational Volatility: The system is designed to pivot mid sentence as humans naturally do.
Latency Discipline: Sub second response time is treated as a core feature, not an optimization.
Memory Design: Past interactions are compressed and retrieved, not stored as full transcripts.
Context Renewal: Each turn rebuilds understanding instead of stretching prompts longer.
Tolan represents the industry's departure from the glowing rectangle as the primary gatekeeper of information. By moving the interface into the background, OpenAI is estimating that the most successful AI is the one we eventually forget is even there. This "Jarvis" moment suggests a future where our devices are no longer apps we open, but partners we simply speak to. The interface is finally disappearing, leaving only the conversation behind.
Anthropic is aggressively moving into the medical sector with the launch of Claude for Healthcare and a major expansion of its life sciences suite. Just days after OpenAI’s recent push, this update targets the entire ecosystem, from patients deciphering lab results to pharma giants managing clinical trials. The initiative attempts to transform Claude into a HIPAA-ready orchestrator for the most fragmented data on earth.
Claude’s new clinical architecture:
Personal Health Connectors: Pro and Max subscribers can now sync medical records and fitness data from Apple Health and Android Health Connect for summarization.
Interoperability Skills: The update introduces FHIR development agents to help engineers bridge the gap between legacy hospital systems and modern AI platforms.
Regulatory Automation: Specialized life sciences tools automate clinical trial protocol drafting and help researchers navigate complex FDA and NIH guidelines.
Compliance Infrastructure: The platform is now fully HIPAA-ready, enabling payers and providers to securely automate prior authorization reviews and insurance claims.
The industry will judge not by benchmarks, but by whether this software earns trust where paperwork, liability, and human lives intersect. Anthropic is careful in how it frames this shift. Claude is positioned as an assistant that organizes, explains, and drafts, while deferring decisions to clinicians and regulators. Still, the ambition is clear. By embedding AI into billing rules, trial systems, and patient records, Anthropic is testing whether models can operate where mistakes carry legal and human consequences.
Visual Field Test is a browser-based test that measures side vision light dot detection. One eye at a time, you look at the screen's center and click when a flash appears. It is not a diagnosis, but it can help you find repeated places where you could not see the light dots, which you can mention at your next eye exam.
Core functions (and how to use them):
One-eye testing: Take the test with your left eye covered, then switch. This allows comparisons instead of guessing.
Test modes: Select test mode (macula, center 10°, 24-2, 30-2, wider fields). Start with 24-2 for a general scan.
Distance setup: Follow the site's screen-size viewing-distance instructions to ensure the dots land correctly.
Results maps: A visual field map shows where you clicked successfully and missed the light dots after the run. Take a screenshot to compare.
Trend monitoring: Save findings for days or weeks to check if the same locations show up as light dots were not visible, which is more valuable than a single run.
Try this yourself:
Take the Standard Field (24-2) test twice today, one per eye, and screenshot both results. Repeat the two tests tomorrow in the same light and time. Put the four screenshots in a folder and mark them “Right eye Day 1/Day 2” and “Left eye Day 1/Day 2.” The same cluster of missed spots in the same position on both days gives you evidence to show an optometrist instead of just saying “my vision feels weird.”
Special highlight from our network
Every week, learn how to use new AI tools, get exclusive discounts, and enter to win free access to them.
