A weekly attempt to isolate real trends from mere conversation — across AI & society, AI & cognition, AI & work, and health AI. Bubble size and signal strength are editorial judgments, not metrics theater.
Big Tech is racing into the exam room while 34 states race to regulate the chatbots already there — and the research literature on what AI is doing to our minds is graduating from anecdote to evidence.
What's bubbling up
Bubble size = signal strength. Click a bubble for the full read.
The timeline
The timeline starts this week — each Friday's edition adds a column, so theme trajectories emerge over time. Hover a dot for the story; click to read the source.
This week's trends
Health AI▲ Rising
Big Tech moves into the exam room
Microsoft announced a partnership with Mayo Clinic to build a healthcare-trained model and a patient-facing assistant inside hospital portals — with Mustafa Suleyman conceding it will take 'many years' to make it trustworthy for high-stakes questions. Hospitals are simultaneously launching their own chatbots as patient-acquisition funnels.
My take The interesting admission isn't the partnership — it's the timeline. The bottleneck was never model quality; it's grounding answers in the patient's actual record. The race is for context, not capability. The well has to be full by default. Related essay →
98 bills across 34 states now target companion AI. Oregon's law creates a private right of action at $1,000 per violation; Vermont banned therapy bots outright; and this week Pennsylvania sued Character.AI after finding multiple chatbots impersonating licensed physicians — complete with fabricated license numbers.
My take Look at what the laws converge on: disclose that you're not human, detect crisis, refer out. Legislatures are effectively mandating epistemic honesty — the 'I know not' that can't be prompted in is being demanded by statute instead. Related essay →
What was vibes a year ago is becoming a research program: a new arXiv paper formalizes 'cognitive divergence' — the widening gap between expanding AI context windows and shrinking human attention — while the Council on Strategic Risks is running a year-long debate series on AI and cognition as a national-security question.
My take The offloading studies keep measuring what we stop doing. The more useful question is architectural: can we design AI use that re-loads cognition — that returns the struggle at the moments struggle builds capacity? Related essay →
US job postings requiring AI skills are up 144% year over year, with wage premiums reported as high as 56% — while Amazon's ~16,000 corporate cuts and survey data showing six in ten companies planning AI-linked layoffs in 2026 tell the other half of the story. Reshaping, not replacement, but unevenly.
My take The wage data measures who can operate the machines. It says nothing about what the work still does for the people doing it — the identity, effort, and pride that the paycheck never captured. That ledger is still unaudited. Related essay →
How this works: each week I sweep research feeds, essays, and the news across these themes, separate durable trends from passing conversation, and write the takes myself. Sources are linked; disagreement is welcome — find me on LinkedIn.