The study: Researchers from Germany, India and the Netherlands presented it at the ACM Web Conference 2026. They analyzed donated chat logs from 80 real ChatGPT users, who got their data through European privacy laws, and found 2,050 memory entries across 65 users.

ChatGPT decides what to remember: 96% of memories were created by ChatGPT on its own, so users had little control over what was logged about them.

What it stores: About 28% of memories contained personal data under Europe’s GDPR rules, and 52% contained psychological insights about the user. Some were assumptions. For example, asking for “bands like Nirvana” was logged as “likes Nirvana.”

Why it matters: Lead author Abhisek Dash warns that people treat chatbots like advisors, friends, doctors and lawyers, so they know everything about you. Combined with AI that tends to validate users, that could deepen political polarization and reinforce biases.

Caveats: OpenAI has updated its memory system since the study, introducing a new version in January. Users can also turn memory off in settings.

The study: Microsoft researchers published a study in Science on Thursday showing their AI diagnostic tool was more than twice as accurate as general physicians at solving complex medical cases.

What’s different: Most medical AI looks at all symptoms, labs and scans at once. Microsoft’s “orchestrator” works step by step, like a panel of experts. It asks about medical history, orders tests, reviews results and updates its theory as it goes.

Real-world testing: Microsoft’s Dr. Dominic King says AI now easily aces multiple-choice medical exams, but real patients never arrive with a neat summary. So the team tested it on 300+ unusual, confusing cases instead, and it also cost less than doctors’ work-ups.

What’s next: Microsoft will test the tool with Beth Israel Deaconess Medical Center in Boston to see how well it can triage patients in emergency room waiting areas.

What it is: Carbon-X is an open-weight, 14.8-billion-parameter model that lets researchers ask questions about DNA in plain English. It was built with Italy’s University of Naples Federico II and China’s Zhongguancun Academy.

How it works: It’s built on Alibaba’s Qwen3-14B and trained on 302 billion tokens of DNA, English text and biology material. It reads DNA in six-letter chunks, so it understands genetic code and words in the same conversation.

What it’s good at: It’s strongest at predicting whether a DNA change is likely harmful and completing DNA sequences. It scored 66.7% on average across five DNA benchmarks, 27 points above the Qwen model it’s based on. It’s weakest at spotting splicing sites (29.8%).

Why it matters: Because it’s open, labs outside big AI companies can run it locally without expensive cloud platforms. It can also act as a DNA specialist for other AI agents.

Limits: It can’t physically sequence DNA, only analyze sequences you give it. It needs a powerful computer, gets worse when generating long DNA fragments, and hasn’t been clinically validated, so it shouldn’t be used for medical decisions.

Badgey Your iPhone becomes a name badge that faces the room

Llama Windows One local AI server for all your Windows apps

Reels Flow Turn ideas and images into AI-powered videos

Deyn Studio AI Video editor & media assisstent

Browser AI History Your AI chat history, all in one place.