3 Human Stories From AI in 2026 Worth Reading

Most AI coverage is about products, prices and promises. These three stories are about people, and each comes from the institution that did the work. We have kept the facts narrow, quoted sparingly and said plainly what each story does not prove.

1. How did a man with ALS use a brain implant to work from home?

On June 15, 2026, UC Davis Health announced results published in Nature Medicine about Casey Harrell, a 47-year-old man with ALS, a condition that left him with arm and leg weakness and speech that is difficult to understand. Harrell is a participant in BrainGate2, a clinical trial of an implanted brain-computer interface, a device that reads signals from the brain. The university worked with colleagues at Brown University and Mass General Brigham Neuroscience Institute.

The system has two parts. A speech decoder turns neural activity linked to attempted speech into text. A second decoder moves a computer cursor. Together, the release says, they let him use a personal computer, including email, messaging and web browsing, and keep in touch with family and friends. AI is the decoding step: software that learns to map patterns of electrical activity to words and movements.

More than 3,800 hours over nearly two years

Home use

Over 183,000 sentences, close to 2 million words

Volume

56 words per minute, rising over time

Average speed

92% of sentences accurate or mostly correct

Self-rated accuracy

Over 99% word accuracy with a 125,000-word vocabulary

Controlled test

The release says this version works at home without researchers present. An earlier study reported 97% word accuracy, but with research staff doing the setup. The release also says the independent computer use supports Harrell's employment. Study lead David Brandman said the work suggests researchers "may have crossed a threshold".

What does this story not show?

• It is one person. The results come from a single participant, so they say little about how well the system would work for others.
• It is investigational. The release states the device may only be used for investigational purposes under federal law.
• The two accuracy figures measure different things. The 99% figure is word accuracy in controlled testing. The 92% figure is Harrell's own rating of whole sentences.
• The team did the hard part. Surgeons, engineers and researchers built and maintained the system. The AI is one component of a long human effort.
• We did not open the paper. Nature Medicine sits behind a login redirect, so every figure here is as UC Davis Health printed it.

2. What happened when a university asked what responsible AI for endangered languages looks like?

Loyola Marymount University reported on June 23, 2026 that Jared Coleman, an assistant professor of computer science and a member of the Big Pine Paiute Tribe, had hosted a three-day workshop on its campus called "AI and Indigenous Language Revitalization". It was funded by a $50,000 National Science Foundation grant awarded to Coleman in fall 2025.

More than 60 invited guests attended: Indigenous community members, computer scientists, linguists, archivists, language teachers and nonprofit leaders. The aim, the university says, was to explore how AI can be built and used responsibly for endangered languages by combining language technology with community-driven data stewardship.

The story is notable for what it chose not to do. The release says participants discussed communities' distrust of outside researchers, rooted in a long history of extractive research, and the way AI jargon causes misunderstanding. It also says community goals vary widely, so no single solution will fit all. Coleman said the field has moved from extractive to community-driven documentation but that AI is still largely extractive. Kiana Maillet, a member of the Lone Pine Paiute-Shoshone Tribe, said outsiders should provide technology, training and support, then let communities lead.

What does this story not show?

• No tool was released and no language was saved. The outcomes listed are a living guidance document and a working group, run with the San Diego Supercomputer Center's CORE Institute Fellowship.
• It does not show that AI helps. It shows that people most affected are setting the terms before the technology is built.
• It is a university's account. We have no independent report of what was said at the meeting.

3. What did a randomised trial find about AI in breast screening?

On January 30, 2026, The Lancet published the full results of MASAI, which its authors describe as the first randomised controlled trial of AI in breast cancer screening. Over 100,000 women at four mammography sites in Sweden were randomly assigned either to AI-supported screening or to standard double reading by radiologists without AI. Each group had about 53,000 women. The trial ran from April 2021 to December 2022. Lead author Kristina Lång is at Lund University, Sweden.

The main outcome was interval cancers: cancers diagnosed after a negative screen and before the next scheduled one, followed for two years. Those are the cancers screening missed or that grew quickly.

That is a 12% reduction in interval cancers, with a similar false-positive rate. First author Jessie Gommers said: "Our study does not support replacing healthcare professionals with AI". At least one human radiologist still read every screen.

What does this story not show?

• It does not measure survival. The outcome is interval cancers over two years, not deaths. Longer follow-up is needed.
• It may not travel. The authors note a single country, one mammography device and one AI system, and moderately to highly experienced radiologists. Race and ethnicity data were not collected.
• Cost is unstudied. The authors call for cost-effectiveness studies.
• We read a trade report. We relied on ecancer's account of the paper, not the paper itself.

What do the three stories have in common?

• People stay in charge. A research team maintains the implant, communities lead the language work, and radiologists still read the scans.
• Each states its limits. One participant, no product, one country.
• The AI is a component. In each case the technology is one part of a larger effort by people.

How can I check a story like this myself?

For the method we use to label evidence, see our methodology. For more on judging study claims, read how to read an evaluation and reading benchmark claims critically. Questions about long-term risks are covered in our guide to AI alignment.

What are the limitations of this article?

We chose three stories from several we found, so the selection is editorial. We read institutional releases and one trade report, not the full papers, and we interviewed no one. Figures can change on correction; if you spot an error, tell us and we will record it on our corrections page.