It is getting more and more difficult to sort out the upside of AI from the downside, Early research, much of it which I reported on here on the blog, focused on the immediate and obvious potential negative impact of AI--on everything from cognition to interpersonal relations. Here is a dozen recent studies that point in the other direction.
 |
| Clker.com |
1. Google's AI "diagnostic assistant," Articulate Medical Intelligence Explorer nearly doubled unassisted clinicians' diagnostic accuracy. In a recent trial, AMIE outperformed physicians on 30 of 32 quality measures. When consulted for a second opinion, it helped identify the right diagnosis (out of the top 10 guesses) far more often — jumping from 34% to 59% accuracy. (Nature.com[nature])
2. In a hospital trial, of the top three diagnoses, AMIE was on target 75% of the time. Furthermore, patients trusted it more, not less, after chatting with it about their conditions, a surprising outcome from the study. (Archive.org[arxiv])
*Two more summarized in This month in Medical AI- August 2026 [linkedin])
3. Harvard's COMPASS model —AI trained to predict how a patient's cancer will respond to treatment. It is designed to flag who will benefit from immunotherapy (a treatment that uses the body's own immune system). It outperformed existing methods by 8.5%.
4. On scans that had already been read as normal, Mayo Clinic's AI caught pancreatic cancer up to three years earlier than it might have been diagnosed. The tool, REDMOD, flagged 73% of cancers “hiding in plain sight” on routine CT scans.
Not sure if the next two are a net positive or not, but anything to help the UK at this point .
5. In the UK, nearly two-thirds of 25-to-34-year-olds have connected with a chatbot for some kind of mental health support. According to Mental Health UK, it's become an accepted, welcome practice in just one generation, Mental Health UK and Censuswide. (Financial Times [ft])
6. Young adults in the UK are apparently trusting AI with their money more than generations before them. A survey by Lloyds Bank found 76% of UK adults aged 18-24 have used AI for personal finance decisions — 20 percentage points higher than an earlier all-adult 2025 benchmark. (Finance Feeds.com [financefeeds])
Good now for physics but maybe not for physicists’ job security
7. AI is now not just crunching data; it is discovering laws of physics. Researchers at Emory University provided an AI model with lab measurements of a strange, electrically charged gas-like substance called "dusty plasma" — found everywhere from in wildfire smoke to outer space. It uncovered a previously unknown process by which particles push and pull on each other, with over 99% accuracy. (Science Daily news [sciencedaily])
8. Forecasting a once-in-a-thousand-years heat wave just got 100 times cheaper to compute. A new AI-assisted method predicted the odds of catastrophic, rare weather extremes using a small fraction of the computing power that traditional climate models require, with nearly identical accuracy. (Phys.org/news [phys])
Good now for software developers and coders, but maybe not for their job security or how fast AI is evolving!
9. AI cut a real coding task's time by more than half. Programmers at GitHub's found using an AI coding assistant finished a task in 1 hour 11 minutes versus 2 hours 41 minutes for those working unassisted — more than twice as fast. (Trace Cohen’s blog [valueaddvc])
10. Software developers using AI are getting nearly twice as much of their work approved and shipped. Google's own research team found individual developers completing 21% more tasks and getting 98% more of their finished code changes approved by teammates after adopting AI tools. (Next wave insight.com [nextwavesinsight])
Good for shoppers and taking legal action against shops
11. Instacart shoppers using AI guidance didn't just work faster — they gained flexibility, taking on a striking 32.5% more shifts across different stores. The gains were biggest for workers who already had substantial experience to combine with the AI's suggestions. (Phys.org/news[phys])
12. Legal AI use has more than doubled in a year. Ninety-four percent of lawyers using AI report real productivity gains, and one industry estimate puts the value at up to $300,000 a year in extra billable-equivalent work per lawyer. (llamaiab.ai [llamalab]) See also my earlier post on the range of services available with Legalshield.
***
None of this potential is guaranteed, though — the same research shows AI's gains evaporate or reverse the moment it's used without structure. Developers who trusted AI blindly on familiar code got slower, not faster, and legal and financial AI use is racing ahead of the safeguards most people have in place. That's the "obvious" downside worth taking seriously, and it's the gap I have been unpacking on the blog.
Note: This post was developed using Perplexity AI, principally for research.
So lawyers are going to get richer because they can bill more work per week--instead of lowering their hourly rates to benefit people in need of legal services? And using AI for mental health support has not worked out well for those families who lost children to suicide after chatting with AI. The Instacart example baffled me until I went to the news item on phys.org to read it, but there I found out "More experienced workers were better able to combine AI recommendations with their own judgment, particularly during more complex shopping trips and periods of heavier workload...."There's a growing assumption that AI levels the playing field by making expertise less important," said study author Benjamin Knight, a researcher at Maplebear Inc. "Our findings suggest something different. AI is most effective when workers have enough experience to understand when to trust its recommendations and when to rely on their own judgment."
ReplyDeleteGrocery stores are getting more value from their experienced employees who use the AI tools: Has it meant any increase for the pay that these employees receive? And since stores can now prepare, sell, and deliver more carts, do customers get any reduction in the costs of their use of this service?
The developments in medicine sound more promising. However an earlier study done after 2022 found that endoscopists who have performed AI-assisted colonoscopies are at risk of being deskilled. (The Lancet: Gastroenterology & Hepatology, 2025, vol. 10, pp.896-903.) But the ability to diagnose pancreatic cancer up to three years earlier than it might have been diagnosed is, indeed, amazing.
Thanks, Anon, for following up on those studies. In the introduction my intention was to add the "at least for some" caveat, as well!
ReplyDelete