https://www.reuters.com/investigations/ai-enters-operating-room-reports-arise-botched-surgeries-misidentified-body-2026-02-09/

[bertdriehuis@infosec.exchange]

@atax1a one rather important distinction that is often lost at reporters (as part of the general public) is whether we're dealing with ML or with LLM. I've seen my share of absolute bonkers implementation of, well, anything, but I have a hard time believing f'ing LLM's entered the operating theatre.

I'm not decided on whether I prefer to die because of an ML model going off the rails, or an old fashioned coding error like the infamous Therac-25. I've seen code for medical software and I'm not optimistic either way.

Frankly, I prefer doctors who don't Google my symptoms during a GP visit, but I'm afraid that is an art that's dying out.

[tsturm@famichiki.jp]

@atax1a We live in the stupidest timeline.

[energisch_@troet.cafe]

@atax1a 🤯

[christianriegel@digitalcourage.social]

@atax1a @bertdriehuis

Thank you! If I'm correct, it's like this:

- I support algorithms. Vital automation.
- Neural nets are - if well trained - tested and efficient algorithms.
- Machine learning is a neural net training itself. Now I'm getting sceptical, demand testing and would hope it's not left unattended to "continue learning".

But all of the above is targeted at maximizing correctness of answers!!

- Then here are the LLMs: targeted at maximizing plausability! SOUNDING correct is the goal.
Oh and here's reddit, learn that. - _-

[ashmire@pagan.plus]

@atax1a Yikes.

[ozzelot@mstdn.social]

@atax1a
Oh... oh fuck.

[drgroftehauge@sigmoid.social]

@bertdriehuis @atax1a Oh, there are absolutely terrible ML implementations out there. Social workers in Denmark got a tool that was supposed to assist them in deciding whether a child should be removed from the home. The strongest feature by far was age of the child (because removing a child is the last thing you try). It was a less than useless linear model.

[wellsitegeo@masto.ai]

@atax1a
Yes. You *are* screaming.

There is a good reason that, after opening the skull under anaesthetic, the patient is normally awake and talking to the surgical team as they hack away inside the brain. And you've just, uh, put your scalpel on it.

Generally the surgeon will be prodding and poking a particular place to cut, *before* cutting, so they can evaluate the effects on the *person* in the wet electric fat. If something produces odd effects, they look for a way around it.

[bertdriehuis@infosec.exchange]

@drgroftehauge @atax1a there are tons of bad models out there, that's a fact. ML is an opaque tool. But an ML model is easier to validate independently. Biases can be shown, and results are reproducible within statistical limits. It is as much a science as statistics are, and those are equally abused in the domain you refer to.

The Netherlands by the way also has its fair share of problematic algorithms based on ML in the social domain. The biggest issue is not ML itself, but the lack of openness and independent validation. If the algorithm were written in a traditional programming language the result would not have been different (and we also have failed examples of those in our governments' recent past).

[jkmcnk@mastodon.social]

@atax1a move fast, break people.