Apropos AI and medicine Surprised the researchers: AI detected 20 per cent more cancer cases Researchers are pleasantly surprised at how many cancer cases artificial intelligence (AI) found. This means that the tumor can be removed earlier. Use Google Translate, if you do not read Norwegian.
The dangers posed by AI may not prevent it from being exceedingly useful in some applications; but successes in one area do not reduce the inherent risks of deploying this technology more widely. To me, the use of such a powerful technology by such fallible, self-interested creatures as we obviously are is like children playing with fire.
1 Like
The “AI detected 20 per cent more cancer” is clickbait. The actual scientific paper (abstract in English) on the study is a bit less dramatic: “AI-supported mammography screening resulted in a similar cancer detection rate compared with standard double reading.” It touts AI not as improving outcomes for patients, but for lowering the workload for radiologists.
There is so much nuance in the cancer-screening business that one must be careful making generalization. For example, it’s possible – when all the follow-up is said and done – that the AI-powered branch of the study just resulted in 20% more invasive biopsies and unnecessary treatment, without any overall improvement in all-cause mortality. Only time (for long-term follow-up) can tell.
And as @omalansky said, potential benefits don’t give us carte blanche to ignore potential harms.
We make sure that purveyors of the technology are held strictly liable for harms when it is abused.
7 Likes
This is a big problem. Government watchdogs rarely do a good job of holding unscrupulous businesses accountable for the harm they do in pursuit of profits. Modern ethics, it seems, gives businesses permission to do anything they can get away with.
2 Likes
Nuance isn’t restricted to cancer screening and I’m very apprehensive about the quality of generative writing based on pattern-matching.
Apparently even human-written statements published by authoritative sources are subject to an echo chamber of garbage in, garbage out. And I expect that the stochastic parrots of generative AI will simply increase the chambers’s volume.
Perhaps this article about “zombie statistics” illustrates the misleading cacophony that we can expect from generative AI writing:
1 Like
Amen to that. I don’t mean to sound condescending in any way, but it does seem to me that many very intelligent people fail to appreciate the scope of the calamity that would ensue if generative AI were allowed to introduce mountains of false information and just plain bad writing into the global reservoir of humanity’s collective knowledge—which thanks to the internet, is truly vast in scale. I am not saying that something like this will actually happen, but it could happen. This being the case, a very cautious approach to the wider deployment of AI technology seems warranted. Personally, I would feel less uneasy about AI if I saw more evidence that the developers and supporters of this technology were being as cautious as I think they ought to be. As the wise man said: “Never assume that the gun in your hand isn’t loaded.”
2 Likes
This discussion sounds like we’ve gone from suggested grammar and spelling corrections to writing new encyclopaedia. Apple seems to be offering the former and if this can demonstrate better grammar and spelling to users it should be applauded.
Whilst there’s no doubt potential for AI to be misleading (based on its source data), I don’t think the ‘sky is falling’ at Apple just yet.
2 Likes
It’s not Apple I’m worried about. This discussion seems to have morphed into a discussion about AI itself, not Apple’s very limited implementation of this technology.
The scenario I wonder about is - what will be taken for fact? Will people have some repository of knowledge about “tested facts” that can be retained and referenced? If a medical student relies on a special “MedAI” for building knowledge, how do we know that facts, and not “hallucinations” are responsible for that bit of info? That includes basic facts of historical names and dates that we now pull from the internet. Once the false information gets into the system, is repeated and then accepted for truth, what will be the reference that something is correct or not.
I am not saying this doesn’t happen now, with one researcher stating one thing and another a different fact. But then you can go through the data and follow the thinking process. AI does not explain how its conclusion is arrived at. It just is.
And I think the term “confabulation” is a better term than “hallucination” for these types of errors (such as when someone with alcoholic dementia will state something and say it in a way that they believe it is true) but that probably won’t catch on.
1 Like
Exactly. And since this is already happening, where are the safeguards needed to prevent this dangerous trend from proliferating?
1 Like
I misphrased my question about the usefulness of AI. I find some AI useful (e.g. the speech transcription of otter.ai, which lets the user review the transcript), However I find other AI useless because it’s error-ridden or unreliable (e.g. Ancestry’s AI searches of old newspaper obituaries which lump together multiple notices in the same column, giving lots of false positives). I was trying to ask if you consider ALL AI to be useful, but reading more carefully I don’t think anyone here does. Unfortunately, too many people outside the tech community so swallow that fallacy.
1 Like
On the upside, Claude 3 Opus probably can’t be gifted bus-sized RVs or vacations by wealthy “friends.”
3 Likes
This combined with the increasing hostility to theoretical and practical expertise in US culture is, in my opinion, a very bad thing. Many important decisions are not best resolved by using ballot measures that force binary choices onto complex issues or by relying on emotions and cursory online searches.
Can you provide a reference for that work? I’m talking about initiation of action—I’d like to see the AI that’s going to email me, offer to write articles for TidBITS, and then do so without any prompting from a person.
(Amusingly, I’m very, very, distantly related to Geoffrey Hinton—my great-uncle Sid Engst was married to Joan Hinton, and Geoffrey and Joan’s common great-great-grandfather was William Boole, of Boolean logic.)
That’s a rather extreme example, but I’d certainly hope that anyone who qualified as a doctor wouldn’t merely pass on the results of AI prompts, just as I’d hope that now they wouldn’t merely be googling my symptoms. The existence of assistive tools doesn’t absolve any professionals of doing their job. And of course, they’d be liable for the consequences of their actions.
That’s not to say that some people won’t cut corners or be bad at their jobs, but that’s true now.
My initial comment was actually, “So yeah, I think I’m on the side of agreeing that if the results are as desired, how you get there may not be that important in most cases,” and was in the context of Ted’s examples of cashiers having difficulty with basic math and in the overall context of AI-driven writing tools.
So while it may be an interesting question to ponder how speculative AI systems might be used or abused by highly trained professionals like doctors or airline pilots, it’s a major jump from where we started.
I agree. There’s no stuffing the AI genie back in the bottle, even with regulation. (Anyone who wants to ignore the regulations will.) My current thinking is that the best approach is cautious, skeptical engagement.
So yes, in the example that Ron shared about doctors writing discharge papers, having an LLM do it feels like a mistake. But one of my hopes for AI in medicine is that an LLM might be able to ingest a patient’s entire medical history and suggest to a doctor diagnoses, unnecessary medications, and treatments that might not be evident to one doctor looking at a single issue in isolation.
This is the AI slop argument, and while I don’t think there’s any benefit to AI slop, I also don’t believe we’re in some rosy-colored present where everything on the Internet is lovingly crafted by experts. Sturgeon’s Law dates from the 1950s, and 90% of everything on the Internet has always been crap. Since the amount of content on the Internet has been effectively infinite for decades, increasing the amount of “effectively infinite” doesn’t really change anything other than making Google’s job harder.
My answer to the AI slop argument is the same as my answer to the infinite Internet problem—avoid algorithms in favor of trusted sources whenever possible. Obviously, that’s tough with search engine discovery, but it works awfully well with social media.
Hah! In another mailing list, I proposed “confabulation” instead of “hallucination” too, since hallucination to me implies a certain wild creativity rather than something that is simply stated as fact incorrectly and without intent to deceive.
2 Likes
The project that Geoffrey Hinton himself referenced in a recent 60 Minutes interview
involved robots that were instructed to score hockey goals but were not taught the rules of the game. They proceeded by trial and error; and whenever an action (or a strategy) on their part was successful, this information was shared with all the layers in the neural net. Little by little, the robots learned how to score hockey goals and win games—all without any further instruction from their human “masters”. According to Geoffrey, neither he nor any other computer scientist knows how they did what they were doing. The entire segment lasts about 13 minutes. Here’s the YouTube link: https://www.youtube.com/watch?v=qrvK_KuIeJk
1 Like
Now I see what you’re saying. I think there’s a big difference between “We don’t know how an AI came up with a solution to a task that it was designed for and prompted to complete” and “AI can do things on their own” (a restatement of “it isn’t really true that AI can’t do anything on its own”).
This isn’t too unusual. It’s actually a fairly standard way to train neural nets for all kinds of things. Start with random weights in the network. Then let it play, giving it positive and negative feedback based on how good the results are.
Variations on this are used for all kinds of tasks where it is possible to automatically determine “right” and “wrong” from results, including image recognition and playing board games. I’ve also seen demonstrations of this for robots self-teaching how to walk or crawl, using positive and negative feedback based on how well it moves in the requested direction.
Yes, it is very impressive that they combined this with robots playing a physical game, but as far as the ML software is concerned, it’s not conceptually that different from other pre-existing experiments in game playing and robotics.
But self-teaching how to play a game based on a hard-coded definition of the game’s rules is not the same as taking independent action (like a vacuum cleaner choosing to learn how to play hockey because it wants to).
1 Like
Yes, indeed. If anybody is curious about the details, you want to look up supervised/reinforcement learning vs. unsupervised learning. The latter is where the AI learns how to play by just telling it what the goal is (put ball into opponent’s goal) and what knobs it has to play with (skate, swing, etc.), without you ever teaching it how to play hockey. In unsupervised learning you don’t need labeled training data (cat or dog), you just need a merit function (the target, i.e. hockey goals).
I understand that AI’s ability to find its own solution to a problem isn’t the same as setting its own goals. Nevertheless, finding its own solution to a stated problem is acting on its own. What is the next-generation AI going to do if the best solution to the problem of global warming is to wipe out the human race? 
Please see my reply to Adam’s last post.