Exactly my point. I agree with everything you wrote. [It’s so nice when that happens. 
]
2 Likes
Alan, when you wrote, “People who can’t write decent English will have even less incentive to try—which may hasten the dumbing down of America,” were you not suggesting that those who can’t write “decent English” are dumb—or at least are “dumbing down” the nation?
American society is my default frame of reference. I was definitely saying that American-born people whose primary (and most likely only) language is English have been “dumbed down” if they can’t write decent English. I was certainly not suggesting that the ability of non-native English speakers to write (or speak) well in English has anything to do with their intelligence. That would be a really stupid thing to say. 
I do not agree that AI is a software tool comparable to Photoshop; and it isn’t really true that AI can’t do anything on its own, as Geoffrey Hinton has demonstrated through his work on neural nets. But if we’re talking about large language models, and if we agree that everything an LLM-based AI produces is based on the work of others, then isn’t it a little absurd to even think about such a system producing a prize-worthy work of literature? A computer-assisted mash-up is not a work of art imho.
I would actually say “yes” to both of those questions – the choice to use assistive technologies for a range of things doesn’t mean they don’t have a range of consequences, including negative ones. Simply because the technology was widely adopted doesn’t mean that it was an unqualified good or that we can’t point out the losses that came with it. AI is rapidly becoming a useful tool, like it or not, and it’s important to be clear about the good and bad consequences of that.
Are you comfortable with a doctor giving you a diagnosis and treating you when they don’t actually know what’s going on and are simply following the prompts of an AI?
The time may come (maybe it’s already here) when AI would give a more informed and accurate diagnosis than the doctor would do on their own. Still, no one (I hope) is suggesting that this means we no longer need to train doctors…or that anyone with a high school diploma can be licensed as a doctor. So…yes…I think doctors should still be trained to “know what’s going on” — but AI may largely replace a lot of what they do. And when AI does what it does…it probably won’t be necessary for the doctor to know exactly how to duplicate all that the AI did to reach its decision. To some degree, it will be a black box and that’s okay. [Doctors should still be prepared to override the AI if they have reason to do so. Again, that’s where their training would come in.]
2 Likes
When airliner pilots rely too much on automation and lose understanding and command of what it is the underlying automation is doing, we end up with AF447 or OZ214. Nobody in their right mind would advocate pilots rely on automation (and that’s even before “artificial” intelligence) without being able to turn it off at a moment’s notice and take over themselves. And that’s irrespective of some guy with a PPL or commanding a 400-pax jet. It’s because we understand what they do calls for actual skill and responsibility without which there is no true accountability.
Yet here we are, cheering on the dawn of a new era where somebody who’s perhaps able to drive a tractor on a country road should appear capable of commanding a jet with 230,000 lbf. When an average athletic person takes anabolic steroids so they can appear able to bat like Ricky Henderson we call that doping and bar them from playing. I see no reason not to cast the same disdain upon ignorant posers who need to rely on AI to grift and cheat their way through life. IMHO there’s no free lunch. If you want it, earn it.
1 Like
How many doctors do you think we’ll be affording once we have AI that gets it right 99% of the time? I don’t think this is about what we think we want, but rather setting ourselves up for a whole bunch of unintended consequences.
The problem is that you don’t know whether the AI is hallucinating and prescribing hydroxychloroquine for COVID. If the doctor doesn’t know either, then you’re probably in trouble.
1 Like
Yes, that’s my point.
(We immediately knew glue on pizza was not a good idea. How about when we don’t?)
Uh, the comment was “how does it matter how we get to the results, as long as we do”? So, yes, a high school student inputting the test results and the designing th treatment is getting to the results, and assigning the treatments. Good luck with the AI getting it right.
“Sorry, the AI hallucinated your heart attack, so the triple bypass we did turns out to be for indigestion. What? Well, I’m not trained to understand the difference.”
Yes, indeed they should – you know, maybe like people should be trained to write well so they understand it and override it in case it’s wrong.
Are you sure AI will become a useful tool? Can we be sure at this point?
Just because people who are trying to sell us technology are saying AI can do everything doesn’t mean it can. About 15 years ago, weren’t people telling that all television would be 3D in a few years ago?
About 5 years ago we were being told that autonomous cars were just around the corner, but they’re a lot further off today, and the AIs doing the driving hasn’t lived up to claims that it would be able to drive everywhere anytime no matter what the road and weather conditions. I’m not saying self-driving cars are impossible, just that they completely replace all cars everywhere. Likewise I think AIs can be useful; I use otter.ai to transcribe interviews, but if I check the transcription because it’s not perfect. But we’re a long way from knowing the limits of AI, and we should be worrying seriously about the massive energy requirements of large language models in a warming world.
One thing AI might offer in medicine is eliminating diseases being dismissed because of rarity. I had a friend die of Malaria (in Sydney) as the Dr dismissed it as a possibility. Despite being told the patient had just returned from a Malaria hotspot, the Dr simply ignored the symptoms as they’d never seen an instance of Malaria.
Quite possible AI would have identified it by the symptoms - without human opinion ‘interfering’.
1 Like
Yes, I am. And yes, we can.
I believe you are over-interpreting what I said. If a person runs a statistical test on a computer without knowing the details of the formula used or how the results were calculated…they might still be able to properly determine how to interpret the results…given proper training. There was never an implication that proper (or any) training was irrelevant.
As to hallucinations, they are not really an issue with statistical tests. As for the medical example, like any technology, if the AI is known to malfunction with a given frequency, that should be taken into account. That’s why, for starters, I indicated that doctors should be prepared to override the AI decision.
Really, we are already doing this in many ways. For example, does the person who does the ultrasound test for a pregnant woman necessarily understand all the technology that underlies how ultrasound works? I doubt it. But they trust that the results are accurate and they (or their supervising physician) are trained how to interpret and make use of the results. Someday (we’re not there yet), AI-generated results will be treated the same way.
Sure, but what happens when, because the technician or doctor doesn’t understand the technology they’re using, they misinterpret the results? This is not an idle thought – just to give one example (not from medicine): in the crash of Air France 447, the fact that the pilots didn’t understand how the technology of the plane worked led them to react in a way that crashed the plane and killed 228 people.
But your ultrasound example is not quite on point. The technician and doctor are still responsible for interpreting the test and determining the results. With AI, it’s highly likely that that role is going to be taken away from them. The AI will read the results and present not only a diagnosis but a course of treatment and the technician or doctor will have no understanding of how to tell if that’s correct or glue in the pizza.
The practice of medicine is chock full of “garbage in, gospel out.” There is a lot of subtlety in interpretation of imaging studies (x-rays), blood pathology, and the like. At the end of a hospital stay, the discharging doctor is required (but often fails) to write up a discharge summary that can reduce a weeks-long hospital stay to a page or two. The x-ray report, the “lab results,” the discharge summary become black-and-white Statements o’ Truth™, and rarely do doctors further down the line check back to make sure they’re accurate. They can be wrong simply because they necessarily omit nuance. They can be wrong because they’re for the wrong patient (happens way more often than you might think). They can be wrong simply because of a human or procedural error.
The most common way for the wrongness to get detected is that something in a report fails a reality test: it just doesn’t seem right. One of my concerns about “AI” in medicine is that LLMs, if nothing else, are designed to create output that appears reasonable. I can see AI being used to write discharge summaries (a chore many docs seem to detest) and it will come up with something that is simply wrong but that passes the sniff test. It likely would take more time to thoroughly check such a summary than to actually write a good one in the first place, so the AI generated BS will become the official history of the patient.
This already happens a lot in healthcare without the “benefit” of AI making the writing sound more plausible.
1 Like
AI-generated BS could very easily become the official history of the human race. There are so many ways for this “miraculous” technology to go wrong that it makes my head spin. 
3 Likes
People in this thread keep posting (or rebutting other posts)…detailing how AI (in areas such as medicine) could go wrong…horribly wrong.
Sure, I agree. There are risks and dangers with using AI. I think everyone here recognizes that. The question is: What do we do about that? Abandon AI altogether? Greatly limit its scope to where there is no chance of significant harm?
Maybe some would prefer that, but I seriously doubt that’s what’s going to happen. The push to use the technology and its potential value are just too great to set aside.
But, as with any new technology…there will be problems that occur and solutions sought. [Just look how car safety has developed…with all its ups and downs…over the decades.]
I remain optimistic that, in the end, AI will be and will be viewed as a positive. If I’m wrong…well, humanity was fun while it lasted.
3 Likes
It developed that way because people were skeptical about the technology when it arrived…and that’s what the people in this thread are doing.