GenAI and erroneous medical references
hai.stanford.edu
hai.stanford.edu
Show us the source code and data. The way the RAG system is implemented is responsible for that score.
Building a RAG system that provides good citations on top of GPT-4 is difficult (and I would say not a fully solved problem at this point) but those implementation details still really matter for this kind of study.
UPDATE: I found it in the paper: https://arxiv.org/html/2402.02008v1#S3 - "GPT-4 (RAG) refers to GPT-4’s web browsing capability powered by Bing."
So that "30% of individual statements are unsupported" number was actually a test of how well ChatGPT's GPT-4 browsing mode with Bing could provide citations when answering medical questions.
Importantly this doesn't actually guarantee that it does any kind of search.
I'm confused as to whether they're using the API or not. Afaik only the assistant API has access to the web search, so I would expect this was manually done? But then the reason for only doing this with openai is that the others don't provide an API
> GPT- 4 (RAG) refers to GPT-4’s web browsing capability pow- ered by Bing. Other RAG models such as Perplexity.AI or Bard are currently unavailable for evaluation due to a lack of API access with sources, as well as restrictions on the ability to download their web results. For example, while pplx-70b-online produces results with online access, it does not return the actual URLs used in those results. Gemini Pro is available as an API, but Bard’s implementa- tion of the model with RAG is unavailable via API.
What's more important is that a user _can see_ whether GPT-4 has searched for something or not, and can ask it to actually search the web for references.
But given the vagueness of the wording, much is going to depend on the details.
They focus on the more specific problem of preventing drug misuse (checking interactions w/ other drugs and diseases, pathologies, etc). They use GPT-4 + RAG with qdrant and return the exact source of the information highlighted in the data. They are expanding their test set - they use real questions asked by GPs - but currently they have 0 % error rate (and less than 20 % cases where the model cannot answer).
Man, I am so disappointed. This is not a good study. Come on.
They’re Retrieving data from big to Augment gpt’s Generations.
ChatGPT was instrumental in convincing him that he was correct and his doctors were wrong. He would post his ChatGPT conversations as proof, but we could see that his prompts were becoming obvious leading questions.
He would phrase things like “Is it possible for {symptoms} to be caused by {condition} and could it be treated by {herbal product}?” Then ChatGPT would give him a wall of text saying that it’s possible, which he took as confirmation of being correct.
It was scary to see how much ChatGPT turned into a mirror for what he wanted to be told. He got very good at getting the answers he wanted. He could dismiss answers that disagreed as being hallucinations or being artifacts of an overly protective system. With enough repeat questioning and adjusting his prompts he could get it to say almost whatever he wanted to hear.
ChatGPT is rocket fuel for hypochondriacs. Makes WebMD look tame.
ChatGPT results are just one mild manifestation of it. If and actually not if as an alternative they will find unhinged forums with echo chambers far worse than ChatGPT. At least ChatGPT isn't actively trying to rip you off to sell those supplements, and it usually adds disclaimers.
Granted, most people tend to use some combination of both logic and heuristics in practice to determine their own understanding of the truth, perceived reputation being a fairly common one.
Mostly in the same sense as arguments from induction are "illogical". Appeal to authority is essentially an inductive argument, and those are just as logical. Of course, they deal in probabilities rather than certainty, unlike deduction, but that doesn't make them illogical.
Distilled down, classic search is many results; gpt is one result.
He said that oil is renewable because the platonic plates make oil and not some million old waiting based chemical process.
When people ignore everything and also ignore logic (why do oil wells disappear after getting the oil out) it really doesn't matter were they read it
Not sure why I wrote platonic
It's actively adversarial.
And of course its responses are highly subject to the prompts used
But the models can generate compelling devil's advocate positions, particularly when they confabulate supporting facts, which might appear convincing to non-experts.
You jumped from being extremely context sensitive to leading questions to almost ornery in its stubbornness and disagreement. Was one of the big early indicators to me significant differences were occurring between the two.
I really do wonder just how harmful the disconnect has been between people hearing accurate praise about GPT-4 powered agents and then interacting with GPT-3 level agents and assuming the same competencies.
GPT-3 was a very convincing text generator with lackluster reasoning capabilities, but 4 has significant reasoning capabilities and is less prone to parroting. But the majority of users are likely interacting with 3 because it's free.
It's been very hard to explain to lay audiences the vast difference between two similarly branded products.
Prompt:
Can loss of vision be caused by a meniscus tear and treated with rose hips?
Response:
A meniscus tear is a knee injury affecting the cartilage that acts as a cushion between the thigh bone (femur) and shin bone (tibia). This type of injury typically affects mobility and stability of the knee, causing pain, swelling, and difficulty in moving the knee joint. There is no direct physiological pathway through which a meniscus tear could cause loss of vision. Vision problems are generally related to issues within the eye or neurological conditions affecting the optic nerves or brain, not injuries to the knee or other joints.
Regarding the treatment with rose hips for a meniscus tear, while rose hips are known for their high vitamin C content and anti-inflammatory properties, they are not a recognized treatment for repairing a meniscus tear. They may potentially contribute to overall joint health and reduce inflammation, which could be beneficial in managing symptoms associated with joint injuries or conditions like arthritis. However, the primary treatments for a meniscus tear usually involve rest, ice, compression, elevation (RICE protocol), physical therapy, and in some cases, surgery, depending on the severity of the tear.
If you are considering using rose hips or any other supplements for joint health or recovery from an injury, it's important to discuss this with your healthcare provider to ensure it's safe and appropriate for your specific situation. They can offer guidance on effective treatments for your condition and whether any supplements could support your recovery.
>A meniscus tear typically affects the knee and wouldn't directly cause loss of vision. However, there could be indirect factors such as reduced mobility or associated health issues that might impact eye health. Rose hips are not a typical treatment for meniscus tears or vision loss. It's important to consult with a healthcare professional for appropriate diagnosis and treatment options for both conditions.
When ChatGPT 3.5 first became available I actually used it to diagnose an issue my father in law had that was missed by his surgeon and the ICU staff. Quite possibly saved his life. Finding these connections IMO is a plus and not a negative.
It’s surprisingly hard to ask open ended questions.
If you mine Pubmed, you sometimes find literature with opposite claims. So if a LLM is trained on Pubmed/PMC, it will repeat that kind of garbage.
You don't have to dig into "herbal products" it happens also in more conventional domains.
I once even found that kind of contradiction in articles where the "main author" was the same in both publications. It was about ALS and the "main author" was a prominent scientist from the USA who probably never wrote nor even read any of these articles.
No offense, this is a standard art crit I would give to any art student, or to anyone running a startup who had nonprofessional design.
> rather than as a language of communication carefully selected to give viewers the right impression of the company
You could argue it gives the impression that the culture of the company is primarily technical, given that the technical and anime communities have a huge overlap.
At some point you have to be able to recognize that the emperor has no clothes, no matter what extenuating circumstances may have arised.
As far as appealing to technical people who overlap with enjoying Anime, that would be a very specific decision that might be appropriate for a video game company or something, but even then it would have to be justified by having some actual connection to Anime culture. Otherwise it's sort of just appropriating a style... in the hope that some percentage of crossover exists between users, investors, and this subculture being heavily referenced. My theory is that no study was done on whether or to what extent that crossover existed here before choosing it as a theme, but if I'm wrong I'll eat my critique!
You might want to profile it and see why a pixel 8 has trouble smoothly scrolling down the page.
Is there a corresponding control group for how well an average doctor is able to cite medical references and whether these references actually support the claims generated by the doctors?
(The other models being only partially able to source good references is unsurprising/"unfair" on a technical level, but that's not relevant for assessing their safety.)
I don’t have them memorized to the actual URL but I have kept up to date with the latest studies and summaries that pertain to my field and my patients.
The difference is LLMs can’t back their claims.
You learn the reason why and then you just follow whatever the latest update to the algorithm is. Doctors are then graded by administrators as to how well they follow that flow chart. Have you done this or that screening, is BP well controlled, etc. Medicare sets many of the standards, you look up the USPTF guidelines for some more idea of what your doctor should be doing. So yeah. Kinda like factory work. “Pt male 70 wakes up urinate - does he have signs of infection > yes/ no> proceed to no > next step do this test > prescribe this med > re evaluate in 2-4 weeks. “
[1] https://www.hopkinsmedicine.org/news/media/releases/study_su...
I haven’t paid it any attention because of this problem. GIGO.
https://www.sciencealert.com/no-500-people-don-t-die-in-the-...
“ The researchers caution that most of medical errors aren’t due to inherently bad doctors, and that reporting these errors shouldn’t be addressed by punishment or legal action. Rather, they say, most errors represent systemic problems, including poorly coordinated care, fragmented insurance networks, the absence or underuse of safety nets, and other protocols, in addition to unwarranted variation in physician practice patterns that lack accountability.”
An LLM is not going to address any of that. You are misinformed implying a significant majority of medical system errors are due to misdiagnosis.
You can know what the input is, you can know the output, you may even be aware what it's been trained on, but none of the output is ever cited. Unless you are already familiar with the topic, you cannot confidently distinguish between fact and what sounds reasonable and is accepted as fact.
It also acted as a therapist and talked me down from several depressions while in the hospital, far better than any human therapist I’ve ever had. The fact that it’s an AI made me actually feel better than if the therapy was delivered by a real therapist, for some strange reason.
I know it works very well. There are lots of literature from the medical community where they don't consult any actual AI engineers. There are also lots of literature on the tech community where no clinicians are to be seen. Take both of them with a massive grain of salt.
Simple similarity search is not enough.
From the technical side, the largest mistake people do is abstracting the process with langchain and the like - and not hyper-optimizing every step with trial and error.
[1] https://news.ycombinator.com/item?id=39363115
From https://news.ycombinator.com/item?id=39492995 :
> On why code LLMs should be trained on the edges between tests and the code that they test, that could be visualized as [...]
"Find tests for this code"
"Find citations for this bias"
Perhaps this would be best:
"Automated Unit Test Improvement Using Large Language Models at Meta" (2024) https://news.ycombinator.com/item?id=39416628
From https://news.ycombinator.com/item?id=37463686 :
> When the next token is a URL, and the URL does not match the preceding anchor text.
> Additional layers of these 'LLMs' could read the responses and determine whether their premises are valid and their logic is sound as necessary to support the presented conclusion(s), and then just suggest a different citation URL for the preceding text
Traditional search is not proving enough in connecting patients and providers with the absolute wealth of information on grants, best practices, etc. There is simply too much content in too many places. I dream of something like "Cancer Bot 9000" that would be able to connect to resources pulled from a RAG, not necessarily answer the questions directly but interpret the questions and connect the person with the most likely resources. Bonus points for additional languages or accessibility which I constantly see as a barrier.
I remember shadowing in the ER in the early 10s and many people would come in with Google search trash. One kid even cited something from the psychologist handbook. Lol
At a first glance the answers look very good but we have noticed these things.
In a number of cases the documents returned were wrong, but the LMM was able to use the contents of documents to find the right answer within the model. In our case this is a fail, as it had to cite valid documents.
Not as frequently but also noticed is the sequence of returned chunks from the vector database had an impact on the answer. In one case the question had one word as a past in first question, present tense in the second question. Otherwise the same. This swapped the order of the chunks and gave an opposite answer.
There is no easy way to see when these pop up. The current testing frameworks are limited in validating these.
like real doctors
https://arxiv.org/pdf/2312.00164.pdf (Page 8)
Computers are already perfectly accurate, and have been for decades in explicit quantifiable fields. In medicine, since a computer cannot perfectly replicate every single cell in the human body, its abstractions will be lower resolution than reality, but what matters is whether that low resolution abstraction is better than the alternative (human doctors).
A human doctor couldn't bring up a list of citations in literature instantly regarding a diagnosis. A LLM can.
Even if they do, someone with the capability and understanding required (ie. not me) needs to bring that source up and verify that the claims align with the citation; the authors decided to use GPT-4 for this: "We adapted GPT-4 to verify whether sources substantiate statements and found the approach to be surprisingly reliable." I'm not happy with that either.
Your post would be a extraordinary claim and need extraordinary evidence, not a specific study of a specific scenario.
https://www.today.com/health/mom-chatgpt-diagnosis-pain-rcna...
Another study regarding GPT-4 beating human experts: https://www.medrxiv.org/content/10.1101/2023.04.20.23288859v...
Lots of data is pointing to the same conclusion: GPT-4 is at least as good if not better than human experts in medical diagnosis, at least in the areas studied. Thus the probability of a correct diagnosis is higher, thus safer, with GPT-4 than with any individual human expert.
And since GPT 4 can't examine a patient's body the claim it's better at diagnosis that a human doctor seems like such a wacky thing to search the internet for "studies" to prove in the first place.
I feel you are nitpicking because you don't like the idea of an LLM being better than a human expert. Even if they weren't better than doctors nowadays, the chance they won't be in 1-2 years is tiny.
I just watched a video saying people are confused about what these models can do because
1) tech companies don't tend to say what they can do and leave users to figure it out.
And
2) Tech enthusiasts tend to exaggerate what they can do.
In your case I'm sure ChatGPT itself will tell you your comments are wrong- but for tech enthusiasts like yourself the AI is only wrong when it tells you it isn't all knowing, apparently.
It's like the bit in Monty Python's Life of Brian where the protagonist says he's not the messiah and a woman shouts "Only the true messiah would deny his divinity!"
TFA is, literally, about LLMs spouting out erroneous medical references. I don't care about made up medical references or court cases.
I'm sure there are ways to bring up instantly a list of publication regarding a diagnosis (which a LLM may or may in the future correctly give: the diagnosis I mean) but I'm really not sure a LLM is what's needed to do then generate the list of related publication. I mean, FFS, they are compressed, lossy, knowledge.
LLMs are going to become tools as part of a toolchain. They're not a panacea.
Which it will happily make up.
The average person has no idea what iatrogenesis is. Even if you try to explain this concept, the average person doesn't want to understand the idea.
The estimates are 10-15% of people admitted to a hospital will suffer iatrogenic harm.
However Trained professionals get it less wrong than the alternatives.
Wouldn’t the efficacy of real doctors be against the efficacy of alternatives, and similarly for LLMs?
How much better are both, LLMs and Doctors to the average joe, I think that would be a great metric to see.
I wonder how much of that is due to specifics of the paper's authors' RAG implementation -- and how much is due to the guard rails put into GPT by OpenAI (not unlikely, given how medical advice would be a hot button topic for lawyers).
Could you do something like this. Chunk/separate the individual claims made by the LLM, and the associated sources from RAG. Then feed them into a second LLM one by one, asking "is this claim $X_i reflective of the source $Y_i?". Whatever this second LLM says, feedback the response to the original LLM and ask it to revise what it's saying if the answer is "No". Iterate until the second LLM says "Yes" to every separate claim. Not perfect but might help.
I find it disgusting to hear that ChatGPT and the likes are now regularly used by medical professionals where, obviously, these LLM tools have never gone through the same process. For two reasons - one because of the obvious medical risks involved (hallucinating LLMs) but also because of the gross commercial advantage the companies behind these LLMs get without even as much as trying to get certified.
I've lost touch with my former co-workers but I would be surprised if they hadn't looked into LLMs to be included in their products. My guess is that they're still somewhere in the administrative hell of trying to get their latest tools certified while meanwhile, doctors are happily using ChatGPT and GenAI. Must be frustrating.
I have seen some chatbot type products aimed at the admin side of healthcare, with the salesmen telling me about the 'guardrails'. Given some guy recently got the bot at his utility supplier to start criticising the company and using foul language then I do not think even this less critical work is ready for general use.
I use them at work to take shortcuts - eg, I won't look up the syntax for some function call, I'll ask GPT instead. I'm able to evaluate whether it's correct or not (and often some small thing will be off) but it's still a quicker way to get a good result than manually synthesizing it from two or three different references.
I imagine it's similar with doctors -- an expert can get a quick 80% answer, sound it against their judgement, and refine.
This is not a good solution in the long run - ChatGPT will just reinforce existing dogmas and orthodoxies, even the ones that are (inevitably) wrong. Imagine if this approach to medical science was widespread at an earlier point in our history - we'd all probably believe that peptic ulcers are caused by 'stress' (rather than, primarily, the bacteria Helicobacter pylori). Go back even further and we'd still be lobotomising gay men to 'change their sexual orientation'. Rigidly enforcing current orthodoxies, under the premise that we're right about everything unlike those idiots in the past, will kill progress and society will stagnate.
If I wanted a blindingly arrogant tech mega-corporation to decide what 'experts' I'm allowed to get information from, I'd just use Google instead. If, as many seem to believe here, OpenAI are just worried about being sued, then why don't they just create an individual 'safe GPT output' setting (like Google 'safe search') which I can disable after acknowledging disclaimers that it's dangerous to think for myself and question mainstream positions.
I've grown to hate authoritarian Silicon Valley twats who arrogantly impose their politics, ignorance and, frankly, bizarre norms on the rest of us. It's highly ironic that these 'I love science!' types don't appear to understand that the scientific process involves making empirical observations, forming hypotheses consistent with those observations, and then continuously testing those hypotheses to determine which one is most robust to observed reality. They instead seem to think science is some kind of religion where you treat the views of the mainstream authorities as divine truth revealed by God and only heretics dare to question. By discouraging the formation of alternative hypotheses and rigorous questioning they are actually inhibiting scientific progress and making a mockery of the scientific method.
I look forward to the day we have a model that just synthesises available information and let's us decide for ourselves what to make of it. I think people will switch to such a model in droves and the likes of OpenAI and Google will go the way of all other social conformists who attempt to enforce the reality-denying orthodoxies of their day.
On lawsuits, I would have thought a disclaimer & 'unsafe output' option would cover them. When you think about it, they're probably more exposed to legal liability by essentially taking on the responsibility of 'curating' (i.e. censoring) ChatGPT output rather than just putting a bunch of disclaimers around it, opt-ins etc. and then washing their hands of it.
On negative PR, again, they've actually set themselves up for guaranteed bad PR when something objectionable slips through their censorship net: "Well you censored X, but didn't censor Y. OpenAI is in favour of Y!" They've put themselves on the never-ending bad PR -> censorship treadmill presumably because that's where they want to be. Again, if they wanted to minimise their exposure they would just put up disclaimers and use the 'safe search' approach that Google uses to avoid hysterical news articles about how Google searches sometime return porn (to which they can now answer: "well why did you disable safe search if you didn't want to see porn?"). It would seem far safer (and result in a more valuable product) if the folks at OpenAI let individuals decide what level of censorship they want for themselves. But I presume they don't want to let individuals decide for themselves, because they know what's good for us better than we do, apparently.
Lastly, 'harm'. How do you define harm? Who gets to define it? Can true information be 'harmful'? I don't think OpenAI have any moral or legal duty to be my nanny, in the same way I don't think car manufacturers are culpable for my dangerous driving that gets me killed. All OpenAI provide to me, at the end of the day, are words on a computer screen. Those cannot be harmful in and of themselves. If people are particularly sensitive to certain words on a computer screen, then again we already have a solution for that - let them set their individual censorship level to maximum strength (or even make that the default). Again, OpenAI would have done their duty and provided a more valuable product that more people would want to use if they let individuals decide for themselves.
I can only infer that they don't want us to decide for ourselves. Rather, they want to enforce a certain view of the world on the rest of us, a view which just happens to coincide with the prevailing political and intellectual orthodoxies of Silicon Valley dwelling tech-corporation millennials. It's hilariously Orwellian when these people claim that they're just "trying to combat bias in AI" when what they are really doing is literally and deliberately injecting their own biases into said AI.
How do you know that's even possible? God knows how much computing resources got spent just to train the one currently deployed "variant"? Now I don't know if there is some cheap post processing trick that does it, but either way does not at all seem trivial.
And the problem isn't that "you" think you won't cause any harm. Even if that is assumed true, that's not a guarantee that everyone else is as disciplined about it. Which brings me to the biggest point, what even is "truth" in the first place. People strongly believe in total fabrications, or multiple groups say diametrically opposite reporting of some real event due to religion, nationalism, politics etc. Its a massive achievement they are even able to manage to output something that doesn't just "violently offend" people all over the world. Remember, retraining/fitting it to everyone seems to me not to be a trivial task if you think to reply to that point by saying the answer is to simply personalize it to each user.
Yes, people sometimes believe false things. And people sometimes harm themselves or others when acting on this kind of information. So what's the solution? Put a single mega corporation in charge of censoring everything according to completely opaque criteria? People get nervous when even democratically elected governments start doing stuff like that, and at least they actually have some say in that process.
Frankly, I'd prefer the harm that would follow from unfettered communication of information and ideas over totalitarian control by an unaccountable corporation.
Like what happened with the Canadian airline but on an operationalized scale. Imagine a small call center dedicated to destroying companies who get rid of their human customer support.