It’s too easy to make AI chatbots lie about health information, study finds
-
There should be a series of AI agents in place when a GPT is used. The agents intake the query and review the output before sending it off to the user.
what makes the checker models any more accurate?
-
what makes the checker models any more accurate?
Possibly, reverse motivation - the training goal of such an agent would not be nice and smooth output, but shooting down misinformation.
But I have serious doubts about whether all of that is feasible, given the computational cost of running large language models.
-
More AI agents /s
its just ai agents all the way down
-
what makes the checker models any more accurate?
The checker models aren’t trying to give you a correct answer with confidence. Their purpose is to find an incorrect answer. They’ll both do their task with confidence.
-
Who verifies the AI agent decisions?
The user. You could have the output include the “conversation” between the agents and validate the decisions. Not perfect, but better. People aren’t perfect either.
-
Well-known AI chatbots can be configured to routinely answer health queries with false information that appears authoritative, complete with fake citations from real medical journals, Australian researchers have found.
Without better internal safeguards, widely used AI tools can be easily deployed to churn out dangerous health misinformation at high volumes, they warned in the Annals of Internal Medicine.
“If a technology is vulnerable to misuse, malicious actors will inevitably attempt to exploit it - whether for financial gain or to cause harm,” said senior study author Ashley Hopkins of Flinders University College of Medicine and Public Health in Adelaide.
Isn't it too easy for the current chatbots/LLMs to lie about everything?
Train it on garbage or in the wrong way, and it will agree on anything you want it to.
I asked DeepSeek about what to visit nearby and to give me some URLs and it hallucinated the URLs and places. Guess it wasn't trained to know anything about my local area.
-
The checker models aren’t trying to give you a correct answer with confidence. Their purpose is to find an incorrect answer. They’ll both do their task with confidence.
the first one was confident. But wrong. The second one could be just as confident and just as wrong.
-
Possibly, reverse motivation - the training goal of such an agent would not be nice and smooth output, but shooting down misinformation.
But I have serious doubts about whether all of that is feasible, given the computational cost of running large language models.
how does that stop the checker model from "hallucinating" a "yep, this is fine" when it should have said "nah, this is wrong"
-
the first one was confident. But wrong. The second one could be just as confident and just as wrong.
Sure but they’re doing opposite tasks. You’re absolutely right that they could be wrong sometimes. So are people. Over time it gets better, especially with more regulation and smarter models.
-
Sure but they’re doing opposite tasks. You’re absolutely right that they could be wrong sometimes. So are people. Over time it gets better, especially with more regulation and smarter models.
opposite or not, they are both tasks that the fixed-matrix-multiplications can utterly fail at. It's not a regulation thing. It's a math thing: this cannot possibly work.
If you could get the checker to be correct all of the time, then you could just do that on the model it's "checking" because it is literally the same thing, with the same failure modes, and the same lack of any real authority in anything it spits
-
opposite or not, they are both tasks that the fixed-matrix-multiplications can utterly fail at. It's not a regulation thing. It's a math thing: this cannot possibly work.
If you could get the checker to be correct all of the time, then you could just do that on the model it's "checking" because it is literally the same thing, with the same failure modes, and the same lack of any real authority in anything it spits
wrote on last edited by [email protected]That’s not how it works though. It would be great if these AI models were deterministic but you can get different answers to the same questions at any given time. Given different input and given different goals, the agents wouldn’t likely fail on the same task when given proper instruction.
The main point is that it’s not going to be correct all the time. And neither is a human.
The regulation comes in when you’re dealing with sensitive information, like health diagnoses. There needs to be some logic in place to stop the models from being so confident with wrong answers that could hurt people.
Realistically, neither of us know what’s gonna work until we try it. Theoretically, verification agents would work.
-
That’s not how it works though. It would be great if these AI models were deterministic but you can get different answers to the same questions at any given time. Given different input and given different goals, the agents wouldn’t likely fail on the same task when given proper instruction.
The main point is that it’s not going to be correct all the time. And neither is a human.
The regulation comes in when you’re dealing with sensitive information, like health diagnoses. There needs to be some logic in place to stop the models from being so confident with wrong answers that could hurt people.
Realistically, neither of us know what’s gonna work until we try it. Theoretically, verification agents would work.
theoretically, they wouldn't, and yes, that is how it works. The math says so.
