RaterHQ vs AI: VA Disability AI Hallucinated in Real Time And Nobody Noticed But Me.

[00:00:00]Hello and welcome to Raider HQ after dark. I'm your host Raider. And before I get started on this video, I want to say that I'm not throwing shade. This isn't a negative video. Um, nobody on this particular topic has been Everyone's been really respectful. As so I'm going to carry that forward in this.

[00:00:18]Now, recently I did a video about artificial intelligence. And my video was used by Claim Raven. They kind of did a one-for-one to try to prove whether or not their program could actually what I assume is could it do it as well as I could.

[00:00:34]And I'll be honest with you, the program did fairly well. It It did okay. Like it's I don't have a problem with it. But as I did point out in the comments section that the the program actually hallucinated and gave inaccurate advice. Which is not this program's fault, right? I'm not I'm not just singling out this one program.

[00:00:58]There's a reason why AI is not good at decision-making because it has a very, very challenging time knowing when to apply the rules and when not to apply the rules.

[00:01:10]So, their program hallucinated.

[00:01:13]And the main thing that it hallucinated on is it it advised the veteran that they could not do a higher-level review based upon the circumstances of my video. Which is factually inaccurate because the main premise was that the Raider didn't catch the fact that it was a Muck Me 38 CFR 3.317 condition.

[00:01:37]Now, the program, what it said to do was do a supplemental claim.

[00:01:41]You can't do a higher-level review because if you tried to claim it as a Muck Me, that would be new evidence and you can't apply new evidence to a higher-level review.

[00:01:51]That fact pattern is true. However, what the computer program hallucinated on is that the veteran has always been an 1117 covered veteran. So, it's not like the new evidence would prove that they were a Southwest Asia serving veteran to qualify for muck me. That's already a record.

[00:02:13]So, what had happened was the rater had done a duty to assist error because they did not consider all of the actual medical opinions or theories for service connection that could technically apply to the rating decision.

[00:02:30]And if you want brass tacks, higher level reviews are almost specifically meant to address duty to assist errors.

[00:02:39]So, the computer program erroneously applied the wrong set of rules at the wrong time.

[00:02:46]Now, in my original video, I had talked about there's two different facets to this that must be true if you want to make a program that gives out the most accurate and efficient evidence more often.

[00:03:02]And half of the equation has to do with the coding, how well it's coded. And, you know, to the Claim Raven owner's credit, he said, "Well, you know what?

[00:03:13]I'll get on this. I'll fix this."

[00:03:15]So, you have to have constant updates, right? So, as things change, as things are found, you have to update them. So, that's one half of the equation to having a good AI program that is efficient and is accurate. As accurate as it can humanly be, right?

[00:03:35]The other facet to it is you have to have it be trained and reviewed by somebody that knows whether or not what it just said was true or false.

[00:03:48]So, in real time, I was able to help the Raven owner actually improve their product. Because up until the point that he fixed it, it was spitting out that same inaccurate evidence for that one specific instance.

[00:04:04]And nobody caught it.

[00:04:07]Because unless you know what right looks like, you can't tell or instruct the program, "Hey, you got that wrong."

[00:04:17]So, there has to be two facets to this.

[00:04:21]There has to be somebody that makes it really well and updates it constantly. And there also should be somebody that is reviewing to see whether or not it's accurate.

[00:04:31]Because unfortunately, reading the 38 CFR is not the same as applying it. Reading the M21 is not the same as applying it. Knowing when to apply it, which one favors more than the other one, those are situations that the AI has to be trained on.

[00:04:52]So, I'm just going to say once again, AI is really good at organizing your evidence.

[00:04:59]Currently, AI is not really good at making decisions or on anything more than like, say, secondaries, increases, newly claimed conditions.

[00:05:13]But the more complex you get it, the more moving parts there are, when 38 CFR might trump the M21 or vice versa, it loses its efficiency.

[00:05:25]Now, one of the things that I'm going to be doing in the future is I'm hoping to be partnering with an AI company to actually build and train to have both sides of that component there. Because you need somebody to say, "No, you're wrong.

[00:05:43]You are not correct, AI." And then tell somebody to fix it. If they're not doing that, then it'll spit something out and you'll say it's true. You'll file supplementals, higher level reviews based upon its its hallucinations, right? Cuz it can get things right a lot of the time. But when it hallucinates, it's to your detriment.

[00:06:03]And the main goal should always be troubleshooting. There should be somebody in there reviewing, beta testing, you know, like just literally trying to break this thing constantly.

[00:06:15]So, that's one of the things that I'm going to be doing here in the near future is I'm trying to line up a kind of a partnership deal to where I can make an AI better. I'm not beating up Claim Raven at all, but I just want to point out the irony of it hallucinating in real time.

[00:06:34]So, if we're going to put money towards something, right? If you the veteran are going to put money towards something, you need to make sure it is hallucinating the least amount of time that it can be, okay? So, if you're paying money, you want to make sure that the information you're getting is true and factual and it's not a hallucination.

[00:06:54]But anyway, guys, that's about all I got for you in this one. Greater out.