# An AI Just Handed Me a Fake $67B Statistic

https://www.youtube.com/watch?v=mKTg53LFrN4

[00:00] Here's something that happened while I was researching this exact video.
[00:04] The AI that I use handed me a number, $67.4 billion, the global cost of AI hallucinations.
[00:12] Specific, confident, even had a source attached, and it was completely fake.
[00:17] None of it was real.
[00:19] If you follow this channel, you know my whole life and business runs through one local folder and Claude, and I do not trust a single AI's built-in web search to tell me what's true.
[00:27] Not before I research, not before I code, not before I build anything in my I core, our membership platform.
[00:34] So, let me show you how I fix that inside the folder.
[00:39] Because here's the thing, AI doesn't fail by going quiet.
[00:41] It fails confidently.
[00:44] Same calm voice for a wrong answer as a right one.
[00:47] And the numbers are worse than most people think.
[00:51] Columbia's Journalism Review tested eight AI search engines.
[00:54] They cited their sources wrong more than 60% of the time, over half.
[00:57] Grok got it
[01:01] wrong on 94% of tests.
[01:03] Now, you'd think, "Okay, that's just a chat model guessing.
[01:05] Give the AI the actual source article and it's fine, right?
[01:09] No, the BBC ran exactly that test.
[01:12] They handed the AI the source.
[01:15] It still misrepresented the news.
[01:18] And this isn't some lab problem.
[01:20] People are already getting sanctioned in court for trusting this.
[01:22] So, back to my $67.4 billion.
[01:24] When my research agent Pax was pulling stats for this video, that number showed up, 67.4 billion in hallucination costs.
[01:35] And a second one, 47% of enterprise users making a major decision on hallucinated content, supposedly from Deloitte.
[01:43] Perfect numbers for video like this.
[01:45] Specific, citation shape, exactly what I wanted on screen.
[01:48] And that's exactly why I don't trust one source, because we checked.
[01:54] The number traced back to a single SEO blog.
[01:57] No study, no Deloitte report, no primary source anywhere.
[02:00] It just
[02:03] evaporated.
[02:03] Think about that.
[02:05] The method called a fake number while researching a video about catching fake numbers.
[02:10] If I trust 1 AI, that 67.4 billion would be on your screen right now presented as fact.
[02:19] So, here's the fix, and it's simple.
[02:19] I don't ask 1 AI.
[02:22] Pax asks the same question twice through two completely independent engines.
[02:28] Pax runs on Claude, but then calls the research via Perplexity, which is built specifically for search.
[02:33] It reads the full page, pulls citation citations, and Brave, which runs its own index of more than 30 billion pages.
[02:41] It's own index, not reselling Google or Bing, generally independent.
[02:47] Two engines, same question, then Pax, the agent running with Claude, compares the two answers.
[02:53] Where they agree, I trust it.
[02:55] Where they disagree, that's the red flag.
[02:58] That disagreement is the signal.
[03:00] That's where the hallucination is probably hiding,
[03:03] that's where I let my AI team dig deeper.
[03:05] But, in order to spot this, you need to have some idea about the topic you're talking about.
[03:11] That's why we keep saying, AI makes the capable faster, but it doesn't make the clueless capable.
[03:15] It's the same rule good journalists have used forever.
[03:18] Two independent sources before you print.
[03:20] There's even research showing that comparing answers like this measurably cuts errors.
[03:25] I just gave that rule to my AI team, and that's why I lean on an engine built for search instead of a chat model bolting on a quick browse.
[03:35] The dedicated engine reads the full page and runs its own index.
[03:37] The chat model peaks at one snippet, then fills the rest from memory.
[03:42] That gap between what it read and what it remembered is where the made up confidence comes from.
[03:48] So, here's how I actually use this on my desk everyday.
[03:50] Before I code, before I build anything in the my I cop location, before I create any content or make business related research, even for my life when I search for the best hotel or holiday
[04:05] destinations.
[04:07] Anytime I'm working in a domain I'm not 100% expert in, I send Pax first because you have to know enough to ask the right question.
[04:14] And then the dual source check is how I kill the false ideas before they cost me anything.
[04:18] That's the whole point of how my team is wired.
[04:20] Larry runs the team.
[04:23] Pax make sure what I build on is actually true.
[04:27] So when you build in your own folder, you're building on something that's been checked, not something an AI just sounded confident about.
[04:32] So that's it.
[04:34] Don't trust one AI's built-in web search.
[04:37] Ask twice through two independent engines and watch where they disagree.
[04:42] The confident answer is not the same as the correct answer.
[04:44] Build that gap into your process and it stops biting you.
[04:49] Quick one for the comments.
[04:51] Has any AI ever handed you a confident completely fake number?
[04:54] I want to read those.
[04:56] This whole video was about the inputs, how I make sure what I build on is actually true.
[05:01] So the real question is what I build once I trust them.
[05:05] That's the next video.
[05:07] I show you how Claude replaced my note-taking app, my planner, even my health apps.
[05:12] Real proof, all of it on screen.
[05:14] Go watch that one next.
[05:17] It's the other half of this story.
[05:17] That's it for this video.
[05:19] I catch you up in the next one.