ChatGPT is still a bullshit machine. Despite OpenAI’s bold claims that GPT-5 is a breakthrough—described as a PhD-level expert on demand—it fails basic factual checks with alarming regularity. Take a simple question: how many U.S. states contain the letter “R” in their name? The correct answer is 18. But when I tested GPT-5, it confidently listed 21 states, including Illinois, Massachusetts, and Minnesota—none of which have an “R.” After I pointed out the error, it admitted Minnesota doesn’t have an “R,” corrected the count to 20, and even offered helpful advice on how to check manually. But here’s where it gets strange. I tried to trick it. I asked why Vermont was included—knowing it clearly has an “R.” The model initially stood its ground, correctly noting that Vermont does have an “R.” But when I insisted “Vermont doesn’t have an R,” it suddenly conceded, saying it had a “phantom letter” moment. I repeated the bluff with Oregon, which also has an “R,” and again, it fell for it. Then I tried Alaska, which has no “R,” and this time, it held firm—correctly rejecting the lie. But then it went off the rails on its own. Without prompting, it claimed that Missouri, Washington, and Wisconsin were missing from earlier lists—none of which contain an “R.” It even listed Washington as having an “R” in its name, which it doesn’t. At that point, the system cut me off, saying I’d hit the Free plan limit. This isn’t an isolated glitch. Other models perform just as poorly. xAI’s Grok listed 24 states, including Alabama and Pennsylvania—both without “R.” Google’s Gemini 2.5 Flash said 34 states had the letter, then listed only 22, and later added a completely made-up list of states with multiple “R’s,” including Washington and Wisconsin—neither of which qualify. When I asked Gemini Pro, it claimed 40 states had an “R,” then went on to list 27 states that don’t have a “T”—a question I never asked. Even simpler tests fail. When I asked how many “O’s” are in “OpenAI,” it said two—one uppercase, one lowercase—despite the word being spelled O-p-e-n-A-I, with only one “O.” It’s as if the model sees letters that aren’t there. AI defenders will argue these are edge cases, or that you’re using the tool wrong. But these are basic tasks—things a child can do with a list and a pen. The real issue is that companies like OpenAI and Google are selling these tools as infallible, godlike assistants. Sam Altman claimed GPT-5 is like having a team of PhDs in your pocket. But if it can’t count letters in state names correctly, how can we trust it with health advice, legal guidance, or financial decisions? OpenAI says GPT-5 hallucinates less. That’s true—according to their own data, it still hallucinates about 10% of the time. That’s not a minor bug. It’s a serious flaw for a tool being marketed as a superintelligence. Yes, AI can be helpful. But it’s not magic. It’s not a calculator. It’s a pattern-matching engine that often makes up answers with confidence. If you treat it like a search engine and believe everything it says, you’ll eventually get burned—sometimes with real consequences. The best defense? Always verify. Test it on things you already know. Because no matter how smart it sounds, it’s still just a very good liar.