2 minute read
If you’ve been messing around with AI since it’s started you’ve hear this. “Yes, it hallucinates, but it will get better in the next update.” So, fast forward to 2026, and the darn chatbots are still making things up. Yes, it’s called hallucination, but it’s really essentially lying. Granted we wanted AI to be human, right? Well, there ya go. People sometimes lie, too.
It doesn’t happen all the time. I’ve seen AI generate some amazing images, great code, and even pretty decent written words. Granted, you can still tell it’s not quite a human behind the wheel in the writing. As for grammar and spelling, I’m sure plenty of us appreciate the help to sound like we have an amazing grasp of the English language because of a little help. Personally, I learned that commas have a bad habit of winding up in the wrong places or sometimes go completely missing without a little assist.
As for images, nano banana creates things I could never even dream of making with Photoshop or Illustrator myself. If you want to read more about that angle, check out my article on AI and the artist designer’s dilemma.
All that said, there are times when these wonderful capabilities get annoyingly mixed in with a stubborn habit of stating things that aren’t true with an air of total and unwavering confidence.
Statistics that just don’t exist
The classic move is the made-up statistic. If you ask about remote work and cities, you might get a confident reference to a Brookings report from 2023 showing “34.12% of downtown retail vacancy tied to hybrid policies.” It sounds real, but it isn’t. The report doesn’t exist, and neither does the number. If you ask for a source, you will get a dead link. Then you might get a second fake citation offered up like nothing happened.
The numbers always sound specific enough to be believable. They use two decimal places, a reputable institution name, and a specific year. That combination gets fake claims into presentations and reports more often than anyone wants to admit.
Frequent apologies become meaningless
When you push back, models tend to fold immediately. You get a big apology, a slight correction, and occasionally a second hallucination delivered in a more humble tone. The sorry sounds genuine, but the underlying issue does not budge. It is less a sign that the model has reflected on its mistake and more a sign that agreeing with you is the path of least resistance.
Factual Lookups shouldn’t be this messy
Sometimes asking for simple facts gets complicated and kind of embarrassing. Try it. Ask for a date of birth with different AI models. Some will be spot on, others with take a guess without even searching. I once had an argument with an AI model using Ollama that said it couldn’t search the internet and insisted it was 2024 because that is when its “training data” ended. I saw it “thinking” with itself and having some sort of trust conflict where it felt I was misleading it.
Searching for yourself is its own adventure. People have asked AI to pull up their professional background and gotten a biography that blends in details from two other people who share their name. This is all presented as legit. If you point it out, sometimes the corrected version is often just as confused or even worse.
Common names are a great way for an AI model to fall off a cliff and suddenly you have the details of a “collage person” the AI has chosen to construct. Less common names present a fresh problem because the model doesn’t have much to work with. It fills the gaps with “inference”, which is a fancy word for guessing, and produces a portrait of someone who doesn’t quite exist.
The Bottom Line
Sure, the AI models have improved on this stuff. Retrieval techniques and better calibration have brought error rates down compared to a few years ago. But these systems are also trusted more and used for higher-stakes work than before. Progress is happening, but so is exposure.
Superconfident tone sometimes delivering a very wrong answer. That combination is still very much on the table in 2026. It is worth keeping in mind the next time a chatbot cites a study you have never heard of. You’re not crazy; it might not exist. Always double-check the work!
Want more AI topics?
- AI Hallucinations in 2026: Still a Problem
- Self-Driving Cars: Ready for Primetime or Still a Highway Pipe Dream?
- SpaceX Bought Cursor for $60 Billion, OpenAI Cut It Off in Days
