No, AI Isn’t Having ‘Hallucinations’ — Turns Out, It’s Worse Than That

Users of artificial intelligence are increasingly reporting issues with inaccuracies and wild responses. Some are even questioning whether it’s hallucinating, or worse, that it has a sort of “digital dementia.”

In June, for example, Meta’s AI chat assistant for WhatsApp shared a real person’s private phone number with a stranger. Barry Smethurst, 41, while waiting for a delayed train in the U.K., asked Meta’s WhatsApp AI assistant for a help number for the TransPennine Express, only to be sent a private cell number for another WhatsApp user instead. The chatbot then attempted to justify its mistake and change the subject when pressed about the error.

Google’s AI Overviews have been crafting some pretty nonsensical explanations for made-up idioms like “you can’t lick a badger twice” and even recommended adding glue to pizza sauce.

Even the courts aren’t immune to AI’s blunders: Roberto Mata was suing airline Avianca after he said he was injured during a flight to Kennedy International Airport in New York. His lawyers used made-up cases in the lawsuit they pulled from ChatGPT, but never verified if the cases were real. They were caught by the judge presiding over the case, and their law firm was ordered to pay a $5,000 fine, among other sanctions.

In May, the Chicago Sun-Times posted a “Summer reading list for 2025,” but readers quickly flagged the article not only for its obvious use of ChatGPT, but for its hallucinated and made-up book titles. Some of the fake titles suggested on the list were nonexistent books supposedly written by Percival Everett, Maggie O’Farrell, Rebecca Makkai and more well-known authors. The article has since been pulled.

And in a post on Bluesky, producer Joe Russo shared how one Hollywood studio used ChatGPT to evaluate screenplays — except not only was the research done by the AI “vague and unhelpful,” it referenced an antique camera in one script. The issue is that there isn’t an antique camera in the script at any point. ChatGPT must have had some kind of digital mental relapse and hallucinated one, despite multiple corrections from the user — which the AI ignored.

These are just a few of the shared posts and articles reporting the strange phenomenon.

AI has been heralded as a revolutionary technological tool to help speed up and advance output, but advanced large language models (LLMs) — chatbots like OpenAI’s ChatGPT — have been increasingly giving responses that are inaccurate, while offering up what it thinks is fact.

There have been numerous articles and social media posts of the tech struggling with more and more users reporting strange quirks and hallucinatory responses from AI.

And the concern might be warranted. OpenAI’s newest o3- and o4-mini models are reportedly hallucinating nearly 50% of the time, according to company tests, and a study from Vectara found that some AI reasoning models seem to hallucinate more, but suggested it’s a flaw in the training instead of the model’s reasoning, or “thinking.” And when AI hallucinates, it can feel like communicating with someone experiencing cognitive decline.

But is the lack of reasoning, the made-up facts and AI’s insistence on their accuracy a real indicator of the tech developing cognitive decline? Is the assumption it has any sort of human cognition the issue? Or is it actually our own flawed input mudding the AI waters?

We spoke with artificial intelligence experts to dig into the evolving quirk of confabulations within AI and how this impacts the overly pervasive technology.

In December 2024, researchers put five leading chatbots through the Montreal Cognitive Assessment (MoCA), a screening test used to detect cognitive decline in patients, and then had the scoring performed and evaluated by a practicing neurologist. The results found most of the leading AI chatbots have mild cognitive impairment.

CEO and co-founder of InFlux Technologies, Daniel Keller, told HuffPost he thinks generalizations about this AI “phenomenon” of hallucinations shouldn’t be oversimplified.

He added that AI does hallucinate, but it is dependent on several factors and that when a model outputs “nonsensical responses” it’s because the data on which models are trained is “outdated, inaccurate or contains inherent bias.” But to Keller, that isn’t evidence of a cognitive decline. And he does believe that the problem will gradually improve. “Hallucinations will become less frequent as reasoning capabilities advance with improved training methods driven by accurate, open-source information,” he said.

Raj Dandage, CEO and founder of Codespy AI and a co-founder of AI Detector Pro, admitted that AI is suffering from a “bit” of cognitive decline, but believes this is because certain more prominent or frequently used models, like ChatGPT, are running out of “good data to train on.”

In a study they performed with AI Detector Pro, Dandage’s team searched to see what percent of the internet was AI-generated and found an astonishing amount of content right now is AI-generated — as much as a quarter of new content online. So if the content available is increasingly produced by AI and is sucked back into the AI for further outputs without checks on accuracy, it becomes an infinite source of bad data continually being reborn into the web.

And Binny Gill, the CEO of Kognitos and an expert on enterprise LLMs, believes the lapses in factual responses are more of a human issue than an AI one. “If we build machines inspired by the entire internet, we will get the average human behavior for the most part with sparks of genius once in a while. And by doing that, it is doing exactly what the data set trained it to do. There should be no surprise.”

Gill went on to add that humans built computers to perform logic that average humans find difficult or too time-consuming to do, but that “logic gates” are still needed. “Captain Kirk, no matter how smart, will not become Spock. It isn’t smartness, it is the brain architecture. We all want computers to be like Spock,” Gill said. He believes in order to fix this program, neuro-symbolic AI architecture (a field that combines the strengths of neural networks and symbolic AI-logic-based systems) is needed.

“So, it isn’t any kind of ‘cognitive decline’; that assumes it was smart to begin with,” Gill said. “This is the disillusionment after the hype. There is still a long way to go, but nothing will replace a plain old calculator or computer. Dumbness is so underrated.”

And that “dumbness” might become more and more of an issue if dependency on AI models without any sort of human reasoning or intelligence to discern false truths from real ones.

Turns out, according to a new study from MIT, using ChatGPT might be causing our own cognitive decline. MIT’s Media Lab divided 54 participants in Boston between the ages of 18 to 39 years old into three groups and had them write SAT essays using ChatGPT, Google’s search engine (which now relies on AI), or their own minds without any AI assistance.

Electroencephalograms (EEGs) were used to record the participants’ brain wave activity and found that, of the three groups, the ones with the lowest engagement and poor performance were the ChatGPT users. The study, which lasted for several months, found that it only got worse for the ChatGPT users. It suggested that using AI LLMs, such as ChatGPT, could be harmful to developing critical thinking and learning and could particularly impact younger users.

Even Apple recently released the paper “The Illusion of Thinking,” which stated that certain AI models are showing a decline in performance, forcing the company to reevaluate integrating present models into its products and to aim for later, more sophisticated versions.

Tahiya Chowdhury, assistant professor of computer science at Colby College, weighed in, explaining that AI is designed to solve puzzles through formulating a “scalable algorithm using recursion or stacks, not brute force.” These models rely on finding familiar patterns from training data, and when they can’t, according to Chowdhury, “their accuracy collapses.” Chowdhury added, “This is not hallucination or cognitive decline; the models were never reasoning in the first place.”

Turns out AI can memorize and pattern-match, but what it still can’t do is reason like the human mind.