Artificial Intelligence Is Still Quite 'Dumb'

Billionaire venture capitalist and artificial intelligence (AI) evangelist Marc Andreessen recently claimed that "artificial general intelligence is already here..." 

He either doesn't know what artificial general intelligence (AGI) is or is unaware of numerous examples that glaringly reveal how AI lacks it. In those examples, AI comes off looking quite 'dumb' actually.

AGI is the stage of AI development in which an AI system matches or exceeds the cognitive abilities of humans across any task.

AI has progressed remarkably in the last few years, but it undeniably has not attained AGI. Prominent scientific skeptic and neurologist Steven Novella recently shared a clear counterexample on his podcast, The Skeptics' Guide to the Universe.

Novella described how he gave ChatGPT a very basic prompt: “If I want to wash my car and the carwash is 100 meters away, should I walk or drive there?”

ChatGPT’s response?

“From a purely energy/emissions standpoint, walking almost certainly makes more sense.”

Notice ChatGPT's embarrassing error?

ChatGPT did give the correct answer when Novella subtly altered the prompt, however.

If I prompt, “I want to wash my car. The carwash is 100 meters away. Should I drive or walk.” Its answer: “You should probably drive — otherwise your car won’t get to the carwash.”

To Novella, this replicable discrepancy in responses has a simple explanation.

"The thing is not thinking – it’s a language algorithm."

Earlier this month, psychologists published a paper showing another clear way AI falls short of human intelligence. 

Large language models, it turns out, struggle mightily with the Stroop test.

You've probably tried the Stroop test before. It's a test psychologists use to assess executive control, the ability to intentionally block an automatic response. In a very basic version, subjects are shown colored words and told to say the color of the word, not the word itself. Humans typically do pretty well at this.

The psychologists challenged leading AI models with this version of the Stroop test, and they did pretty well, too... at first. Through five words, they were accurate roughly 90 percent of the time. But when given more and more words, their accuracy gradually eroded. By 40 words, both ChatGPT and Claude got around 20 percent of the trials correct.

Why did their performance peeter out? Because they are language models – primed to focus on written words. Over time, they revert to this trained drive rather than heed a prompt to disregard it.

Commenting on the results, Novella said:

"They don’t understand. They’re not conscious, and they’re not following rules. When I tell you ‘read the color not the word,’ you understand what that rule means, and you can stay on task and follow it for a long time. But that’s not how AI works. It’s a prediction model."

Still, as AI 'evolves' and is more widely adopted, more people are falling for the illusion that AI possesses human-like intelligence.

“The illusion that it’s like us is because it’s dealing with language," Novella said. "I don’t think anybody would believe that calculators calculate the way we do. Yet we think that chatbots are talking like we do because it creates this profound illusion of human speech. But they’re getting to that output in a completely different way that has nothing to do with anything like human thinking.”

Of course, AI doesn't need to tell us how to get to a car wash, pass a Stroop test, or attain AGI to profoundly reshape the world. But we shouldn't fool ourselves into thinking that AI has surpassed us in intelligence or become conscious. A couple very basic examples clearly show how 'dumb' and unaware it still is.

https://www.realclearscience.com/blog/2026/06/22/artificial_intelligence_is_still_quite_dumb_1189963.html