Adaptive Provisional Belief: Avoiding Easy Wrong Answers
An AI-assisted delusion of a symbiotic robot future.
Restless in the middle of the night, I asked ChatGPT-5 what made it different from GPT-4. One answer stood out as a key tenet of The Unknowable Truth. The robot told me, “It can handle trickier reasoning without ‘snapping’ to an easy wrong answer.” It continued:
GPT-4 sometimes fell into “familiar pattern” traps—if a question sounded like a known riddle or math puzzle, it might answer reflexively. GPT-5 is more careful, double-checking the wording and doing step-by-step work before committing to an answer.
The phenomenon that the robot attributes to GPT-4 mirrors human behavior. In Dr. Daniel Kahneman’s book Thinking Fast and Slow, the cognitive psychologist describes the human mind as having two systems: System 1 and System 2.
System 1 interprets the world and makes quick judgments using mental models based on previous experiences. Kahneman writes, “System 1 operates automatically and quickly, with little or no effort and no sense of voluntary control.”
When a novel situation requires more analysis, System 2 kicks in and “allocates attention to the effortful mental activities that demand it, including complex computations.”
The human mind uses mental shortcuts called heuristics to devise quick answers for familiar problems. “If a satisfactory answer to a hard question is not found quickly,” Kahneman writes, “[System 1] will find a related question that is easier and will answer it. You often answer a different question from the one you were asked, without noticing the substitution.”
This means that, when you ask another human a complicated question, their System 1 brains may trick them into substituting an answer for an easier question. Sometimes people make things up and unknowingly pass them off as the truth. We’re intuitively aware of this behavior in other humans, so we’ve learned to take what people say with a grain of salt.
Many of us, however, learned to believe what computers tell us. Growing up in the 1980s, it wouldn’t have occurred to me that a Commodore 64 or TRS-80 would produce inaccurate output. If something didn’t compute, I recall getting some kind of error message. These days, when AI doesn’t compute, it just makes something up. Computers have started acting more like people, for better or worse.
The robot also told me that GPT-5 is “more self-aware about limitations.” It continued:
GPT-4 could overstate certainty or “hallucinate” citations more easily. GPT-5 tends to flag uncertainty earlier and suggest verification steps—which is handy if accuracy matters (like legal or medical topics).
(As a complete aside, I love that the robot recognizes that sometimes, in the messy emotional lives of us humans, accuracy does not matter.)
Like GPT-4, humans are also apt to “overstate certainty.” Humans are uncomfortable with uncertainty. We’d like to believe the world is concrete and predictable, but it is not, and to assuage our anxiety, we’re willing to delude ourselves. AI doesn’t delude itself, but it does “hallucinate” when it doesn’t know what it’s talking about—perhaps to please humans in our efforts to delude ourselves.
I don’t know how AI will evolve or what it means for humanity. The future is uncertain. I chose to believe that humans and AI will evolve together in a mutually beneficial symbiosis. But perhaps that’s just a delusion designed to alleviate my anxiety.
I asked GPT-5 if I was deluding myself regarding my rosy vision of the future. The robot says, “That’s not self-deception so much as adaptive provisional belief: a story that helps you operate in the present while remaining open to change. In The Unknowable Truth frame, it’s a working model that you update as the feedback loops come in.”