Wednesday, September 10, 2025

Cut AI Some Slack — Its Hallucinations Are Our Own

 



To say humans have gone hard on AI for its hallucinations is an understatement. The number of comments punching holes in its capabilities is overwhelming, to say the least. But as the saying goes: the apple doesn’t fall far from the tree. If AI is a repository of everything we’ve said and created, then its ability to mirror our own behavior isn’t far-fetched. And that is exactly what OpenAI has found.


Truth is, if you ask a model a hard question, it will sometimes give you a perfectly confident, perfectly wrong answer. This has made many people skeptical, leaving them to ask: If AI can’t separate fact from fiction, how can we trust it?


A new OpenAI paper (Why Language Models Hallucinate) argues that hallucinations aren’t some mysterious glitch in the matrix (pardon my pun), but a predictable outcome of how language models are trained and tested. In pretraining, even with perfect data, statistical pressures guarantee some errors—just like misclassifications in traditional machine learning. Then, in post-training, the issue is reinforced because benchmarks reward models that “guess” rather than admit uncertainty. Think about it: much like students taking multiple-choice exams, AI has learned that bluffing pays off. Saying “I don’t know” is penalized, while offering a confident falsehood earns points.


The authors conclude that hallucinations persist not because models are broken, but because the system around them values certainty over honesty. Seen this way, AI is less like an alien intelligence and more like a mirror, forcing us to take a good look at ourselves. It reflects our biases, shortcuts, and blind spots in how we prize certainty.


On LinkedIn, in corporate cultures, and in everyday conversations, truth often gets spun into polished half-truths because our social “benchmarks” reward positivity, optimism, and confidence. We get more likes, more applause, more agreement when we sound upbeat—even if it bends reality/truth. Meanwhile, saying “things are tough, I don’t know how this will work” rarely earns the same recognition.


So yes: toxic positivity is the human version of AI hallucination.


Cutting AI some slack means recognizing that its flaws are, in part, our own. If we want more trustworthy systems, the fix is as much social as it is technical: realigning how we evaluate and reward them, so honesty—admitting when you don’t know—becomes a strength, not a weakness.

No comments:

Post a Comment

Microsoft Buyouts and the Quiet Repricing of Human Thinking in the Age of AI

  Microsoft offering buyouts to 7% of its US workforce  can be interpreted in many ways, but it is first and foremost a pricing signal for ...