A Hacker News thread asking whether the hallucination problem in AI has been solved produced one of the sharper reframes I’ve seen on the subject. According to Hacker News, a commenter cut through the panic with a distinction most of the debate skips right past: the machine vision systems that have run factory floors for 30-odd years are not the same “AI” that invents fake case law in a chatbot. Lump them together and the hallucination question becomes unanswerable, because you’re asking one question about two completely different machines.
That’s the myth worth killing. “AI” gets talked about as a single thing that either does or doesn’t make things up. It isn’t.
Two Machines, Two Failure Modes
The commenter’s example is blunt. Industrial vision systems have spotted misaligned labels, badly folded paper, and wrong items on high-speed conveyor belts for decades. They fail, but they fail in narrow, measurable ways. Feed one a picture of a cat and it returns a probability, not a confident paragraph of fiction.
Large language models are different animals. They generate fluent, plausible text whether or not the underlying claim is true. The failure isn’t a bad reading on a sensor. It’s the model doing exactly what it was built to do: predict the next likely token, facts optional.
Here’s why the difference matters:
- Vision classifiers operate in a bounded problem space. Cat or not-cat. Label straight or crooked. You can measure error rates and set thresholds.
- LLMs operate in an open-ended space. There’s no fixed “correct” output to check against for most prompts, which is exactly why hallucination is so hard to pin down.
- Confidence signals differ. A vision system hands you a score. An LLM sounds equally sure when it’s right and when it’s inventing.
Why This Lands Now
The industry keeps shipping products that blur these categories on purpose. “AI-powered” gets slapped on everything from spell-check to autonomous targeting, and buyers assume the reliability of a 30-year-old vision pipeline transfers to a brand-new language model. It doesn’t.
The Hacker News discussion actually started somewhere darker: the ethics of autonomous targeting systems, using the example of vision-guided devices that trigger on feral cats while letting native animals walk by. The point that stuck with me wasn’t the ethics. It was that a well-bounded vision system can be held to a real accuracy standard, while an LLM’s “accuracy” is a moving target we’re still learning to define.
That gap is the whole ballgame for anyone deploying this stuff.
What Practitioners Should Actually Do
If you’re building or buying, stop treating hallucination as one problem with one fix:
- Name the machine. Is this a bounded classifier or an open-ended generator? Your risk controls should be completely different for each.
- Demand error rates for narrow systems. Vision and classification tasks have measurable accuracy. If a vendor can’t give you numbers, that’s your answer.
- Wrap generators in guardrails. For LLMs, assume confident fiction is the default failure. Retrieval, citation checks, and human review aren’t nice-to-haves, they’re the product.
- Watch the “AI” label. When a pitch borrows the reliability of one category to sell another, push back.
So, has the hallucination problem been solved? For bounded vision tasks, it was largely a solved engineering problem before the current AI wave even started. For LLMs, no, and framing it as one binary question is part of why progress feels slippery.
The useful move isn’t waiting for a magic fix. It’s getting precise about which machine you’re holding, and building for how that specific machine breaks. More detail and the full thread are at the original source.