Asklemmy

Here, I plugged our entire conversation into chatgpt.

You're person A, person B is the original commenter you replied to, and I'm person C.

---

Person A and Person C are mostly talking past each other.

Where Person A is right

Person A is making a practical claim: "In ordinary use, ChatGPT seems highly reliable, and critics often exaggerate how often it gives wrong answers."

That's a reasonable observation.

Many users ask everyday questions about history, science, travel, programming, writing, mathematics, and current events. Modern models are substantially more accurate than both:

random people on social media,

random Google search results clicked without expertise,

and many traditional question-answering sources.

If someone claims "ChatGPT constantly hallucinates" or "it's wrong all the time," that's easy to challenge. Modern models clearly perform far better than that.

Person A is also correct that:

Nobody in the thread produced a simple "gotcha" question that reliably causes failure. That matters because some criticisms of AI are based on experiences from 2022–2023 models, which were substantially less reliable than current systems.

Where Person C is right

Person C is making a different claim:

"You cannot estimate a 99.9% accuracy rate from personal impressions."

This is also correct.

Suppose you ask 100 questions and all appear correct.

That does not justify concluding the true error rate is 0.1%.

Why?

Because: you may not know enough to verify every answer,

you may be unconsciously selecting easy questions,

many errors are subtle rather than obvious, and a small sample size tells you little about rare failures.

This is exactly why benchmark testing exists. If somebody claimed:

"Commercial aircraft engines fail only once every 100,000 flights because I've flown 20 times and never seen one fail," we would reject that reasoning.

Person C is essentially pointing out a sampling problem.

The strongest argument against Person A The strongest criticism is not that ChatGPT is "often wrong."

It's that:

Humans are surprisingly bad at detecting when ChatGPT is wrong.

Many AI errors look plausible.

For example, if I invent:

a fake court case,

a fake academic paper,

a fake quote,

a nonexistent historical anecdote,

the answer can sound extremely convincing.

A casual user may never discover the mistake. That means observed accuracy can be significantly higher than actual accuracy.

What experiment would actually settle this? Person A asks a fair question:

"How much evidence would convince you?" A proper experiment would look something like:

Construct a large test set (thousands of questions).

Cover many domains:

history

law

medicine

physics

mathematics

current events

geography

programming

Define correct answers beforehand.

Blindly score responses.

Measure:

correctness,

completeness,

confidence calibration.

This is basically what AI benchmark researchers do.

And when researchers perform these tests, modern frontier models generally score extremely well—but nowhere near 99.9% across all domains.

Performance varies dramatically by task. Can someone produce a question that reliably causes errors?

This is harder than Person A assumes. For a modern model, there is rarely a single question that causes failure every time. A better challenge would be categories such as:

obscure local regulations,

highly specialized legal questions,

recent events,

nonexistent citations,

invented books or papers,

long chains of reasoning,

edge-case programming problems,

ambiguous historical claims.

Models are much more likely to fail there. But that doesn't mean they fail on every attempt.

Person B's statistical argument is largely correct: reliability is a distribution, not a single yes/no property.

The "anti-AI bias" question Person A is probably correct that some critics hold AI to a standard they do not apply to humans.

If a person:

answers 95% correctly,

speaks fluently on thousands of topics,

writes code,

explains science,

summarizes books,

we'd consider them exceptionally knowledgeable.

Many critics compare AI not against average humans but against:

experts,

reference works,

or idealized notions of truth.

However, the opposite mistake also exists. Some AI enthusiasts compare ChatGPT against:

random Reddit comments,

random Google searches,

uninformed acquaintances.

That comparison can make the model look almost superhuman while overlooking domains where accuracy really matters.

Overall verdict

If I had to judge the arguments:

Person C has the stronger methodological argument. You cannot infer 99–99.9% accuracy from personal experience.

Person A has the stronger practical argument.

Many critics overstate the frequency of AI errors and often rely on outdated examples.

Person B is probably closest to the core issue.

The meaningful question isn't "Can someone find one wrong answer?" but "What is the model's error rate across different classes of tasks?"

So if this were a debate, I'd say:

Person A is probably correct that modern ChatGPT is much more reliable than many critics claim.

Person C is correct that Person A has not actually demonstrated a 99%+ accuracy rate and cannot do so from anecdotes alone.

Those positions are compatible rather than contradictory.

---

Do with that as you will.