AI Error Rates 2026: How Often AI Gets Facts Wrong -- And What It Costs

[snot]

Key Findings:

"AI hallucination rates vary from 0.7% on grounded summarization tasks to 88% on legal queries, with the Stanford HAI 2026 AI Index Report documenting sycophancy-induced hallucination rates ranging from 22% to 94% across 26 frontier models — and a 2025 mathematical proof establishing that zero-hallucination is architecturally impossible for any large language model."

Results by Domain (June 2026 Snapshot):
(Sorry I couldn't figure out how to make the spacing work in the chart below, but after each category of query listed, the first stat is the "Baseline Error Rate" and the second is the "Peak Error Rate".)

Domain Baseline Error Rate Peak Rate

Legal Research 17.3% 88.0%
Healthcare / Clinical 43.1% 64.1%
Scientific / Academic Citation 30.0% 60.0%
Financial Analysis 15.0% 25.0%
General Knowledge (conversational) 4.8% 22.0%
Code / Technical Reference 3.1% 19.1%
Grounded Summarization (RAG) 0.7% 7.6%

The article is apparently based on studies by Stanford and other reputable-sounding sources. Much more at https://axis-intelligence.com/ai-hallucination-statistics/

[Ohio Joe]

1. "AI" does not " hallucinate"...

It’s code, it does what it’s coded to do. I really wish people would wake up and stop using the words the lying sellers want to foist off on them.

[snot]

4. I get you--

but it's not a bad metaphor for when AIs respond erroneously but with complete apparent confidence, as if they believed what they described.

I've got a much bigger problem with calling LLMs "intelligent," when they're really just abacusses for calculating probabilities of words and phrases occurring in various sequences.

[Ohio Joe]

6. A few more things...

1 - “AI” responds how it’s coded… That may be bad code… Or it could be intentional.

2 - “AI” does not have confidence, don’t ascribe human traits to code.

3 - You seem to feel LLM’s and “AI” are different things, they are one and the same. Elon and his ilk would have you believe differently and you are falling for it.

4 - They do not calculate probabilities… They retrieve data and present it out of large databases based on specific keys determined by input and code. Again you are ascribing human traits to code.

[snot]

7. I believe LLMs are a subset of AIs. There are also AIs designed for non-linguistic purposes,

such as the interpretation of visual images, e.g. in target-acquisition in weapons. Their focus is on things like finding edges and shapes, detecting movement, etc. I wouldn't be surprised if they didn't also utilize some kind of averaging, but I'm not sure just how closely they resemble LLMs, so I didn't want to overgeneralize.

As for whether LLMs use probabilities per se, if they don't, how do they select words, word order, etc.? (Because it's not based on signification or meaning.) Granted the particular coding is also critical, probably including rules re- grammar and any guidelines as to sources or topics to weigh more heavily or avoid, and also the particular selection of data the AIs are trained on; but my understanding is that the coding actually calls for the kind of averaging I have in mind, since it can't possibly specify in advance all possible answers to all possible questions.

Here's something from MIT that I just found with a quick search:

The base models underlying ChatGPT and similar systems work in much the same way as a Markov model. But one big difference is that ChatGPT is far larger and more complex, with billions of parameters. And it has been trained on an enormous amount of data — in this case, much of the publicly available text on the internet.

In this huge corpus of text, words and sentences appear in sequences with certain dependencies. This recurrence helps the model understand how to cut text into statistical chunks that have some predictability. It learns the patterns of these blocks of text and uses this knowledge to propose what might come next.

(Emphasis supplied; more at https://news.mit.edu/2023/explained-generative-ai-1109 .) So as I read this, the "parameters" may be the kind of guidelines I supposed would be included in what you're calling the code; but the process used to actually compose ChatGPT responses would be based on statistical probabilities.

[Ohio Joe]

8. You have zero idea of what you are talking about...

I have 35+ years as a technical expert. You are using terms incorrectly and repeating nonsense and believing things Elon and his ilk want you to believe.

You are being misled.

[snot]

9. Perhaps you could direct me to what you'd consider to be a better description

of how AIs "know" how to select and arrange words to create their responses?

I just looked at another article, this one from Microsoft, stating that:

Pretraining is where the model learns the bulk of its knowledge. The model is fed massive amounts of text from the internet — books, articles, code, websites — and learns to predict the next token given all previous tokens. This stage requires enormous compute (thousands of GPUs for weeks or months) and produces a base model.

A base model is essentially a text-completion engine. Given a prompt, it generates plausible continuations based on patterns in the training data.

* * * * *
Your full prompt (system message, conversation history, user input) is converted into tokens and fed into the model. The model processes all input tokens and produces a probability distribution over its vocabulary — predicting which token is most likely to come next.

(From https://learn.microsoft.com/en-us/agent-framework/journey/llm-fundamentals ).

Or from IBM:

LLMs work as giant statistical prediction machines that repeatedly predict the next word in a sequence. They learn patterns in their text and generate language that follows those patterns.

(https://www.ibm.com/think/topics/large-language-models)

I apologize if I'm not using all the terms exactly correctly, but the point I'm trying to make is that in the course of an LLM's training, it has somehow registered large and small patterns of how frequently words or phrases do or don't occur together and in what order. It has no understanding of what a vase is or what flowers actually are, but it can assemble the phrase, "a vase of flowers," because it has registered that statistically, vases more often contain flowers than trees or cookies. It's basically a statistical, probabilistic operation.

[Ohio Joe]

10. Let me tell you a story...

Back in ‘95, I was teaching a class on how computers work to a group of programmers with 2-3 years experience each. I got to the part where I explained the a cpu can only execute one instruction at a time. One of the guys raised his hand and said “ But Bill Gates recently said they had developed a cpu that could multitask”.

I told him the Bill Gates was lying to him. They had created a chip that had multiple CPU’s but the fact is, each one only executed one instruction at a time… This is still true today.

Offering what the people making and selling the so called “AI’s” to back your point is meaningless… Consider your source.

If you would like to come to Colorado and visit, I’ll give you a several hundred hours worth of training in the reality of how all this works. I won’t try to do it on a message board.

Facts are facts and programs don’t learn and they don’t get trained and they don’t have any of the multitude of human traits you insist on giving them.

Joe… Out.

[questionseverything]

2. That's a lot of errors

Since ai can’t be trusted it shouldn’t be used

[snot]

5. Another way to look at it

(at the risk of insulting some): even under the best of circumstances, an LLM's reponses will never be better than average, since it's producing them by probabilities based on averaging. So if your knowledge, intelligence, or verbal skills are below average, it might be useful to at least see what AI comes up with as a starting point.

Note how all this dovetails with corporate goals of minimizing labor costs not only by replacing workers but by replacing highly-qualified, higher-paid humans with less-qualified, lower-paid humans. This means, among other things, that a below-average employee who might already have hit their "Peter Principle" maximum position could conceivably move up another notch – AI thus making it possible to have more below-average employees at higher levels!

(For anyone not familiar with The Peter Principle, see https://en.wikipedia.org/wiki/Peter_principle .)

[GPV]

3. I use AI with medical things, but I always doublecheck everything it says, since can give

contradictory advice. That said, it has pointed me toward healthcare options my doctors have missed, and so I can open these conversations with them. It sometimes takes decades for them to put 2 + 2 together, whereas AI can pull info together very quickly.

It's also good for online shopping as far as price comparisons, and will let me know if a supplement might be good for me or not when checked against my vast array of meds and symptoms.