Can AI Make Mistakes? Who Checks the Work in 2026

Can AI make mistakes? Yes, and it makes them confidently. Every major AI tool ships with a disclaimer saying so, and the companies getting real value from AI in 2026 are not the ones with the best prompts. They are the ones with a review process. AI produces the draft; a human, a checklist, or even a second AI checks the work before it touches a customer.

That answer sounds obvious. What surprised us is how much money is now being spent on it.

About 720 people a month search the exact phrase "AI responses may include mistakes," according to our July 2026 DataForSEO pull. That is not a search trend. That is the disclaimer under Google's own AI answers, typed back into Google by people who wanted to know if the machine was serious.

It was serious. And this month, three very different companies showed us what they are doing about it.

Can AI make mistakes even when it sounds confident?

Yes. AI can be wrong and confident at the same time, and that is the whole problem.

A language model does not look facts up. It predicts what a good answer probably looks like, which means a wrong answer arrives in the same fluent, assured tone as a right one.

OpenAI's own SimpleQA benchmark made this measurable: its top models answered fewer than half of the short factual questions correctly (OpenAI, 2024). The models have improved since, but no vendor claims zero. That is why the disclaimer exists.

For a small business, the tone is the trap. A junior employee who is unsure sounds unsure. AI never does. Confidence is not accuracy, and nothing in the output tells you which one you got.

So the practical question is not whether AI makes mistakes. It is who catches them before your customer does.

What does an AI mistake actually cost a business?

Real money, and sometimes a courtroom.

If you are wondering whether AI can make mistakes that cost actual dollars, the clearest answer comes from a real case. In 2024, Air Canada's website chatbot invented a bereavement refund policy that did not exist. A customer relied on it, the airline refused to honor it, and a Canadian tribunal ordered Air Canada to pay, ruling the company was responsible for what its own chatbot said (Moffatt v. Air Canada, 2024).

That ruling answered the liability question for everyone: the business owns the mistake, not the AI. Your chatbot's promise is your promise. The wrong price your AI assistant quoted is your price. "The AI said it" has already failed in court.

For the small businesses we work with in AI consulting, the risky mistakes are rarely exotic. They look like a wrong service price in a follow-up email, invented business hours in a Google profile post, a citation to a study that does not exist, or a confident answer to a customer question the AI had no data for.

Each one is small. Each one is public. And your customers are better at spotting them than you might think: 83% of consumers say they can detect AI-generated messaging (Averi.ai). The catch has to happen on your side of the send button.

Which raises the question the big companies just answered.

How do big companies check AI work in 2026?

They build verification layers instead of trusting single outputs.

Three examples from this month alone show the pattern at three different scales.

DoorDash runs juries on its AI. Its engineering team published how it validates AI-generated food data with what it calls LLM juries: multiple model passes that check and refine each other's output before anything ships (DoorDash Engineering, July 2026). When the data involves allergens, one confident wrong answer is not an acceptable failure mode.

Programmers are redesigning code around review. Jacquard, a new programming language that topped Hacker News in July 2026, is built on the assumption that AI writes the code and humans judge it. The language's whole design goal is making the review step fast and reliable.

Microsoft measured its own rollout. The company published a study of its early 2026 internal deployment of AI coding agents, treating adoption as something to measure and verify rather than assume (arXiv, July 2026).

Different industries, same conclusion. Here is what that pattern looks like at enterprise scale versus yours:

Verification layer	What DoorDash-scale looks like	What your business needs
Second pass on output	Juries of multiple AI models	One human read before send
Rules for what ships	Automated validation pipelines	A one-page checklist
High-stakes red lines	Allergen data never ships unverified	Prices, promises, and policies never ship unverified
Learning from errors	Model evaluation datasets	A running log of catches

The budgets differ by six zeros. The idea does not: the value is not in generating the work, it is in trusting what ships.

And that idea scales down further than most owners expect.

How does a small business build an AI review process?

You need five habits, not a jury of AI models.

Most teams can install them in an afternoon.

Give every AI lane an owner. If AI writes your follow-up emails, someone owns follow-up emails. Unowned output is unchecked output, and it will ship a mistake eventually.
Use the two-pass rule. Never send an AI draft in the same sitting it was generated. Generate, step away, then read it once as the customer would. The mistake that was invisible while you were prompting is usually obvious when you are reading.
Make facts earn their place. Any number, name, date, or claim in AI output either gets traced to a source or gets cut. This single rule removes most of the embarrassing failure modes, because AI mistakes cluster around invented specifics.
Draw red lines. Some things never ship without human review: prices, availability, legal or medical wording, refund promises, anything with a customer's name on it. Write the list down. The Air Canada case is what a missing red line costs.
Log every catch. Keep a running note of the mistakes you caught. Within a month you will know your tools' failure patterns, and the log becomes the training doc for the next person, or the checklist for the next tool.

We hold ourselves to the same gate. Every post on this blog, including this one, moves through an AI-assisted pipeline that is not allowed to publish anything. A human reads the draft and taps approve, or it does not ship. That rule has caught real errors, and it costs about four minutes per post.

The five habits work because they match how AI actually fails. Which is worth understanding for one more minute.

Why does AI make mistakes at all?

Because it predicts, it does not look up.

AI generates answers by predicting likely words, not by consulting a database of facts. So can AI make mistakes even as the models improve? Yes, by design it can, because prediction has no built-in fact check. When the prediction lands on truth, it looks like knowledge. When it lands beside the truth, it looks exactly the same. The industry calls these misses hallucinations, and every vendor, from OpenAI to Google to Anthropic, documents them as a known limitation.

Two things make it worse in business use. AI does not know what it does not know, so it fills gaps instead of flagging them. And it optimizes for a satisfying answer, which is not the same thing as a correct one.

None of that makes AI a bad tool. It makes AI a fast, tireless, occasionally wrong junior employee. Businesses already know how to manage one of those: you review the work, you keep the stakes matched to the trust earned, and you never let the intern quote prices unsupervised.

We wrote about the input side of this in AI slop is a planning problem. This is the output side: garbage out is survivable, unreviewed out is not. If the review burden feels like it erases the time savings, it does not, and AI still needs you breaks down where the saved time actually goes.

FAQ

Can AI make mistakes with a good prompt?

Yes. Better prompts reduce mistakes, they do not eliminate them.

A prompt shapes what the AI attempts; it cannot give the model facts it does not have or make it flag its own gaps. Prompt quality is the input control. A review step is the output control, and businesses need both, the same way a good brief does not remove the need to proofread.

How often does AI make mistakes?

Often enough that no vendor will publish a zero.

Error rates vary widely by task: summarizing a document you provided is far more reliable than answering open factual questions, where OpenAI's SimpleQA benchmark showed even top models scoring below 50% (OpenAI, 2024). The honest planning number for a small business is not a percentage. It is "assume every output can be wrong, and size the review to the stakes."

Who is responsible for AI mistakes?

The business that published the output, not the AI vendor and not the AI.

That is not theory; a Canadian tribunal ordered Air Canada to honor a refund policy its chatbot invented (Moffatt v. Air Canada, 2024). Courts, customers, and platforms all treat your AI's words as your words. If a mistake would embarrass you coming from an employee, it carries the same weight coming from your AI.

Does AI make more mistakes than humans?

It makes different mistakes, faster.

A tired employee makes sloppy errors that look like errors. AI makes fluent, confident errors that read like facts, and it can make them at volume. Humans also catch their own mistakes at a rate AI cannot match, because AI does not know what it does not know. That is why the fix is pairing them: AI for speed, a human for judgment.

Can AI check its own work?

Partly, and big companies now do this at scale.

DoorDash validates AI output with juries of multiple AI passes (DoorDash Engineering, July 2026), and asking a model to critique a draft does catch real errors. But every layer of AI checking still ends at a human red line for high-stakes output. Self-review raises the floor. It does not remove the need for an owner.

What should a small business never ship without human review?

Anything involving money, commitments, or a customer's name.

That means prices, quotes, refund and warranty language, business hours, legal or medical wording, and replies to upset customers. These are the outputs where one wrong sentence costs real dollars or real trust. Everything else can move faster, which is exactly why the red-line list has to be written down.

Sources

DoorDash Engineering, "Building Food Metadata with LLM Juries" (July 2026)
Moffatt v. Air Canada, British Columbia Civil Resolution Tribunal (2024)
OpenAI, SimpleQA factuality benchmark (2024)
Microsoft, arXiv study of early-2026 internal AI coding agent rollout (July 2026)
Jacquard programming language (July 2026)
Averi.ai, consumer detection of AI-generated messaging
Kindly Creative, DataForSEO keyword pull (July 2026)

The tools will keep getting better, and the disclaimer will keep being true at the same time. Both things have been true for three years now. So the question that actually separates businesses in 2026 is not "can AI make mistakes?" That one is settled. It is the quieter one: when your AI is wrong next week, in front of a customer, with your name on it, who catches it?