AI tools write fluent, confident answers, and sometimes those answers are simply wrong. A made-up statistic, a policy that doesn’t exist, a citation to a case that was never decided. Businesses have learned this the hard way: in 2023 a U.S. court sanctioned lawyers who filed a brief citing cases ChatGPT had invented, and in 2024 a Canadian tribunal held Air Canada responsible for a refund policy its website chatbot made up. Understanding why AI makes mistakes is the key to catching them before your customers do.
Why AI gets things wrong
Reason 1
It predicts, it doesn’t look up
Language models generate the most likely next words based on patterns. A plausible-sounding answer and a true answer usually match, but not always. That’s a “hallucination.”
Reason 2
Its knowledge has gaps and an end date
Models learn from data up to a cutoff, and know little about your business, local rules or anything recent unless you give them sources.
Reason 3
It aims to please
AI tends to give an answer rather than admit it doesn’t know, and may follow a leading question to a wrong conclusion.
Reason 4
Numbers are a weak spot
Calculations, dates and unit conversions can slip, especially in long answers. Reasoning models and built-in calculators help, but you still need to check.
Reason 5
It can misread your documents
With long files, AI may skip sections, mix up details or blend in outside knowledge. That’s why citations to specific passages matter.
Reason 6
Vague prompts, vague answers
Missing context forces AI to guess. Clear instructions, sources and examples cut errors dramatically.
Spot the AI mistake
Each answer below is the kind of thing AI writes every day, and each contains one error. Click the sentence you think is wrong.
Prompt: “Summarize the benefits of upgrading our office Wi-Fi.”
Upgrading to a modern Wi-Fi standard can improve speed and reliability for busy offices. Newer access points handle more devices at once, reducing slowdowns during peak hours. Most offices see a return of about $4,000 per gigahertz within the first year. Better coverage can also eliminate dead zones in conference rooms.
Customer asks your website chatbot: “Can I return an opened item?”
Thanks for reaching out! We want you to be happy with your purchase. Under our 90-day Total Satisfaction Promise, opened items can be returned for a full refund, no receipt needed. You can start a return from the Orders page of your account.
Prompt: “We spend $2,400 a month on supplies. What would a 5% savings be?”
Your current monthly supply spend is $2,400. A 5% reduction is a meaningful savings target for most small businesses. That works out to $240 per month, or $2,880 per year.
Prompt: “Is there a legal precedent on employers monitoring work email?”
In the U.S., employers generally have broad rights to monitor company-provided email systems. Policies that clearly notify employees of monitoring strengthen the employer’s position. The leading case is Henderson v. Pacific Data Systems, decided by the U.S. Supreme Court in 2019, which settled the question nationally. State laws vary, so check local requirements.
How to catch mistakes: match the review to the risk
Review level finder
Seven habits that cut AI mistakes
- Give it sources. Paste or attach the document, policy or data you want it to use, and tell it to answer only from those.
- Give it permission to not know. Add “If you’re not sure, say so” to your prompts.
- Ask for citations, then check them. Request the exact passage or link for each claim, and open the important ones.
- Break big tasks into steps. Outline first, then sections. Errors hide in long, one-shot answers.
- Check the numbers separately. Use a spreadsheet or calculator for anything financial.
- Use a second pass. Ask AI to critique its own answer, or check it with a different tool, for important work.
- Ground customer-facing bots. Chatbots should answer from your approved content, with a handoff to a person. See our CustomGPT guide.
Review the text below before I publish it. List every factual claim, number, name, date, quote, link and citation as a checklist. For each, rate how confident you are that it's accurate (high/medium/low), explain why, and tell me exactly what I should verify and where. Do not rewrite the text. Text: [paste]
Recommended gear
Gear that helps
- AI Snake Oil (book)
A clear-eyed look at what AI can and can't do. - Co-Intelligence (book)
How to get value from AI while checking its work.
Links go to Amazon. As an Amazon Associate, Outshined.io earns from qualifying purchases, at no extra cost to you.
Frequently asked questions
Will AI stop hallucinating eventually?
Newer models make fewer mistakes, especially when connected to search or your documents, but the way they generate text means errors can’t be ruled out entirely. Plan on reviewing important output for the foreseeable future.
Who’s responsible when AI gets it wrong?
You are. Courts and regulators have consistently treated AI output as the business’s own statements. That’s the lesson of the Air Canada chatbot case. Build review into your process, and cover it in your AI policy.
Which AI tool makes the fewest mistakes?
It varies by task and changes with every model update. Accuracy depends more on how you use the tool: giving sources, asking for citations and reviewing output. See our assistant comparison for each tool’s strengths.
Related: Is AI safe for your business? · AI writing for business · AI glossary
