Use AI to research faster, not as your source of truth.

AI can turn a broad question into a useful outline, suggest search terms, summarize a long document, or explain unfamiliar language. That makes it a valuable research assistant. It does not make the answer on the screen a verified fact.

A generated response is assembled to fit the prompt and the patterns the system has learned. It may be accurate, partly accurate, out of date, or confidently wrong. The safest way to use it is to let AI help with the process while a person remains responsible for the conclusion.

Start with AI as a map, not the destination

A good first prompt can help define the territory. Ask the tool what questions should be answered, which terms belong in a search, what assumptions need testing, and what kinds of original sources may exist. This is especially useful when you are entering a subject you do not know well enough to search efficiently.

Then leave the generated answer and open the sources. For a law, read the statute, regulation, or published decision. For a technical claim, look for the product documentation, standard, or research paper. For a company policy, use the company’s current published terms or a direct answer from someone authorized to provide it.

• Use AI to break a broad topic into smaller questions.
• Ask for search terms, alternative explanations, and claims that need verification.
• Open the original source instead of relying on the AI summary of it.
• Check the date, author, jurisdiction, product version, and surrounding context.
• Compare important claims with another independent and appropriate source.

A citation is not proof that a source exists

Generative AI can produce a title, author, case name, quotation, or web address that looks entirely believable but does not lead to the claimed material. NIST calls this kind of confidently presented false content a confabulation. Its Generative AI Profile also warns about automation bias, the tendency to give an automated result more trust than it has earned.

Source: NIST AI 600-1, Generative AI Profile

The failure is not always obvious. A real source may be summarized incorrectly. A quotation may be altered. An older rule may be presented as current. A source from another state, country, software version, or industry may be applied to the wrong situation. Verification means checking what the source actually says and whether it answers your question.

The dealership chatbot story proves the point in an unexpected way

In 2023, a user manipulated a chatbot on a Chevrolet dealership website into agreeing to sell a Tahoe for one dollar and calling the exchange legally binding. The screenshots spread quickly, but the dealership did not complete the sale. It was a viral chatbot incident, not the tribunal decision that is sometimes blended into retellings of the story.

Source: AI Incident Database record for the Chevrolet dealership chatbot

This is an important research lesson on its own. The dealership, the chatbot, the one-dollar offer, and the legal language are real parts of the reported story. The formal legal dispute came from a different chatbot case. A plausible retelling can blend separate events unless someone follows the trail back to the records.

A real decision about incorrect chatbot information involved Air Canada. In Moffatt v. Air Canada, a customer relied on the airline website’s chatbot for guidance about a bereavement fare. The British Columbia Civil Resolution Tribunal found that Air Canada had not taken reasonable care to ensure the chatbot was accurate and ordered it to pay damages, interest, and fees. It was a Canadian small-claims tribunal decision, not a universal rule for every chatbot, but it shows why a business cannot treat automated customer communication as consequence-free.

Source: Moffatt v. Air Canada, 2024 BCCRT 149

Human review has to test the substance

A person can review an AI answer and still miss the problem if the review only checks grammar and tone. The reviewer needs enough subject knowledge, or access to someone who has it, to test the claims that matter. The higher the consequence, the stronger that review should be.

In Mata v. Avianca, lawyers submitted court filings containing nonexistent opinions and fake quotations generated by ChatGPT. The federal court explained that using an AI tool for assistance was not inherently improper. The problem was abandoning the responsibility to verify the filing, including after the questionable cases were challenged. The court imposed sanctions.

Source: U.S. District Court sanctions order in Mata v. Avianca

Match the review to the consequence

Not every AI-assisted task needs the same process. Brainstorming five headline ideas is different from interpreting a contract, recommending a medical treatment, changing a network, or sending a price to a customer. Decide what can be used as a draft, what needs source checking, and what requires review by a qualified professional.

Keep a simple research trail

Save the links or documents that support the final work. Note when they were checked and which claims they support. If an AI tool helped create an outline or summary, preserve enough of the process to understand what was generated and what a person verified. This makes later updates easier and gives another reviewer something more useful than “the chatbot said so.”

• Separate generated suggestions from verified findings.
• Prefer primary and authoritative sources for consequential claims.
• Record uncertainty instead of asking the wording to sound more certain.
• Do not paste confidential or regulated information into an unapproved service.
• Name the person responsible for approving the final result.

AI is useful precisely because it can help people move through information faster. Its limitations do not erase that value. They tell us where to place the guardrails: use the tool to explore, organize, compare, and draft, then verify the parts that the business will rely on.

RIKR can help your team choose approved tools, protect business information, and build a review process that fits the work. The goal is not to make every AI question complicated. It is to keep a useful assistant from quietly becoming the only source in the room.

Related: using AI without exposing business data

Explore RIKR managed IT support