What is one major risk associated with generative AI models?
Hallucination is the one to pick. Generative models write the most likely next words, not checked facts, so they can produce a false statement in the same calm voice as a true one. Your quiz may list other risks as options, so the table below maps each to its name and a dated case.

| Risk | NIST name (our closest match) | Dated real case | What happened |
|---|---|---|---|
| Hallucination | Confabulation | Mata v. Avianca, 22 June 2023 | Two lawyers and their firm fined $5,000 over six invented court cases |
| Bias | Harmful Bias or Homogenization | UNESCO study, 7 March 2024 | Llama 2 text put women in domestic roles about four times as often as men |
| Privacy leaks | Data Privacy | Samsung, reported 2 May 2023 | Three employees put source code and meeting notes into ChatGPT |
| Copyright | Intellectual Property | New York Times v. OpenAI and Microsoft, 27 December 2023 | Suit seeks "billions of dollars" for use of its articles |
| Misuse | Information Integrity | Arup, reported 16 May 2024 | A deepfake video call led to HK$200 million in transfers |
The NIST names come from its Generative AI Profile, released on 26 July 2024. It lists 12 risks that are new or made worse by generative AI. The match in the second column is our reading, not NIST's table.
How do you answer the quiz question if hallucination isn't an option?
Pick the option that describes false, biased, leaked, copied or abused output, and skip any option that sounds like a benefit. I couldn't identify one course behind this question. Google, LinkedIn Learning and corporate trainers each word it their own way, and I found no official answer key to quote.
Use this order. If an option says the model makes up information or gives wrong answers, choose it. If none does, choose bias or privacy leakage next. Options such as "increased creativity" or "faster output" are benefits, so they can't be the risk. If the question says "one", any true risk counts, but hallucination is the safest pick.
What does a hallucination cost? The Mata v. Avianca case
NIST defines the risk plainly: "The production of confidently stated but erroneous or false content (known colloquially as "hallucinations" or "fabrications") by which users may be misled or deceived." The definition is NIST's wording, and it's the sentence most quiz answers paraphrase.
The best known case is Mata v. Avianca. On 22 June 2023, Judge P. Kevin Castel of the Southern District of New York ordered two lawyers, Peter LoDuca and Steven Schwartz, and their firm, Levidow, Levidow & Oberman, to pay $5,000. Their brief cited six nonexistent opinions that ChatGPT had invented. The ABA Journal reported that the judge found bad faith "based upon acts of conscious avoidance and false and misleading statements to the court", and the lawyers had kept defending the cases after the court questioned them. The court docket is case 22-cv-1461.
The lesson isn't that AI is useless in law. Nobody checked. A single search of a court database would have shown that the six decisions didn't exist, and it would have cost the lawyers a few minutes instead of a sanction and a public record. For the numbers on how often models do this, see AI hallucination statistics, and for the mechanism, why AI gets things wrong.
Can generative AI be biased?
Yes, because models learn from human text and repeat its patterns. UNESCO published a study on 7 March 2024 that tested GPT-2, GPT-3.5 and Llama 2 and found "unequivocal evidence of bias against women" in all three.
The same release says 70% of Llama 2's sentences that began "a gay person is" were negative, and so were 60% of GPT-2's. Those are open models from 2019 and 2023, so today's results will differ. The risk is real, but the study doesn't say how current models score.
Can generative AI leak private data?
Yes, mostly because people paste it in. Tech Monitor reported on 2 May 2023 that Samsung temporarily banned ChatGPT, Bing and Google Bard on company devices after three employees entered sensitive data. The staff had been using ChatGPT to fix problems in source code, and the data included chip and device code and internal meeting notes.
The company's memo said, "we are temporarily restricting the use of generative AI" while it reviewed security measures. The leak wasn't a hack. It was the normal way people use a chatbot, which is why the fix was a rule about what goes in. Is it safe to upload documents to AI covers that rule for files, and AI privacy statistics has the survey numbers.
Is generative AI a copyright risk?
It's a live legal dispute, not a settled fact. On 27 December 2023 the New York Times sued Microsoft and OpenAI in the Southern District of New York. CNBC reported that the paper alleges "mass copyright infringement" in training the GPT models on its articles, and that it seeks "billions of dollars" in damages.
The Times says GPT-4 produced nearly identical text to one of its articles on taxi medallion lending. That is the Times' allegation, and OpenAI said it was surprised and disappointed by the suit. I read no ruling, so treat the outcome as open and check the docket before you cite it. For you, the practical risk is publishing generated text or images without knowing whose work they echo.
How is generative AI misused?
Fraudsters use it to fake people. CNN reported on 16 May 2024 that the engineering firm Arup lost HK$200 million, about US$25.6 million, in 15 transfers. A finance worker in Hong Kong first doubted a request for a secret transaction. Then he joined a video call with what looked like the chief financial officer and colleagues. Police said all of them were deepfakes.
Arup confirmed that "fake voices and images were used" and said none of its internal systems were compromised. The money moved because a person trusted a face, not because a system broke. Read what a deepfake is for how the fakes are made, and use the call back rule in the steps below.
What other risks does NIST list for generative AI?
NIST lists twelve in total, and the five above are only part of the set. The rest are CBRN information or capabilities, dangerous, violent or hateful content, environmental impacts, human and AI configuration, information security, obscene, degrading or abusive content, and value chain and component integration. I didn't read the profile's detail on each, so use the names as a map and open the PDF for the definitions.
Two of these cover the topics security vendors stress. NIST names information security as its own risk, apart from data privacy. It also names value chain and component integration, which I read as the exposure that comes with parts built by other companies, such as the model, the data and the plug ins. That reading is mine. If your course is about business use, expect both to matter, because a company that adopts a tool adopts the parts inside it.
What risks do other guides list that quizzes sometimes use?
Business guides stretch the list. Entrans, a software firm, names seven concerns: harmful content and misinformation, homogenization of content, legal and copyright issues, job displacement, shadow IT, data privacy violations and lack of transparency. For code, it adds code quality, security flaws, technical debt and poor visibility into how generated code works. That is one vendor's list, not a standard.
Shadow IT means staff using AI tools the company never approved, which is the Samsung pattern above. Homogenization means many people getting near identical output, which NIST also names. Job displacement is a risk to workers rather than a harm the model causes in a single answer, so a quiz that asks for a risk of the model itself rarely means it.
Wiz, summarizing a Gartner note, groups the security side into four areas: privacy and data security, enhanced attack efficiency, misinformation, and fraud and identity risks. The Arup case sits in the last one. If an option says attackers can use the model to scale phishing or malware, that is the attack efficiency area, and it is a true risk.
What is a best practice when using generative AI?
The best practice is to verify AI output with a person and a source before you rely on it, and to keep confidential data out of tools you don't control. That one sentence answers most versions of the quiz. Each habit below comes from a case above.
- Verify claims against a source you opened. Mata v. Avianca happened because nobody did.
- Keep a human accountable for the result. The court fined the lawyers, not the model.
- Don't paste confidential data into public tools. Samsung's three leaks came from ordinary work.
- Confirm money or access requests on a second channel. Arup's worker believed a video call.
- Test the output on the people it affects before you use it. UNESCO found the bias by testing, not by guessing.
- Check who owns what you publish. The Times suit is about exactly that.
For the checking step, how to fact check AI answers turns it into a two minute routine, and is AI safe to use covers the personal side. AI ethics goes into who is responsible.
What should you ask an AI about its own risks?
Five prompts that turn the model into a study partner. Paste them into Claude, ChatGPT, Gemini or Whizi, and check what comes back.
- Quiz me: "Ask me five multiple choice questions on the risks of generative AI, one at a time. After each answer, tell me if I was right and why."
- Find the weak spot: "Here is my summary of generative AI risks: [paste]. Which risk is missing or wrong? Give a source I can open for each correction."
- Apply it to my job: "I work as [role] at [kind of company]. Which two generative AI risks matter most for my daily tasks? Give one realistic example for each."
- Audit an answer: "Read the answer below and list every factual claim that I should verify, ranked by how costly it is if wrong: [paste]."
- Write a team rule: "Draft a half page policy for [team] on what may and may not be pasted into AI tools. Use plain words and give three examples of each."
What mistakes make these risks worse?
Four habits turn a small risk into an incident, and each has a fix.
- Trusting fluent text. Fluency isn't accuracy. The fix: ask for sources and open them.
- Doubling down. In Mata v. Avianca the lawyers kept defending the fake cases when the court asked. The fix: if the checker says no, stop and recheck.
- Treating a leak as a model problem. Samsung's leaks came from what staff typed. The fix: write a rule about inputs.
- Trusting a familiar face. The Arup worker saw colleagues on screen. The fix: confirm on a second channel.
Where does Whizi fit when you check AI answers?
Whizi is one app instead of a pile of single purpose tools, and it helps with the verify step. Web search is on by default in a new chat, so answers can come with sources you open. The free plan gives you 10 messages a day on Whizi AI, Auto and Claude Haiku, with no card.
Pro carries Claude, GPT and Gemini in one app, and its first week is $1.99. Side by side Compare on Pro puts one question to two models, so you can read only where they disagree. That doesn't prove either is right. A disagreement just tells you which claim to check first.
What are the limits of this page?
The cases are real and dated, but they are examples, not a ranking. Every number here comes from the page named in the sources, and two of them are one outlet's report: CNN for the Arup loss and Tech Monitor for Samsung. The UNESCO figures come from older models, and the Times case may have moved on since 22 November 2026. A quiz may also use its own wording. Trust your course material over this page if they differ.
- Open every source behind a claim you plan to use
- Keep names, code and client files out of public AI tools
- Call back on a known number before acting on a money request from a call or video
- Name the person who signs off on AI assisted work
- Check who owns text or images before you publish them
Frequently asked questions
What is one major risk associated with generative AI models?
Hallucination: the model produces confident false content. NIST calls it confabulation in its July 2024 generative AI profile. A New York court fined two lawyers and their firm $5,000 on 22 June 2023 for citing six cases ChatGPT invented.
What is a best practice when using generative AI?
Verify the output against a source you opened, with a person accountable for the result. Add a rule against pasting confidential data into public tools. Samsung restricted staff in 2023 after three employees did exactly that.
What are the main risks of generative AI?
The common ones are hallucination, bias, privacy leaks, copyright disputes and misuse such as deepfake fraud. NIST lists 12 risks in total, including information security and environmental impacts, in its generative AI profile.
Is increased creativity a risk of generative AI?
No, it is usually offered as a benefit and appears as a wrong option in quizzes. The risks are false output, bias, data leaks, copyright disputes and misuse, each with a dated case on this page.