Generative AI can create legal exposure for the business that deploys it, not just for the model provider. The exam guide names five risks: intellectual property infringement claims, biased model outputs, loss of customer trust, risk to end users, and hallucinations. This lesson explains what each looks like in practice and the usual ways to reduce it. It describes general risks, not legal advice; AWS's own responsible AI guidance tells teams to consult their legal counsel about their obligations.
Intellectual Property Infringement
A model learns from large amounts of existing text, images and code. Its output can sometimes closely resemble a protected work, a logo or someone's distinctive style, which can lead to claims that the output infringes copyright or other rights. The business that publishes the output may be the one a rights holder pursues.
Ways to reduce the risk include reviewing outputs before publishing, avoiding prompts that ask for a named artist's or brand's work, and choosing providers that stand behind their models. For example, AWS offers uncapped intellectual property indemnity for the outputs of generally available Amazon Nova models, which protects customers against third-party claims such as copyright infringement about those outputs, under AWS's service terms. The indemnity has conditions: it excludes trademark claims, and it doesn't apply if you disable the model's available filters.
Biased Model Outputs
If a model's outputs treat groups of people differently, for example wording job ads that discourage older applicants or recommending different terms by gender, the business can face discrimination complaints and regulatory action. Bias usually comes from the training data or from how the model is used. Testing outputs across different groups before launch, and monitoring after launch, are the main defenses.
Hallucinations and End-User Risk
A hallucination is a confident but false output. When people act on it, they can be harmed: a customer follows wrong instructions, a patient gets wrong health information, or a real person is described saying something they never said. The business that deployed the system may be responsible for that harm.
AWS's guidance for Amazon Nova is that customers should check outputs for accuracy, and that any workflow making consequential decisions needs human oversight, testing and safeguards suited to the use case. Grounding answers in trusted sources and using Bedrock Guardrails' contextual grounding checks also reduce hallucinations.
Loss of Customer Trust
Every risk above can end the same way: customers stop trusting the company. A chatbot that invents policies or an image tool that produces offensive content can damage a brand far beyond the original mistake. Being open about where AI is used, and fixing problems quickly, protects trust.
| Risk | Typical mitigation |
|---|---|
| IP infringement | Output review, provider indemnity |
| Biased outputs | Testing across groups, ongoing monitoring |
| Hallucinations | Grounding, guardrails, accuracy checks |
| End-user harm | Human review for consequential decisions |
| Loss of trust | Transparency and fast correction |
On the Exam
Identify which risk a scenario describes: copying or resembling protected work is IP; unequal treatment of groups is bias; invented facts are hallucinations; harm to the person relying on the answer is end-user risk. For consequential decisions, the expected mitigation is human oversight, not a model setting.