Exam-style
A marketing team wants to publish AI-generated images and AI-generated copy on the company's public website, replacing a photo shoot and a freelance copywriter. Legal will not sign it off. Their worry is that a third party turns up later claiming a published image or paragraph reproduces work the company had no right to use, and asks who pays. Which set of measures addresses the risk legal has described?
Reveal the answer
C. Choose a model whose provider offers IP indemnity for generated output, read the model's licence and acceptable use policy before selecting it, ground generation in content the company already owns, and keep a person reviewing every asset before it publishes
The risk legal described has a name in the AI Practitioner material: it sits under the legal risks of working with generative AI, and specifically under intellectual property infringement claims. A third party asserts that published output reproduces their protected work, and the argument is about who carries the liability. Four measures on the table are real controls that close real risks, and none of them closes that one. Content filters in Amazon Bedrock Guardrails block harmful categories, which is safety and toxicity; a paragraph can be entirely inoffensive and still be somebody else’s paragraph. A contextual grounding check scores an answer against the source material it was given, which addresses veracity and hallucination; a faithfully grounded sentence lifted from a source the company had no right to use is still an infringement. Encryption with AWS KMS keeps the material confidential in storage and transit, which has nothing to say about a claim over content the company published on purpose. Watermarking and its detection API establish provenance, letting anyone confirm that an image came out of a model, which is useful for disclosure and for the loss of customer trust that arrives when readers discover machine-written work sideways. Provenance answers “did a model make this”. An infringement claim asks “does this reproduce my work, and who pays”, which is a different question. That leaves the fourth set, and its four parts do four different jobs. Provider IP indemnity is a contractual promise to defend and cover the customer against third-party IP claims arising from generated output, subject to conditions such as not disabling the provider’s own filters; it moves part of the liability off the company and it is the only measure here that does. Licensing and the acceptable use policy are read before selection rather than after, because they set what may be done with the output commercially and what may not be generated at all, and they differ sharply between hosted proprietary models and open-weight ones. Grounding generation in the company’s own product photography, spec sheets and prior copy reduces how much the model is left to invent from its training data. The human review step catches the recognisable logo, the famous character and the borrowed slogan, and it is also where end user risk gets caught, since a person is the last chance to stop a generated claim that could hurt someone acting on it.
Q. Marketing wants generated images and copy on the public site. Legal is worried a third party will claim the output reproduces their protected work. Which controls answer that?
A. Pick a model whose provider indemnifies you against IP claims on generated output, read its licence and acceptable use policy before you select it, ground generation in content you already own, and keep a person reviewing assets before they publish.
Why? Name the risk first. Legal has described intellectual property infringement claims, one of the legal risks of working with generative AI, and the argument is over who pays. Contractual indemnity is the only control on the table that moves liability. Licensing sets what you may do with the output commercially, and it differs sharply between hosted proprietary models and open weights, so it belongs in the model choice itself rather than in a later review. The other four options each close something else. Guardrails content filters block harmful categories, which is safety. A contextual grounding check scores an answer against its sources, which is veracity, and a faithfully grounded sentence can still belong to someone else. KMS encryption keeps material confidential, which says nothing about content you published deliberately. Watermarking and its detection API prove an image came out of a model, which helps with disclosure and with loss of customer trust, and does not answer a claim of ownership. Human review is where end user risk gets caught as well, because a person is the last chance to stop a generated claim somebody might act on.