Nvidia’s Nikki Pope on AI Risks, Copyright and Access

For companies deploying generative AI, the hard questions are already practical: who bears responsibility when a model produces harmful falsehoods, how should firms handle copyrighted material, and what checks fit a system’s level of risk? Nikki Pope, who leads Nvidia’s Trustworthy AI Initiative, argues that existing rules provide a starting point. The work ahead is applying them to systems whose outputs can affect people in new ways.

Pope’s comments came as a copyright dispute between The New York Times, OpenAI and Microsoft brought training data and generated content into sharper focus. The Times alleges its material was used to train AI models and that the systems can produce content resembling its articles. The case illustrates the questions at stake; it does not settle them. Pope says the outcome will depend on how claims are argued under existing intellectual property rules.

Risks emerge in everyday deployments

For Pope, copyright is only one part of the exposure. She points to product liability, privacy and other established areas of law as relevant to AI. A company adopting a large language model still has to consider what information enters the system, what it produces and how people may rely on those outputs.

False statements about real people show why those questions can become difficult quickly. Pope cited a 2023 incident in which a chatbot falsely accused a law professor of sexual assault, describing an imaginary trip and pointing to a Washington Post article that did not exist. If an affected person brought a defamation claim, responsibility could be contested: would it rest with the company that deployed the chatbot, the person who prompted it, or another party? Pope framed that as an open question.

The risks also extend to workplace tools. A model that drafts recommendation letters or helps screen applicants can reproduce or amplify bias. Pope referred to research in which chatbots used different descriptors for men and women: men were characterized as “listeners” and “thinkers,” while women were associated with “beauty” and “grace.” Amazon had earlier stopped using an AI résumé screening tool after it showed bias against women applicants.

These examples make evaluation a deployment concern, not just a model development task. Companies using AI in recruitment need to assess whether outcomes disadvantage people based on characteristics such as gender or race. A tool that appears to save time can still create a harmful pattern if nobody checks how it performs across groups.

Pope also pointed to the Federal Trade Commission’s consumer protection work. The commission has emphasized that existing rules still apply to AI and warned that companies may run afoul of consumer protection requirements if they fail to be transparent about data collection. For businesses, the practical message is that adding a model to a product does not remove obligations around privacy and clear handling of information.

Access and language shape who benefits

Pope expects future debate to focus on which communities can take part in AI development and benefit from its use. That includes looking for bias in training data, but it also means asking whether people can use tools in the languages they speak and sign.

She noted that more than 7,000 spoken or signed languages are used around the world. Making AI accessible across that range, she argued, cannot be reduced to translating English outputs. Communities also need the means to build tools in their own languages, including sign languages.

That goal runs into a resource constraint: building models is expensive and requires skilled developers. If only well-funded teams can create and adapt AI systems, the cost of participation could widen existing gaps in access to technology. Pope sees training as one part of the response—both helping workers use AI and developing the skills to build tools. Public education also matters, so people can weigh potential benefits against risks.

Match safeguards to the use

For businesses concerned about copyright, Pope’s advice begins with understanding the rules that apply where they operate. Intellectual property standards vary by region. As one example, the US fair use doctrine can allow limited use of copyrighted work as a quotation or sample; Pope said there is no equivalent doctrine in the EU. Firms also need to consider brand consequences when a use is challenged, even if the alleged violation has not been established.

On safety, her starting point is to treat AI as a product that must meet the safety requirements for its category. A self-driving car, for example, must satisfy the requirements that apply to cars and may need additional checks for detecting pedestrians and cyclists. The appropriate safeguards depend on what a system does and the harm a failure could cause.

Pope described several practices that can help teams find problems before release: assessing and measuring bias, using benchmarks to track performance, and conducting penetration tests or red-team exercises. Guardrails on prompts, outputs and topics can also constrain a system’s behavior. None of these steps substitutes for judgment about the use case. A recommendation engine for a streaming service carries a different risk profile from one that helps guide decisions about a medical procedure; the latter calls for more demanding safeguards.

She also sees value in certification frameworks. A shared baseline could help customers understand whether a product has met minimum safety expectations and give developers a clearer target. For companies, such benchmarks could make evaluations easier to compare, provided the tests reflect the system’s actual purpose and users.

Responsible AI starts with present harms

Pope defines responsible AI around the effects that products and systems have on individuals, communities and society. The meaning of the term will continue to evolve, but she challenges a narrow focus on speculative scenarios in which people lose control of AI. Her concern is that serious harms are already possible when people lack access to AI’s benefits or when systems omit their communities from the data used to build them.

“An AI that treats black kidney patients adversely compared with white kidney patients poses a potential threat to the lives of black kidney patients.”

Her point is that disparities involving race, gender and other characteristics can have direct consequences in high-stakes settings. The discussion of long-term AI risk should therefore include present-day questions of access, representation and unequal treatment. For teams shipping products, that means testing the system against the people it will affect, selecting safeguards based on the stakes, and revisiting those checks as use changes.

Pope’s view puts the burden on concrete practice: apply existing requirements, examine data and outputs for bias, and make the safeguards proportional to the task. Those steps will not resolve every dispute around AI, but they give organizations a basis for identifying exposure before a system reaches users.