Francesca Rossi on IBM’s Approach to AI Ethics, Bias and Generative AI

For Francesca Rossi, responsible AI depends on decisions made throughout development and deployment: what data a system uses, how teams assess its behavior, and what role it plays in people’s lives. Rossi, an IBM Fellow and the company’s AI Ethics Global Leader, argues that these questions need to be built into ordinary engineering and business processes. They can’t be left until after a system causes harm.

In an interview with Techopedia, Rossi described IBM’s centralized approach to AI ethics, its work on open tools for fairness and explainability, and the challenges posed by generative systems. Her account offers a company perspective rather than an independent assessment of IBM’s products, but it lays out a practical framework: assess risks for each use, make governance part of product development, and help teams recognize how human assumptions can shape technical choices.

Shared principles, applied to each use

Rossi’s roles extend beyond IBM. She is a GPAI expert and Steering Committee member, serves on the IEEE Executive Committee’s AI Ethics initiative, co-chairs the World Economic Forum’s GFC on AI for Humanity, and co-authored the ACM TechBrief on generative artificial intelligence. Her research focuses on embedding human values in AI systems.

At IBM, a centralized internal governance structure sets common principles across the company’s global operations. Teams then assess the risks of particular AI solutions delivered to clients, asking whether those systems align with the principles. Rossi presents that assessment as a starting point for an organization’s work: it makes the desired direction explicit before teams decide how to build and deploy a solution.

One principle is that AI should augment human intelligence. Rossi says that doesn’t rule out automation. Rather, she describes AI as a way to extend people’s capabilities in areas such as creativity, problem-solving and analysis. That distinction matters in product design: a system’s value depends partly on whether its role supports human work and whether people understand how to use its output.

Another principle concerns data ownership. Rossi says IBM treats data as belonging to the people or organizations that generated it. In its business-to-business work, she says, client data and insights from one engagement aren’t reused for another client. That commitment speaks to how a provider handles customer information; it doesn’t, by itself, resolve every question about data provenance, permission or the conditions under which models are trained.

IBM’s stated focus also includes transparency, explainability, robustness and privacy. With generative AI, Rossi says, the company is paying greater attention to responsible content generation and its societal effects. These principles cover different parts of a system’s lifecycle: what information it handles, how its behavior can be understood, and what risks arise when people rely on generated content.

Bias can enter at every development stage

Rossi describes bias as a problem that can emerge in training data, testing and other development decisions. Patterns in the data may reflect relationships that people haven’t noticed, while decisions made by developers can carry their own assumptions. The result may be a system that treats groups differently, even when no one intended that outcome.

That makes bias assessment a team responsibility as well as a technical one. Rossi recommends bringing together developers with different backgrounds, genders and areas of knowledge. A varied group can compare perspectives and question decisions that might otherwise pass without discussion. Diversity alone doesn’t guarantee fair results, but it can help teams identify assumptions that a less varied group might miss.

IBM’s response includes education and design thinking sessions intended to help staff connect abstract ethical concerns to daily work. Rossi says many programmers were accustomed to focusing on technical performance without considering social effects. The aim, as she describes it, is to make questions about bias and discrimination part of how developers understand their own choices, rather than a topic reserved for specialists.

IBM has also released tools for teams working on these problems. Its AI Fairness 360 toolkit, donated to the Linux Foundation, is open source and offers tools for examining fairness in AI systems. Rossi says IBM has also released toolkits related to explainability, fairness, robustness and privacy.

Open access can benefit researchers and academic groups with fewer resources than large companies, as well as teams building commercial systems. Shared tools may broaden participation in testing and mitigation work. They don’t remove the need for careful evaluation: teams still have to choose suitable methods, interpret results and decide what trade-offs are acceptable for a particular application.

Generative AI brings content risks and uneven stakes

Rossi identifies unreliable output as a key challenge for generative AI. These systems can produce material that appears plausible but is false, and may generate content that conflicts with the values expected in a particular setting. The risk is not limited to text: systems that generate realistic images or video can also make it harder to tell synthetic material from authentic media.

How much those failures matter depends on the application. Rossi contrasts a movie recommendation with a suggestion about treatment for a patient. Both could rely on similar technical foundations, but the consequences of a poor recommendation differ sharply. Her point is that organizations need to evaluate systems in the context where people will use them, rather than treating a model’s general capabilities as a complete measure of its risk.

This distinction also shapes Rossi’s view of oversight: attention should focus on what a system is used for, as well as on the underlying technology. The same model or technique may appear in low-stakes and high-stakes settings, so one assessment won’t necessarily fit both. For organizations, the practical question is how to identify the system’s users, the decisions it informs and the harm that could follow from an error.

Governance belongs alongside data and model development

IBM’s Watsonx platform is one example of how the company wants clients to organize that work. Rossi describes three components: Watsonx.data for data, Watsonx.ai for building machine-learning approaches and solutions, and Watsonx.governance for the company’s AI ethics work. Her emphasis is on the relationship among them: governance should be integrated into development, alongside data preparation and model building.

That approach makes governance a routine part of a project rather than an optional review at the end. Teams can then consider risks as they choose data, develop a system and prepare it for use. The interview doesn’t provide performance figures, pricing or deployment details for Watsonx, so it supports a description of the platform’s structure and intent, not a comparison of its technical results with competing products.

Rossi pushes back on the idea that ethical work necessarily slows commercial progress. She argues that shortcuts can create damage and disruption, while responsible development can support more durable innovation. She refers to studies suggesting that companies with this attitude are more successful, but the interview gives no study names, methods or figures. That claim is therefore best read as her view, rather than as a quantified result that can be assessed here.

Education and collaboration are part of deployment

To help organizations and individuals understand AI’s limitations, IBM develops internal learning materials and works with Coursera on educational modules covering topics such as AI ethics and explainability. Rossi also points to a risk atlas included in Watsonx, which turns information about risks in content generation into an online resource clients can search.

These efforts address a practical gap: teams may adopt a system without understanding where it can fail or how its outputs should be checked. Training and risk references can help make those limits visible, though their usefulness depends on whether people consult them and apply the information to the specific system they’re using.

Rossi’s broader argument is that responsible development requires deliberate collaboration among companies, researchers, educators, standards organizations, and people who build or use AI. She says the field has grown from a small community focused on AI ethics into a much wider discussion. The next step, in her view, is to act before problems arise, rather than relying on fixes after deployment.

Open tools and shared learning can widen access to both AI development and risk mitigation. Rossi’s final point extends beyond organizations: individuals also make choices about how they use technology and what effects those choices may have. For teams adopting generative AI, her framework leads to a concrete task: define the intended use, identify who could be affected, and make risk assessment and governance part of the work from the start.