Trust in AI depends on privacy, security and governance working together. At COLLIDE, leaders from TD Bank, TelevisaUnivision, Vertex and Truist discussed LLM privacy risks, AI ethics boards, the limits of explainability and why companies should build the cake before adding generative AI as the cherry.
Key takeaways
- Use only the data you need. Data minimization and consent are the foundation of AI privacy, the panel said.
- Public LLMs carry hidden privacy risk. Fogelson said models trained on the open internet may contain exposed private data about non-customers.
- Ethics boards set the principles. Gorricho said AI ethics boards must align use cases with company values and push those principles into the culture.
- Explainability has limits with LLMs. Keep LLMs behind a human for critical decisions and check the references behind generated answers.
- Build the cake first. Get data, governance and privacy right before adding generative AI on top.
How do organizations earn trust when AI systems handle sensitive customer data? At COLLIDE in Atlanta, a panel on responsible AI tackled security, privacy and ethical governance. Speakers were Juan Gorricho, Head, Executive Product Owner and Data Strategy, AMCB Data as a Service at TD Bank; Sergey Fogelson, VP, Head of Data Science at TelevisaUnivision; and Ahmad Haider, Senior Director, Data and Advanced Analytics at Vertex. Anqi Zou, Head of Fair & Responsible Banking Analytics at Truist, moderated.
Zou said the panelists' careers spanned more than a dozen industries, so the lessons apply widely.
How do you build trust in AI systems that handle sensitive data?
Haider started with the data: collected under the right standards, clean, as free of bias as possible and covered by a privacy and security framework. He also recommended letting users give feedback on AI outputs.
Gorricho warned against using all available data for every purpose. "Don't overuse data that is not needed," he said. Transparency about how a model makes decisions matters too. In banking, he said, black-box models aren't an option because decisions must be explained, and he argued companies outside banking should hold themselves to the same standard.
Fogelson took a different view for customer-facing generative AI. Any clunkiness signals to customers that an LLM is behind the experience and raises their guard, he said. Trust comes "through very, very long flawless execution."
What are the biggest privacy risks with LLMs?
Fogelson said the largest models are trained on nearly all the text on the internet, including private information that was exposed unintentionally. A company deploying a generative AI tool built on that data could end up violating the privacy of people who aren't even its customers. "If the data you're training it on is clean, you'll be fine," he said. "If it's not, well, good luck."
Gorricho stressed consent and urged organizations not to jump "from zero to 100" on personalization and micro-segmentation. He said he had read that more than 150 data privacy regulations are in process worldwide. Haider said laws such as GDPR and HIPAA make a privacy framework essential for deciding what personal, health and payment data can be shared.
Zou summed up the theme as data minimization: use only the data fit for the purpose, with consent built into the pipeline.
How should organizations govern AI to avoid ethical pitfalls?
Gorricho said mature organizations already have data councils, and AI needs the equivalent. At Visa, he said, an AI ethics board evaluated use cases and models against company principles. A misfiring AI tool can damage reputation, and "the risk is way higher" than the potential value, he said.
He recalled Disney's MagicBands project from his time there. When critics asked why the company was collecting location data, including on children, he said, the CEO answered that the data was meant to enhance the guest experience and that guest trust came first. Principles set by an ethics board must be pushed down in ways that are easy to digest so they become part of the culture, Gorricho said.
Haider added continuous risk assessments, audits and checks on generated content for misinformation, since AI can't yet be used without human review. Zou pointed to ongoing monitoring of performance and of data and concept drift.
Can you explain how a generative AI model reaches its answers?
Fogelson said classical machine learning offers approximate interpretability, but that "goes away" with LLMs, which can give different outputs for the same input. Companies should either make the AI invisible or tell users plainly that it is an experiment with variable results, he said.
Gorricho suggested placing LLMs behind the scenes, such as helping call center staff find information faster, with a trained human still deciding what to tell the customer. For critical processes, he said, there is rarely a need for a sophisticated black-box model. Haider described referential checking: providing sources for generated answers and verifying that they are accurate.
How do you secure generative AI applications?
Fogelson called an LLM with outside access "a completely unsecured client" that could reach all of an organization's information, so it needs very strict bounds. Haider recommended a security engineer on the development team from day one, a team responding to alerts and threat modeling of data, compute, connected systems and users. For vendor tools, he said, third-party risk reviews should check questions such as whether query data is stored on a model provider's servers.
Gorricho cited cases where chatbots leaked data or gave customers wrong advice, and said governance must explain to employees why they can't paste proprietary code or documents into public tools.
How do companies balance AI innovation with privacy?
Gorricho said trouble starts when technology is deployed "for technology sake." Haider recommended small teams that prioritize use cases offering competitive differentiation, paired with responsible implementation.
Fogelson put it in baking terms: companies need to "make the cake first" before adding generative AI as the cherry on top, because reputational damage can sink a brand. His closing advice: "Learn from the mistakes of other companies," he said, "because you won't have time to learn from your own."
Discuss responsible AI with peers at COLLIDE 2027
COLLIDE, the largest AI conference in the Southern U.S., brings more than 1,000 enterprise data and AI leaders to Atlanta. The 2027 conference takes place Wednesday, October 13, 2027, at the Sandy Springs Performing Arts Center. Registration details are coming soon.
Frequently asked questions
How do companies build trust in AI?
The panel pointed to clean, properly collected data, bias mitigation, user feedback, transparency about how models decide and, for customer-facing generative AI, long flawless execution.
What are the privacy risks of large language models?
Fogelson said large models trained on internet text may contain unintentionally exposed private information, so a company could violate the privacy of people who aren't its customers.
Can generative AI outputs be explained?
Fogelson said true interpretability for LLMs isn't available yet; Haider suggested referential checking, where answers cite sources that are verified for accuracy.
How should companies secure generative AI applications?
Put a security engineer on the team from the start, threat-model data, compute, systems and users, restrict outside access and run vendor tools through third-party risk review.
Watch: Building Trust in AI: Addressing Security, Privacy, and Ethical Governance in Responsible AI on YouTube
Event: COLLIDE AI Conference · More coverage: COLLIDE AI Conference insights
