The Data Privacy Day 2025 keynote recording is now available!
The keynote presentation by Carnegie Mellon University Prof. Sauvik Das about privacy in the age of AI covered a wide range of AI impacts on privacy, including risks to personal data and opportunities to better protect privacy.
Das began by acknowledging that, “People are pretty apprehensive about how rapid advances in AI will and are already impacting personal privacy.” He went on to present a taxonomy of the risks AI exacerbates and, in some cases, creates, including:
- Facial recognition technologies and other surveillance.
- Information distortion, such as disinformation, misinformation, and deepfakes.
- Questions over consent to using personal data to train large language models (LLM).
- Data insecurity caused by new AI-enabled attack vectors for compromising data.
- Identification through linking of data points to particular individuals, including real-time identification.
- Exposure risks like the creation of non-consensual intimate imagery and use of AI to expose “primordial” information (things we are socialized to keep to ourselves)
- The use of AI to classify people into groups around race, gender, sexual orientation, or likelihood to be criminals; based on the debunked pseudosciences of physiognomy/phrenology.
AI technologies have meaningfully changed privacy, and we need better systems to help practitioners recognize the risks that they produce.
— Sauvik Das
The speaker offered an assessment of how prepared practitioners who create and use AI are to face these challenges. He juxtaposed motivators for AI developers to care about privacy, such as differentiating their product from a competitor and a sense of social responsibility against limitations caused by a focus on generic privacy concerns rather than AI-specific risks. Ultimately, he summarized the AI landscape as presenting more inhibitors to good privacy practices than motivators towards it. ”AI technologies have meaningfully changed privacy, and we need better systems to help practitioners recognize the risks that they produce.”
To counteract this grim assessment, Das demonstrated privacy-protecting technology solutions powered by AI:
- A tool that can help developers understand privacy issues in their work by analyzing the project concept, generating use cases, and suggesting where privacy can be improved and whose privacy might be particularly affected by that project.
- Imago Obscura: An AI-powered image editing copilot that helps identify and mitigate privacy risks in images.
- Privacy Mirror: An AI assistant that helps identify privacy risks when asking for online information or advice, particularly around sensitive subjects where the advice seeker wants to avoid disclosing personal information for fear of downstream risks.
- A browser extension for identifying potential privacy risks in posts.
- Sketch Based Access Control: Uses simplified language while still capturing nuances for individuals’ privacy preferences. The model takes a sketch, produces rules or policies, and also presents counterfactuals to help tune the rules.
U-M School of Information Prof. Florian Schaub moderated the well-attended event, and participants presented the speaker with many thoughtful questions in the post-presentation discussion.
A recording of the keynote is now available on the Privacy in the Age of AI event page on Safe Computing.
Take the Privacy Portrait Quiz
Are you a Guardian, Free Spirit, Skeptic, Pragmatist, or Wildcard? The U-M Privacy Portrait Quiz gives you a set of questions to answer and, based on your score, reveals your privacy portrait and customized privacy advice. Your results are private, of course; you can choose to share them with others or keep them just for you.
Privacy@Michigan, a collaboration between ITS and the School of Information, hosts events and activities intended to raise awareness, promote best practices, and provoke thought and conversation on privacy topics broadly relevant to our community members and society at large. Browse Past Privacy@Michigan events on Safe Computing.
