Say Something Else: Rethinking Contextual Privacy as Information Sufficiency

A team of researchers from Cornell University and Google DeepMind has introduced a novel framework called "Say Something Else" (SSE) for contextual privacy in AI-generated communications, as detailed in a research paper published on arXiv on April 10, 2026. This work addresses the critical challenge of preventing large language model (LLM) agents from oversharing sensitive user information by expanding beyond traditional privacy methods of suppression and generalization. The research was motivated by the growing reliance on AI assistants to draft messages and the fundamental disagreement among users about what constitutes private information in different contexts. The core innovation of the SSE framework is the introduction of a third, more flexible strategy: substitution. Unlike suppression, which removes sensitive details entirely, or generalization, which replaces them with broad categories, substitution allows the AI to replace sensitive information with plausible, contextually relevant, but non-sensitive alternatives. For instance, instead of omitting a user's specific medical condition or vaguely labeling it as a "health issue," an SSE-equipped agent might substitute it with a different, benign reason for an appointment that maintains the conversational flow while protecting the original sensitive data. This approach recognizes that complete omission can break the coherence of a conversation, while over-generalization can render messages uninformative. The researchers developed their framework by first formalizing the problem of contextual privacy for multi-turn dialogues, moving beyond the limited evaluation of single, isolated messages. They constructed a comprehensive taxonomy of privacy strategies and created a benchmark dataset to test their model. The paper argues that effective privacy preservation must be contextual, meaning the AI must understand the specific conversation, the relationship between participants, and the user's personal privacy preferences. The evaluation demonstrated that the SSE framework, by employing substitution, could generate more natural and useful responses while maintaining a higher degree of privacy protection compared to models limited to suppression and generalization alone. This research represents a significant step toward more trustworthy and user-aligned AI assistants. By framing privacy as a problem of "information sufficiency"—providing just enough relevant information to achieve the communicative goal without revealing sensitive details—the work shifts the paradigm from simple redaction to intelligent, context-aware rewriting. The findings highlight the nuanced balance required in AI-mediated communication between utility, naturalness, and privacy, setting a new standard for how future language models should handle personal data in dynamic social interactions.