Google DeepMind’s CALLIOPE drops embeddings to map AI chats privately
Google / Wikimedia Commons (Public domain)
A new Google paper claims LLM-generated keywords beat embedding-based clustering for building topic maps under strict differential privacy
Google researchers think they have found a better way to see the big picture without reading anyone’s diary. A paper released on October 20, 2026, by teams at Google Research and Google DeepMind introduces CALLIOPE, a framework that builds topic maps from conversation logs while keeping formal differential privacy guarantees intact.
The key result: CALLIOPE outperforms its predecessor, URANIA, and the gap shows most clearly when privacy settings are at their strictest.
What CALLIOPE actually does
The paper is titled “CALLIOPE: Private Insights into AI Use, Without Embeddings.” It is listed on arXiv as 2610.28797, and Edith Cohen is among its authors.
The goal is a hierarchical topic map. Picture a table of contents for a giant pile of unstructured text, with broad themes at the top and narrower subtopics nested beneath.
Earlier approaches typically get there by turning text into embeddings, which are numerical representations of meaning. Similar documents end up near each other, and a clustering algorithm groups the neighbors.
CALLIOPE skips that step entirely. It works within what the researchers call the discrete text domain, meaning it handles words rather than vectors, and it relies on large language model reasoning to do the sorting.
The pipeline runs in three broad stages. First comes abstraction, where an LLM writes summaries and pulls out keywords. Next, the system synthesizes a robust topic schema. Finally, it assigns documents to topics.
Privacy controls apply across the whole process, and the framework is described as offering end-to-end differential privacy.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
Beating URANIA where it struggled
URANIA, introduced in 2025 under arXiv 2506.04681, was the earlier effort to bring formal differential privacy to this problem. It did so through private clustering built on embeddings.
In differential privacy, the budget controls how much any single person’s data can influence the output. A smaller budget means stronger protection, and generally more noise blurring the results.
According to the paper’s experiments, this is exactly where CALLIOPE pulls ahead, delivering better topic quality and stronger cluster coherence than URANIA.
The researchers go further still. They claim CALLIOPE performs comparably to, or better than, CLIO, a non-private baseline from 2024 described by Tamkin et al.
From CLIO to CALLIOPE
Anthropic’s CLIO system, released in 2024, helped establish the idea of using LLMs to analyze conversation logs at scale.
CLIO’s privacy protections, however, were heuristic. They relied on sensible rules rather than mathematical proof.
URANIA tried to close that gap by adding formal guarantees, but its embedding-based design paid a price in usefulness under strict settings. CALLIOPE keeps the formal guarantees and throws out the embeddings, which is the paper’s central bet.
The naming committee, for what it is worth, has been consistent. All three systems borrow their names from the Greek muses.
What this means
For AI developers, the stakes are practical. Knowing how people use a chatbot shapes product decisions, safety work, and research priorities, yet the raw logs are among the most sensitive data these companies hold.
The methodological shift is arguably the bigger story. Moving away from embeddings suggests LLMs can serve not just as the subject of analysis, but as the analysis engine itself, reasoning over text directly while staying inside formal privacy limits.
The performance comparisons come from the researchers’ own experiments, and independent replication will be the real test of the claims, particularly the assertion of parity with a non-private baseline.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.

