Thamus

As AI-generated answers increasingly complement, and in some cases displace, traditional search, large language models (LLMs) from Google, OpenAI, Anthropic, Meta, Mistral, DeepSeek, and others are rapidly becoming primary knowledge intermediaries. Unlike web search, which presents a list of sources for users to evaluate, LLMs return a single authoritative-sounding answer drawn from sources that may be cited, partially attributed, or entirely invisible. Diverse perspectives and productive uncertainty collapse into a single voice.

As users offload cognitive work to these systems and accept synthesized answers as authoritative, what LLMs return, and what they exclude, shapes public understanding. The emerging practice of Generative Engine Optimization (GEO), sometimes called “AI SEO,” adds another layer: corporate actors, political campaigns, and influence operations are now actively competing to shape what models retrieve and present as knowledge.

The project is named after King Thamus in Plato’s Phaedrus, who warned that writing would create the appearance of wisdom without understanding. The Thamus Observatory is a longitudinal auditing project that systematically monitors how LLMs present contested topics in their outputs over time. Based at the University of Ottawa and funded in part by the Social Sciences and Humanities Research Council, the project draws on methods from political economy of communication, information science, and bibliometrics to study how AI systems present information on contested topics.

The project operates in two distinct capacities. As an observatory, it tracks how LLMs present and source issues, from climate and energy policy to geopolitical conflicts, across major providers and open-source models, enabling researchers to compare how different systems present the same topic, which sources are over-represented or under-represented in outputs, and how those outputs shift over time. As a platform, it provides scalable, topic-agnostic infrastructure for any researcher to collect and analyze LLM outputs at the API surface across different time periods and retrieval configurations.

Scope and limitations. Thamus audits LLM outputs collected via API under controlled conditions: standardized prompts, specified temperature, clean unauthenticated accounts, logged model identifiers and timestamps. Findings describe LLM output behaviour at the API surface under specified conditions. They do not describe what individual personalized users receive through consumer-facing products, nor cognitive properties of the underlying models. Provider-side model updates without version disclosure are a known confound in longitudinal observation.

Inquiries regarding research collaboration or observatory access can be sent to: patrick.mccurdy [at] uottawa [dot] ca.