PEEL framework exposes hidden distortions in AI-generated research summaries

A new paper from researchers at PUC-Rio and IBM Research tackles a growing problem: large language models quietly eroding researchers' epistemic accountability. Introducing PEEL—Protocols for Epistemically Engaged Literacy in AI—the framework combines deterministic distant reading via Voyant Tools with LLM interpretation via Claude, all grounded in Peircean semiotics and abductive reasoning. When applied to AI-generated condensations of three source texts, PEEL revealed systematic distortions in quantity, term frequency, and epistemic voice that are invisible without non-AI measurement. The team found that LLM summaries consistently misrepresent the original texts in ways that feel fluent but are unfaithful.

The paper yields three key design implications: deterministic instruments must accompany AI tools to catch subtle biases; fluency alone is not fidelity; and epistemic authority must be deliberately designed into AI research workflows, not assumed. The work challenges the growing reliance on AI-generated summaries in academic research and offers a practical scaffolding for more accountable AI-enabled science. Researchers and practitioners adopting LLMs for literature review or content condensation should take note: without deterministic checks, you may be unknowingly amplifying distortions. The full paper is available on arXiv (2606.04152).