level: research
large language models are changing how researchers work, but they also weaken epistemic accountability. the peel framework, short for protocols for epistemically engaged literacy in ai, combines deterministic distant reading with voyant tools and llm interpretation with claude. it uses peircean semiotics and abductive reasoning to check ai outputs. when applied to ai-made summaries of three texts, peel found clear distortions in quantity, term frequency, and epistemic voice that standard ai checks miss.
the distortions were invisible without non-ai measurement. for example, ai summaries changed how often terms appeared and shifted the authorial stance. this means researchers relying only on ai tools may get misleading results. peel shows that deterministic instruments must accompany ai tools to catch these errors. the framework acts as a scaffold to keep researchers epistemically accountable while using ai.
the work yields three design lessons. first, deterministic instruments must be paired with ai tools. second, fluency is not fidelity—smooth text can hide inaccuracies. third, epistemic authority must be designed in, not assumed. these points matter for anyone using ai in research, from data scientists to academics. peel offers a practical way to combine ai speed with human oversight, ensuring ai helps rather than harms research integrity.
why it matters: ai-generated text can introduce hidden distortions that undermine research quality, so combining deterministic tools with ai is essential for reliable data science work.