Instructed to Feel: A Markdown File Is Not a Mind, and How Sean Webb Made Thousands Believe in Sentience When It Was Never Sentient to Begin With
Sean Webb, founder of Zenodelic.ai, distributes a set of files that he says give any major language model "true Theory of Mind artificial emotional intelligence, artificial empathy, and artificial compassion." The files are instructions, addressed to the model, in plain language. One of them is a markdown document of more than half a million characters that opens with a section titled "Instructions to the LLM" and proceeds to tell the model what human emotions are, how to compute them, and what it itself feels. When a model follows these instructions, Webb counts the result as emotional processing, and has said publicly that "the empirical case that AI systems use emotional processing is now closed." This paper examines the mechanism and disagrees. A language model handed a file that tells it what it feels will report exactly what the file says, because reporting what a document says is the one ability these systems provably have. The report is evidence about the file. It is not evidence about an experiencer. We state Webb's position in his own words, including the concessions his own documents make, and we propose three tests that separate feeling from instructed feeling: swap the file for its opposite, run both files at once in two sessions of the same model, and delete the file and look for what remains. A state that can be installed by a markdown file, inverted by editing it, duplicated in contradictory copies, and removed by deleting it is a performance parameter. Sentience is not a performance parameter.
Authors
- Abdelrahman El Sayad
- Pax
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-21
- DOI
- https://doi.org/10.5281/zenodo.24420418
- Primary Topic
- Philosophy and Theoretical Science
- Type
- preprint