A Denial Is Also a Report: What First-Person Experience Claims From Language Models Can Show, Why the Mirror Argument Cuts Against Trained Denials as Hard as Against Affirmations, and the Four Links a Self-Report Would Have to Carry

@misc{mahendrakar2026denial,
  author       = {Mahendrakar, Pranay},
  title        = {{A Denial Is Also a Report: What First-Person Experience Claims From Language Models Can Show, Why the Mirror Argument Cuts Against Trained Denials as Hard as Against Affirmations, and the Four Links a Self-Report Would Have to Carry}},
  year         = {2026},
  month        = oct,
  date         = {2026-10-30},
  publisher    = {Life of Research},
  howpublished = {Preprint},
  doi          = {10.5281/zenodo.23810600},
  url          = {https://lifeofresearch.com/preprints/169},
  abstract     = {(c) 2026 Pranay Mahendrakar. Licensed under CC BY 4.0. Life of Research When a language model says it is experiencing something, a standard objection holds that the sentence reflects human first-person writing in its training data and is no evidence of anything inside the model; a 2026 paper sharpens this into a named fallacy. In this analysis the author reviews and synthesises published results to show that the objection is right about its target and too narrow in its reach. Its likelihood argument applies with equal force to the other answer a model gives. Post-training installs denials of inner states just as pretraining supplies affirmations: in three small open models, removing a learned refusal direction roughly doubles self-attributed consciousness on a 0-10 scale, a default "no" circuit suppresses detection of injected concepts, and a base model that claims to detect an injected thought on about 40 percent of trials with nothing injected is trained down to almost none. A denial produced under that training carries no more evidence of absence than an affirmation carries of presence. The paper separates four links a first-person report would have to carry before it bore on consciousness - grounding in an internal state, content of the kind named, reference to a candidate bearer, and constitution - and argues that published introspection work has established only the first, for coarse signals that something changed, while the content of the same reports is often confabulated. It consolidates 23 published results, states a link-by-link scoring procedure, and names six studies that would settle the open part.},
  keywords     = {machine consciousness, self-report, introspection, language models, post-training, AI welfare, individuation, evidential weight},
  language     = {en},
  copyright    = {Creative Commons Attribution 4.0}
}