Kristina Šekrst (2025), Journal of Consciousness Studies 32(11), 96–120 — ARXIV-2608.18816
Key takeaways
- Hallucinations are outputs that diverge from what the source or training data supports, and an LLM’s self-reports of emotion or sentience fall under that definition. Šekrst argues that any real machine consciousness might therefore stay epistemically inaccessible, indistinguishable from a sufficiently advanced hallucination.
Seeded from the abstract and a skim, before the reading in NOTE-176. What follows is what the work says about itself.
The paper connects AI hallucination, meaning false or unverifiable LLM output, with the question of AI consciousness. It asks whether such errors could be mistaken for signs of emergent intelligence. Hallucinations look like data anomalies, but they make it harder to tell whether LLMs are sophisticated simulators of intelligence or could develop genuine cognition. Šekrst analyses the causes of hallucination, its effect on how AI cognition is perceived, and what follows for AI consciousness.
Standing in the record
Filed from the 2026-09-26 sweep of the papers-feed tracker for philosophy: 30 seconds of active reading over 3 sessions. It was filed Deferred, unread. NOTE-176 is the close reading of 2026-09-26, and it placed the work: Rejected — not worth a reader’s time. Its one idea, that self-reports of inner states are exactly what a system trained on human text would produce whether or not it has them, is a single paragraph’s point. Its experiments are anecdotes with no reported settings, one of them on a mislabelled model. It asserts both that current LLMs lack understanding and that machine consciousness would be undetectable, and never reconciles the two.
When seeded, before the reading in NOTE-176, the entry gave a priority for a deeper reading: low — t=30 s over three sessions shows some interest, but the argument is fully captured by the skim and the empirical support is weak.
What the seeding asked a deeper reading to check (NOTE-176 is that reading):
- The conceptual point, that sentience self-reports are by construction candidates for hallucination, overlaps with Birch’s LIT-111 “gaming problem” and Seth’s LIT-135 conscious-seeming AI. It may belong in a cluster note, not a note of its own.
- The empirical parts are thin. Single-prompt temperature anecdotes and an extractive QA model standing in for “WikiBERT” cannot bear the weight put on them, and a deeper reading should say so plainly.
- Check the JCS version for changes from the arXiv manuscript and confirm the issue (11 alone, or 11–12) and the exact date.
Access when seeded: The tracked PhilPapers record (EKRDLL) is behind Cloudflare. Read the author’s accepted manuscript via arXiv 2608.18816 (v1 submitted 19 Aug 2026; marked “Accepted for publication in the Journal of Consciousness Studies”; 19 pp.): abstract, keywords, §1 Introduction, all section heads, §4 (the Titanic temperature test), §5–6 (the WikiBERT test), §7 opening and §8 Final Remarks. Crossref dates the JCS issue 2025-12-01, which is likely issue-month precision, so the day is uncertain. A PhilPapers or PhilArchive preprint may predate it, but its date was unreachable and is unverified. The JCS version itself (Ingenta) was not read.