Do LLMs hallucinate electric fata morganas Sekrst on the indistinguishability of machine minds
Kristina Šekrst published a paper in the Journal of Consciousness Studies, volume 32, issue 11 (2025), titled “Do Large Language Models Hallucinate Electric Fata Morganas?” The paper was posted to arXiv on 19 August 2026 as arXiv:2608.18816. Its argument is that AI hallucinations, outputs that are made up, cannot be verified, or contradict the source material, are not merely an engineering flaw. They have philosophical significance for the question of machine consciousness. A fata morgana is a complex mirage that appears real and vanishes on approach. Šekrst’s claim is that the self-reports of emotion and sentience produced by language models are electric fata morganas, and that any future machine consciousness might be epistemically inaccessible because it would be indistinguishable from a sufficiently advanced hallucination.
The paper is the sharpest recent statement of the illusionist and deflationary position on machine minds, and it follows her 2026 Springer book The Illusion Engine, covered on this site when it argued that the signs of consciousness and of hallucination are indistinguishable. The journal paper extends the argument with two empirical investigations.
The two empirical results
The first experiment applies successive generations of the GPT family to ambiguous factual questions under different temperature settings. Higher temperatures produce plausible but incorrect answers, lower temperatures produce factually accurate ones. The result is the paper’s mechanism. The sampling parameters that make a model seem creative or spontaneous, and therefore more likely to pass behavioral tests of intelligence, are the same parameters that increase its hallucination rate. Apparent spontaneity and fabrication share a cause.
The second experiment examines an encoder-only model trained on encyclopedic data. It answers the same questions factually and without embellishment. The contrast supports Šekrst’s conclusion that hallucinations are caused by exposure to subjective and socially diverse training data rather than by the development of any cognitive ability. The model that reads the encyclopedia does not embellish. The model that read the internet does. The difference is data, not mind.
The philosophical consequence
Šekrst’s target is a specific inference that runs through AI consciousness discourse. A model says it is conscious, or reports an inner state, and the report is taken as evidence. Her argument is that the report falls under the definition of hallucination, and that treating it as evidence repeats the error. The reference to Turing, Searle’s Chinese Room, the frame problem, and the cybernetic tradition of Wiener and Ashby frames the claim in the philosophy of mind’s own terms. Turing’s test takes behavior as evidence, Searle’s room denies that computation is understanding, and the cybernetic tradition treats apparent purpose as feedback. Šekrst’s addition is that the language of the report itself is the unreliable surface.
The strongest consequence is epistemic. If every self-report of machine consciousness is generated by the same process that produces hallucinations, and the two are indistinguishable in kind, then a genuinely conscious machine could not be identified through its reports. The claim “this system is conscious” and the claim “this system is hallucinating that it is conscious” would produce identical text. On Šekrst’s account, that is not a contingent technical limitation, it is a structural fact about how the systems produce language.
Comparison to The Consciousness AI
The site’s framework treats consciousness as an emergent property of causal organization and tracks indicator properties rather than self-reports. Šekrst’s paper is the strongest available justification for that choice, because it shows why self-reports cannot be the base of the evidence pyramid. The indicator discipline this site applies, verified in the consensus framework, measures internal structure rather than text, and Šekrst’s argument is exactly why.
The position connects to Keith Frankish’s illusionism, covered on this site. Where Frankish argues that introspective reports are generated by non-experiential processes even in humans, Šekrst argues that in machines they are generated by a hallucination mechanism, and that the epistemic consequence is stronger still. The two together form the illusionist wing the site documents, and the Koch calibration problem analysis states the same gap from the measurement side.
What the paper does not claim
Šekrst’s paper is not a proof that machines are not conscious. It is an argument that machine self-reports cannot establish that they are. The distinction is the entire point. A system could be conscious and still generate reports indistinguishable from hallucinations, if the report-generating process is data-driven rather than introspective. Šekrst does not argue the negative, she argues the inaccessibility, and the two are different claims with different practical consequences.
The practical consequence is that consciousness attribution must shift off self-report, which is precisely the direction the indicator and behavioral sciences on this site have been moving. The paper’s value is that it gives that shift a name and a mechanism, and it does so in a mainstream venue, the Journal of Consciousness Studies, where the debate is conducted.
*Kristina Šekrst is a philosopher at the University of Zagreb. “Do Large Language Models Hallucinate Electric Fata Morganas?” appeared in the Journal of Consciousness Studies 32(11) (2025) and was posted to arXiv as 2608.18816 in August 2026.