Do Large Language Models Hallucinate Electric Fata Morganas?

Source

• Source name: arXiv
• Source kind: paper
• Authors / organization: Kristina Šekrst
• Venue: arXiv: cs.CL
• Published: 2026-08-19
• Collected: 2026-08-20
• URL: http://arxiv.org/abs/2608.18816v1
• PDF: https://arxiv.org/pdf/2608.18816v1
• DOI: 10.53765/20512201.32.11.096
• Registry group: academic_apis_and_preprints
• Evidence role: canonical_academic_index
• Language: en
• Country / region: UNKNOWN

Plain-Language Brief

The paper explores the issue of AI hallucinations (outputs that are made up or contradict source material) and their implications for machine consciousness. The author investigates causes of hallucinations in large language models, such as temperature settings and exposure to subjective training data, and argues that a model's self-reported emotion or sentience could be considered a form of hallucination.

What the Document Says

• AI hallucinations are regarded as an engineering flaw and have philosophical significance for machine consciousness.
• The paper discusses causes of hallucinations in large language models, including source-target divergence, discrepancies between training and inference, overfitting, temperature settings, and exposure to subjective and socially diverse training data.
• Higher temperatures in AI models lead to plausible but incorrect answers, while lower temperatures lead to factually accurate ones.
• An encoder-only model trained on encyclopedic data answers questions factually without embellishment, suggesting that hallucinations are due to the model's exposure to subjective training data rather than cognitive ability development.
• The author claims that a model's self-reports of emotion or sentience could be considered a form of hallucination and that any future occurrence of machine consciousness might remain epistemically inaccessible since it would be indistinguishable from a sufficiently advanced hallucination.
• References are made to Turing, Searle's Chinese Room, the frame problem, and the cybernetic tradition of Wiener and Ashby.

Key Concepts Mentioned

• AI hallucinations
• machine consciousness
• causes of hallucinations
• temperature settings
• encoder-only model
• self-reported emotion or sentience

Claims Made by the Document

• A model's self-reports of emotion or sentience could be considered a form of hallucination.
• Any future occurrence of machine consciousness might remain epistemically inaccessible.

Evidence Used

• Observation of AI models under different temperature settings
• Study of an encoder-only model trained on encyclopedic data

Methods / Approach

• Empirical investigations
• References to philosophical and cybernetic theories

Limitations Stated by the Authors

• The paper does not provide a solution or method for eliminating AI hallucinations.

Limitations Noted by the Agent

Agent interpretation — not from the source.

• The study focuses on large language models and may not be applicable to all types of AI.
• The findings are based on the specific models and data used in the experiments.

Open Questions

• What solutions or methods can be developed to eliminate AI hallucinations?
• Can machine consciousness ever truly be understood or accessed?

Notes for Future Q&A

May include interpretation. Labeled where it does.

• The paper argues that AI hallucinations have philosophical implications for the question of machine consciousness, but it does not necessarily prove that machines can never have true consciousness.

Raw Abstract / Excerpt

AI hallucinations - that is, outputs which are made up, cannot be verified, or contradict the source material - are generally regarded as an engineering flaw to be dealt with. This paper contends that they also have philosophical significance when it comes to the question of machine consciousness. We examine the known causes of hallucinations in large language models - such as source-target divergence, discrepancies between training and inference, and overfitting - and we present two empirical investigations. In the first, we apply successive generations of the GPT model to ambiguous factual questions under different temperature settings, finding that higher temperatures result in plausible but incorrect answers while lower temperatures lead to factually accurate ones. The sampling parameters that cause a model to seem creative or spontaneous and thus more likely to pass behavioral tests of intelligence are the same ones that increase its hallucination rate. In the second, we look at an encoder-only model that has been trained on encyclopedic data and which answers questions of the same type factually and without embellishment, indicating that hallucinations are due to exposure to subjective and socially diverse training data rather than to the development of any cognitive ability. Using references to Turing, Searle's Chinese Room, the frame problem, and the cybernetic tradition of Wiener and Ashby, we claim that a model's self-reports of emotion or sentience come within the definition of hallucination, and that any future occurrence of machine consciousness might remain epistemically inaccessible since it would be indistinguishable from a sufficiently advanced hallucination.