Is Hallucination a Hallucination?
Yucong Duan, Lei Yu, Yingbo Li, Haoyang Che
International Standardization Committee of Networked DIKWP for Artificial Intelligence Evaluation(DIKWP-SC)
World Artificial Consciousness CIC(WAC)
World Conference on Artificial Consciousness(WCAC)
(Email: duanyucong@hotmail.com)
Abstract
The term "hallucination" has been widely adopted in the field of artificial intelligence (AI) to describe instances where models generate outputs that are not grounded in the provided data or real-world knowledge. This usage borrows from the psychological concept of hallucination in humans, which refers to perceptions without external stimuli. In this article, we critically examine the appropriateness of the term "hallucination" when applied to AI systems, exploring the conceptual differences between human cognitive phenomena and machine-generated outputs. We argue that the term may be misleading and propose alternative terminology that more accurately reflects the underlying mechanisms in AI systems.
Introduction
Advancements in artificial intelligence, particularly in natural language processing and generative models, have led to remarkable capabilities in text generation, translation, and conversational agents. However, these systems sometimes produce outputs that are nonsensical, factually incorrect, or unfaithful to the input data—a phenomenon commonly referred to as "hallucination" in the AI community. The adoption of this term draws a parallel between human perceptual experiences and machine-generated errors.
In human psychology, hallucinations are vivid sensory experiences occurring without external stimuli, often associated with mental health disorders or substance use. These experiences are subjective and rooted in complex neural processes. In contrast, AI "hallucinations" result from computational processes within models trained on large datasets. This article examines whether the term "hallucination" is conceptually appropriate for AI systems, considering the fundamental differences between human cognition and machine operations.
Understanding Human Hallucinations
Human hallucinations involve sensory perceptions without corresponding external stimuli, encompassing visual, auditory, olfactory, gustatory, or tactile experiences. They are often symptomatic of neurological or psychiatric conditions, such as schizophrenia, or induced by psychoactive substances. The underlying mechanisms are complex, involving aberrant neural activity and neurotransmitter imbalances. Importantly, hallucinations are intrinsically linked to consciousness and subjective experience.
AI "Hallucinations": Mechanisms and Causes
In AI systems, particularly large language models (LLMs) like GPT-3 and GPT-4, "hallucinations" refer to outputs that deviate from expected or accurate responses. These may include:
Factual Errors: Generating incorrect information not supported by the training data.
Incoherent Responses: Producing nonsensical or logically inconsistent statements.
Unfaithful Summarization: Summaries that misrepresent the content of the source material.
These issues arise due to several factors:
Training Data Limitations: Gaps or biases in the training corpus can lead to incomplete knowledge.
Model Overgeneralization: The probabilistic nature of language models may generate plausible but incorrect continuations.
Lack of World Modeling: Absence of a structured understanding of the real world or factual consistency.
Unlike human hallucinations, AI outputs are not the result of conscious experience but are generated through mathematical transformations and pattern recognition in data.
Comparative Analysis
Subjectivity vs. Objectivity
Human hallucinations are subjective experiences, deeply personal and not directly observable by others unless communicated. AI outputs are objective artifacts that can be examined and evaluated by users for correctness and coherence.
Consciousness and Intent
Humans possess consciousness and intentionality; their hallucinations are unintended and often distressing. AI systems lack consciousness and intentions; any "hallucination" is an unintended byproduct of algorithmic processing without awareness or experience.
Underlying Processes
The neurological processes leading to human hallucinations involve complex brain functions, neurotransmitters, and neural circuits. In AI, the generation of incorrect outputs is due to model architecture, training data properties, and inference algorithms.
The Case for Terminological Precision
Using the term "hallucination" for AI systems may be misleading for several reasons:
Anthropomorphism: It anthropomorphizes AI systems, attributing human-like cognitive processes where none exist.
Conceptual Confusion: It conflates subjective human experiences with objective computational errors.
Communication Clarity: It may obscure the technical causes of AI errors, hindering effective communication among practitioners and with the public.
Alternative Terminology
To enhance clarity and precision, alternative terms are suggested:
"Generation Errors": Emphasizes that the issue arises during the output generation process.
"Fabrications": Indicates that the model is creating information not present in the input or training data.
"Unfaithful Outputs": Highlights the lack of fidelity to the source material or intended response.
"Model Deviations": Reflects departures from expected or correct outputs due to model limitations.
Implications for AI Development
Recognizing the differences between human hallucinations and AI output errors has practical implications:
Improved Diagnostics: Clear terminology aids in diagnosing and addressing the root causes of errors.
Public Understanding: Avoiding misleading terms enhances public comprehension of AI capabilities and limitations.
Ethical Considerations: Precise language supports responsible communication about AI systems, preventing overestimation of their cognitive abilities.
Conclusion
While the term "hallucination" has been adopted in AI to describe certain erroneous outputs, it is conceptually misaligned with its original psychological meaning. The differences in subjectivity, consciousness, and underlying mechanisms between humans and AI systems warrant a reevaluation of this terminology. By adopting more accurate descriptors, the AI community can improve communication, focus on addressing the technical causes of errors, and avoid anthropomorphizing machines. This shift supports the responsible development and deployment of AI technologies.
References
Berrios, G. E. (1995). The history of mental symptoms: Descriptive psychopathology since the nineteenth century. Cambridge University Press.
Marcus, G., & Davis, E. (2020). GPT-3, Bloviator: OpenAI's language generator has no idea what it's talking about. MIT Technology Review.
Maynez, J., Narayan, S., Bohnet, B., & McDonald, R. (2020). On faithfulness and factuality in abstractive summarization. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 1906–1919.
Turing, A. M. (1950). Computing machinery and intelligence. Mind, 59(236), 433–460.
Vaswani, A., et al. (2017). Attention is all you need. Advances in Neural Information Processing Systems, 5998–6008.
Acknowledgments
The author thanks colleagues in the fields of cognitive science and artificial intelligence for their insightful discussions on the conceptual frameworks of human cognition and machine learning. Their contributions have been invaluable in shaping the perspectives presented in this article.
转载本文请联系原作者获取授权，同时请注明本文来自段玉聪科学网博客。
链接地址：https://wap.sciencenet.cn/blog-3429562-1451788.html?mobile=1
收藏
