Deanthropomorphising NLP: Can a language model be conscious?
PloS one January 1, 2024 Matthew Shardlow, Piotr Przybyła 8 citations
A pretrained large language model (LLM) such as LaMDA or ChatGPT cannot be conscious or sentient. Analyzing the Transformer architecture through Integrated Information Theory shows no basis for consciousness. Claims of sentience reflect a tendency to use anthropomorphic language in NLP reporting, not genuine advances. The discussion is an opportunity to evaluate progress in language modeling and its ethical implications, with background provided for non-specialists.