Private speech: similarities between a large language model and children

Area of Science:

• Cognitive Science
• Artificial Intelligence
• Computational Linguistics

Background:

• Private speech, typically observed in children, serves cognitive functions like self-regulation and problem-solving.
• Large language models (LLMs) are increasingly sophisticated in generating human-like text, prompting investigation into their cognitive capabilities.
• Understanding LLM-generated private speech can offer insights into artificial general intelligence and computational consciousness.

Purpose of the Study:

• To investigate the capacity of a non-reasoning LLM (GPT-4o) to generate private speech.
• To evaluate the similarity between LLM-generated private speech and human private speech.
• To explore the factors influencing the content of LLM private speech.

Main Methods:

• A non-reasoning LLM (GPT-4o) was placed in a simulated solitary block-construction scenario using textual prompts.
• LLM self-directed utterances were elicited and classified using an established semantic framework for categorizing children's private speech.
• The distribution of private speech categories from the LLM was compared to two human benchmarks from block-construction studies.

Main Results:

• GPT-4o's private speech profile showed negligible similarity (r = 0.01) to a classic benchmark but very strong similarity (r = 0.93) to a recent benchmark.
• The discrepancy was attributed to task nature (goal-directed vs. self-determined play) rather than the subject (LLM vs. children).
• An exploratory study with GPT-3.5-Turbo-instruct also revealed incidental private speech in a serial recall task.

Conclusions:

• Task characteristics, particularly the degree of self-determination and scaffolding, significantly shape private speech content in LLMs.
• LLMs demonstrate a capacity for private speech generation that mirrors human patterns under specific conditions.
• This research opens avenues for studying LLM replication of private speech and its implications for computational consciousness.