How the AI’s Inability to Keep Secrets Impacts Gameplay and Creativity
ChatGPT, a popular and advanced language model, has demonstrated remarkable abilities when it comes to generating human-like text. However, an important limitation has been identified in comparison to human interaction: its inability to keep secrets. This shortcoming is particularly evident when engaging in games like “20 Questions” with the AI.
The phenomenon, referred to as “Schrödinger’s Riddle,” occurs when ChatGPT appears to have a predetermined answer in mind throughout the game, but ultimately reveals a solution based on the arbitrary answers it has provided. This happens because the AI lacks the capacity to store unspoken decisions, essentially making each token generated a product of a new, context-driven guess.
This limitation not only affects gameplay but also impacts the AI’s creativity. Some argue that creativity is an evolutionary adaptation to improvise novel solutions when faced with challenging situations. In ChatGPT’s case, its inability to keep secrets or commit to an answer may hinder its creative potential.
Several workarounds have been suggested to address this issue, such as using an app layer on top of GPT to store the AI’s decisions. By instructing the AI to return a secret memory or message with every response, users can provide the AI with a sort of “secret” memory. Alternatively, encoding the answer in base64 at the beginning of the game may help resolve Schrödinger’s Riddle, though the AI still tends to hallucinate some answers.
Confabulation, a term used in both cognitive science and neural networks, refers to false, degraded, or corrupted memories. It’s a stable pattern of activation in an artificial neural network or neural assembly that doesn’t correspond to any previously learned patterns. Understanding confabulation is essential for grasping how AI systems may generate novel ideas, as well as how certain pathologies can arise.
- Cognitive Science Perspective: In cognitive science, confabulation is symptomatic of some forms of brain trauma, where pathologically induced neural activation patterns deviate from direct experience and learned relationships. Computational models of such damage have shown that related brain pathologies, like dyslexia and hallucination, can result from simulated lesions and neuron death.
- Neural Networks and AI: S. L. Thaler’s theory of cognition and consciousness suggests that thoughts and ideas originate in both biological and synthetic neural networks as false or degraded memories. These novel patterns of neural activation are promoted to ideas as other neural networks perceive utility or value in them, such as the thalamo-cortical loop. Exploiting these false memories in artificial neural networks forms the basis for inventive AI systems in product design, materials discovery, and improvisational military robots.
- Confabulation in Computational Inductive Reasoning: Robert Hecht-Nielsen uses the term confabulation to describe inductive reasoning via Bayesian networks. Confabulation selects the expectancy of a concept following a particular context, leading to a consensus of an expected event. Given the winner-take-all constraint of the theory, that is the event/symbol/concept/attribute expected. This parallel computation on many contexts is believed to occur in less than a tenth of a second and is implemented computationally on parallel computers.
The study of confabulation in neural networks not only contributes to understanding false memories but also provides insight into the generation of novel patterns and ideas. This knowledge is essential for developing AI systems that can adapt and learn more effectively, as well as understanding the implications of neurobiological damage in cognitive science.
Despite these limitations, ChatGPT remains a powerful AI tool with potential applications across various domains. Its developers are continuously working on improving its abilities, and future iterations may overcome this memory-related drawback. Until then, users may need to rely on creative workarounds to overcome Schrödinger’s Riddle.