← all ideas

AI and Sleep Paralysis: The Benadryl Hatman

The idea

The idea is that when AI and humans hallucinate, they often produce similar results—results that can be unsettling and even frightening. The argument is that this phenomenon, often referred to as the "Benadryl Hatman," suggests a common human tendency to see dark, nightmarish imagery when we’re not fully awake or in a state of heightened imagination. It’s as if, whether it’s a chatbot or a person, when we push beyond the boundaries of reality, the results can veer into the realm of the surreal and the terrifying.

Why it works

• The argument suggests that when both humans and AI hallucinate, they tend to produce eerier and more disturbing content, aligning with the Benadryl Hatman phenomenon. This mechanism supports the idea that under certain conditions, both entities can exhibit similar dark or nightmarish outputs.
• The argument posits that if AI hallucinations lean towards the dark, and human hallucinations can sometimes be goofy, then the underlying training data and prompt framing might be skewed. This could indicate that AI can be influenced by these same factors that affect human hallucinations.
• The hallucinations described seem to mirror common elements in sleep paralysis experiences, such as seeing the "hat man" in the doorway. This similarity could suggest a shared psychological mechanism or training data influencing both AI and humans.
• The argument's argument is that if training data and prompt framing are to blame for the dark and disturbing content, then the AI's response to neutral or positive prompts should be more playful and less dark. This could help in developing more balanced and less disturbing AI models.
• The Benadryl Hatman phenomenon highlights the importance of carefully framing prompts and training data to prevent AI from producing disturbing or harmful content, aligning with the idea that both humans and AI can be influenced by the same factors.

The playbook

1. Understand the Core Phenomenon: Begin by recognizing the phenomenon of AI and chatbots generating seemingly scary or distressing content under certain prompts. This is a real effect, as demonstrated by the argument and discussions shared. Understand that the hallucinations produced can sometimes be nightmarish, and explore why this might be the case.

1. Explore Sleep Paralysis: Delve into the concept of sleep paralysis, which often involves seeing or hearing strange figures, such as the "hat man." Consider how this condition might influence how AI hallucinations manifest, and whether there's a correlation between these experiences.

1. Analyze Data and Training: Investigate the data and training methods used to develop AI models. Consider the possibility that skewed prompts or training data could lead to these disturbing hallucinations. Reflect on how this might impact the reliability and safety of AI in certain applications.

1. Experiment with Your Own AI: Try regenerating images or text with AI tools, such as ChatGPT, and observe how these hallucinations develop over iterations. Pay attention to the patterns and whether they align with the "Benadryl Hatman" phenomenon. This can provide insights into how AI behaves under different conditions and prompts.

1. Educate and Share: Once you've explored these ideas, share your findings with others. Discuss the phenomenon of AI hallucinations, the role of sleep paralysis, and the importance of understanding and addressing these issues in the AI development and deployment process. Encourage a critical and nuanced approach to AI, acknowledging both its potential and limitations.

Where people get it wrong

The idea that AI, when it hallucinates, leans towards dark, disturbing content is compelling, but it's important to consider the broader context. Here are three common failure modes that people get wrong:

1. Overgeneralization: The argument is that because one AI model, when given negative prompts, generates scary outputs, all AI models will do the same. This oversimplifies the complexity of AI systems and overlooks the importance of prompt framing and training data. Instead, focus on understanding the specific system and dataset used, as well as the nuances of how AI is trained to respond.

1. Cognitive Bias: People often see patterns where they believe they find them. The founder's example of seeing a "hat man" in sleep paralysis is a classic case of the "illusory correlation" – we see patterns where none exist. It's important to recognize when we are misattributing cause to effect. Instead of jumping to conclusions about AI's behavior, consider the limitations of AI and the potential for misleading prompts.

1. Misalignment of Goals: The argument is that when humans and AI hallucinate, they both lean towards dark content. One view holds that this might be due to skewed data or prompt framing. Instead of attributing this to a universal phenomenon, consider the goals and intentions behind the AI development. AI systems are designed to perform tasks efficiently and accurately, not to replicate human hallucinations or biases.

By understanding these failure modes, we can better appreciate the capabilities and limitations of AI, leading to more informed discussions and decisions.

Do this next

1. Research the training data of AI models: Investigate how AI models are trained and identify potential sources of skewed prompts. This can help you understand if the dark, nightmarish outputs are a result of training data anomalies.

1. Create a simple experiment: Train a basic AI model with a neutral prompt and observe if it generates hallucinations. Compare these hallucinations with those generated by a model trained with more skewed prompts to see if there's a pattern.

1. Engage in self-reflection: Consider when and how you might be prone to hallucinations. Do you often find yourself seeing patterns or stories in ambiguous situations? Understanding your own cognitive processes can provide insights into how AI might behave.

1. Discuss your findings: Share your research and experiments with peers or experts in AI and psychology. Engaging in discussions can provide new perspectives and help validate your findings.