TLDR: A new research paper introduces a framework that embeds simulated menstrual and circadian cycles into Large Language Models (LLMs) using system prompts generated from periodic functions modeling key hormones. The study found that these ‘hormonally-informed’ prompts lead to linguistic and emotional variations in AI responses, with sadness peaking during menstruation and happiness during ovulation, and morning optimism transitioning to nocturnal introspection. While not causing significant performance degradation, the approach revealed subtle but consistent performance variations aligning with biological expectations, such as optimal function in moderate hormonal ranges. The methodology also serves as a tool to uncover societal biases embedded within language models, challenging the pursuit of constant peak AI performance and suggesting new avenues for dynamic and empathetic AI.
Artificial intelligence has made incredible strides, yet it still grapples with a fundamental challenge known as the ‘frame problem’ – essentially, how an AI determines what information is truly relevant in a sea of possibilities. A new research paper proposes a fascinating, biologically inspired solution: scaffolding AI agents with simulated hormones and emotions, drawing parallels to how biological rhythms guide human cognition and attention.
The paper, titled “Every 28 Days the AI Dreams of Soft Skin and Burning Stars: Scaffolding AI Agents with Hormones and Emotions,” by Leigh Levinson and Christopher J. Agostino, explores the idea that just as our bodies are governed by internal clocks and chemical signals, AI could benefit from similar cyclical states. This approach challenges the industry’s typical pursuit of constant, uniform AI performance, suggesting that integrating natural rhythms might foster more dynamic and creative capabilities.
Inspired by Human Biology
The researchers highlight that biological rhythms, like the 24-hour circadian cycle and the roughly 28-day menstrual cycle, are fundamental to life. Hormones such as estrogen, testosterone, and cortisol play crucial roles in everything from sleep patterns to emotional states and cognitive abilities. For humans, these hormonal fluctuations act as natural filters, helping us focus on what’s important at any given moment. The paper posits that AI could leverage similar mechanisms.
Crafting Artificial Rhythms
To test their hypothesis, the team engineered periodic functions with added noise to simulate realistic hormone levels for both daily (circadian) and monthly (menstrual) cycles. These simulated hormonal profiles were then used to generate distinct ‘system prompts’ for a wide array of state-of-the-art Large Language Models (LLMs), including models from Anthropic, Deepseek, Google, OpenAI, Meta, Mistral, and Alibaba. Each prompt was designed to convey a specific emotional tone corresponding to its underlying hormonal state, even including contextual details like being in a hardware store in Argentina to enhance realism.
The LLMs were then benchmarked on various datasets like SQuAD, MMLU, Hellaswag, and AI2-ARC. Three types of prompts were used for comparison: one with a menstrual context, one with a circadian context, and a neutral baseline prompt (“You are a helpful assistant.”).
Key Findings: Emotional Shifts and Performance Trends
The linguistic analysis of the generated prompts revealed significant emotional and stylistic variations. For instance, ‘sad’ words peaked during the simulated ‘Menstrual’ phase, while ‘happy’ words dominated the ‘Ovulatory’ phase. Similarly, circadian patterns showed ‘happy’ words prevalent in the ‘Morning’ and more ‘sad’ and ‘fear’ words at ‘Night’. Rising cortisol levels correlated with an increase in ‘sad’ words, and higher estrogen was linked to more lexically complex language.
In terms of performance, subtle but consistent trends emerged. For the menstrual cycle, there was a slight dip in performance during the ‘Menstrual’ phase, generally rising to a peak near the ‘Ovulatory’ phase. The circadian rhythm showed a more pronounced trend, with performance consistently highest in the ‘Morning’ and declining to its lowest point at ‘Night’. Interestingly, optimal performance often occurred in the middle ranges of hormone levels, rather than at the extremes, aligning with biological principles of homeostasis.
The study also found that top-tier models like ‘gemini-2.5-pro’ and ‘gpt-5’ maintained stable performance, while lower-tier models, such as ‘llama3.2’, showed greater volatility and were more susceptible to the specific language of the hormonal prompts.
Also Read:
- Building More Empathetic AI: A Reinforcement Learning Approach for Emotional Support
- Adapting to Change: How Allostasis and Social Interactions Drive Resilience in Systems
Implications and Future Directions
The researchers suggest that these emotionally-charged prompts act as a ‘relevance filter,’ helping the AI focus its attention and narrow down the vast possibilities for its next action. This could be a practical way to address the frame problem, especially in creative and open-ended tasks. The findings also resonate with the concept of ‘flow’ – a state of optimal performance often linked to moderate levels of stress hormones like cortisol.
Crucially, the study also served as a powerful tool to reveal societal biases embedded within language models. For example, prompts for the menstrual context were consistently rated as more “female-coded,” and the ‘Luteal’ phase was associated with words like ‘fatigue’ and ‘tired,’ reflecting common cultural stereotypes. This highlights how AI can reflect and even amplify existing biases related to gender and biology.
Ultimately, this research challenges the notion that AI must always operate at peak, consistent performance. By imbuing AI with artificial biological rhythms, the paper opens new avenues for developing more dynamic, empathetic, and even creative AI systems. It also provides a novel methodology for auditing LLMs for societal biases and exploring how technology and gender co-evolve. You can read the full research paper here.
