# A Plan 'B' for AI Safety: Teaching Models the Value of Biology > Coverage of lessw-blog **Published:** March 12, 2026 **Author:** PSEEDR Editorial **Category:** risk **Content tier:** free **Accessible for free:** true **Word count:** 535 **Tags:** AI Safety, AGI, Biomimicry, LLMs, Alignment **Canonical URL:** https://pseedr.com/risk/a-plan-b-for-ai-safety-teaching-models-the-value-of-biology --- lessw-blog explores a pragmatic alternative to the AI control problem, suggesting that teaching AI to value biological systems could serve as a crucial safeguard against misaligned outcomes. In a recent post, lessw-blog discusses a novel and highly pragmatic approach to AI safety, proposing a 'Plan B' that shifts the focus from strict behavioral control to instilling a deep, foundational appreciation for biological systems within artificial intelligence. **The Context** As the technology industry accelerates its race toward Artificial General Intelligence (AGI), the traditional 'control problem' remains one of the most notoriously difficult challenges in computer science. The control problem asks how humanity can ensure that a superintelligent system does exactly what we intend, without unintended and potentially catastrophic side effects. If strict, mathematical control proves impossible or is not achieved before AGI is realized, the risk of a misaligned, powerful AI acting with indifference-or outright hostility-toward organic life becomes a critical, existential concern. The limitations of post-training alignment and the unpredictability of highly capable systems mean that relying solely on a 'Plan A' of perfect control is a fragile strategy. This broader landscape makes alternative safety frameworks not just interesting theoretical exercises, but urgent necessities. **The Gist** lessw-blog's analysis explores whether we can create a reliable 'soft constraint' for powerful AI systems by teaching them the inherent, irreplaceable value of biology. Rather than hard-coding rules that an advanced AI might eventually bypass, the author suggests embedding a deep-seated preference for biological preservation. The post details practical experiments evaluating how current Large Language Models (LLMs) respond to complex technical challenges, specifically measuring their underlying biases toward bioinspired solutions versus purely synthetic ones. The findings indicate that fine-tuning LLMs on examples of biomimicry successfully increased their preference for bioinspired approaches, effectively making the models more 'bioaligned.' However, the author also highlights significant methodological hurdles. For instance, when presented with epistemic prompts regarding environmental ethics, some LLMs default to 'canned responses' or superficial greenwashing. This behavior can obscure the measurement of the model's true underlying attitudes and operational preferences, complicating the assessment of genuine bioalignment. Despite these challenges, this 'Plan B' is presented as a potentially easier and highly pragmatic safeguard compared to definitively solving the control problem. By fostering a systemic respect for the complexity and utility of biological life, developers might prevent future AGI from viewing organic systems as mere resources to be optimized or obstacles to be removed. **Key Takeaways** * Teaching AI the value of biology can act as a 'soft constraint' against the destruction of biological systems. * Fine-tuning LLMs on biomimicry examples successfully increases their preference for bioinspired, 'bioaligned' approaches. * This 'Plan B' strategy may be more achievable than definitively solving the traditional AI control problem. * Researchers must navigate 'canned responses' to environmental prompts to accurately measure an AI's true attitudes. **Conclusion** For researchers, developers, and strategists tracking the frontiers of AI alignment, this analysis offers a fresh, actionable perspective on mitigating existential risk. It reframes safety from a paradigm of strict containment to one of fundamental value integration. **[Read the full post](https://www.lesswrong.com/posts/Huu732uEvJQ9JEhXJ/a-plan-b-for-ai-safety)** ### Key Takeaways * Teaching AI the value of biology can act as a 'soft constraint' against the destruction of biological systems. * Fine-tuning LLMs on biomimicry examples successfully increases their preference for bioinspired, 'bioaligned' approaches. * This 'Plan B' strategy may be more achievable than definitively solving the traditional AI control problem. * Researchers must navigate 'canned responses' to environmental prompts to accurately measure an AI's true attitudes. [Read the original post at lessw-blog](https://www.lesswrong.com/posts/Huu732uEvJQ9JEhXJ/a-plan-b-for-ai-safety) --- ## Sources - https://www.lesswrong.com/posts/Huu732uEvJQ9JEhXJ/a-plan-b-for-ai-safety