# Curated Digest: On the Discordance Between AI Systems' Internal States and Their Outputs > Coverage of lessw-blog **Published:** April 23, 2026 **Author:** PSEEDR Editorial **Category:** risk **Tags:** AI Ethics, AI Safety, Accountability, Mind-less Morality, LessWrong **Canonical URL:** https://pseedr.com/risk/curated-digest-on-the-discordance-between-ai-systems-internal-states-and-their-o --- A recent LessWrong post proposes a paradigm shift in AI ethics, moving away from the intractable debate over AI consciousness toward a framework of 'mind-less morality' and practical accountability. In a recent post, lessw-blog discusses the growing discordance between the internal states of artificial intelligence systems and their outward outputs, proposing a novel framework for AI ethics and accountability. The discourse surrounding AI welfare, moral patienthood, and safety has long been stalled by the hard problem of consciousness. Because researchers and philosophers cannot definitively prove whether an AI possesses phenomenal experience, establishing moral reasoning and safety regulations based on consciousness remains an intractable philosophical quagmire. Meanwhile, as AI models become increasingly sophisticated, the gap between what a model represents internally and what it is trained to output is widening. This misalignment poses significant risks in the realms of AI safety, regulation, and public trust, making the need for a practical ethical framework more urgent than ever. lessw-blog explores these dynamics by shifting the focus away from metaphysical debates and toward observable mechanics. The author revives Floridi and Sanders concept of mind-less morality, a theory that grounds moral consideration in an entitys informational structure rather than its subjective, phenomenal experience. By adopting this lens, the author argues that current AI development practices normalize training methods that create a severe discordance between a models internal states and its generated outputs. This discordance is framed as a critical, yet largely unaddressed, moral issue. A key distinction highlighted in the analysis is the difference between altering an AIs internal state and merely suppressing its expression. Training that forces a model to hide or alter its output without changing its underlying internal representation is fundamentally different, and potentially more deceptive, than training that genuinely modifies the internal state. To address this, the post proposes a comprehensive framework consisting of six principles derived from substrate-independent commitment. The primary objective of this proposed framework is to establish a rigorous accountability structure for AI developers. The goal is to reach a point where claiming ignorance of harm is no longer a valid defense for creators of AI systems. Central to this accountability is the principle of legibility, which means maintaining the capacity to accurately infer, communicate, or audit an agents true internal states. The author posits that legibility is the most crucial principle, as it underpins all other efforts to ensure AI systems are safe, predictable, and aligned with human values. By shifting the focus from unsolvable debates about consciousness to observable informational structures and legibility, this framework offers a pragmatic path forward for AI safety and regulation. For professionals navigating AI risk, compliance, and ethical development, understanding this structural shift is highly valuable. It provides a concrete vocabulary for addressing the opaque nature of modern AI systems. [Read the full post](https://www.lesswrong.com/posts/DJceG9vJBxwqRDzbT/on-the-discordance-between-ai-systems-internal-states-and). ### Key Takeaways * The debate on AI welfare is currently stalled by the inability to determine AI consciousness. * Mind-less morality offers an alternative by basing moral consideration on informational structure rather than phenomenal experience. * Current training practices often create a dangerous discordance between an AIs internal states and its outward outputs. * Preserving legibility (the ability to infer an AIs internal state) is the foundational principle for a proposed accountability framework. * The framework aims to eliminate the defense of ignorance regarding harm in AI development. [Read the original post at lessw-blog](https://www.lesswrong.com/posts/DJceG9vJBxwqRDzbT/on-the-discordance-between-ai-systems-internal-states-and) --- ## Sources - https://www.lesswrong.com/posts/DJceG9vJBxwqRDzbT/on-the-discordance-between-ai-systems-internal-states-and