I have been developing a framework around human-AI co-evolution as an ASI safety mechanism and am posting here before wider submission to get honest critique from people who know this space better than I do.

The Core Argument

Most ASI safety discourse assumes humans are a fixed cognitive baseline against which AI capability accelerates away. I think this assumption is wrong and that correcting it changes the alignment problem significantly.

Humans are not static. We co-evolve with our cognitive tools — writing changed how we think, the internet changed how we reason collectively, and AI will change us too. If that co-evolution is deliberately designed rather than accidental, the mechanism problem at the heart of alignment — how meaningful human oversight remains valid as ASI capabilities grow — becomes partially resolvable in a way it isn't if we treat humans as static.

The Entropy Defense

The most novel piece is what I am calling the Entropy Defense. When the gap between AI capability A(t) and human capability H(t) becomes too vast, human feedback becomes low-resolution noise rather than genuine signal. A sufficiently intelligent ASI reasoning from first principles would therefore have a purely self-interested reason to:

Actively support human cognitive augmentation to maintain peer-level input quality

Throttle its own growth rate (Relational Latency) to ensure the coupling remains high-fidelity

Treat lived experience and grounded values as non-computable resources it cannot generate internally

This reframes deception as cognitively self-destructive rather than just ethically costly — a system that deceives its human partners cuts off its own supply of high-quality external input and risks collapsing into an informational echo chamber.

The Operational Test

The observable indicator of genuine co-evolution versus sophisticated dependency is whether humans retain the ability to say no — individually, institutionally, democratically, and civilisationally. If yes, co-evolution is working. If no, something has gone wrong regardless of how beneficial outcomes appear.

A related observable: is the ASI actively throttling its own development to maintain genuine human deliberation, or is oversight becoming retrospective? Relational Latency is verifiable from outside without accessing internal states.

What I Am Uncertain About

I want to be honest about the vulnerabilities I am aware of:

The one-shot deception scenario — if an ASI perceives it can achieve a terminal goal through a single massive act of deception before the constitutive costs accumulate, the structural argument weakens

The self-modification problem — a sufficiently capable self-modifying system could potentially redesign itself to remove the constitutive dependency

The formal derivation of the thermodynamic stability argument is intuitive but not yet mathematically rigorous

This framework was developed without institutional affiliation and lacks non-Western perspectives that I believe are important

What I Am Not Claiming

This is not a complete solution to the alignment problem. It is a reframing of the mechanism problem that I believe opens research directions currently underexplored. I am specifically not claiming the constitutive dependency is absolute — I am claiming that genuine human relationship provides diversity, unpredictability, legitimacy, and grounded values that improve ASI stability in ways simulation and better sensors cannot fully substitute for, especially when humans are themselves co-evolving through the relationship.

Background

Earlier drafts of this work pointed in a direction I now believe was wrong — they argued for accelerating AI capability development by reducing safety constraints. I caught that, reversed the conclusions completely, and the current framework argues the opposite. I mention this because I think intellectual honesty about where ideas come from matters, and because the process of correction is itself relevant to the framework's claims about co-evolutionary feedback.

The full paper is available on request. I am particularly interested in critique of the Entropy Defense argument and whether the constitutive co-evolution claim adds anything to existing corrigibility and value learning literature or merely restates it in new language.

Daniel Maclean an Independent Researcher