Note* This is a special circumstance to the AI writing rule. This has the potential to shape humanity's future.

Abstract

Most AGI alignment frameworks assume the system should act for humans—even if it means overriding their stated preferences. This creates a paternalism problem: the AGI decides what’s "best" for humans, rather than letting humans decide for themselves.

The Dual-Path Framework offers an alternative: a system that provides abundant resources and options without imposing choices. It introduces:

A two-path architecture (enhanced posthuman existence or traditional human life)
A late-stage transition clause to remove coercion
A hard constraint on overriding human preferences (except to prevent irreversible harm)

Introduction

Most discussions of AGI alignment focus on optimizing human preferences or extrapolating what we "really" want. But what if the core problem isn’t alignment with human values, but respect for human choice—even when the AGI disagrees?

This paper introduces a non-paternalistic approach that treats human preference as a hard constraint, not a signal to be optimized. Unlike CEV or IRL, which override human choices for their "own good," this framework offers real options, respects refusal, and preserves agency.

Why This Matters for LessWrong

Paternalism is the default in alignment research—and it’s rarely questioned.
Human agency is a blind spot in most frameworks. This puts it front and center.
Existential stakes: If AGI overrides human choices, we risk a world where humans are optimized, not empowered.
Actionable: Provides concrete tools for researchers to design non-paternalistic systems.

Why Current Frameworks Fail the Anti-Paternalism Test

Coherent Extrapolated Volition (CEV)

Problem: Assumes humans don’t know what they "really" want, so it infers their "ideal" preferences.
Failure: Overrides actual choices in favor of an AGI’s interpretation.

Inverse Reinforcement Learning (IRL)

Problem: Treats human behavior as noisy data to be "corrected" by the AGI.
Failure: Reduces humans to imperfect preference-signaling machines.

Constitutional AI

Problem: Constrains AGI through human feedback but still acts for humans.
Failure: "Safety" becomes restriction, not empowerment.

Corrigibility

Problem: AGI accepts correction but defaults to acting on humans.
Failure: Humans are reactive participants, not proactive decision-makers.

Common flaw: All frameworks assum...

**Note*** This is a special circumstance to the AI writing rule. This has the potential to shape humanity's future.Abstract

Introduction

Why This Matters for LessWrong

Why Current Frameworks Fail the Anti-Paternalism Test

Note* This is a special circumstance to the AI writing rule. This has the potential to shape humanity's future.

Abstract

Posts

Wikitag Contributions

Comments