Sign in
← AI PulseSep 9, 2026

Deep · research · Single-source brief

Moral Self-Correction Capability in RLHF Models Emerges at 22B Parameters

Anthropic Research found that large language models trained with reinforcement learning from human feedback can avoid harmful outputs when instructed, a capability observed to emerge around 22 billion parameters.

By Illumora Editorial

Source · Sep 9, 2026, 7:28 PM · On Illumora · Sep 9, 2026, 7:38 PM

Media from the primary source — shown here so you can stay on Illumora.

Rewritten from one allowlisted primary — not independent enterprise reporting. Lanes →

Brief drafted by Illumora’s editorial model from the linked primary source. Ops desk reviews flagged pieces. How we write →

Read the source →Anthropic Research — Moral self-correction in large language models
Save

Anthropic Research investigated the capacity of large language models (LLMs) to "morally self-correct" by avoiding harmful outputs when explicitly instructed. The research provides evidence supporting this hypothesis across three distinct experiments, each exploring different aspects of this self-correction ability.

The capability for moral self-correction is observed to emerge in models at approximately 22 billion parameters. This capacity generally improves with increases in both model size and the extent of reinforcement learning from human feedback (RLHF) training.

Key Points

  • RLHF-trained models can avoid stereotyped or biased outputs when given specific instructions.
  • The capability for moral self-correction emerges at 22 billion model parameters.
  • This capability typically improves with increasing model size.
  • It also typically improves with increased RLHF training.
  • Language models at this scale can follow instructions.
  • Language models at this scale can learn complex normative concepts of harm, such as stereotyping, bias, and discrimination.

Context

According to Anthropic Research, the ability for moral self-correction in LLMs stems from two core capabilities that models acquire at a certain scale. First, they gain the capacity to follow instructions. Second, they develop the ability to learn complex normative concepts related to harm, including stereotyping, bias, and discrimination. These combined capabilities allow models to adhere to instructions aimed at preventing morally harmful outputs.

Why It Matters

This research indicates that specific training methodologies and model scales contribute to a model's ability to align with ethical guidelines. For builders, understanding the parameter threshold and the role of RLHF in enabling moral self-correction is crucial for developing more robust and ethically aligned AI systems, potentially reducing the incidence of biased or harmful outputs through instruction-based interventions.

What To Do

  • Note that moral self-correction is observed to emerge at 22 billion parameters.
  • Consider the role of RLHF training in developing models capable of adhering to ethical instructions.
  • Observe how model scale influences the ability to follow instructions and comprehend normative concepts of harm.