Training
RLAIF
Reinforcement Learning from AI Feedback — alignment training where another AI model, not a human, provides the preference signal used to fine-tune the target model.
01 ——
In plain English
RLAIF (Reinforcement Learning from AI Feedback) replaces the human-rater step of RLHF with a strong AI judge. Instead of paying humans to compare responses, the lab uses a separate model to score which response better fits the principles they want the trained model to learn.
Why labs use RLAIF:
- Scale — an AI judge can produce millions of comparisons cheaply
- Consistency — no inter-rater disagreement
- Speed — overnight runs instead of multi-week human campaigns
- Constitutional pairing — combine RLAIF with a constitution to get reproducible alignment
Trade-offs vs RLHF:
- The AI judge inherits its own biases — garbage in, garbage out
- Catastrophic blind spots if the judge is bad at the criterion
- Still typically combined with human feedback at the most important checkpoints
Where it sits in the pipeline: Most modern post-training stacks use a mix — SFT, then DPO/RLHF for the highest-stakes behaviours, then RLAIF to scale alignment across the long tail of edge cases.
02 ——
Related terms
RLHF
Reinforcement Learning from Human Feedback — the training technique that teaches AI models to be helpful, harmless, and honest.
DPO
Direct Preference Optimization — a simpler alternative to RLHF that fine-tunes a model directly on preference pairs, no separate reward model required.
Constitutional AI
An Anthropic-pioneered training method that teaches a model to critique and rewrite its own outputs against a written set of principles (a constitution).
Fine-tuning
Further training a pre-trained AI model on your own data to specialise it for a specific task or style.
Post-training
Everything done to a model after pretraining — fine-tuning, RLHF, DPO, safety training — to turn a raw base model into a usable product.
Alignment
The challenge of making AI systems behave in ways that match human values and intentions — not just their literal instructions.
Back to glossaryLast reviewed May 2026