Generative AI

RLHF Isn't About Scale Anymore. It's About Who You Can Get to Do It.

June 22, 2026 • By Dr. Elena Rostova • 6 min read

For the last few years, the story of RLHF (Reinforcement Learning from Human Feedback) was largely a story about scale: how many human raters can you recruit, how fast can you process feedback, how cheaply can you label preference data. That story is changing. Foundation models have become good enough to handle routine pre-labeling and first-pass evaluation on their own. What's left for human involvement — and what's becoming genuinely valuable — is judgment on the cases models can't confidently resolve themselves: edge cases, subjective calls, and regulated domains where a wrong answer carries real consequences.

This is a meaningful shift in what "data services" actually means for anyone providing RLHF and human feedback work, and it changes who is valuable to have on a project.

From "How Many Raters" to "Which Raters"

The old RLHF operating model treated human raters as largely interchangeable at scale — recruit enough people, give them clear rating guidelines, and throughput determined how fast you could generate preference data. That model still works for genuinely routine tasks. But as models improve at handling routine judgment calls themselves, the tasks still requiring human input are, almost by definition, the harder ones: ambiguous ethical tradeoffs, nuanced factual disputes, domain-specific correctness questions in medicine, law, or finance, and subjective quality judgments where reasonable experts might disagree.

For this kind of work, rater interchangeability breaks down. A generalist rater without domain background can't reliably evaluate whether a model's medical explanation is subtly wrong, or whether a legal summary omits something that matters. The value has shifted from "can we get enough raters" to "can we get the *right* raters" — people with genuine subject-matter credentials, not just training on a rating rubric.

Where This Shows Up Most

A few domains illustrate this shift clearly:

Regulated and safety-critical domains. Medical, legal, and financial content evaluation increasingly requires raters with actual credentials or professional background in the field, not general-purpose crowdworkers following a rubric. Getting this wrong doesn't just produce a worse model — it can produce a model that confidently gives incorrect specialized advice, which is a much bigger problem than an obviously wrong general-knowledge answer.

Frontier model alignment and evaluation. As models get more capable, distinguishing a genuinely better response from a worse one on complex, open-ended tasks increasingly requires evaluators who can actually assess the quality of complex reasoning, code, or analysis — not just check surface-level correctness.

Red-teaming and safety evaluation. Finding genuinely novel ways a model could produce harmful or unsafe output increasingly rewards creative, adversarially-minded specialists over large pools of generalist testers running through a fixed checklist.

Culturally and linguistically specific evaluation. Judging whether a model's output is appropriate, accurate, and natural-sounding for a specific dialect, region, or cultural context requires evaluators from that specific background — not a generalist rater working from a translated rubric.

What This Means for the Economics of RLHF Work

This shift changes the underlying cost and value structure of RLHF services in a few specific ways:

Volume-based pricing models are giving way to expertise-based pricing. When the value proposition was "a large pool of raters providing throughput," pricing naturally centered on volume. When the value proposition is "access to credentialed specialists who can make judgment calls generalists can't," pricing reasonably shifts toward reflecting that specialized expertise, similar to how expert consulting is priced differently than generic labor.

Recruitment and vetting become the actual differentiator. The hard part of this work is no longer managing large-scale rater logistics — it's building and maintaining a genuine network of credentialed, verified domain experts across the specialties clients need, and having a reliable way to route the right task to the right expert quickly.

Smaller, higher-quality panels replace large generalist pools for many tasks. A project might need a panel of twenty genuinely qualified medical evaluators rather than two thousand generalist raters — a very different resourcing and management challenge than traditional crowdsourced labeling.

What Foundation Models Still Can't Do Here

It's worth being precise about the boundary, since it's easy to overstate how much models can now handle independently. Foundation models are increasingly capable of:

They remain much weaker at:

This is precisely the boundary where human expert involvement remains not just useful but necessary, and it's the boundary that's shrinking the *volume* of human RLHF work needed while raising the *bar* for the humans still involved.

What This Means for Data Services Providers

For an organization providing human feedback and RLHF services, this trend suggests a clear strategic direction:

The Bottom Line

RLHF hasn't become less important as models improve — the work has become more concentrated on genuinely hard judgment calls, and the people needed to make those calls well have become more specialized and harder to substitute. Organizations still operating on a "recruit as many raters as possible" model are optimizing for a version of this work that's shrinking. The providers building genuine domain-expert networks, with reliable ways to match the right specialist to the right task, are positioning themselves for where this work is actually heading.