What "AI training" for physicians actually means

The phrase sounds misleading at first. You're not tutoring anyone. You're not explaining medicine to a model the way you might to a medical student. Instead, you're acting as a domain expert that an AI lab can point at its outputs and ask: is this right? is this safe? what's missing?

In practice, the work looks like a mix of:

  • Evaluating medical accuracy. Reading an AI-generated differential, treatment plan, or imaging interpretation and grading it against what a competent physician would produce.
  • Ranking model outputs. Two or three responses to the same clinical prompt; you pick the best and say why. This preference data is what "reinforcement learning from human feedback" (RLHF) actually runs on.
  • Writing evaluation rubrics. Defining what "good" looks like for a given clinical task, in enough detail that the model can be scored against it at scale.
  • Identifying failure modes. Finding the places where a model hallucinates a drug interaction, invents a guideline, or misses a safety-critical step.
  • Structured feedback. Short written rationales that explain why an answer is wrong or incomplete, not just that it is.

None of it involves patient care. Nothing touches a chart with a real name on it. There's no licensure risk in the way clinical moonlighting carries it. It is, functionally, consulting work that happens to be paid by the hour.

The platforms, ranked by what physicians actually report

Pay ranges vary wildly by specialty, platform, and project. The numbers below are ranges that have been consistently reported by physicians working on each platform as of early 2026. Treat them as guidance, not guarantees.

Mercor

$130–$400/hr

Probably the most commonly cited platform in this space. Entry-level clinical roles land in the $130–$175/hr range; specialty-specific radiology work has been reported up to $400/hr; experienced annotators report $225–$275+/hr. Physician sentiment is strongly positive overall — predictable onboarding, legitimate projects, consistent payment.

Handshake AI

$170–$250/hr

Tighter range, cleaner onboarding. Physicians describe Handshake as the most "professional feeling" of the bunch — clear expectations, straightforward scopes, responsive coordinators. A good first platform if you want the work to feel like a real consulting engagement.

Outlier

$50–$120/hr + bonuses

Base pay runs lower than peers, but bonus periods can push effective hourly rates meaningfully higher. Best for physicians who are efficient, can hit throughput targets, and are comfortable with variable income. Not the first platform to try, but a useful second one for filling gaps.

Medcase

$140+/hr

Competitive with Mercor, sometimes slightly above. Reports include $140/hr for family medicine projects. Physicians flag it for clean onboarding and clear expectations — a good candidate if Mercor's queue is quiet.

Micro1

Highly variable

The outlier in both directions. Historical highs near $200/hr; recent reports of projects posted under $50/hr. Worth watching — but go in with eyes open, and read the rate before accepting any specific engagement.

What you should actually expect

  • Work is episodic. You don't clock in. Projects come in waves; some weeks you'll have hours of available work queued up, other weeks nothing. Plan around that.
  • Efficiency equals earnings. Most platforms pay hourly, but cap projects at a budget. Physicians who produce cleaner, faster, more structured feedback get more work routed to them.
  • Quality audits are standard. Your outputs are spot-checked. Consistent low scores mean fewer assignments; consistent high scores often unlock higher-paying tracks.
  • Taxes are on you. This is 1099 contractor work. Quarterly estimated taxes, separate bookkeeping, deductions for a home office if applicable. Budget for it upfront.
  • No patient care, no prescribing, no licensure touch points. Which means for most employed physicians, this work sits comfortably outside the scope of non-compete and moonlighting clauses. Check your specific contract, but it's usually cleaner than clinical moonlighting.

Is this training your replacement?

The honest answer: no, and the question itself slightly misreads what's happening. Models are getting better at well-bounded tasks. They are nowhere close to doing what a physician does in a clinic — the interleaved reasoning across history, exam, imaging, labs, prior context, patient preference, and real-time decision making under uncertainty.

The people shaping how AI shows up in medicine over the next five years are, overwhelmingly, the physicians willing to engage with it now. Teaching a model what "unsafe" looks like is how the safety rails get built. Sitting it out just means they get built by people with less clinical expertise than you.

How to start

  1. Pick one platform to begin with. Handshake AI or Mercor are the two most-recommended starting points.
  2. Finish onboarding end-to-end before trying a second platform — each has its own style guide and quality bar.
  3. Block a consistent weekly window (even 2–4 hours) and treat it like moonlighting. Consistency is what gets you routed higher-paying projects.
  4. Track your effective hourly rate honestly — including unpaid training time — for the first month. Drop platforms that don't clear your threshold.

If you'd rather skip the account-by-account hunt, Rounds lets you build one physician-verified profile, set your own rate, and get matched to consulting work across these platforms. Build your Rounds profile →