Physicians Are Getting Paid $100–$300/Hour to Train AI: The Complete Guide to the Medical Annotation Side Gig (2026)

The reason this market exists is simple scarcity economics. Every frontier AI lab — OpenAI, Google DeepMind, Anthropic, Microsoft, and the healthcare AI companies building on their models — needs clinical judgment to train and validate medical AI, and clinical judgment cannot be crowdsourced. As Mercor, one of the largest expert platforms, states plainly: "There aren't many board-certified physicians available to rate AI-generated clinical summaries, and the companies that need them know it." A generalist doing basic data labeling earns $12 to $25 per hour. A physician doing expert medical evaluation earns five to ten times that — for the same reason your clinical time has always been expensive: the credential is scarce, the judgment took fifteen years to build, and the downstream cost of getting medical AI wrong is enormous.

This guide extends our Pharma and Biotech Consulting guide into territory no physician finance site has covered — and it is written to be the reference document for physicians entering this market.

What the Work Actually Is: The Four Task Types

"Training AI" is a vague label covering four distinct kinds of work, and understanding the differences matters because they pay differently, demand different skills, and vary in how engaging physicians find them.

1. Evaluation case writing (prompt authoring). You write realistic clinical scenarios in your specialty — the questions, cases, and edge situations a medical AI should be tested against. A nephrologist might construct a case involving drug dosing in stage 4 CKD with a subtle interaction; a pediatrician might write triage scenarios distinguishing benign from can't-miss presentations. This is the work the SF Standard profile describes: writing prompts about medical situations in your specialty, then grading the model's responses. It rewards the same skill as writing good board-exam questions, and physicians with question-writing or medical education backgrounds tend to excel and clear quality bars quickly.
2. Response grading (RLHF and rubric-based evaluation). The model answers medical questions; you score the answers — for factual accuracy, clinical appropriateness, safety, completeness, and adherence to detailed rubrics — or rank multiple candidate responses against each other. This is the highest-volume task category. It is structured, repetitive, and asynchronous — which is precisely why it fits between cases, on a post-call afternoon, or during a slow locums stretch.
3. Red-teaming clinical AI. You deliberately try to make the model fail: eliciting dangerous dosing advice, testing whether it can be manipulated into inappropriate recommendations, probing for bias across patient demographics, finding the prompts where a confident-sounding answer is clinically wrong. This is genuinely skilled adversarial work, and it is not a fringe activity — peer-reviewed work in medicine notes that red-teaming is a recognized and now federally mandated practice in AI development, that it remains underdeveloped in healthcare specifically, and that clinician evaluators independent of the model's creators are essential to minimize conflict of interest. Translation: the field has formally concluded it needs external physicians to do exactly this, and the labs are paying accordingly.
4. Benchmark and dataset development. Longer-form projects building the evaluation sets against which medical models are publicly measured — curating question banks, writing gold-standard answers, adjudicating disagreements between other reviewers. These projects tend to run weeks rather than hours, sometimes with minimum weekly commitments (a representative Mercor project posting specifies 15 hours per week for 3–4 weeks), and function more like short consulting engagements than gig tasks.

What It Actually Pays in 2026

The verified rate landscape, from most conservative to most aggressive:

What moves you up the range: board certification (verified during onboarding — the platforms check), subspecialty scarcity (the same supply-demand logic that drives the specialty premiums in our expert network coverage — a general internist is more replaceable than an interventional cardiologist or pediatric geneticist), demonstrated quality (platforms track acceptance rates and route better-paying projects to higher-scoring contributors), and platform selection — which deserves its own emphasis: the single fastest way to raise your rate is to move from generalist platforms to expert-focused ones. Mercor's own guidance says it directly: a credentialed expert earns more on a platform built for experts than competing in a general queue.

The realistic earnings math:

• 5 hours/week at $150/hr: $3,000/month — $36,000/year
• 10 hours/week at $150/hr: $6,000/month — $72,000/year
• 10 hours/week at $250/hr (in-demand subspecialist): $10,000/month — $120,000/year

Now the honest volatility caveat that no platform's marketing page leads with: this is project-wave work, not salaried work. Platforms including Outlier are known for droughts between projects; projects pause or cancel without notice; and steady 10-hour weeks require either a hot project or presence on two to three platforms simultaneously. Treat the annualized figures as what sustained engagement produces — not a guarantee any single platform will deliver. The experienced-contributor playbook is consistent: diversify across 2–3 platforms, maintain high quality scores, and treat this as variable income in your planning, exactly as our Physician Side Income guide treats every 1099 stream.

The Platforms: Who's Actually Hiring Physicians

Getting accepted: expect credential verification, an unpaid or modestly paid assessment (know your time risk before starting — some platforms' assessments run 1–3 hours with acceptance rates as low as 10–20 percent on generalist tracks, though physician acceptance on expert tracks is meaningfully higher because the bottleneck is credentials, not writing samples), and a ramp period where early quality scores determine your project flow. Apply to two or three platforms in parallel rather than sequentially — the onboarding lag is the longest part of the pipeline.

Scam awareness, briefly but seriously: the growth of this market has spawned impersonation scams. No legitimate platform ever asks for payment, banking passwords, or "registration fees" during application. Apply only through official URLs, never through links in unsolicited messages.

The Compliance Section: Five Landmines, All Avoidable

This is the section that separates doing this side gig well from doing it in a way that jeopardizes your primary job — and it is the section no platform's recruiting page will walk you through.

1. Your employment contract's outside-activities clause. As covered throughout our Physician Contract Red Flags guide, many physician employment agreements contain language along these lines: "Physician shall devote full professional time and attention to Employer and shall not engage in any other professional or commercial activities, paid or unpaid, without prior written consent." The trap physicians fall into is assuming "this is nonclinical, so they won't care." As one contract-review resource puts it bluntly: they might care a lot if the activity uses your medical expertise — and AI evaluation work is, definitionally, the commercial use of your medical expertise. Read your clause. If it requires notice or consent, give notice and get consent in writing before onboarding, not after. For most physicians this is a five-minute email to a medical director; skipping it converts a compliant side gig into a contract breach.
2. Institutional conflict-of-interest policies. Academic physicians face a second layer: institutional COI disclosure requirements and outside-activity day caps, exactly as described in our pharma consulting guide. AI training work is new enough that many COI offices have no category for it — disclose it anyway, described plainly ("compensated evaluation of medical AI outputs for [platform]"), and let the office classify it. An undisclosed activity discovered later is a problem; a disclosed one almost never is. One additional academic-specific wrinkle: if your institution is itself developing or licensing clinical AI, evaluating a competitor lab's models could raise a genuine conflict — surface it proactively.
3. PHI: the absolute rule. Never — under any circumstance, in any prompt, case, or example — use real patient information. Every clinical scenario you write must be synthetic: composite, invented, de-identified beyond recognition. Pasting real case details into an AI training platform is a HIPAA disclosure to a third party with no BAA, full stop. This should be obvious, and the platforms prohibit it too, but the failure mode is subtle: a physician writing a "realistic" evaluation case reflexively reaches for the memorable patient from last month. Build cases from textbook physiology plus invention, never from your panel. Relatedly, never input your employer's confidential information — internal protocols, order sets, pricing, proprietary pathways — into platform tasks; confidentiality obligations written before generative AI still cover prompt inputs, and "no one told me typing it into a platform counted as disclosure" is not a defense your employer's counsel will accept.
4. Platform NDAs and monitoring software. You will sign an NDA — typically covering which lab you're working for, which models you're evaluating, and everything you see in the tasks. Honor it the way you'd honor any consulting confidentiality obligation, including in physician Facebook groups where "which platform pays what" threads get specific. Separately, know before onboarding that several platforms install productivity-monitoring software — Mercor uses Insightful, Invisible uses Hubstaff — that tracks activity during logged task time. The universal advice from experienced contributors: use a dedicated device, or at minimum a dedicated user profile, and never a hospital-issued laptop, which likely violates your institution's device policy on its own and puts monitoring software on a machine that touches clinical systems. A $400 laptop is a deductible business expense (see the tax section) and cleanly separates the two worlds.
5. The scope boundary. This work involves no patients, no clinical care, no doctor-patient relationship — which is exactly why no malpractice coverage question arises and why it's cleaner than moonlighting. Keep it that way: if any platform task ever drifts toward reviewing real, identified patient cases for care decisions, or anything resembling asynchronous telehealth, that is a different activity with licensure and liability implications, and it's the point at which you stop and reassess. Evaluation of synthetic cases and model outputs sits safely outside clinical practice; keep the boundary bright.

The Tax Treatment: Why This Income Is Cheaper Than You Think (For Attendings)

Every dollar from these platforms arrives as 1099 self-employment income — no withholding, reported on Schedule C, with the full mechanics covered in our Locum Tenens Tax guide. Three points deserve specific emphasis here.

• The self-employment tax quirk that favors attendings. Self-employment tax is nominally 15.3 percent — but the 12.4 percent Social Security portion applies only up to the annual wage base ($184,500 in 2026), and W-2 wages count against that base first. An attending whose W-2 salary already exceeds $184,500 — which is nearly every attending — pays no Social Security tax at all on AI-training income, only the 2.9 percent Medicare portion (plus the 0.9 percent additional Medicare tax above $200K single / $250K MFJ). Practical translation: for a typical attending, this income carries roughly a 3.8 percent payroll-tax load instead of 15.3 percent — making it structurally cheaper than the same dollars earned by a resident, whose $65K salary leaves the full 12.4 percent Social Security portion applying to every side-gig dollar. Residents can absolutely do this work (with program disclosure — see the FAQ), but should budget the heavier tax load.
• The Solo 401(k) unlock. Genuine self-employment income makes you eligible to open a Solo 401(k) — and for a physician already maxing an employer 403(b), the Solo 401(k)'s employer contribution space (roughly 20 percent of net self-employment earnings) is additional tax-advantaged room that exists only because this side income exists. A physician earning $50,000/year from AI platforms can shelter roughly $10,000 of it — and a custom Solo 401(k) document opens the Mega Backdoor Roth mechanics on top. Pay quarterly estimates on the rest; the underpayment penalty math is unforgiving at physician marginal rates.
• Deductions that are actually legitimate here: the dedicated device recommended above, a proportionate share of home internet, a home office if a space is used regularly and exclusively for the work, and platform-related professional expenses. Keep it honest and documented — the amounts are modest, but they're real, and the dedicated-device deduction neatly funds the compliance best practice.

How This Compares to Every Other Physician Side Income

The distinctive position AI work occupies: expert-network-adjacent rates with survey-level scheduling flexibility and zero clinical liability — and unlike pharma engagements, none of it appears in the Open Payments database, because the payers are AI labs, not drug or device manufacturers. The tradeoffs are the volatility already described and the repetitive nature of high-volume grading work, which some physicians find meditative and others find deadening — the SF Standard's profiled internist landed on the first side, describing it as genuine breathing room from hospital work. One underrated non-financial return: physicians doing this work are developing hands-on fluency with exactly the AI systems arriving in their EHRs — the systems whose compensation implications we covered in Who Gets Paid for the AI Scribe? — and that fluency is career capital in a decade where it will be scarce among clinicians.

Frequently Asked Questions