Do doctors and lawyers become AI contractors? Is a hourly wage of $400 considered high pay or a high exploitation of knowledge?

Opening Dialogue: High-知izens Enter the AI Gig Economy

Shiwen: This week, Bloomberg Law reported that some doctors and lawyers in the US are starting a new gig providing training data for AI models, with pay rates reaching up to $400 per hour. This sounds counterintuitive.

Yongliang: The key insight is that the AI training data industry is becoming more hierarchical, and this process is happening rapidly. On one side, there are low-paid generalist annotators, while on the other, high-paid experts are emerging. Bloomberg Law’s report on October 16 exposed this phenomenon, with the headline reading “Doctors and Lawyers Are Joining AI’s $400-an-Hour Gig Economy.” LinkedIn also listed data annotators as one of the fastest-growing jobs in the US this year. My judgment is that model companies are using a gig economy approach to extract professionals’ tacit knowledge for use in models. I will explain these points individually below.

Shiwen: $400 per hour sounds high, but some argue it represents “knowledge exploitation,” as they see it, which doesn’t fit the definition of “knowledge monetization.”

Yongliang: The real issue is that this controversy is more interesting than the numbers themselves. It combines two questions: how much value professional time actually has, and whether what they sell is knowledge itself or a license to reproduce knowledge. I will share my perspective as the technical director for 17 years, with experience managing teams of over 70 people at peak times. I won’t give a standard answer; instead, I will break down several key contradictions.

Shiwen: Let’s start with “what’s happening?”

Q1: Who is involved in AI training gigs now?

Shiwen: Let’s align the facts first. What does a high-知izens gig provider for AI training data look like?

Yongliang: Bloomberg’s report provided some numbers. Civil rights lawyers can earn up to $400 per hour; condensed matter physics experts around $170; software engineers around $120. The gig platform Mercor reported an average hourly pay of $110. These figures come from the platforms or reports themselves, not third-party audits, so we need to judge them carefully. Mercor did not disclose the coverage rate, so whether “average $110” represents most cases is uncertain.

Shiwen: This is very different from what traditional perception defines as “data annotation work.”

Yongliang: The industry is using two terms to distinguish between these types of people. Low-paid workers are called annotators, who perform generalist judgment; high-paid workers are called experts, including doctors, lawyers, engineers, mathematicians, accountants, and scientists. The report says these experts are “cherry on top,” meaning that model companies know most the hard and tiring work is done by low-paid workers, but the final “taste” depends on professionals. This creates a structure with two large ends and a smaller middle.

Shiwen: What do they specifically do?

Yongliang: From the information disclosed in the report, they mainly provide professional domain judgment, proofreading, case annotation, and help models understand complex scenarios’ edge cases. For example, if a medical case is not accurately annotated by the machine, doctors are called in to judge; if a contract’s legal consequences are unclear, lawyers are asked to annotate it. The essence of this is to turn the judgment accumulated by professionals over the past decade or two into samples that models can learn from.

Q2: Is a hourly wage of $400 expensive?

Shiwen: At $400 per hour, it’s almost over two thousand yuan in RMB. For many people, this is considered “sky-high gigs.”

Yongliang: My core judgment is that if we only look at the hourly rate, it can be misleading. $400 is the peak, not the norm. According to reports, civil rights lawyers earn at this rate, physical experts $170, software engineers $120, and Mercor averages $110. The large gap between these numbers shows that professional barriers and market scarcity determine prices. But I must add: these figures are from platform descriptions or reported sources, not my statistics. This is my estimation, not data.

Shiwen: So from a company’s perspective, is $400 expensive?

Yongliang: From the perspective of model companies, it might not be. The cost of training a large model can be in the millions of dollars. A single incorrect case can cost far more than a few hundred dollars. Hiring a senior lawyer to annotate a case for a few hours is more reliable than having a group of ordinary annotators guess for a year. But on the other hand, for professionals, the income structure isn’t stable: gig platforms can shut down accounts at any time, and task volume is unpredictable. Reports mention that many low-paid annotators face account suspension and unstable income. Expert gigs have high income ceilings, but they don’t guarantee long-term stability.

Shiwen: So high pay doesn’t equal stability.

Yongliang: My core judgment is that high pay masks the instability of labor relations. This is similar to the old problem in platform economy: attracting people with high unit prices without taking employment responsibilities. For doctors, lawyers, and engineers who traditionally rely on stable professional identities, this is actually a degradation of their status. They sell not just professional opinions, but judgment samples that can be infinitely replicated.

Q3: Is this knowledge monetization, or is it exploitation?

Shiwen: So returning to the initial controversy, is this knowledge monetization, or is it exploitation?

Yongliang: My core judgment is that both statements are true, but they occur at different stages. Knowledge monetization happens at the moment of transaction: I spend time, you pay hourly wages, and both parties are voluntary. But knowledge exploitation occurs after the transaction: the model solidifies my judgments into parameters that can be called infinitely, and what I get is just a one-time fee. The problem is that the value after “solidification” is completely disproportionate to the compensation I receive.

Shiwen: Is it possible for professionals to also gain extra benefits, like fame or connections?

Yongliang: Reports don’t mention authorship, reputation, or long-term cooperation. So the value distribution is clear: professionals sell their one-time labor, while platforms and large model companies buy scalable knowledge assets. Under this structure, “monetization” seems more like a one-time transaction, while “exploitation” seems like a transfer of future income rights.

Shiwen: Does that mean all professionals shouldn’t do this?

Yongliang: My core judgment is that we can’t simply draw conclusions. For some people, spending spare time doing a few hours of professional annotation can earn considerable money; for others, it’s like training a system that might replace them in the future. Personal choices may vary, but platforms cannot模糊 their responsibilities. At least they should make participants aware of what models data will be used for and whether it will be retrained. Transparency is the baseline.

Q4: How should companies evaluate this AI training data procurement?

Shiwen: You have managed technical teams before. From a corporate perspective, how should we evaluate the procurement of such expert freelancer data?

Yongliang: The key judgment is that the most common mistake companies make is treating expert data as “more expensive outsourcing.” Over 17 years of technical management, I’ve seen many teams only consider the unit price and not the total cost. The real value of expert data is helping you identify the boundaries of models in specific scenarios. If you simply ask experts to fill in answers according to your templates, it’s no different from hiring low-paid workers; but if you can get experts to point out that “the machine misunderstood this part” or “there’s no standard answer for this case,” that’s truly valuable data.

Shiwen: How do you proceed?

Yongliang: I require teams to do three things. First, clarify the specific problem you want to solve with expert data, not just say “improve quality” in general terms. Second, design a feedback loop so that experts’ judgments are not just samples but a continuous basis for calibrating models. Third, record the decision-making process of experts, not just the final labels. Labels are static; the decision-making process is what models really need to learn. This is my experience, not statistics.

Shiwen: Are there any pitfalls?

Yongliang: The biggest pitfall is treating experts as “human correct answers.” Experts make mistakes, and there can be disagreements among them. If teams lack evaluation of expert consistency, it’s easy to solidify expert biases into models. I once established an internal rule: Any project that relies on external expert judgments must include at least two independent sources for cross-validation. This rule sounds cumbersome, but it prevents one person’s opinion from being taken as industry truth.

Q5: What does this mean for ordinary professionals?

Shiwen: Do people who aren’t doctors or lawyers need to care about this too?

Yongliang: The key judgment is that they do. “High-knowledge freelancers” is just one specific phenomenon; the essence of this is that the AI industry chain is breaking down more and more human knowledge into tradable data units. A doctor’s diagnostic logic, a lawyer’s contract judgment, an engineer’s debugging intuition—today they’re broken down, and tomorrow they might become your industry experience. LinkedIn listing data annotators as the fourth fastest-growing job shows this trend is spreading.

Shiwen: What lessons does this offer ordinary people?

Yongliang: There are three levels. First, any professional knowledge that can be broken down into clear judgment criteria becomes a candidate fuel for models. This means “how much you know” is no longer the key; what matters is “where the judgment criteria are vague and where they need to be weighed.” Second, high-priced freelancers seem attractive, but they also train a system that can replicate your core abilities. Third, companies will increasingly need people who understand both business and AI training logic to act as “knowledge translators” between business and models.

Shiwen: So do you see this as an opportunity or a risk?

Yongliang: The key judgment is that it’s an opportunity for individuals but a risk for the industry. The opportunity lies in how professionals can monetize knowledge in a way not seen before; the risk is that this monetization may be irreversible. Once your judgments are learned by models, you’re no longer “the only person who knows this” but “one of the first people to teach models.” This identity shift is much more important than hourly numbers.

Epilogue: Maps Do Not Equal Territories

Shiwen: A final note.

Yongliang: In one sentence, AI training gig workers are turning the judgment of high-skilled professionals into data. A hourly wage of $400 might seem expensive, but the real cost is whether you understand what you’ve sold. Professionals’ time can be bought out, but when knowledge is absorbed by models, it’s no longer time that’s traded—it’s the value that should belong to you in countless future calls.

Shiwen: In one sentence, summarize this issue?

Yongliang: In the AI gig economy, a $400 hourly wage isn’t the most expensive. What’s truly costly is not yet realizing what you’ve sold.

Shiwen: I’d like to share this with everyone. See you next time.

[Technical Depth]

Yongliang: This point deserves more discussion. The AI training data market is rapidly becoming segmented. At the bottom are large amounts of low-cost general knowledge annotation, primarily done by global gig workers. In the middle layer are “semi-professional annotations” for specific tasks that require certain judgment skills. The top tier is what this report focuses on: “expert annotations” or “expert feedback.”

Yongliang: This segmentation corresponds to two stages of model training. The pre-training phase uses a large amount of publicly available text, with relatively low quality requirements for individual samples. In the post-training or fine-tuning phase, high-quality, difficult samples become crucial, and the value of professionals is amplified here.

Yongliang: There’s another overlooked technical point: the quality of expert data depends not only on the expert themselves but also on task design and acceptance criteria. If the task is too narrow, experts can only choose between “right or wrong,” and their judgment won’t be maximized. If the task is open enough to allow experts to explain why a case is ambiguous or where the boundaries lie, then the collected data can truly improve the model’s robustness. This is why I always emphasize that when companies purchase expert data, what matters is the “decision process,” not just “final labels.”

Yongliang: Finally, with the development of multimodal and agent technologies, the demand for expert data will become more specialized and frequent. If professionals use gig work as a side job, there’s no problem. But if it’s their main career transition path, they need to consider whether the judgment they accumulate is becoming increasingly scarce or being standardized and replicable. This judgment is more important than any hourly wage offered by any platform.

👇 Do you think $400 AI training gig workers offer more opportunities or greater risks? Share your thoughts in the comments section. We’ll discuss some ideas next time. Follow me for honest workplace insights in the AI era, updated weekly.