Behind every AI system is a workforce you never see.
The language models that write fluent prose were shaped by millions of human judgments about what "good" writing looks like. The vision systems that recognize objects learned from billions of human-drawn bounding boxes. The robots that navigate warehouses trained on demonstrations from human operators. These people, the annotators, evaluators, teleoperators, and demonstrators, are the invisible foundation of the entire field. And the industry that depends on them has largely treated them as interchangeable inputs: anonymous workers cycled through platforms with no training, no growth, and no voice.
We think this is a mistake, and not just ethically. The quality of AI is inseparable from the quality of the data it learns from, and the quality of the data is inseparable from the conditions of the people who create it. That is what we mean by People-First AI: building AI data infrastructure on a foundation of human dignity, not despite what's good for the technology, but because of it.
Churn destroys exactly what hard tasks need
The dominant model optimizes this labor for cost. Workers are recruited through platforms and treated as fungible: if one person stops working tomorrow, another fills the gap, and nobody notices what left with them. That works passably for simple tasks. If you need to classify whether an image contains a cat or a dog, the criteria are clear and any attentive person produces the same result.
But the tasks that matter most for modern AI aren't like this. Evaluating whether an AI response is helpful requires judgment about context, audience, and purpose. Determining whether content is harmful involves nuanced understanding of culture, intent, and potential misuse. Demonstrating physical tasks through teleoperation demands skill that takes months to develop. Training safety-critical systems requires evaluators who understand the principles behind the guidelines, not just their literal words. These tasks require expertise, consistency, and context, and a workforce built to churn systematically destroys all three.
Retention, consistency, engagement: the mechanisms
The connection between treatment and output quality isn't abstract. It operates through specific, measurable mechanisms.
Retention. Complex AI data work benefits enormously from experience: an evaluator who has spent months learning a company's safety principles and calibrating on edge cases produces better judgments than someone seeing the task for the first time. When workers churn constantly, that expertise never accumulates. The economics are stark: when contracted security screeners at San Francisco International Airport received a wage increase from $6.45 to $10 per hour, turnover dropped from 95% to 19%, and employers reported immediate improvements in performance, morale, and service. The workers hadn't changed. The conditions had.
Consistency. If the person who labeled Monday's data interprets guidelines differently than the person who labeled Tuesday's, the model learns that sometimes the rules are X and sometimes they're Y. This shows up directly in model behavior: responses that flicker between interpretations, safety boundaries enforced unevenly. Stable teams that have reached shared agreement about what "helpful" and "harmful" mean produce judgments that form a coherent whole rather than a statistical average of divergent opinions.
Engagement. When people feel valued and see a future in their work, they notice things they weren't asked to notice, flag ambiguities instead of making arbitrary choices, and ask whether their judgment is right rather than merely defensible. Costco pays higher wages than competitors and sees 17% turnover compared to Walmart's 44%, and gets higher productivity, better service, and lower theft for it. Workers who feel like stakeholders act like stakeholders. That discretionary effort is exactly what complex AI data work requires, and exactly what an anonymous, churn-by-design model discourages.
Training amplifies whatever the workforce produces
These dynamics matter for any work. They matter more for AI, because models don't just learn the intended lessons; they learn everything the data teaches, including the inconsistencies and errors nobody consciously encoded.
Train a language model on preference data from a churning workforce and it learns a preference function that is the average of divergent judgments: a blurred compromise reflecting values no individual actually holds, producing behavior that seems erratic because the training signal was. In safety-critical annotation for autonomous vehicles, a workforce with constant churn means there are always novices, so the error rate never drops to what an experienced, stable team achieves, and those errors propagate into the model, the vehicle, the world. In teleoperation, a skilled operator produces smooth trajectories that are easy to learn from; a novice produces jerky, hesitant movements that teach the wrong lessons.
A small percentage of errors in the data becomes a much larger percentage of errors in the model. This workforce isn't just creating a product; it's shaping systems that will make millions of decisions.
What we're building
At Avala, we call our workforce "coworkers" rather than contractors. Most of them work flexibly, from wherever they are, and are paid for accepted work rather than for hours. Flexibility is not the problem. Anonymity and churn are. So we build the things the anonymous model leaves out.
We start with fair compensation, pay that is above the norm for the work and the place, because this is where a lot of "people first" rhetoric falls apart. It costs more; we accept that tradeoff because the investment pays off in quality, retention, and consistency, and because it's right regardless. We invest in training that goes beyond the task to why the work matters and how it connects to the AI systems it shapes, because context translates directly into better judgment calls. We build paths to more skilled work, from annotation to review to leading a team or training others, because the possibility of growth changes how people relate to their work and how long they stay. We keep experienced coworkers on the projects they know, so expertise accumulates on a customer's data instead of resetting with every batch. And we listen: the people doing the work see what managers and engineers miss, and their flags change our guidelines and our tools.
The research says this isn't a tradeoff
Worker welfare is usually framed as a cost you bear despite the bottom line. The research consistently shows the opposite. MIT Sloan professor Zeynep Ton, studying retail operations, found that companies with generous labor budgets had better-performing employees, more effective operations, stronger sales, and higher profits, while companies that cut labor costs ended up undermining the profits they were trying to protect. Wharton research found that every dollar increase in retail payroll could produce monthly sales increases of four to twenty-eight dollars. Even in gig contexts, Yale and UCLA researchers studying Uber drivers found that a 10% wage increase led to 20% more hours worked, a response four times larger than passengers' sensitivity to price changes.
For AI data work, these dynamics are amplified: the link between worker quality and output quality is tighter, the cost of inconsistency higher, the returns on expertise larger. Teams focus on architectures, scaling laws, and training techniques, and those matter. But underneath it all is data, and underneath the data are people. Investing in those people isn't a distraction from building good AI; it's a prerequisite.
People-First AI
We chose the name because the order matters. Put people first, treat them with dignity, invest in their growth, keep experienced people on the work they know, and better AI follows. This isn't people first despite what's good for AI. It's people first because of what's good for AI. The ethics and the effectiveness point in the same direction.
We're not claiming to have figured everything out; we'll make mistakes along the way. But being explicit about our commitments creates accountability and invites others to hold us to our own standards. If you're working on AI and thinking about where your data comes from, look closely at how the people creating it are treated, because it directly affects the quality of what you're building. The invisible workforce behind AI doesn't have to stay invisible, and it doesn't have to be treated as interchangeable. There's a better way, and it produces better results.
That's what People-First AI means to us.
If you're interested in the research on worker welfare and performance, Zeynep Ton's work at MIT Sloan and the Good Jobs Institute is an excellent starting point. And if you're thinking about these issues in your own AI data operations, we're always happy to share what we've learned.
