Ghost Mail: The Gig Economy Job Where You Answer Emails Nobody Actually Sent

There's a particular kind of loneliness in answering a message from someone who doesn't exist. Not the loneliness of being ignored — the other kind. The kind where you're fully engaged, genuinely thinking through a response, maybe even feeling a little empathy for the person on the other end, and then you remember: there is no person. There never was. The whole thing was generated by a language model at 3am on a server farm in Virginia, and your job is to tell the machine whether it did a good job pretending to be human.

Welcome to one of the stranger corners of the modern gig economy — AI training annotation, specifically the slice of it dedicated to synthetic communication. It's a real job. Thousands of people are doing it right now through platforms like Scale AI, Remotasks, and a constellation of smaller contractors. And it raises questions that go way beyond whether the pay is fair.

The Mechanics of a Non-Existent Conversation

Here's how it typically works. A company developing a large language model — or fine-tuning an existing one — needs to teach it how to handle email correspondence. Customer service threads. Professional back-and-forth. Sensitive HR situations. Scheduling chains. The model generates thousands of synthetic examples: fake emails, fake replies, fake entire inboxes worth of fictional dialogue between fictional people at fictional companies.

Then a human annotator steps in. Their job might be to rate the tone of a generated response on a scale of one to five. Or to flag whether a synthetic reply sounds passive-aggressive. Or to actually write a better version of what the AI produced, so the model can learn from the correction. Sometimes they're asked to evaluate whether a fake customer complaint was handled with sufficient empathy — even though the complaint itself was fabricated, the customer never existed, and the product being complained about isn't real.

The annotator does all of this for somewhere between $10 and $25 an hour, depending on the task complexity and platform. Then they log off. The data gets fed back into the model. The loop closes.

You're Not Just Reading. You're Teaching.

What makes this stranger than your average data labeling gig — tagging images of stop signs, say, or transcribing audio clips — is the recursive nature of the work. When you annotate a photo for a self-driving car dataset, you're helping a machine understand the physical world. When you annotate synthetic emails, you're helping a machine understand you. Your sense of appropriate tone. Your instinct for when a sentence sounds off. Your cultural read on what counts as professional versus cold.

The AI isn't just learning how to write emails. It's learning how humans judge emails. Which is a different and arguably more powerful thing to know.

Some researchers have started calling this "preference learning" — the process of training models not just to produce outputs, but to produce outputs that humans will evaluate favorably. The annotators are essentially encoding their own aesthetic and social sensibilities into the model's behavior. And because annotator pools tend to skew toward certain demographics — English-speaking, US or Philippines-based, college-educated — the preferences being baked in aren't exactly a representative sample of all human communication.

The Philosophical Trap Door

There's a feedback loop buried in all of this that's worth sitting with for a second. The goal of these training programs is to build AI systems that can handle email correspondence so well that humans won't need to do it themselves. The method for achieving that goal is to hire humans to interact with fake emails. So the labor that makes human labor obsolete is itself human labor. You're being paid to train your replacement. That's not a metaphor. That's literally the job description.

Annotators interviewed for various tech-labor investigations have described a creeping uncanniness to the work. One contractor described spending an afternoon evaluating a synthetic thread between a fictional HR manager and a fictional employee going through a fictional divorce, rating each message for appropriateness and warmth. "It started to feel real," she said. "Like I was actually invested in how this fake situation resolved. And then I'd submit the batch and move on to the next one."

That emotional investment — that flicker of genuine human response to something entirely fabricated — is exactly what the model is trying to capture. The annotator's empathy is the data point.

The Market Logic of Digital Ghosts

From a pure business standpoint, the practice makes obvious sense. Real email data is messy, legally complicated, and ethically fraught. You can't just scrape corporate inboxes to train your model — privacy laws, trade secrets, and basic liability make that a nightmare. Synthetic data sidesteps all of that. Generate your own emails, own the data, control the variables, scale infinitely. It's clean. It's cheap. It's completely untethered from the actual texture of how real people communicate.

Which is part of why the resulting systems sometimes feel slightly off — capable but hollow, fluent but weirdly frictionless. They've been trained on human judgments about synthetic language rather than on language that emerged from actual human need. There's a difference between an email written because someone was genuinely frustrated and an email designed to simulate frustration convincingly enough to pass a human eval.

Whether that gap matters — whether the downstream systems will ever close it — is genuinely unclear. Some argue the models are already good enough that the distinction is academic. Others think the hollowness compounds over time, that training on synthetic data creates systems that are getting better at performing communication rather than actually doing it.

Signal Without Source

What the phantom inbox economy really represents is something Void Signal has been circling for a while: the growing infrastructure of the fake real. Systems that look like human activity, feel like human activity, and are evaluated by humans — but aren't, at any point in the chain, actually generated by human experience or need.

The annotators aren't the villains here. They're just people taking gigs in an economy that doesn't offer many alternatives. But the work they're doing is quietly shaping the next generation of communication tools, and the philosophical weight of that rarely comes up in the task description.

Somewhere in the training data for a major AI assistant, there's a synthetic email about a fake quarterly report, rated 4.2 out of 5 for professionalism by someone who logged off immediately after and never thought about it again. That rating is now permanent. It's part of how the model learned what good writing sounds like.

Nobody sent that email. Nobody received it. And yet it exists, in some form, shaping every message the model generates going forward. A ghost that learned to speak by watching us react to ghosts.