This past spring, when Harrison Shao (C \u201829) would get back to the Quad after a long day of classes, swim practice, Pottruck sessions, club meetings, or any of the normal exploits of an overcommitted freshman, there was one more thing he made sure to do: train artificial intelligence.<\/p>\n
Whenever he got the chance, Harrison opened the AfterQuery Experts Slack channel and assumed his place in a massive, virtual assembly line manufacturing AI training data.<\/p>\n
When we talked, Harrison had just wrapped up a project testing an AI model\u2019s ability to read research papers on a variety of topics, from biology to finance. He had grilled AfterQuery\u2019s client model on the experiments and results of over 50 biology papers, and he was part of a team processing what he supposed to be thousands of academic articles. Once he received the auto\u2013generated answers to his questions, he graded the model\u2019s response based on another worker\u2019s \u201crubric\u201d and sent the results to yet another team for review.<\/p>\n
In addition to training AI, Harrison helped make marketing materials for AfterQuery Experts, the company\u2019s contract worker\u2013focused branch, allowing the startup to attract more AI trainers generating data in their fields of expertise.\u00a0<\/p>\n
Often, this marketing was aimed at his fellow undergraduates. AfterQuery, a startup that reached a $3.2 billion valuation<\/a> after just eighteen months in business, has placed a heavy emphasis on recruiting college students since its inception. The startup has an especially strong presence at Penn\u2014many of its contract workers and full\u2013timers come from the university, as do two of its co\u2013founders.\u00a0<\/p>\n College students have always been eager to pick up part\u2013time jobs and earn some pocket money, but this new kind of gig work is especially fitting for today\u2019s hyper\u2013online, AI\u2013pilled economy.\u00a0<\/p>\n Of course, the catch is glaringly obvious. College students aiming for entry\u2013level white\u2013collar jobs are extremely vulnerable to automation, and here they are, training AI to replace them.<\/p>\n Like many Penn students, Harrison learned about AfterQuery Experts from a peer: in this case, a friend at Duke.<\/p>\n \u201cOver winter break, I just kind of scrolled through LinkedIn, looking at different startups,\u201d he says. His friend had already assumed a part\u2013time role at AfterQuery Experts, and Harrison came across one of his posts advertising a growth marketing position at Experts that paid an impressive $40 per hour.<\/p>\n The role seemed right up Harrison\u2019s alley\u2014in the fall, he had worked as a growth intern making Instagram reels and other user\u2013generated content for AI startups hoping to grow their brands. Following the post\u2019s instructions, he commented with his email and received a message inviting him to apply.<\/p>\n Harrison sent in his resume, answered a few short questions on the online portal, hit submit, and quickly forgot about his application as the holiday season picked up. A few months later, however, he was notified that his application was accepted, and he could begin onboarding.<\/p>\n \u201cNo interview,\u201d he notes, despite the job\u2019s generous pay. \u201cI just got the offer, so it was surprising.\u201d<\/p>\n Despite applying only to a growth analyst role, Harrison realized he could also earn some quick cash by training AI for the company. He soon became part of AfterQuery Expert\u2019s sizable workforce\u2014though it\u2019s unclear how many users are active, the Experts Slack channel currently includes over 15,000 members.<\/p>\n AfterQuery\u2019s rapid growth comes at a time when AI progress demands input from subject matter experts\u2014a relatively recent development.<\/p>\n Before the AI boom, data annotation was a niche and relatively unattractive job. The first AI trainers<\/a> might have spent their days categorizing movie reviews as positive or negative or labeling all the food and drinks in a refrigerator, often for single\u2013digit wages. This annotated data, together with the vast troves of internet data OpenAI scraped to train their models, formed the basis for ChatGPT\u2019s shocking fluency when it launched in 2022.\u00a0<\/p>\n When training previous Large Language Model (LLM) generations, it was possible for lay annotators to judge whether an English response was coherent or not. But the ability to sound confident didn\u2019t reflect true accuracy or expertise, leading to AI models\u2019 infamous hallucinations.\u00a0<\/p>\n For AI to do the work of real\u2013life experts, they need to be trained by those experts. A doctor, for instance, could grade an AI model on its medical accuracy, comparing the AI\u2013generated responses to her own until their answers gradually began to match. Done at a large enough scale, this sort of training improves an LLM\u2019s accuracy by orders of magnitude across a variety of high\u2013skill trades, from medicine to finance to poetry.<\/p>\n To attract working professionals, AI training companies often offer generous pay to so\u2013called \u201cdomain experts.\u201d These wages, buoyed by venture capital funding and a high demand for training data, have led thousands of consultants, software engineers, screenwriters, and researchers to put their skills at the service of LLMs.<\/p>\n Sometimes, these trainers are employees or freelancers supplementing their other income streams; sometimes, they\u2019re laid\u2013off workers<\/a> training AI to do their old jobs.\u00a0<\/p>\n In AfterQuery Experts\u2019 case, the workforce also includes students attending high\u2013ranking colleges, eager to do work related to their major and earn some easy cash.<\/p>\n At Experts, as in the rest of the data annotation industry, workers are staffed on projects with unrelated names\u2014Egg, Red, or Super Red, for example\u2014and perform one narrow task to help train an unnamed model.<\/p>\n \u201cI’m not really sure which companies I work with, but I know they just have some huge contracts,\u201d says Joanne Lin (W, E \u201829), a student who used to work for Experts.<\/p>\n In her role as an \u201cundergraduate finance expert,\u201d Joanne had two types of tasks: coming up with finance questions that would stump an AI model and reviewing others\u2019 questions to ensure they made sense and fit the project\u2019s needs.<\/p>\n For one project, Joanne grilled the client model on various financial documents from companies listed on the NASDAQ. She asked the model to perform complex retrieval and financial calculations, or forecast trends in the company. At the time, the project paid $40 a question, with a $1,000 bonus every 12 questions.<\/p>\n Putting together a good question was not always easy, especially given Joanne\u2019s relative inexperience. \u201cI only know very basic finance terms, and AI also knows basic finance terms,\u201d she explains. \u201cI guess it took around one to two hours for one question for me, which is why I, like, rage\u2013quitted.\u201d\u00a0<\/p>\n When she worked for Experts, the young data annotator made most of her money reviewing other workers\u2019 questions for $50 an hour, at a minimum rate of six questions per hour.\u00a0<\/p>\n After onboarding in the spring, Harrison wondered if the recommended ten hours per week of work was a hard cap. \u201cI asked, like, \u2018Is ten hours the maximum?\u2019\u201d he recalls. \u201c\u2018Can you go over?\u2019\u201d<\/p>\n In fact, there was no limit to the amount Harrison could earn, so long as he was staffed on a project. After spring break, he decided to take full advantage of this and devoted himself to the \u201cdark grind.\u201d<\/p>\n \u201cOver break, I felt like I didn’t really do anything. I didn’t make any money,\u201d he says. \u201cSo there was, like, four days in a row where I skipped all my classes \u2026 [and] worked eight hours every day.\u201d<\/p>\n \u201cI\u2019m not gonna sit in [class] when I have more important things going on,\u201d he continues. After making $1,300 in less than a week, it was hard to justify doing much else.\u00a0<\/p>\n \u201cI was actually gonna rush \u2026 a frat,\u201d he explains, \u201cbut then I got the job, so I was like, \u2018I need to make bank.\u2019\u201d<\/p>\n Like all good things, AfterQuery began as a Y Combinator<\/a>\u2013backed startup.\u00a0<\/p>\n In 2025, founders Spencer Mateega (W, E \u201825), Danny Tang (W \u201825), and Carlos Georgescu left behind prestigious job offers in investment banking, Big Tech, and quantitative finance to found their own company in San Francisco. Once there, the trio worked out of their two\u2013bedroom in Dogpatch and explored multiple AI automation\u2013related ideas. Eventually, they realized current models weren\u2019t trained to do the work of industry professionals.<\/p>\n Inspired by this very obstacle, they founded AfterQuery and secured a place in YC\u2019s Winter 2025 class of startup founders. Their goal was to compile human\u2013made datasets to train AI in industries like finance, software engineering, and law. The startup\u2019s introductory post<\/a> on the Y Combinator LinkedIn page announced that \u201cif AI is one day to replace jobs entirely, it can\u2019t just be good, it needs to be near perfect.\u201d\u00a0<\/p>\n AfterQuery did not respond to requests for comment on this article.\u00a0<\/p>\n In short order, the co\u2013founders began hiring project leads to procure contracts with frontier AI labs and oversee contract workers under AfterQuery Experts. By April, they had raised $50 million in a Series A funding round, with a total valuation of $500 million. Five months later, that valuation has more than sextupled, making AfterQuery the fastest\u2013ever unicorn to come out of Y Combinator.<\/p>\n At the same time, undergraduates hired by Experts as growth associates and campus leads across top universities began recruiting other students as contract workers. This past spring, the Experts LinkedIn page announced roles for college software engineers and finance students. Applicants could reply with their emails to be considered\u2014combined, the posts raked in almost 2,500 comments. Another LinkedIn post from a former Penn campus lead garnered over 2,200 responses from prospective AI trainers.<\/p>\n Perhaps because two of the three co\u2013founders were Penn students, the startup has had an especially strong presence at the University. Advertisements for contract worker roles have appeared on the Wharton Investment Trading Group\u2019s newsletter, and many current students\u2014including the 2029 class president\u2014have created annotated data or recruited classmates for the company.<\/p>\n While it\u2019s hard to judge exactly how many Penn students and recent grads do contract work for the startup, Joanne believes the number is quite high. \u201cI think a lot of them are Penn students,\u201d she says. \u201cI know I got a lot of my friends on there.\u201d<\/p>\n When I talk to Pritika Kharkwal (W \u201829), an undergraduate finance student, she tells me that she used many of the same skills she learned from her part\u2013time private equity internship to create finance training data for Experts.\u00a0<\/p>\n \u201cWe create deliverables, and then kind of see if AI can do that,\u201d Pritika explains. She might, for example, create an Excel sheet with sample data to calculate discounted cash flows\u2014a staple task for entry\u2013level finance workers.<\/p>\n \u201cExperts will be asked to submit deliverables, and then \u2026 their goal is, like, bridge that gap [between humans and AI].\u201d As she speaks, a shadow comes over her expression. \u201cThey’re gonna train their model to perform to that capability \u2026 so you’re kind of what they’re measuring.\u201d\u00a0<\/p>\n After a short pause, she completes the thought. \u201cSo we\u2019re basically becoming more replaceable, but, like, oh\u2014I just realized.\u201d<\/p>\n As we sit in the basement of the McClelland dining hall, she leans forward with her elbows on her knees and places her hands over her temples, staring blankly into space.<\/p>\n \u201cThis is gonna be a dark article you’re writing,\u201d she says.<\/p>\n The future of employment and AI is a sensitive subject for college students and fresh grads. Across every profession AfterQuery is training AI for\u2014software engineering, finance, law\u2014entry\u2013level tasks are the most easily automated. Even as undergraduates learn the skills historically necessary to secure a job offer, some of them are passing along that knowledge to Claude, ChatGPT, and other frontier models.<\/p>\n \u201cThe hypothesis is \u2026 AI can currently do the types of things that recent college grads would do at an entry\u2013level position,\u201d explains Chris Callison\u2013Burch, professor of computer and information science. The professor has become an authority on AI\u2014in the fall, he teaches a class on the subject which draws in over 500 students. He says that the aspiring software engineers in his lecture hall occasionally express concerns about being automated away.<\/p>\n \u201cI think computer science is on the front lines for this, because AI writing software has become incredibly good,\u201d Callison\u2013Burch says. \u201cIt’s become much harder as a recent college grad to get your first job at a big technology company, and so I feel like the concern is grounded.\u201d<\/p>\n As models extend their expertise to other white\u2013collar industries, the professor fears that entry\u2013level positions like paralegals \u201cwill go the way of the lamplighter.\u201d<\/p>\n \u201cThere’s lots of these studies on careers that are exposed to AI, and all of them are the types of things that we universities are preparing people for,\u201d Callison\u2013Burch says.\u00a0<\/p>\n Indeed, the tasks that interns and fresh graduates perform, like note\u2013taking, slide\u2013making, document\u2013compiling, or appointment\u2013scheduling, are the most easily completed by an AI agent. Normally, this work justifies a junior employee\u2019s existence while buying them time to advance. But as companies automate these tasks away, many are predicting an abysmal junior\u2013level job market in the near future.\u00a0<\/p>\n This is already the case for many of the most \u201cAI\u2013exposed\u201d occupations.\u00a0 Take software engineering, in which employment of young workers (ages 22\u201325) is 19% below<\/a> where it would have been if growth had kept pace with less\u2013exposed sectors. Indeed, many companies are more than willing to hire fewer entry\u2013level employees as AI systems improve\u2014in a recent poll<\/a> of over 600 recruiters, the Graduate Management Admission Council found that one\u2013third of respondents are replacing entry\u2013level jobs with AI.<\/p>\n Callison\u2013Burch, who testified in front of Congress in 2023 on the \u201creal possibility that generative AI may be able to replace a large number of white\u2013collar jobs,\u201d sometimes finds himself imploring company leaders not to pull that trigger.<\/p>\n \u201cIf an AI system is capable of making your employees twice as productive, you can make a decision,\u201d he says. \u201cDo you want your company to produce more, or do you want your company to have half as many employees?\u201d<\/p>\n