Selling Fear to Job Seekers

Opening

Dear reader, “AI is screening out your resume.” That was the most-shared sentence on LinkedIn and in career communities over the past month. The source is a single paper published by researchers at Stanford University. The paper is real. It analyzes 4.2 million real hiring records — the largest empirical study ever conducted in this field. But as the paper spread across social media, the context got stripped away, one line at a time. Here is the bottom line: what the paper found was a flaw in one game-based hiring tool; what the internet manufactured was a fear narrative that “the entire AI hiring system is rejecting you.”

What genuinely floored me while working through this paper and preparing this newsletter was that the self-styled “job-search consultants” and “career consultants” were the ones selling this fear. (Not all of them, of course.) What makes it even more absurd is that most of them have no professional knowledge of the industries or job functions in question — in many cases, no actual employment experience at all. They were simply running fear marketing on people desperate to land a job.

What the Paper Actually Found

This past May, Stanford HAI researchers presented a paper at FAccT¹ 2026 titled “Algorithmic Monocultures in Hiring,” built around the concept of algorithmic monoculture². The hypothesis: when many companies use the same hiring-AI vendor, one model’s bias can propagate across the entire market.

The data the researchers secured to test this hypothesis is impressive. 4.2 million applications, 3.4 million applicants, 156 companies, 1,746 jobs. Four years (2018–2022) of real data processed by a single hiring vendor.

But once you learn which vendor it was, the story changes character.

The tool is pymetrics. It is not a system that analyzes resumes. Applicants play 12–16 online games, the tool measures traits like risk tolerance, processing speed, and planning ability, and then it outputs one of two verdicts: “recommend” or “not recommend.” Most of the games are identical regardless of the job. Whether you are applying to manage a warehouse or work as a financial analyst, you play the same games — and about 42% of applicants get a “not recommend.”

How this tool is trained is the real problem. For each company, the model is trained on at least 50 employees currently holding that job as “good examples,” and randomly selected people as “bad examples.” The “good examples” are not high performers. They are simply whoever happens to be sitting in that seat right now. And the “bad examples” are not people who failed at the job — they are arbitrary, unrelated profiles. What this tool ultimately learns is “how much do you resemble the existing employees,” not “can you do this job well.”

Before this study, pymetrics ran its fairness audits by pooling all of a company’s applicants together. Pooled that way, Black applicants pass at 52.5% and white applicants at 58.3% — which clears the adverse impact³ threshold under US employment law.

The researchers nailed this precisely. US employment law (Title VII) requires evaluation job by job, not company-wide. So they took apart the 1,746 jobs one by one, and found that roughly 11% of jobs were operating to the disadvantage of Black applicants. And roughly 26% of all applications from Black applicants were concentrated in exactly those jobs.

This is a valid finding. It demonstrates, with 4.2 million real records, that averages can hide discrimination. Even New York City’s AI hiring audit law (Local Law 144)⁴ instructs auditors to pool their data — exactly the trap the paper identifies. For any company running hiring tools, the lesson is clear: switch to job-level audits.

So far, so good. The problem is what came next.

What the Paper Did Not Say

The emotional core of this paper is the “algorithmic blacklist” — the fear scenario that because the same model is used across many companies, one rejection means rejection everywhere. But the researchers’ own data does not support this scenario.

84% of applicants applied to exactly one job. More than 95% applied to two or fewer. The number of people who applied to 10 or more was 522 — 0.02% of the total. For the “rejected everywhere” nightmare to operate, you would have to apply to multiple companies using the same vendor — and almost nobody did.

Even more decisive is the simulation the researchers ran themselves. They took 1,000 applicants and ran them through all 495 pymetrics models. The result? Not a single person was rejected by every model. Even the worst-off applicant received a “recommend” for 52 jobs. The researchers acknowledge this themselves in the paper’s results section.

The paper’s first sentence opens with “more than 90% of US employers use hiring algorithms.” The introduction cites HireVue as being used by more than 60% of Fortune 100 companies. Sounds terrifying, right? Yet HireVue is never analyzed in this paper — not once. The sole subject of the study is pymetrics. HireVue is built on structured interviews and job simulations, a completely different product category from pymetrics’ game-based assessments. Placing HireVue’s market dominance in the introduction was rhetoric designed to inflate the scale of the fear, unrelated to the actual analysis.

The comparison baseline deserves scrutiny too. To compare pymetrics’ results against traditional human hiring, the paper pulls in Kline et al.’s 2022 study (an experiment with 83,000 resumes). In that study, hiring decisions were statistically independent, whereas pymetrics shows a pattern of clustered rejections. But the Kline study covered only entry-level jobs in the United States, while the pymetrics data spans every seniority level worldwide. The city with the most applications is not New York — it is London. When two datasets differ in country, seniority, and measurement method, there are far too many uncontrolled variables to line them up side by side and conclude “the algorithm did it.”

And there is one fact nearly every news report missed. This study contains no data validating whether pymetrics actually predicts job performance. The researchers admit this themselves in the limitations section. Set aside the tool’s bias: no one has ever verified whether this tool achieves the basic purpose of hiring — “picking good employees.” Two failures are stacked on top of each other: bias and invalidity.

In fact, the paper’s limitations section is remarkably honest. The authors write outright that because the results come from a single game-based tool, they are hard to generalize to other types of AI hiring tools. They also concede that they cannot know whether rejected applicants would have become good employees, that the comparison with human hiring processes is not a clean control group, and that these results do not prove illegal conduct. All of it, acknowledged directly.

This is a careful piece of research. The problem is not the paper — it is what happened afterward.

The Supply Chain That Turns Fear into a Product

Trace the path from paper to headline, and you can watch the context fall away one step at a time.

Step 1, the paper: “In the game-based assessment tool pymetrics, roughly 11% of jobs show adverse impact by race when analyzed job by job. Caveat: limited generalizability.”

Step 2, the university press release: “Clear racial disparities found in an AI hiring tool.” The name pymetrics survives, but the “game-based” context fades.

Step 3, news articles: “Stanford study exposes massive racial bias in AI hiring.” Here pymetrics disappears, and “AI hiring tools” becomes the subject. The “90% of companies use them” rhetoric from the paper’s introduction gets positioned as if it were the core finding.

Step 4, influencers: “AI is filtering out your resume. Stanford proved it.” The limitations section vanishes entirely. A careful hypothesis becomes established fact.

Step 5, career coaches: “To survive the AI era, you need to learn this.” Here, fear becomes a product.

This structure is already operating in Korea, too. Self-proclaimed “career coaches” and adult-education platforms are stoking FOMO under the pretext of AI training. “Fall behind on AI and you’re obsolete.” “Without this skill you’ll lose your job within 3 years.” What these narratives share is that they stack urgency on top of unverified premises.

The most desperate people are the target. The unemployed, job seekers, people weighing a career change — that is who is being sold financially burdensome courses and coaching. The fear that “AI will reject you” and the hope that “learn AI and you’ll survive” are two sides of the same coin. Both are narratives that strip out the verification and keep only the emotion.

Turn the questions this paper asked of pymetrics back around, and the structure comes into view. Of pymetrics it asked: “What was this trained on?” “Can it predict performance?” You can ask career-education products exactly the same things. “What is the evidence behind this curriculum?” “Have you ever measured post-course employment rates or salary changes?” Just as pymetrics cloned the profiles of existing employees, fear marketing clones unverified premises. Different tool, same mechanism.

That is why nobody quotes the paper’s limitations section. Quote it and the fear shrinks — and when the fear shrinks, the product stops selling.

Oswarld’s Take

Reading this paper, I found myself paying more attention to the distribution structure than to the hiring tool.

There is a pattern I have seen countless times while building GTM strategies. When fear appears in a market, the product that promises to relieve that fear sells fastest. The problem is that nobody verifies whether the product actually relieves the fear. When the purchase decision comes from dread, validating effectiveness gets pushed to the back of the line.

Just as pymetrics skipped the question “does this tool predict performance?”, Korea’s AI career-education market is skipping the same question. “Does taking this course actually raise your odds of getting hired?” “Did this coaching provide real help with a career transition?” How many services can answer those questions with data?

As I see it, the paper leaves two real lessons. First, when choosing a hiring tool, ask “what was it trained on?” and “does it predict performance?” If the answers are “existing employees” and “we don’t know,” that is not science. Second, put the same questions to anyone selling fear. “What is the evidence behind this training?” and “Have you ever measured its effect?” Any service that cannot answer those two questions is standing on the same structure as pymetrics.

Closing

This paper is worth reading — especially through the limitations section. The finding that pooled audits can hide job-level discrimination is a lesson hiring practitioners can apply immediately.

But the “AI rejects you everywhere” narrative is not something the paper created. It was created by people who read no further than the abstract, and distributed by people who turn that fear into product. Whether it is a hiring tool or a career course, remember one thing. The moment the question “does this actually work?” gets skipped, what you are buying is not a solution — it is anxiety.

💬 Have you ever used an AI hiring tool or a career coaching service? If you have a “this actually helped” or “this was FOMO merchandising” story, share it in the comments.

References & Further Reading

Primary sources

• Bommasani, Bana, Creel, Jurafsky & Liang, “Algorithmic Monocultures in Hiring,” FAccT 2026. : The central paper of today’s newsletter. Be sure to read the limitations section (Section 7: Limitations) alongside it — the authors’ honest self-scrutiny is striking.
• Stanford HAI, “Q&A: Algorithmic Monoculture in Hiring,” 2026. : An interview in which the 3 authors explain their motivation, methodology, and limitations in their own words. Easier reading than the paper.
• Placementist, “Fear-farming the Stanford AI hiring study,” 2026. : A systematic analysis of the paper’s 6 structural limitations, and a key reference for today’s newsletter.

Background

• Kline, Rose & Walters, “Systemic Discrimination Among Large U.S. Employers,” Quarterly Journal of Economics, 2022. : The 83,000-resume experiment the paper uses as its “human hiring is biased too” baseline.
• DLA Piper, “Critical audit of NYC’s AI hiring law signals increased risk for employers,” 2026. : A legal analysis of why NYC Local Law 144’s audit regime was judged “ineffective.”

The author, Kwangseob Ahn, is a professor of business administration at Sejong University and lead consultant at OBF (Oswarld Boutique Consulting Firm). He teaches statistics and data analysis — including business data management and business analytics — at the university, while leading GTM strategy and AI strategy consulting in the field, designing the interface between technology and business. He has published academic research on memory architecture for AI dialogue systems (HEMA), and runs Daily Arxiv, a project curating global AI papers every day. He completed a master’s program at Korea University’s Graduate School of Technology Management and its KMBA. He is the author of “Homo Brainless: The People Who Outsource Their Thinking.”

Footnotes

1. FAccT (Fairness, Accountability, and Transparency): An academic conference on fairness, accountability, and transparency hosted by ACM. It is one of the most influential venues in the field studying the societal impact of AI and algorithms. ↩

2. Algorithmic Monoculture: A state in which many organizations depend on the same or similar algorithms for their decisions. The principle is the same as in agriculture — plant a single crop variety and one pest can wipe out the whole harvest. ↩

3. Adverse Impact: A standard in US employment law triggered when the selection rate for a particular racial or gender group falls below 80% of the rate for the highest-selected group. Also known as the 4/5 rule. ↩

4. Local Law 144: New York City’s AI hiring audit law, in effect since 2023. It requires employers using AI-based hiring tools to undergo an independent bias audit once a year, but a 2025 audit by the NYC Comptroller’s office judged it “ineffective.” ↩