Two things became true at roughly the same time. Job seekers started using language models to write applications, and employers started using them to sort applications. The coverage tends to treat these as one story about AI in hiring. They are two stories with very different evidence behind them.
On the candidate side, the evidence is encouraging
A field experiment on an online labor market with nearly half a million job seekers offered half of them algorithmic writing assistance on their resumes. Those who received it were hired 8 percent more often, and when hired earned 10 percent higher wages, with no significant change in how satisfied employers were afterward (van Inwegen, Munyikwa and Horton, NBER Working Paper 30886, published in Management Science).
The mechanism the authors propose is worth holding onto, because it explains what kind of help works. Better writing does not signal ability. It removes friction between what you did and what a reader can perceive. The assistance corrected grammar, word usage and tone in text the candidate wrote. It did not invent experience.
On the employer side, the evidence is a warning
Researchers tested whether text embedding models, the technology underneath a lot of automated resume matching, rank candidates fairly. They ran a resume audit across nine occupations using more than 500 real resumes and 500 job descriptions, varying only the names attached (Wilson and Caliskan, AAAI/ACM Conference on AI, Ethics and Society, 2024).
The models significantly favored White-associated names in 85.1 percent of cases, and favored female-associated names in only 11.1 percent. Intersectionally, resumes with names associated with Black men were disadvantaged in up to 100 percent of cases. The authors also found that document length and how frequently a name appears in training data influenced selection.
Why this is an industry problem and not a candidate tactic
Nothing in that finding suggests an action a candidate should take. Names are not a lever, and treating them as one would be both offensive and useless.
What it does tell you is where the system is unreliable, and that has consequences for how much weight to put on any single automated outcome. If a matching model can rank identical resumes differently based on a name, then the ranking it produces for you is a noisy signal about your fit rather than a verdict on it.
It also matters legally, which shapes how employers use these tools. Automated screening is a selection procedure, and under the Uniform Guidelines a selection rate for any group below four-fifths of the highest group's rate is generally treated by federal enforcement agencies as evidence of adverse impact (29 CFR 1607.4(D)). An employer that hands ranking to a model without monitoring outcomes is accumulating exposure, which is one reason serious employers keep humans in the loop.
What follows for your search
Use assistance to clarify, not to generate
The measured benefit came from correcting real text. Generated experience is a different product with no supporting evidence and an obvious failure mode, which is the interview where you are asked to walk through something you did not do.
Treat automated scores as information, not judgment
A match score tells you how closely your document resembles a posting. That is useful for spotting a gap you can fix. It is not a measurement of whether you can do the job, and given what the bias research found, it is not a clean measurement of anything.
Build the routes that bypass ranking
Every path into a company that does not run through an automated queue is worth more than it used to be. A conversation with someone inside the organization is not subject to embedding similarity, and an internal referral enters the process at a different point than a portal submission does.
This is not an argument for abandoning applications. It is an argument about the ratio. If every hour of your search goes into a channel where the first filter is a similarity score with known defects, you are concentrating your effort in the part of the system with the least reliable signal.
The honest summary
AI is currently better at helping you explain yourself than it is at deciding who deserves attention. Use it for the first and keep your expectations low for the second, including when the second is being used on you. Both findings come from the same year and the same broad technology, which is a reminder that a tool being useful in one role says nothing about whether it is trustworthy in another.
In practice that means writing from real material you have kept, which is what the achievement library holds, tailoring it to one posting at a time with the resume optimizer, and reading the resume score as a diagnostic rather than a grade.
For how the pieces fit together, see how it works and the plans page.
References
- van Inwegen, E., Munyikwa, Z., and Horton, J. J. Algorithmic Writing Assistance on Jobseekers' Resumes Increases Hires. NBER Working Paper 30886, published in Management Science. nber.org/papers/w30886
- Wilson, K., and Caliskan, A. (2024). Gender, Race, and Intersectional Bias in Resume Screening via Language Model Retrieval. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, 7(1), 1578 to 1590. arxiv.org/abs/2407.20371
- Uniform Guidelines on Employee Selection Procedures, 29 CFR 1607.4(D). govinfo.gov
Disclosure: This article is published by MyJobsSearch and reflects our analysis and commentary on career strategy and job search tactics. It is for informational purposes only and does not constitute professional career, legal, or financial advice. It does not guarantee employment outcomes, interview offers, or job placements. Career results depend on many factors specific to your situation, industry, and geography. Consult a career coach or qualified professional for personalized guidance.
| Where AI is applied | What the research measured | Result |
|---|---|---|
| Helping a candidate clarify their own resume text | Hiring rate against a randomized control | 8% more hires, 10% higher wages |
| Employer satisfaction with those hires | Post-hire ratings | No statistically significant change |
| Ranking resumes by similarity to a job description | Name-varied resume audit, nine occupations | White-associated names favored in 85.1% of cases |
| The same ranking, gender | Same audit | Female-associated names favored in 11.1% of cases |
AI is currently better at helping you explain yourself than it is at deciding who deserves attention.
MyJobsSearch