Hi, I'm Charles Costa, researcher, strategist, and founder of Lexora Labs®. Each week, I dig into research on AI, knowledge management, and the future of work so you don't have to.
EDITOR'S NOTE
Last issue was about test scores that stop meaning what you think they mean. This week the score is attached to a person. Employers now list AI fluency as a requirement, but most cannot say what it means or how they would test for it.
Candidates, meanwhile, are using AI to polish every application, and the AI screens sorting those applications cannot always explain why one person moved forward and another did not. One of those screens is already in court. The question running through this issue is what hiring should actually be looking for.
THIS WEEK'S READING:
Companies Want AI Fluency, But Are Struggling to Get It | Tech.co / TestGorilla
Why It Matters: A TestGorilla study reported by Tech.co found that 95% of employers now list AI fluency as a hiring requirement. Among managers who prefer AI-fluent candidates, 59% say they have already made a bad AI hire, and 37% set the bar at tool awareness, meaning a candidate knows the tools exist. One caveat is worth keeping in view: TestGorilla sells hiring assessments, so a study concluding that companies struggle to evaluate skills lines up neatly with its business.
Even with that caveat, the pattern is useful. "AI skills" means nothing without context. What matters is how a person handles the tasks that are relevant to the role, and then whether they used AI appropriately along the way. That calls for rethinking hiring, moving away from looking at the tools someone has used and toward looking at how they approach a problem. It is plausible that a 37% tool-awareness bar is related to the 59% reporting bad hires. The study does not show that one causes the other, but the two numbers sit uncomfortably close together.
If employers can't define fluency, the applications landing on their desks are not making the job any easier.
The AI Labs Are Selling Your Applicants a Perfect Résumé | State of AI
Why It Matters: The piece reports that AI screeners tend to prefer text written by AI, and it cites a study in which GPT-4o picked its own rewrite of a résumé 82% of the time. In other words, the same models candidates use to polish their applications may also favor that polish when they sort them.
The 82% figure is a preprint result, and it remains unproven in real hiring. Employers already knew résumés weren't working well before any of this, so the open question is not whether self-preferencing can happen in a test setup but whether the screens companies have actually deployed behave this way. It is not purely theoretical, though, because the Workday case further down in this issue shows that AI screens are already making real-world decisions about real applicants.
If the screen might be favoring its own writing, the next question is whether anyone can explain what it decided.
Why It Matters: Dice's argument comes down to one line: "a score is not an explanation." A ranking tells a recruiter who came out on top without telling anyone why, and even tools that passed a bias audit can drift over time. A related study, Malin et al.'s "Rejected by an AI?" in Frontiers in AI (2025), found that giving people an explanation for a rejection raised how fair they judged it, whether an AI or a human made the call. That was a vignette study of 921 participants reacting to scenarios rather than people going through real hiring, so it speaks to perception more than outcomes.
Automatic rejection is not new. ATS knockout rules already auto-reject candidates, but they are explainable, because the rule says exactly what it checks for. Transparency in AI models, such as open weights, would let companies tune their scoring so it works more like those knockout rules, with criteria someone can actually inspect. Either way, it is not set and forget.
Explainability is a design question. In one case, it has become a legal one.
Why It Matters: In Mobley v. Workday, a federal judge allowed an age discrimination claim under the ADEA over Workday's AI screening to proceed as a collective action. The judge acknowledged that the group could potentially reach "hundreds of millions" of people and said that is "not a basis for denying notice." Workday says the claims are without merit. These are allegations, and there has been no finding of liability.
I am not an HR or legal expert, and I am not going to predict how this case ends. My take is simpler than that: candidates should know what they are being evaluated on.
Not every finding on AI in hiring points the same direction, and the last study is the most rigorous of the five.
Does AI Cheapen Talk? | Management Science
Why It Matters: Cowgill, Hernandez-Lagos, and Langburd Wright published this peer-reviewed study in Management Science in 2026. In a 2023 experiment on Prolific, writers with real experience in one of two fields wrote cover letters, and in a second setting startup pitches, both in the field they knew and in one they didn't, first without ChatGPT and then with it. Screeners with experience in the domain then tried to guess who was the real expert. With ChatGPT, letter quality rose about 4% and the spread in quality fell 54%, so the letters started to look alike. Screening accuracy fell 4-9%, and screeners said they would pay 8% more for extra information. Accuracy actually improved for senders from non-English-speaking countries. It is a controlled study, not a real hiring pipeline.
AI is democratizing hiring. Language differences may still show in an interview, but the candidate gets a real chance to sell themselves instead of a closed door. Screening accuracy does drop 4-9%, and that should be said plainly. Hiring decisions, though, can lean on more accurate data points like interviews and skills tests rather than resting on the letter alone.
Taken together, the fluency survey, the self-preferring screens, the unexplained scores, the lawsuit, and the experiment describe one pattern. Employers are asking for a skill they can't define, receiving applications that look more alike, and sorting them with tools that can't always say why.
Companies need to evaluate how a person approaches the problem, not simply the tools they use, especially since AI can overcomplicate things. AI is not a substitute for being a subject matter expert.
Know someone hiring for AI fluency without a way to test for it? Forward this to them.
Have a take on any of this? Reply directly. I read every response.
-Charles