Applicants and employers are stuffing the same keyword into opposite ends of the hiring funnel. The résumé is losing its signal value, and LinkedIn is selling the fix to both sides.
PwC measures a 62% wage premium on job ads that require AI skills. Adding “AI” to a LinkedIn profile takes a few seconds and costs nothing. That gap is the entire story: a priced signal that anyone can produce for free will get produced until it stops carrying information. US members added AI terms to their profiles at a 70% higher rate last year, employers put AI in the title of 8.3% of distinct job listings, and only 1.7% of American job postings ask for a specific AI skill by name. The money is moving away from the résumé and toward whoever can sell verification, and right now that is Microsoft, whose agentic hiring tools inside LinkedIn Talent Solutions already run above a $450 million annual rate.
Art Davis lost his job as a systems analyst this spring. He has sent out 100 applications and landed zero interviews. Somewhere in that process, an AI writing tool he was using to draft résumés and cover letters suggested he add “AI-powered automation” to his LinkedIn profile, so he did. He had been using AI at work since last year without ever thinking to name it.
Read that sequence again. The machine that generated his résumé also told him which words to put in it. Every other applicant using a similar tool is receiving similar advice from a similar model trained on similar job ads. The advice is correct, it is free, and it works right up until the moment the recruiter on the other end also buys a machine.
What Happened
The Wall Street Journal reported on August 13 that millions of job seekers are giving their work histories an AI makeover, some by earning real certifications and some by dropping buzzwords into their profiles. Tampa career coach Denise Bitler told the paper that both sides are doing it, applicants and employers alike.
LinkedIn’s own data shows the share of US users adding AI-related terms to their profiles rose 70% last year. The stock is still small. Roughly 1.2% of members list AI skills, which works out to about 16 million people against the 1.3 billion member base LinkedIn discloses. In some occupations the practice has become normal already: 13% of product managers, 10% of researchers and 7% of consultants say they have such skills.
Stanford economists Nick Bloom and Gideon Moore, working with Lisa Simon and Caelan Wilkie-Rogers at workforce analytics firm Revelio Labs, pulled archived LinkedIn profiles from before 2023 and compared them with the live versions. People have gone back and inserted AI into job titles that never mentioned it at the time. Profiles now carry 30% more AI references than their historic versions. Bloom allows that some of this corrects the record, since plenty of people used the technology and never bothered to say so, but given the volume he concluded that a lot of it looks like AI washing.
The same study found the opposite move happening with terms that have gone out of fashion. References to remote work and to diversity, equity and inclusion have both come down. Workers are editing their pasts in both directions, adding what pays and deleting what does not.
Demand for the training is real and it is enormous. Balaji Padmanabhan, who runs the Center for Artificial Intelligence in Business at the University of Maryland, launched a free AI course last May expecting perhaps 500 sign-ups. He got 62,000, ranging from laid-off engineers to former federal workers.
The Backstory
The 8% figure that circulated all week deserves a closer look, because three different numbers are being used interchangeably and they are not the same number.
Indeed’s Hiring Lab counted 822 distinct US job titles containing AI in the first quarter of 2026, which is 8.3% of all titles carrying at least five postings. That is up from 264 titles in 2022 and a dip to 159 in 2023. It measures the variety of job titles, not the volume of jobs. Indeed’s AI Tracker, which measures the share of actual postings mentioning any AI keyword, hit a record 4.2% at the end of 2025 and sat near 5% at the end of February 2026. Tighten the definition once more and the Federal Reserve Bank of Atlanta found that postings requiring at least one named AI skill made up 1.7% of the US market in 2024, about 628,000 listings, up from 0.5% in 2010.
Each step toward a stricter definition roughly halves the number. The version that gets quoted is the loosest one on the list.
Concentration matters too. Indeed’s own firm-level work found AI adoption clustered in the largest employers, and its Australian data showed two-thirds of AI-related postings coming from 1% of employers. Whatever the aggregate share, most of the hiring sits with a small number of companies.
Then there is the tell that undercuts the panic entirely. Indeed reported in April that job seeker searches for AI roles, while up elevenfold since ChatGPT launched, still account for under 1% of all searches on the platform. Run the comparison on identical keywords and there are fewer searches than postings. Workers are adding the word to their profiles without going looking for the work. That is camouflage, not a career change.
Underneath all of it sits an application flood that predates the keyword. LinkedIn was taking 9,000 applications a minute in January 2025 and 11,000 a minute by June, a 45% jump in a year. An HR consultant quoted by the New York Times pulled a single remote posting after it drew more than 1,200 applications, and she was still working through them three months later.
Employers have noticed what that does to the document. Willo, a candidate screening vendor, surveyed more than 100 hiring professionals for its 2026 report and found that 37% still count credentials and learning history among the most reliable indicators of talent. Four in ten are moving away from résumé-first hiring and about one in ten has largely replaced the résumé with skills tests and scenario work. Three-quarters run into AI-generated applications on a regular basis. The sample is small and Willo sells the alternative, so weigh it accordingly, but the direction matches what recruiters have been saying out loud for two years.
The Plan
Each side of this market has a plan, and each plan involves buying software.
Microsoft is selling employers the filter. On the April earnings call, Satya Nadella said agentic products inside LinkedIn Talent Solutions that automate sourcing, screening and message drafting had passed a $450 million annualized run rate. LinkedIn revenue for that quarter came in at $4.83 billion, up 12%, and grew another 12% in the June quarter. Microsoft also sells the applicant side, through Premium tiers that tell members whether they match a posting and recommend profile changes that improve the odds.
The credential vendors are selling workers the keyword. Coursera booked 8 million generative AI enrollments in 2025, up 195%, on $757.5 million of revenue, and has since agreed to combine with Udemy in a deal implying roughly $2.5 billion of equity value. Google launched its first AI Professional Certificate on the platform in February, completable in ten hours, free for US small businesses. Certifications in this category run from nothing to several thousand dollars in tuition, on the same freemium ladder that online course platforms have used for a decade.
The screening vendors are selling the interview. HireVue, owned by Carlyle since 2019, is one of them, and its chief evangelist Dina Taylor told the Journal that people should list AI skills even while they are still developing them. Her framing was that saying yes and figuring it out later is a reasonable way to operate. From a vendor whose product exists to test claims candidates make, that is a coherent position: the more claims enter the funnel, the more testing the funnel needs.
The Business Model Angle
Three things are happening at once here, and only one of them is a labor story.
The market is pricing a word, not a skill. PwC’s 62% premium comes from more than a billion job advertisements across 27 countries, and it measures what employers offer on postings that require AI skills. It does not measure what people who claim AI skills go on to earn. A posted premium attached to a token that costs nothing to produce is an arbitrage, and arbitrages close through inflation of the token rather than through supply of the underlying asset. The 70% surge in profile edits is the arbitrage closing. PwC’s own series shows the premium climbing from 25% in 2024 to 57% in 2025 to 62% now, which means the closing has not caught up yet, which is exactly why the editing continues.
The résumé is repeating what happened to the web page. Keyword density used to be a ranking signal for search engines, until producing keywords became free and publishers stuffed them until the signal died. Google repriced around link graphs and then around behavior, and the value migrated to the company that owned the ranking system rather than to the publishers who had optimized for the old one. Hiring sits at the same juncture. The text layer no longer separates candidates, so screening moves to graph signals and behavioral evidence: who you have worked alongside, what you have shipped, how you perform on a work sample. The company holding the employment graph collects the repricing, and Microsoft has been charging for it since April.
The toll booth takes money from both sides of the fight. Workers buy certificates, résumé tools and Premium subscriptions to get past the filter. Employers buy sourcing agents, screening software and video assessments to hold the filter. Neither purchase raises the number of good matches, and both scale with the volume of noise. Compare the run rates: LinkedIn’s agentic hiring line alone, at over $450 million, sits above half of what third-party analysts size the entire AI-recruitment software category at for 2025, which mostly tells you the third-party sizing is wrong. The category is bigger than the trackers think because the incumbent platform is booking most of it.
Here is the number that frames the whole thing. In 2025, LinkedIn took roughly 11,000 job applications a minute. Coursera booked its 8 million generative AI enrollments at a rate of about 15 a minute. That is around 720 applications for every one course enrollment, and an enrollment is not a completion. The claim layer is scaling roughly three orders of magnitude faster than the skill layer. Put differently, one platform’s single year of AI course sign-ups equals about half the entire global stock of LinkedIn members who claim an AI skill at all.
The Risk
The strongest counterargument is that the premium is real and the workers are right. PwC’s 62% and Lightcast’s separately measured 28% posting premium use different methods and point the same way. If employers pay materially more for AI-skilled roles, then adding the term is a rational response to a price signal rather than a fraud, and the AI washing framing amounts to scolding people for reading the market correctly. Bloom’s own caveat cuts this way too: some of those retroactive edits fix an omission rather than invent a qualification.
The second risk is that the platform’s incentives are not what the arms-race thesis assumes. LinkedIn sells recruiter satisfaction. If applicant quality collapses, Talent Solutions churns, which is why the company has been shipping match tools that suppress low-fit applications rather than encourage them. A referee paid by the round still needs the fighters to come back next season.
The third risk is the direction of the hiring market itself. BLS put job openings at 6.5 million in December 2025, the lowest since 2017. LinkedIn cut about 875 roles in May, roughly 5% of its workforce, in the same quarter it reported 12% revenue growth. A toll booth on a shrinking road still shrinks, and Talent Solutions bookings have tracked hiring weakness before.
The fourth is timing. Skills-based assessment has been announced as the résumé’s replacement roughly once a decade since aptitude testing in the 1950s, and the résumé keeps surviving because it is cheap to read and legally defensible. Willo’s own data has 59.6% of teams still leading with it.
Quick Questions
Does adding AI to a résumé actually work? For the applicant, sometimes, and the cost of trying is close to zero, which is why volume keeps rising. For the market, it degrades the signal each time it happens. Art Davis added the term and has yet to get an interview after 100 applications.
Is AI washing on résumés the same thing companies do? The mechanics rhyme. A University of Florida study built separate measures of corporate “AI talk” from earnings calls and “AI walk” from employee résumé data, and found that past talk did not predict future AI hiring. Markets rewarded the rhetoric short term and penalized it later.
Who makes money from the confusion? Microsoft, through LinkedIn on both sides of the transaction. Coursera, Udemy and the certificate issuers on the supply side. HireVue, Willo and the assessment vendors on the demand side. None of them is paid on match quality.
Should employers stop reading résumés? Not yet, but stop treating keyword presence as evidence. The screening question worth asking is what the candidate produced, not what the candidate listed. A structured work sample costs less than the interviews you waste on candidates who tested well against your keyword filter and nothing else.
What would fix the signal? Anything the applicant cannot fake for free: verified work history through the employment graph, portfolio artifacts with provenance, or paid assessments. Every one of those is a product someone is already selling.
The Business Model Analyst Take
The résumé became an ad unit, and ad units get keyword-stuffed until somebody changes the ranking system. Search went through this between 2003 and 2012 and came out the other side with a value chain where the platform owning the ranking algorithm captured almost everything and the publishers optimizing for it captured a shrinking slice. Hiring is midway through the same transition, and the party holding the graph is the same kind of party that won last time.
For operators, two practical conclusions. If you hire, your screening stack is now selecting for tool access rather than ability, so move budget from filtering claims to testing output. A structured work sample is a cost line, and it is smaller than the cost of a mis-hire in a market where three-quarters of your inbound is machine-assisted.
If you build, the pattern worth studying is the one Microsoft is running: sell the noise to one side, sell the filter to the other, and price both off the volume of the conflict rather than the quality of the resolution. It is a wonderful business to own and a miserable one to compete against, which is why the interesting startups in this category are not building better filters. They are trying to build the thing that makes filters unnecessary, which is proof of work that a candidate cannot manufacture in a text box.
