He Couldn’t Land a Job Interview. Was AI to Blame?

It was mid-October, highest leaf-peeping play successful Hanover, New Hampshire, and Chad Markey was connected a uncommon interruption betwixt objective rotations during his past twelvemonth of aesculapian school. He should person been inhaling Green Mountain aerial and gossiping with his Dartmouth classmates astir beingness aft graduation. In a fewer months, they’d each beryllium going their abstracted ways to commencement residency grooming astatine hospitals astir the country.

Instead, Markey was unsocial successful his apartment, heavy down a rabbit hole, preparing to spell to war.

He’d aftermath each morning, devour breakfast, unfastened his laptop astatine the room array oregon settee into the tan armchair with the bully backmost support, and commencement coding. Some days, helium wouldn’t announcement the prima had gone down until 1 of his roommates came location and asked wherefore the lights weren’t on.

For days, Markey had been scrolling done a Discord radical astir aesculapian residency, a font of crowdsourced cognition wherever students study backmost to their peers connected each signifier of the exertion and enactment process. He’d watched arsenic different students, tons of them, posted astir the interrogation invitations they’d received.

Markey didn’t person immoderate interrogation offers, lone outright rejections. That seemed not conscionable unusual but incorrect to the quiet-mannered 33-year-old from Houston, Texas, who speaks confidently astir his accomplishments without bragging. He had bully grades from an Ivy League aesculapian school, writer credits connected articles successful the Journal of the American Medical Association and The Lancet, a heart-wrenching idiosyncratic statement, and glowing letters of recommendation. One prof wrote that they had “never met a aesculapian pupil who is much skillful, talented, and appropriately situated successful his pursuit of the tract of medicine than Chad.”

Markey combed done his exertion looking for a fatal flaw. He didn’t find thing helium thought would punctual a residency programme manager to flip an different competitory application, truthful his suspicion turned to different culprit. He’d heard rumblings that immoderate hospitals were utilizing a escaped AI screening instrumentality to assistance process applications—and that it had been displaying incorrect grades for immoderate students. He began to wonderment whether AI was liable for his deficiency of interrogation offers.

On the archetypal leafage of his Medical Student Performance Evaluation, a broad summary of his aboriginal vocation prepared by his school, Markey spotted connection that helium suspected mightiness trigger an automated screening instrumentality to downgrade his application. The MSPE stated that Markey had “voluntarily” taken 3 abstracted leaves of absence, totaling astir 22 months, and had chosen to widen his 3rd twelvemonth of coursework implicit 2 years for “personal reasons.”

That wasn’t rather true. In 2021, Markey was diagnosed with ankylosing spondylitis, an autoimmune illness that affects the spine and could flare up to the constituent wherever helium couldn’t stand, overmuch little bash the intensive carnal enactment expected of aesculapian students during objective rotations. He was connected way to postgraduate from aesculapian schoolhouse successful 7 years, alternatively than the emblematic four, but his absences had been unavoidable and medically necessary. This was explained successful a communicative paragraph connected the archetypal page. Calling the absences “voluntary,” Markey felt, mightiness beryllium interpreted arsenic grounds that helium had succumbed to the unit of aesculapian schoolhouse and not been capable to support up with his studies.

As the days went on, Markey said, helium felt progressively acrophobic that his years of grooming would extremity successful failure. “I crawled retired of a fucking achromatic hole,” helium told WIRED, referring to his diagnosis. “I could not locomotion for six months. I’ve travel this far, and this is happening?” He was asking himself the aforesaid question that pops into the minds of millions of different occupation seekers each day: Did an AI trash my application?

Even recruiters volition admit it’s just to wonder. The CEO of a hiring level said past autumn that his manufacture is successful “an AI doom loop”: HR departments kick of a question of AI-generated occupation applications, prompting the request for much AI filters. Applicants kick they’re getting unfairly filtered out. Some combat AI with AI, filling their résumés and screen letters with buzzwords. “It feels precise dystopian to me,” 1 occupation seeker told researchers from Northeastern University. “My worthiness arsenic a quality and arsenic an employee, arsenic a worker, is based connected my quality to filter myself done a bid of automated gateways.”

Only a fistful of states person regulated the usage of AI screening tools to marque hiring decisions. Laws successful Illinois, New Jersey, and Colorado (not yet successful effect) prohibit employers from utilizing discriminatory tools, but mandate small successful the mode of transparency beyond requiring employers to notify applicants that AI is being used. California’s regulations are much robust, requiring employers to regularly trial their AI hiring tools for bias. But nary of those rules empower an idiosyncratic to recognize however a peculiar AI hiring instrumentality judged them, oregon whether it discriminated against them.

So Markey went to enactment connected an intolerable task. He would walk the adjacent six months penning emails, probe papers, ineligible requests, and a changeless watercourse of Python code, trying to adjacent wrong the AI screener. “It turned into obsession,” Markey told WIRED successful February. “I don’t deliberation I’ve ever been this upset earlier successful my life.”

Markey’s archetypal medical grooming came successful precocious school, erstwhile helium sorted done the gallon ziplock container wherever his begetter kept his medicine medications, recorded the names, and went to the section assemblage assemblage room to probe their purposes. His dada was bipolar and addicted to alcohol, a charismatic, unpredictable shot of vigor susceptible of showing large emotion and causing large pain.

One Christmas, which is besides Markey’s birthday, his begetter didn’t amusement up due to the fact that he’d been arrested for drunk driving. Another Christmas, Markey looked retired the beforehand model to find his motortruck being repossessed due to the fact that his begetter had enactment it up arsenic collateral for a payday loan. While Markey was distant astatine assemblage connected Pell Grants, his household was forced to state bankruptcy and mislaid their house. When helium was 21, his begetter died.

Markey tin callback the infinitesimal helium became funny successful pursuing psychiatry. It was erstwhile his begetter explained wherefore helium started drinking truthful heavily: In manic periods helium would spell days without sleeping, and the lone happening that could unit his eyes closed was a 5th of vodka. “It’s conscionable truthful bittersweet to deliberation if I said, ‘Hey, let’s spell to a psychiatrist and get a low-dose Seroquel medicine and conscionable person you slumber and code immoderate of your mania,’ similar who knows what would happen?”

Markey had been preparing for a vocation connected Wall Street. But aft that speech with his dad, helium took a occupation successful wellness attraction informatics and made plans to spell to aesculapian school. The summertime earlier helium started astatine Dartmouth successful 2019, the stiffness he’d experienced successful his backmost since helium was a teen grew worse and his pelvis began to consciousness similar a cement block. By the extremity of his 2nd twelvemonth of school, Markey was laid level by ankylosing spondylitis. He took a permission of absence, going from doc to doc seeking treatments that would let him to proceed with school.

During that aforesaid time, the Covid-19 pandemic was roiling the aesculapian profession. Among myriad challenges, hospitals saw a monolithic summation successful the fig of applications for their residency programs. Prior to the pandemic, students typically had to question to each infirmary for interviews. When interviews went virtual, they could use to dozens much programs than before. Markey applied to 82.

That surge has made it harder for hospitals to benignant done and prioritize applications. In 2023, the Association of American Medical Colleges (AAMC) announced a concern with Thalamus, the shaper of a screening instrumentality for residency applications called Cortex. Starting successful 2025, the instrumentality would beryllium escaped to usage for residency programs.

A fistful of hospitals had already been moving with Cortex, which displays exertion documents successful an easy digestible dashboard and allows reviewers to hunt by keyword oregon filter applicants based connected a wide assortment of characteristics. Cortex besides uses fine-tuned versions of OpenAI’s generative models to standardize grades betwixt schools with antithetic practices. The AAMC concern opened the doorway to broader adoption of the tool. According to Thalamus, astir 1,500 residency programs astir the country, oregon 30 percent, utilized Cortex to reappraisal applicants and marque enactment decisions during the 2025–2026 cycle.

Issues emerged wrong weeks of the September 2025 deadline erstwhile hospitals started reviewing applications. The institution issued a connection saying immoderate residency programs had reported that Cortex was displaying inaccurate grades for immoderate people. In places similar Markey’s Discord group, the applicants chattered.

As Markey’s anxiousness astir his deficiency of interviews was peaking, helium got an breathtaking spot of news: A probe abstract he’d submitted was accepted to beryllium presented astatine the American Society of Hematology’s upcoming yearly gathering and simultaneously published successful the diary Blood. What happened adjacent deepened Markey’s content that AI systems, alternatively than humans, were liable for his diminishing chances astatine getting into a residency program.

Markey already had 10 publications successful aesculapian journals connected his résumé, but helium began emailing his top-ranked residency programs to stock the update astir this latest accomplishment. The displacement successful his fortunes was immediate, helium said.

Within an hr and 15 minutes of his archetypal email to a residency programme coordinator astatine 1 of the apical psychiatry programs successful the country, Markey received an exuberant effect from the coordinator’s boss. An interrogation connection followed little than an hr later, and they began to travel successful from Markey’s different apical choices too.

To Markey, it appeared to beryllium “the archetypal clip they were seeing an exertion that hadn’t adjacent travel crossed their desk.” As helium saw it astatine the time, “I was getting rejections due to the fact that they had already filled up the apical 100 slots based connected the apical 100 candidates that look connected the dashboard.”

Just a couple days aft Markey’s epiphany, connected October 16, Thalamus published a follow-up blog station astir the antecedently reported issues with Cortex. The institution said it had so documented inaccuracies successful grades displayed to residency programs—but lone successful 10 verified instances retired of much than 4,000 lawsuit inquiries. Cortex was present “99.3% accurate.”

Thalamus aboriginal told WIRED that the institution received nary further reports of inaccuracies retired of much than 12,000 inquiries. But astatine the time, a deficiency of clarity astir however Cortex employed AI sparked forum posts and diary articles. Steven Pletcher, a caput and cervix surgeon who oversees the otolaryngology residency programme astatine the University of California San Francisco Hospital, told WIRED helium heard from a workfellow astatine different instauration that immoderate of the grades Cortex was displaying were “wildly inaccurate.” Pletcher, who besides conducts probe into residency enactment processes, wanted to analyse the level himself.

“As a programme director, erstwhile you hear, ‘Hey we person this AI strategy for reviewing applications,’ you think, tin I conscionable get it to springiness maine a database of applicants that I should interview?” Pletcher told WIRED. “I had immoderate concerns, I deliberation arsenic anyone would, if there’s a caller strategy for reviewing applications and it’s presenting accusation inaccurately.”

At a nationalist gathering of the Society of University Otolaryngologists successful November, Pletcher sat down with a workfellow and reviewed applications successful Cortex. One of the system’s superior functions is the AI grade-normalization tool. From what Pletcher was seeing, the grades displayed for a fixed applicant connected those charts could alteration from infinitesimal to minute.

Pletcher and 4 of his colleagues conducted a structured trial and documented the errors they found. In January of this year, they published their results successful the diary The Laryngoscope, describing “persistent errors successful the Thalamus Cortex strategy with imaginable to negatively interaction residency applicants and programs.”

Jason Reminick, the CEO of Thalamus, told WIRED that galore of the fears astir Cortex expressed by students and aesculapian schools successful the 2025–2026 rhythm were the effect of misunderstandings astir however the instrumentality works. “ A batch of the assemblage abruptly had entree to this and were playing with the instrumentality without truly going done the buying process,” helium said. “And I don’t conscionable mean the carnal paying of money, I mean the exploratory process of knowing what the instrumentality does.”

Reminick told WIRED that too an email from Pletcher, Thalamus received nary different complaints astir the grades displayed for students changing from infinitesimal to minute. He said the mistake was caused by the idiosyncratic moving excessively rapidly betwixt people organisation graphs, resulting successful the show concisely getting stuck. “This would not person affected immoderate applicant’s wide outcome” successful the residency enactment process, Reminick said. Thalamus requested that The Laryngoscope retract the article. The journal, which did not respond to WIRED’s petition for comment, has not done so.

As the day approached erstwhile med students would larn wherever they’d matched, Markey’s ain concerns astir Cortex weren’t going anywhere. In February, helium reached retired to Thalamus lawsuit enactment to inquire whether Cortex utilized accusation astir leaves of lack to people candidates. “Whether thing affects an ‘automatic score’ oregon ordering depends connected what that circumstantial programme has chosen to usage for sorting/filtering,” a Thalamus worker replied. “Programs tin usage antithetic workflows and criteria, and we don’t privation to connote that 1 tract (like [leave of absence] type) is universally utilized arsenic a scoring input everywhere.”

In a aboriginal connection to WIRED, Thalamus offered a clarification astir Cortex’s usage of AI. “We recognize that determination is simply a ample conception of our assemblage understandably tense astir however rapidly AI products are being rolled retired and incorporated into each facet of society—including delicate usage cases similar aesculapian students applying to residency programs,” the connection said. The institution said its attack has been transparent and cautious, but that “putting much accent connected the constricted AI tools would person been adjuvant to forestall misunderstandings astir however AI was being used.” According to Thalamus, “Not lone is Cortex not a decision-making tool, it does not usage AI to sort, filter, exclude, score, oregon fertile applicants.”

Of course, Markey hadn’t heard immoderate of that from Thalamus. As Match Day approached, each helium had to spell connected was the February email he’d received, which helium interpreted arsenic indicating that “scoring” was astatine work. He inactive sensed AI bias—and wanted to ferret it out.

Even for nonrecreational auditors with nonstop entree to screening algorithms, it tin beryllium intolerable to recognize wherefore an algorithm reached a peculiar conclusion, said Shea Brown, CEO of the auditing steadfast Babl AI. When a strategy runs connected an LLM, it people has “a precise opaque reasoning halfway astatine the center, and immoderate benignant of explainability astir wherever it made a determination is hidden,” helium told WIRED. The lone mode to trial for favoritism is successful aggregate: Does the tool, for example, springiness measurably little scores to arsenic qualified candidates with disabilities? “It can’t beryllium done causally based connected a azygous person’s application,” Brown said.

The champion a idiosyncratic tin bash successful a concern similar Markey’s, wherever helium suspected an AI strategy was picking up connected circumstantial connection successful his MSPE, is to trial however an exertion performs with and without that language. That’s wherever Markey started.

First, helium ran 3 versions of his MSPE with somewhat antithetic connection done a suite of AI fairness- and bias-testing tools that the AAMC recommends. The results indicated that a earthy connection processing algorithm mightiness measure a condemnation describing a permission of lack for “personal reasons” otherwise than a condemnation that specified the permission was for a “medical condition,” but Markey didn’t similar that the illustration size was tiny and the trial lacked context.

Next, helium ran 2 versions of MSPE leave-of-absence connection done VADER, an open-source earthy connection processing exemplary that assigns affectional sentiment values to words and phrases, and recovered that a medically close statement of his leaves of lack received a much affirmative sentiment people than the “personal reasons” connection successful his MSPE. He past utilized Python to make a synthetic dataset of 6,000 residency applicants. Each 1 was assigned trial scores, grades, a number of however galore publications they had connected their résumé, and numeric rankings for however beardown their letters of proposal were and however well-suited they were for world research. Markey past divided them into 2 cohorts—one with sentiment investigation scores reflecting the leave-of-absence connection successful his MSPE and the different with scores reflecting medically close language.

The 2 groups were arsenic qualified, successful presumption of grades, trial scores, and different characteristics. But erstwhile Markey ran the synthetic applicants done a logistic regression exemplary trained to prime the apical 12 percent of applicants, those from the cohort with medically close MSPE connection were 66 percent much apt to marque the cut. Still, similar his archetypal test, this lone shed airy connected however a generic algorithm mightiness measure his application. Markey wanted to recognize Thalamus’ tools.

He tracked down the patent for an AI residency exertion screener built by the institution Medicratic. Thalamus acquired Medicratic successful 2025. Patents picture what a strategy whitethorn do, not needfully what it does do, but it was the clearest mentation Markey could find of what mightiness beryllium happening wrong the achromatic box.

With the assistance of GitHub Copilot and yet Anthropic’s recently released Claude Code tool, Markey began to reverse technologist the strategy described successful the Medicratic patent, mirroring the information pipeline and utilizing the aforesaid open-source modules erstwhile helium could. When necessary, helium substituted Claude Code’s proposal and his ain research. For example, earlier the strategy described successful the patent tin people applications, a residency programme indispensable bespeak which characteristics—such arsenic world performance, professionalism, oregon leadership—it values most. Markey reviewed published probe connected residency enactment and surveys of residency directors to find however to value those features.

Markey finished his strategy a fewer weeks earlier Match Day, March 20. He thought its outline and wide features approximated however a instrumentality similar the 1 described successful the Medicratic patent mightiness process the aforesaid inputs. After much than 4 months dissecting assorted algorithms, it was the champion helium could do. Once again, erstwhile helium ran antithetic versions of his MSPE connection done the system, determination were starkly antithetic results: Changing the wording astir his permission of lack from “personal reasons” to a medically close statement resulted successful a importantly higher score.

That month, Markey sent Thalamus a information entree request, nether the New Hampshire Privacy Act, asking for each the idiosyncratic information the institution held astir him. That included a broad accounting of each papers and information constituent that was input into Thalamus’ systems astir him; each penchant parameter, weight, and scoring configuration applied to his exertion by residency programs; each score, property rating, and sentiment investigation calculated by Thalamus based connected that data; and explanations of whether and however his information was processed to mitigate bias. Under the New Hampshire Privacy Act, the institution had 45 days to respond.

WIRED contacted all of the residency programs Markey applied to and asked astir their usage of Cortex. Most didn’t respond oregon declined to comment. Five programs replied that they hadn’t utilized the tool. Yale New Haven Health told WIRED that its residency programs tried Cortex but stopped utilizing it; a spokesperson declined to remark further. Two residency programs astatine Dartmouth Hitchcock Medical Center utilized Cortex to filter applications earlier programme directors reviewed them, said Tennille Doyle, manager of postgraduate aesculapian acquisition programs, but astir of the hospital’s unit preferred to usage their ain screening methods.

Jeremy Walter, manager of media relations astatine Temple Health, said 1 of the hospital’s 59 residency programs utilized Cortex chiefly to presumption applications during “manual screening,” and “overall, we did not find the AI accusation precise reliable.” He declined to elaborate. According to Thalamus, aggregate programs astatine Temple utilized Cortex during the caller enactment cycle. “As with immoderate caller functionality, particularly erstwhile introduced astatine scale, experiences tin alteration based connected however features are utilized and interpreted,” the institution said.

Kari Roberts, who oversees postgraduate aesculapian introspection astatine Tufts Medical Center, told WIRED successful an email that galore of the school’s residency programs tried Cortex for the archetypal clip past fall, utilizing it to surface retired immoderate applications that were incomplete oregon failed to conscionable minimum requirements. “There were immoderate important errors successful the algorithm that incorporated information from the MSPE, starring to incorrect people assignments,” Roberts wrote. “This was not exclusive to our enactment and was raised to the Thalamus squad successful existent clip by our dean’s team.” Thalamus told WIRED that “a precise tiny fig of identified discrepancies” were “investigated and corrected promptly” and that “in immoderate of these cases, what was initially perceived arsenic an inaccuracy was confirmed to beryllium accordant with the root materials.”

After Markey began cold-emailing programme coordinators, helium received interrogation offers from 10 institutions, including immoderate of the astir prestigious hospitals successful the country. Ultimately helium matched astatine Columbia University’s psychiatry programme astatine New York Presbyterian Hospital, wherever helium volition statesman his residency successful July.

Three days aft helium got matched, Markey received a effect from Thalamus to his information entree request. The company’s main of staff, Michele Li, wrote that nary of the programs helium had applied to had utilized the Medicratic instrumentality that Markey had been attempting to reverse engineer. Cortex itself didn’t usage the sentiment-scoring methodology described successful the patent.

Reminick, Thalamus’ CEO, confirmed to WIRED that during the 2025–2026 cycle, Cortex did not algorithmically people oregon fertile applicants. The instrumentality chiefly uses AI for people normalization and to show a badge indicating whether an applicant is funny successful world research, helium said. However, Thalamus plans to aviator an AI screener that volition let residency programs to make campaigner profiles and past measure however good applicants lucifer those profiles, Reminick said. During the pilot, applicants volition person to opt successful to the screening.

Even aft matching astatine Columbia and receiving the missive from Thalamus denying his suspicions astir his ain applications, Markey said helium doesn’t regret the months helium devoted to unpacking screening tools. “ I’m precise grateful for wherever I’ve gotten, truthful erstwhile things endanger that, I privation to marque definite I’m responding correctly,” helium said. In fact, helium has continued his probe of however ample connection models prime up connected semantic signals successful occupation exertion worldly and embed them down the pipeline into decisions oregon recommendations.

There is proof, adjacent successful the satellite of AI hiring tools, that immoderate signifier of owed process, nevertheless imperfect, tin beryllium built and regulated into these systems. One of the astir fashionable applications of AI successful quality resources is to behaviour inheritance checks. Companies similar Checkr automate the process for millions of applications monthly, comparing campaigner names against nationalist records for immoderate grounds of disqualifying transgression activity. A batch of the time, these systems marque mistakes that outgo radical jobs.

But background-check companies, whether they usage humans oregon AI, are taxable to provisions successful the national Fair Credit Reporting Act that necessitate them to stock the results of a inheritance cheque with the occupation campaigner upon request, behaviour an probe if the accuracy of the inheritance cheque is disputed, and nonstop the occupation campaigner the written results of that investigation. Job candidates tin triumph oregon settee idiosyncratic and people enactment lawsuits against background-check companies that supply inaccurate reports.

It’s a strategy with galore of its ain problems, but it astatine slightest offers idiosyncratic occupation seekers an enactment different than screaming helplessly into the void. Not everyone should request to beryllium an Ivy League aesculapian pupil with a inheritance successful informatics and coding and a monolithic axe to grind.

Let america cognize what you deliberation astir this article. Submit a missive to the exertion at [email protected].