They No Longer Ask If You Can Code. They Ask Where the AI Goes Wrong
GitHub is replacing the résumé at AI-native US companies, while the EU AI Act opens a 17-month compliance window for hiring. The junior developer market is already reshaping itself in the numbers.
In short:
AI-native tech companies in the US have stopped reading résumés: they scout on GitHub and X instead, and in interviews they no longer ask candidates to solve exercises, they ask them to judge where artificial intelligence gets it wrong.
The junior developer job market is already reshaping itself in the numbers: developers aged 22 to 25 are down 19% from the late-2022 peak, while the 41-49 age bracket grew 14% over the same period.
In Europe, the AI Act imposes strict controls on any system used to screen candidates, with a 17-month window before the toughest deadlines hit. Companies hiring with AI have time to comply, just less than they think.
Le’ale Addison is 22, holds a freshly minted computer science degree, and has already moved through Amazon and KPMG before finishing school. Two years ago, when she sat through technical interviews, she shared her screen and solved exercises while an evaluator timed every line she typed. Now the questions are different: how would you use AI for this task, and how would you notice if it gave you a wrong answer. “Those questions weren’t there before,” she told Business Insider.
It’s not an isolated anecdote. It’s a snapshot of a hiring process rewriting itself from the ground up, at the exact moment when 75% of résumés submitted in the US today never reach a human recruiter at all.
The interview that no longer exists
Business Insider’s investigation, part of its “The Great Coding Reset” series by Ana Altchek and Shubhangi Goel, describes a shift that at Replit already has an operational name: X has become, in the words of the company’s chief people officer, the “main medium” for recruiting. Not a side channel. The main one. At Cognition, hiring isn’t owned by a department at all: every employee scouts, evaluates, and proposes candidates.
Xavier Contreras runs data engineering at a New York hedge fund. Two years ago his interviews were pure coding challenges: write this function, optimize this algorithm, prove you know the syntax. Today he evaluates two entirely different things: systemic thinking, and the candidate’s ability to spot where an AI-generated output is quietly wrong. LeetCode, the platform that for a decade served as the mandatory rite of passage for anyone wanting a developer job, has been demoted to a first filter. Not the test anymore. Just the waiting room before the real one.
The reason is easy to state and hard to sit with: when a model writes syntactically flawless code in three seconds, syntax stops being the skill worth a salary. What’s worth paying for now is the judgment applied to what the model produces.
The numbers behind the shift
The Replit and Cognition cases aren’t cover-story exceptions. 85% of US employers already use skills-based hiring instead of relying on degrees or traditional résumés. 87% of companies filter résumés with AI tools before a human ever sees one, which is exactly why 75% never reach a human recruiter at all.
There’s a less-discussed side effect: in-person interviews climbed from 24% in 2022 to 38% in 2025, not out of nostalgia, but because AI-assisted cheating in remote interviews became serious enough that companies wanted to look candidates in the eye again. 71% of hiring managers admit AI has made it harder, not easier, to actually assess technical skill.
Candidates feel it directly. According to HackerRank’s 2025 Developer Skills Report, 74% of developers struggle to find a job even as hiring in the sector rises. 66% would rather take practical, job-relevant tests than the usual abstract puzzles, 77% openly say current assessments don’t reflect the skills the job actually requires, and 42% cite test prep itself as their main obstacle. Meanwhile 97% of developers already use AI tools, 82% inside their daily workflow, and 40% are seriously considering leaving their job within a year. The hiring system and the reality of daily work have stopped talking to each other.
Who’s disappearing, and who’s cashing in
Laurie Voss, npm’s co-founder, combined ADP and Stanford Digital Economy Lab data with US Bureau of Labor Statistics figures in an analysis worth reading in full at seldo.com. What comes out of it is precise, and uncomfortable precisely because of that.
Developers aged 22 to 25 are down 19% from the late-2022 peak. The 41-49 bracket grew 14% over the same period. Looking at BLS occupational categories: “computer programmer,” the role that writes code to spec, is down 16% in a single year. “Data scientist,” the role that decides what to build before writing a line of it, is up 12%. “Systems analyst” is up 4.4%.
Developer jobs aren’t disappearing: total employment grew 10% between 2022 and 2025. What’s changing is what gets paid, and who gets paid to do it.
Computer science graduates now face a 6.1% unemployment rate, higher than many humanities degrees, upending decades of assumptions about which major “pays off.” And while junior openings shrink, GitHub logged 36 million new accounts in the past year, more than one per second. The platform where you prove you can actually do the work is exploding at the exact moment the document that used to describe that same capability in words is losing its grip.
What companies are actually evaluating now
An Anthropic study of roughly 400,000 Claude Code sessions, run between October 2025 and April 2026, puts numbers on what Xavier Contreras was describing in plain words: it isn’t coding ability that determines how much autonomous work a model can do, it’s the domain expertise of the person guiding it. Expert users trigger action chains twice as long, 12 steps versus 5, and get five times more output per instruction. The division of labor that emerges is almost surgical: users make roughly 70% of the planning decisions, the model handles roughly 80% of the execution.
Translated into interview terms: nobody’s asking anymore whether you can write a function. They’re asking whether you can decide what to build, and whether you can tell when the machine building it for you has taken a wrong turn. That shift, from executor to orchestrator, is exactly what I built the From User to Orchestrator course around: not another prompting course that expires in six months, but a program designed to stay current by construction, because the gap that matters now isn’t technical anymore. It’s judgment.
The gap nobody wants to admit
There’s a more cynical reading of these numbers, and it comes from Robert Half and Orgvue’s research: 32% of US hiring managers have already rehired for roles they had replaced with AI, and 55% of AI-linked layoffs were later judged a mistake. This isn’t the story of a clean automation that swaps humans out without friction. It’s the story of companies that cut before understanding what they were actually buying, now paying for that haste twice: once for the layoff, once for the rehire.
Europe is playing a different game
Anyone hiring in Europe, or hiring people who live there, is looking at a different board. The AI Act classifies as “high-risk” any system used for recruitment, résumé screening, ranking, or candidate evaluation. Once the obligations fully apply, every rejection will need an objective explanation, every system will need a risk assessment, and companies will need human oversight, technical documentation, and continuous monitoring in place. The penalties aren’t symbolic: up to 15 million euros or 3% of global annual turnover for ordinary violations, up to 35 million or 7% for prohibited practices. And the rule applies to any company whose AI outputs affect people living in the EU, regardless of where the company is headquartered.
On June 29, 2026, the Council of the European Union gave final approval to the so-called Digital Omnibus, following Parliament’s endorsement on June 16, as reconstructed by Gibson Dunn. The delay applies only to the high-risk obligations: standalone Annex III systems, which explicitly include recruitment, now shift to December 2, 2027, while AI embedded in already-regulated products shifts to August 2, 2028. What wasn’t delayed matters just as much: the prohibited-practices ban has been in force since February 2, 2025, the AI literacy obligation for staff has applied since the same date, penalties have been enforceable since August 2, 2025, and on August 2, 2026 transparency obligations kick in regardless, as Warden AI documents: chatbots will need to disclose they’re bots, AI-generated content will need to be labeled, and national supervisory authorities will be fully operational.
It’s not a free pass. It’s a window: 17 months to get the most sensitive piece, candidate selection, in order, while transparency and oversight are already live now. Anyone running hiring inside a company that uses AI anywhere in that process should start mapping today where the process touches the “high-risk” perimeter, not in 17 months. Redesigning a screening process from scratch under deadline pressure costs far more than doing it calmly now. That’s exactly the kind of work I help companies with: if your hiring process already leans on AI in any part of it and you don’t have a clear map of your exposure yet, it’s worth talking about it before the window closes.
There’s a parallel gap on the US side worth naming too: 275,000 open AI-related positions against 81,000 tech workers laid off in Q1 2026 tell the same story from the other direction. The people getting cut and the people getting hired increasingly aren’t the same people, and the skills gap between them is wider than most severance packages account for.
Reality Check: no GitHub-based screening process will make hiring fairer. It will just make visible whoever had the time, energy, or luck to build a decent public profile, and invisible whoever spent the last two years quietly shipping proprietary code nobody will ever see. The 2026 interview doesn’t only measure skill. It measures who could afford to document it.
The old deal was simple: study, get a degree, write a clean résumé, someone calls you. That deal is breaking apart piece by piece, not through an official announcement but through hundreds of quiet decisions made by recruiters who stopped opening PDF attachments. The new deal, still taking shape, seems to say something else: show what you can do somewhere anyone can verify it, and get ready to explain not what you can write, but what you can judge.
If this changed how you see things, a restack is the most direct way to put it in front of someone who needs to see it too.
None of this means much until you apply it to your own case, though. Knowing the rules changed isn’t enough. The real question is what you do tomorrow morning, with the résumé you already wrote and the GitHub profile you probably haven’t looked at in months.
Here’s what’s inside:
The prompt that interviews you before rewriting your résumé. Two turns: the right questions first, then a rewrite anchored strictly to confirmed facts. Because an inflated résumé is worse than a weak one, and an experienced recruiter can tell from across the room.
The 30-second test. How to have AI reread your résumé as the human recruiter who genuinely has half a minute to decide whether to call you.
The 90-second GitHub audit. What a scout actually sees when they open your profile, which negative signals you’re sending without knowing it, and the 3 fixes with the best effort-to-impact ratio.
The one-weekend plan. How to make the work you’ve already done legible, without faking activity you never did.
The 2026 interview simulator. Five questions on how you use AI and where you catch it going wrong, with a final scorecard across 4 dimensions and the answer a strong candidate would have given in your place.
These three tools are the interview they’ll give you in six months. Might as well run it yourself now, while getting it wrong costs nothing.
