Young Workers' Employment Plummets 16% in AI-Exposed Occupations

The Open Roles at Cua

Cua, a three-person startup from Y Combinator's Spring 2025 cohort, builds cloud desktops for AI agents. The company's public listings on the Y Combinator Work at a Startup board show three engineering positions: Founding Engineer, Infra & Agent Systems ($100K–$150K, 0.50%–0.75% equity); Research Intern, Summer 2026 ($8K–$9.25K/month, 0.25%–0.50% equity); and Founding Technical GTM Lead ($150K–$200K, 0.50%–1.50% equity). The founder, Francesco Bonacci, previously worked at Xbox and Microsoft AI.

The Founding Engineer role tasks the hire with turning "research prototypes into production-ready infrastructure, powering everything from secure agent runtimes to cross-cloud container orchestration." That scope covers the full stack of problems that appear when moving an agent from a lab notebook into a multi-tenant cloud: isolation, latency, deterministic execution, observability, and the orchestration layer that lets a fleet of agents share compute. The "founding" prefix is literal: these hires will shape architecture decisions that usually fall to a VP or CTO.

Cua is not hiring for the model layer; it's hiring for the computer layer. The listings emphasize systems depth (Linux namespaces, cgroups, seccomp, virtualization primitives (KVM, Firecracker, gVisor), Kubernetes operators, custom schedulers, and cloud provider APIs (AWS Nitro, Azure Hyper-V, GCP gVisor)) over model breadth. A researcher who has published at top conferences but never debugged a kernel panic in a nested container will face a steeper learning curve than an engineer who has scaled Kubernetes clusters or built sandboxing layers at companies like Replit.

What Gets You Past the Screen

Candidates who clear modern AI-driven screens share a recognizable playbook — one built on reverse-engineering how the software actually reads. The tactics are mechanical, not mysterious.

Keyword extraction, not guesswork

The most consistent finding across NYU's 2024 guidance, ORISE's 2024 formatting guide, and 2025–2026 LinkedIn practitioner posts: AI screeners match resumes against the job description's exact terminology. They do not infer. If the posting says "Snowflake," writing "cloud data warehouse" fails. If it says "SAFe agile transformation," "Agile delivery" scores lower. The workaround is mechanical: copy two to five target postings into a large language model, ask for repeated phrases and required tools, then embed those exact strings (Python, TensorFlow, LangChain, Kubernetes, Terraform, Docker, Power BI, Tableau, SQL, SAFe) into your resume's skills section, headline, and bullet points. NYU's researchers advise a 60–85 percent keyword match; a 100 percent match signals copying and gets down-ranked.

Format for the parser, not the eye

Overdesigned resumes (columns, graphics, tables, icons, text boxes, colored fonts) choke ATS parsers. The mployee.me 2026 guide and ORISE both recommend a clean single-column layout with standard headings ("Work Experience," "Skills," "Education"), simple fonts (Arial, Times New Roman), and .docx unless the employer explicitly requests PDF. Image-based PDFs are the most common silent killer. A software engineer in Austin went from zero callbacks to five in a week after stripping tables, converting to .docx, and adding a plain-text skills block the parser could read, according to a 2026 LinkedIn case study published by the platform's editorial team.

Bullets the algorithm can score

"Responsible for tasks" and "hardworking team player" are invisible. The 2026 LinkedIn prompt library shows how to rewrite every bullet into the XYZ formula: [Action] + [Metric] + [Business outcome]. "Managed social media" becomes "Planned a 90-day content calendar that grew weekly reach by 58%." "Improved site speed" becomes "Improved page load time by 74% by optimizing image assets and caching." The same source provides prompts for baseline-to-after lifts, time-and-cost saved, scale-and-scope, quality-and-reliability, and speed-and-throughput, each forcing a number the screener can index. Recruiters using AI filters search on those numbers: "reduced delivery time by 40%," "reviewed 10K+ posts daily," "cut manual review time by 60%."

One role, one resume

The 2026 LinkedIn "five rules" post is blunt: one CV for every application no longer works. ATS scores on keyword density per posting, and every company describes the same role differently. Candidates who pass modern screens build a master resume, then tailor a version per application — mirroring the JD's job title, echoing its "Who we're looking for" section in the qualifications block, cutting irrelevant roles entirely. The 2025 LinkedIn TPM-at-Amazon example: "Led ambiguity-heavy programs across 4+ teams using Agile/Scrum" and "Owned program execution tied to customer-facing delivery and ops efficiency" — language lifted directly from the posting.

Demonstrate AI literacy, not just domain expertise

A 2025 LinkedIn analysis of trust-and-safety hiring notes the shift: companies want professionals who work with AI tools, not just understand them. Resumes that survive now show that fluency: "Used AI-powered detection tools to review 10K+ posts daily, improving accuracy from 70% to 95%," "Designed prompts for LLM-based policy violation detection, reducing manual review time by 60%," "Analyzed AI moderation tool performance across 15 global markets, identifying bias patterns and improving detection accuracy by 40%." The pattern: name the tool, name the metric, name the outcome.

Prepare for the AI interview layer

Many companies now include an AI-mediated interview. YouTube's 2026 breakdown of AI interview platforms (HireVue-style) identifies three essentials: practice one-way video responses out loud (Western Carolina University research shows this reduces verbal tics and increases clarity), build answer "story buckets" (challenge-action-result, failure-learning, cross-functional collaboration) so each response stays structured and keyword-dense, and speak the job posting's skills explicitly. "I excel at time management, attention to detail, and organization" feels robotic to a human but is exactly what the audio scorer weights. The system typically allows one to three retakes; use them.

The meta-shift

Writing for the robot that guards the door is now the price of admission. "Stop writing your résumé for yourself, and start writing it for the robot that guards the door," said a recruiter with 25 years' experience in a 2026 LinkedIn post. "Your value isn't determined by a keyword scanner. Keep putting yourself in rooms where humans, not filters, make decisions." Candidates who clear these screens treat the AI as a known, testable system: they feed it the exact tokens it expects, in a format it can parse, with metrics it can score, and they rehearse the spoken version for the next gate. The door opens; the work starts after.

Fairness, Transparency, and the Market

Stanford researchers who studied 3.4 million applications across 1,700 job postings put it bluntly: AI screening tools combine three properties that should not co-exist in high-stakes decision-making. They are pervasively adopted, highly consequential, and opaque to the public. Ninety percent of U.S. employers now use these tools, most relying on the same few third-party vendors. That concentration (what researchers call "algorithmic monoculture") means a single algorithmic flaw can replicate across hundreds of companies at once.

Derek Mobley experienced that replication firsthand. The African American, disabled applicant over forty applied to more than one hundred positions through Workday's platform. He was not hired for a single one. His applications were rejected within minutes of submission — fast enough to suggest automated flagging rather than human review. Mobley sued Workday in February 2023, alleging the company's algorithmic screening tools discriminated based on race, disability, and age. A federal judge later expanded the case to include applicants screened by HiredScore AI, another Workday tool, and ruled that AI tools can function as an "agent" of the employer. Arshon Harper, another African American applicant, reported 149 rejections from Sirius XM Radio's AI hiring system.

The Stanford study quantified what Mobley and Harper lived. Twenty-six percent of Black applicants and 15 percent of Asian applicants applied to positions where the AI system discriminated against their racial group. Had the AI recommended Black and Asian candidates at the same rate as the most-favored group (typically white applicants), 40,000 more applications would have advanced. Ten percent of applicants who submitted four applications were rejected from every single one. People who apply to multiple positions screened by the same vendor are more likely to be rejected from all of them than if companies made decisions independently.

Age and gender biases compound the problem. In October 2025, Stanford researchers tested AI resume-screening tools with identical candidate data and found older male candidates received higher ratings than both female candidates and younger candidates. A separate VoxDev study published in May 2025 showed AI tools systematically favored female applicants over Black male applicants with identical qualifications. The World Economic Forum reported that in 85 percent of AI-driven hiring decisions, recruiters followed AI recommendations without questioning fairness or accuracy. Brookings researchers ran an experiment where human subjects screened resumes alongside racially biased AI models. When the AI reinforced race-occupation stereotypes (favoring white candidates for high-status roles), respondents selected majority-white candidates 90.4 percent of the time. When the AI contradicted those stereotypes, respondents still selected majority-non-white candidates only 90.7 percent of the time — a statistically insignificant difference. Without AI, the split was roughly even.

The mechanism is not mysterious. AI tools learn from historical hiring data shaped by decades of exclusion. Amazon scrapped its own recruitment tool after discovering it penalized resumes containing the word "women" — as in "women's chess club captain" or "women's college." HireVue's speech recognition algorithms, used by more than 700 companies including Goldman Sachs and Unilever, disadvantaged non-white and deaf applicants. Even when explicit protected characteristics are removed, the models identify proxy variables: graduation years for age, ZIP codes for race, employment gaps that disproportionately affect women who took parental leave. A 2022 study found 61 percent of AI recruitment tools trained on biased data replicated discriminatory patterns. A 2023 survey found only 17 percent of training data sets used in recruitment were demographically diverse.

The "black box" problem deepens the concern. Developers frequently cannot explain how their algorithms interpret training data. Large language models predict plausible word sequences, not truth — they generate text that appears reasonable but is accurate only by coincidence. There has been a historic lack of transparency about what data models are trained on. The EEOC and DOJ have warned employers they cannot outsource responsibility to AI tools. Courts are signaling that both employer and technology provider can be held liable. The "we're just using a vendor's tool" defense is eroding.

Regulators have started responding. New York City's Local Law 144, effective July 2023, requires annual independent bias audits for automated employment decision tools. California's Civil Rights Council regulations, effective October 2025, make it unlawful to use any automated-decision system that discriminates based on protected traits. Colorado's AI Act takes effect June 2026, regulating "high-risk" AI systems that influence hiring. Illinois House Bill 3773, effective January 2026, prohibits AI use that results in bias against protected classes and requires notification when AI is used in employment decisions. Texas's TRAIGA, also effective January 2026, bans intentional discrimination but rejects disparate impact as a standalone basis for liability. Virginia passed similar regulation in 2025, though Governor Youngkin vetoed it citing burdens on smaller firms.

The patchwork creates compliance whiplash for companies operating across states. But for applicants, the signal is clear: the system that screens them is itself being screened — by courts, by legislatures, by researchers with millions of data points.

The arms race is visible on both sides. A 2025 Reddit survey of job seekers reported 75% now use AI to write resumes and cover letters. A Reddit post from a hiring manager stated 83% of companies will use AI to screen resumes by the end of 2026. Candidates mass-apply to hundreds of roles with near-zero effort; AI tools tailor every application to the job description automatically. Recruiters respond by automating the filter. A 2025 Reddit post from a recruiter reported 21% of companies already let AI reject candidates with zero human review. Hiring managers report a surge in spammy, low-effort applications (90 percent say they've seen it), and 88 percent claim they can spot AI-written materials, though blind tests show 75 percent cannot.

The vendor landscape is consolidating around voice and video agents. Hivemind AI runs screening calls that ask follow-up questions and send back summaries; candidates often don't realize they're talking to a bot. Mercor's "Melvin" conducts voiced interviews that reference specific resume details. Zappyvue and similar platforms handle pre-screening so recruitment teams skip the volume sift and focus on later rounds. "We stopped human screening calls when we realised we'd rather spend the time giving more later-round interviews," a recruiter at a Series B fintech wrote on Reddit in March 2026.

Over 65 percent of recruiters have already implemented AI, primarily to save time (44 percent) and improve candidate quality (25 percent), while cutting time-to-hire by 30–50 percent, per a 2025 McKinsey survey.

The stakes are high and the evidence base thin. Stanford researchers flag three facts that "really should not coexist": AI affects high-stakes decisions, adoption is pervasive, and independent research is extremely limited. Employment of 22- to 25-year-olds has dropped 16 percent in the most AI-exposed occupations — before significant organizational redesign has even occurred. "When an employer posts a job today, they're investing in someone for three to five years, and they're anticipating what AI will handle in that period," the Stanford team wrote. "Young people are bearing the brunt of that uncertainty."

The market is reacting. Skills-based hiring (competitions, work samples, portfolios) is gaining traction as a way to make the application irrelevant. Some employers are retreating to "safe" signals: degrees, brand names, referrals. Networking sidesteps the low-trust anonymous system entirely. And a new class of tools promises explainability: scoring every candidate against the actual job description with evidence attached and exportable for audit. But "every AI hiring vendor says their product is 'explainable,'" said a founder of a stealth HR-tech startup on LinkedIn. "Almost none of them mean the same thing."

Cua's process sits inside this momentum. Candidates adapt because the screen demands it. The result is a hiring loop where both sides optimize for the algorithm, and the human signal gets harder to find.

The Cloud Desktop Waits

The agents Cua is building infrastructure for will one day run on the cloud desktops its founding engineers ship. They will click, type, navigate file systems, juggle windows across operating systems — the same digital labor the company's screening systems now perform on the humans trying to build it. The screen filters for systems depth. The work demands it. But the loop closes only when a human sits down at the keyboard the agent cannot reach.

Working in AI? Zero G Talent tracks the openings: see every open Databricks role, browse AI jobs, openings at Anthropic and Harvey AI, and the people building the field.