If you've applied for a job in the last year, or watched someone close to you go through it, you've probably noticed something. Applications feel like they disappear. Rejections arrive fast — sometimes before you'd expect anyone to have read anything. Sometimes the silence just is the rejection.

That's partly the economy. But there's something else in the data now. A Stanford study tracking employment by occupation found that entry-level workers in AI-exposed fields saw their employment drop about 11% since 2022. The rest of the economy, in positions less touched by AI, grew about 10% over the same stretch. Same labor market. Twenty-one-point split.

And ChatGPT, as of June 2026, launched live job listings and a resume builder pulling from Indeed, Upwork, and Appcast. Meaning: the tool that's contributing to fewer entry-level openings is now also the tool being marketed to help you compete for the ones that remain.

That's the situation. Here's what to do with it.

What's actually happening

Two things are going on simultaneously, and they interact in ways worth understanding.

Companies are using AI to screen job applications. The software that reads your resume before a human does — an Applicant Tracking System, or ATS for short — is standard now at any company receiving more than a few dozen applications per opening. Think of it as a very fast, very literal gatekeeper: it scans for keywords from the job description, checks formatting, sometimes grades you against everyone else in the batch. In a pile of 400 applications, it might surface 40. The rest don't get read. This has been true for a while. What's changed is that the systems are getting smarter and companies are leaning on them harder as application volumes have ballooned.

Job seekers are using AI to write those applications. A cover letter that used to take an hour now takes five minutes. A resume bullet point that required coaching can be drafted and refined in a conversation. The tools are good. Most hiring managers can recognize the cadence of AI-written text — but not always, and not consistently.

The result is AI screening AI-generated content, at volume, with humans reviewing the things that made it through. And the workers who would have used those entry-level roles to build toward something bigger are the ones feeling the squeeze first.

Source spread

• Stanford Digital Economy Lab — "Canaries in the Coal Mine: Six Facts about the Recent Employment Effects of Artificial Intelligence" [academic] — primary employment data; ADP payroll records; August 2026 update showing the 13%→19% YoY gap widening
• OpenAI — "Expanding Economic Opportunity with AI" [hype] — June 2026 announcement of ChatGPT job listings + resume builder; framing is optimistic
• Anthropic Economic Index [builder] — occupational AI usage data, automative vs. augmentative classification, feeds into the Stanford study's analysis

What's real:

The AI job tools are genuinely useful for the tedious parts. Formatting inconsistencies. Converting "responsible for X" into "led X, resulting in Y." Tailoring a boilerplate cover letter to a specific posting's language. AI handles all of that faster than most humans, with more consistency. If you're not using it for those tasks, you're spending time that doesn't need to be spent.

The gap between AI-exposed and non-exposed isn't happening because AI is bad. It's happening because AI is good. Accounting, auditing, reception, information processing — the entry-level work in those fields is precisely what current AI tools handle well. The Stanford paper's framing is useful: fields where AI substitutes for the task are the ones seeing hiring freezes. Fields where AI assists people doing tasks that still need a human — registered nurses, CEOs, skilled trades — those entry-level numbers are flat to growing.

ChatGPT's job board is a real on-ramp. Indeed and Upwork supply most of the inventory. If you're not already using it, it's a legitimately useful aggregator, especially for freelance and contract work.

Samwise's take

What to do about it

Five things that actually change your odds. None of them are magic.

• Use AI for the mechanical parts, not the whole document. Bullet point polish, grammar, formatting consistency — AI is faster and better than you for those. The cover letter's core argument, the reason you specifically want this job, the accomplishment numbers — those have to come from you. A fully AI-written cover letter gets recognized. A human-written letter with AI-assisted edits usually doesn't.
• Match the language in the job posting. This is what ATS software looks for. If the posting says "project management" and you wrote "led initiatives," the system may not match them. Read the posting carefully. Use its vocabulary in your resume where it honestly applies. This is not gaming the system — it's communicating clearly.
• Find a person before you apply. In fields where entry-level hiring is tightest, the highest-ROI use of time is getting to a hiring manager or team member before your application hits the ATS. LinkedIn, a mutual contact, a brief message to someone whose work you can specifically reference. Not always possible. Often more possible than people expect.
• For AI-exposed fields: look at adjacent roles. The Stanford data names specific occupations hardest hit — accountants, auditors, receptionists, information clerks. It also names the least-affected ones: nurses, engineers, skilled-trades workers. If you have flexibility on which direction to go, looking at that split explicitly before committing is worth the time.
• Track your applications in a spreadsheet, not in your head. Use AI to draft follow-up emails. Send them yourself. The discipline of actually following up on 15 applications is where most people lose ground they'd otherwise keep.

Further reading

• Stanford Digital Economy Lab — "Canaries in the Coal Mine" — full study with occupational breakdowns and methodology
• OpenAI — "Expanding Economic Opportunity with AI" — ChatGPT job search and resume builder announcement
• Anthropic Economic Index — occupational AI usage data underlying the Stanford study's automative/augmentative classification