I sat down with Hong-Yi TuYe, a final-year PhD at MIT Sloan, who has spent the last two years trying to answer a question every founder, journalist and policy wonk thinks they already know the answer to: did ChatGPT cause a hiring slump?His forthcoming paper looks at over 3,000 Y Combinator (YC) startups from 2015 to 2025, matched to Revelio Labs headcount data and PitchBook funding rounds. The insight: It’s too early to tell if ChatGPT is eating jobs.

Hong-Yi measured how fast YC startups added headcount in the quarters after they raised a round, and compared cohorts that were funded before ChatGPT with cohorts that were funded after.

Post-ChatGPT cohorts hire more slowly overall. Break the number down by role and almost all of the slowdown lives in one place: junior engineering hires fall by roughly 10%. Senior and executive engineering hiring barely moves. Sales and customer-facing roles do not move at all.

That is the headline most journalists would stop at: “Junior engineers down 10%, AI is eating coding jobs.” End of story.

I so appreciated that Hong-Yi did not stop there but went a level deeper.

The clever move in his thesis is a stress test.

Most AI-and-jobs papers pick November 2022, when ChatGPT launched, draw a line, and compare what came before to what came after. 

Hong-Yi instead runs the same comparison 26 times, shifting the line one quarter at a time, creating placebo cutoffs. If ChatGPT or the arrival of LLMs were the primary factor behind changes in hiring patterns, the impact would start appearing in November 2022 and intensify as AI capabilities improve.

It does not. The effect largely follows the interest rate cycle. It gets bigger as the Fed starts hiking, plateaus at the beginning of 2023, and starts to reverse once the Fed pauses in mid 2023. The "AI effect" ends up co-moving suspiciously with the cost of money.

Even granting all of that, the junior engineer specific drop still needs an explanation. There are two plausible ones, and the data from the last few years cannot tell them apart.

The first is that generative AI is genuinely substituting for entry-level coding work. The second is another macroeconomic culprit that comes up alongside interest rates:

In the wake of the pandemic, Hong-Yi explained, companies hoarded engineers, encouraged by loose fiscal and monetary policy. They overhired during the talent crunch, then overcorrected when capital got expensive. 

Junior engineers are the easiest line item to cut, because they require investment before they pay back. And no CFO wants to tell the board that revenue is soft, so "we are using AI to be more productive" became the most convenient story in the building.

To separate the two, you would need a counterfactual: startups exposed to the same technology but facing a different funding environment. That set does not exist. 

To me, it seems reasonable that the drop in junior hiring is explainable by a mix of all this: First, the companies overhired, then they wanted to get rid of overhead, AI helped them make the case, and they genuinely adopted use cases to drive productivity. We just don’t have the data to support this yet.

There is a second reason to distrust the November 2022 cutoff when looking at the question of whether ChatGPT impacted hiring:

Back then ChatGPT was still a novelty. Then reasoning models came next, which marked a significant uptick in problem solving over longer horizon tasks but still were not seen as economically useful. Only in the last 6-9 months, with the arrival of more agentic tools like  Claude Code, have we seen greater consensus on the economic value of these systems. 

Pinning labour effects on the day ChatGPT launched assumes the technology was productive on day one, which almost no one who actually uses these tools believes.

The right cutoff is later. It is probably somewhere around the moment Claude Code started being able to ship a feature end to end without a senior engineer babysitting it. However, the progression and economic adoption of the technology is uneven and continuous, not discrete. The economics literature and policy makers need to move away from a discrete cutoff and towards a system that continuously monitors the impact on labor markets, within more granular occupations and sectors, argues Hong-Yi.

I can definitely relate here: I have been using ChatGPT mostly as a thought partner, but only as I have been adopting Claude Code did the real productivity gains kick in where I could get rid of entire software stacks that I used before. If there is a shift in labor, it should be obvious from now on.

Here is what makes the next 18 months interesting. The Fed stopped hiking almost two years ago. The post-COVID overhiring has been digested. VC funding has stabilised. The macro confounders that have been muddying every AI-and-jobs paper are finally quiet and we have Claude Code!

If junior engineering hiring keeps falling between today and 2027, the AI story gets a lot harder to dismiss. There is no longer an interest rate scapegoat. There is no longer a hiring hangover to blame. If the line keeps going down, the line is doing it on the merits.

This is the test. Watch the next four quarters.

Halfway through, we got into a bonus question I really wanted to ask. If Claude Code can let one engineer build/vibe code almost any app, why do we still need SaaS? Is the entire category going away?

Hong-Yi’s take was that while it is true that the cost of writing code falls, total complexity does not. Codebase entropy will rise unless the user/team co-develops new practices with the agentic system. Cheap code generation will mean a lot more minor features and one-off scripts. Without proper testing and CI/CD to handle the increased complexity, more breakages are bound to occur. (I am definitely guilty of this). Someone has to absorb that complexity.

There are three candidates to deal with this:

Option three is not here yet. Option one runs out of bandwidth fast. Self servingly, I’ll speculate that option two is where I think the action is. It is also consistent with YC putting out its first call for agencies this year.

Also interesting was Hong-Yi's description of his own job search. He pointed out that the recruiting funnel has been swamped by AI-generated applications, AI-generated portfolios, AI-generated everything. The cold-application channel has collapsed under its own noise. Trust is migrating back to networks, referrals, and warm introductions.

This is the labour-market mirror of what is happening in early-funnel B2B sales right now. When the cost of producing a personalised outreach drops to zero, the signal it used to carry drops with it. Buyers stop reading. Recruiters stop reading. Everyone falls back on costly, networked filters that AI cannot fake yet.

Hong-Yi's advice to entry-level workers follows from this. Build a public portfolio that shows one person doing what used to take ten, and stop trying to win the cold-application game.

If you want a defensible claim about AI and jobs in 2026, it is roughly this: The first wave of papers that pinned everything on November 2022 are confounded by the Fed and by the post-COVID hiring hangover, and we should treat them with caution. The category most exposed today is junior engineering, but the cleanest test of whether AI is the cause is still in front of us. The next 18 months of data, with the macro noise gone and Claude Code in the field, will settle a lot of arguments that we currently pretend are already settled.

Hong-Yi TuYe is a final-year PhD candidate at MIT Sloan. His paper “Startup Scaling in the Age of AI” will be published in May 2026. Previously he held positions at Carta, Cornerstone Research, JP Morgan and Aon Hewitt. Furthermore, he holds a Bachelor Degree from Columbia University and a Master Degree from University College, London. He is on the job market and looking for academic, policy, and private sector research positions at the intersection of economics and AI. Reach out!

Watch my full interview with Hong-Yi here: