A Harvard team claims AI killed junior jobs.
They tracked 62 million workers and found that companies posting “AI integrator” jobs saw junior headcount drop 9% within six quarters.
But they missed something fundamental…
The junior hiring crash started in 2022, before many people knew what “prompt engineering” was. A few months after the Fed jacked rates from 0% to 5.25%.
I’ll show you why this matters: We’ve blamed the wrong culprit for entry-level job losses three times in 25 years. Each time, a generation’s careers were permanently scarred. And we’re about to do it again.
TL;DR
Did AI cause the junior hiring crash?
No. AI didn’t trigger the collapse. The hiring drop started before AI entered the workflow; AI added some pressure only from late 2024 onward. Macro factors—rate hikes, post-COVID correction—were the main drivers.
If not AI, what did wipe out junior roles?
It was a messy cocktail: high interest rates, economic uncertainty, and firms over-hiring, then slamming the brakes. The study ignores how big shifts in inflation and recession fears shaped these cuts, despite trying to account for rate changes.
Why do tech companies' C-suites say “it’s AI” when it isn’t?
Because blaming AI is easier PR than saying “our board forced us to cut.” Most CEOs don’t know what AI can actually do—job cuts are about cost, not capability.
Is this a temporary dip or a permanent shift?
If companies just use AI as a crutch to stay lean, this could become the new normal for entry-level jobs. Unless something changes, fewer juniors now means fewer seniors later, period.
Is it just tech and US firms, or everyone?
No, this is broader. The drop hit multiple sectors—everything from finance to marketing. Wherever early-career tasks are routine and budget-sensitive, juniors felt the axe.
AI is a convenient excuse now. But it’s not the reason. By the way, you want to save this one first, I promise it’s worth your time.
Shall we?
Here’s a video version:
The Harvard Study (In Plain English) Part 1.
Now, you might say:
But Jing, their data is interest rate adjusted!
That’s correct.
However, there is a huge difference between the interest rate and inflation. These are correlated but not always aligned (not immediately),
Rewind a bit…
Let me walk you through what the Harvard researchers actually did, because understanding the method helps you see why I feel skeptical and am scrutinising their conclusions.
The Data
They used LinkedIn-based resume data covering:
285,000 U.S. firms
62 million workers
150 million employment records (2015-2025)
200 million job postings (2021-2025)
Each worker’s position has a standardized seniority level: entry, junior, associate, manager, director, executive, senior executive.
They collapsed it into two groups: juniors (entry + junior) and seniors (associate and above).
Identifying AI Adoption Specific Roles
This is the clever and also the first problematic part.
Instead of asking “Did you adopt AI?” (which gets you PR nonsense), they looked at LinkedIn job postings. Specifically, postings seeking “GenAI integrators”, roles explicitly tasked with implementing generative AI into company workflows.
The process:
Flag LinkedIn job postings containing terms like “ChatGPT,” “large language model,” “prompt engineering,” “retrieval-augmented generation,” “LangChain,” etc.
Then they further fed flagged postings to a large language model to identify GenAI integrator roles, filtering it down to 131,845 postings (0.066% of all postings).
What do those “GenAI integration” postings actually mean?
The image attached includes two examples of the GenAI Integration roles. The highlight shows how these roles are already beyond the Grads’ capability.
The GenAI integrator jobs tracked by this study are not entry-level, even when labeled “junior.” “GenAI integrator” in their definition requires actual implementation work—prompt engineering, product workflow design, and application development.
For this Junior Product Manager example, it requires experience integrating with GenAI security and safety products, not a recent graduate.
If your company has product managers (many confuse them with project managers—there’s an enormous difference), you’ll know that being successful as a product manager absolutely requires work experience.
This distinction matters because, by their definition, they are likely failed to capture the broadest graduate/ junior labor market.
The Harvard Study (In Plain English) Part 2.
This Harvard study undoubtedly used some really sophisticated methods.
Unfortunately, the sophistication ≠ correct.
They employed a few main statistical approaches to make their case, and each one made an ungrounded assumption when you understand what’s really happening underneath.
I won’t go through all the figures; let’s just focus on the two main statistical operations.
Finding #1:
Junior employment dropped 9% at AI-adopting firms starting Q1 2023, while senior employment climbed.
Finding #2: Using triple-difference analysis (comparing juniors vs seniors, adopters vs non-adopters), they found a 10% relative drop in junior hiring.
So they claim
“Beginning in 2023Q1, the coefficients drop sharply, reaching roughly a 10 percent decrease after six quarters.”
The Fatal Timeline Problem
Here’s what destroys their argument:
March 2022: Fed starts hiking rates from 0% (we will cover this more later)
November 2022: ChatGPT launches
Q1 2023: Harvard sees “AI impact” begin
The study assumes firms posting AI jobs in 2023 are different because they have adopted AI.
But what if they were already different in ways that made them vulnerable to rate hikes?
Consider two companies in 2022:
Company A (Future AI Adopter):
Venture-funded startup
Burned cash assuming 0% rates forever
Heavy debt load
When rates hit 5%: Funding dries up → slash juniors → post “AI transformation” jobs to look innovative
Company B (Non-Adopter):
Profitable, boring business
Much less debt, cash reserves
Never needed cheap money
When rates hit 5%: Business as usual
This team is comparing apples to oranges.
The “AI adopters” could very possibly be those firms most exposed to the rate shock.
They didn’t cut juniors because of AI; they cut juniors because they were hemorrhaging cash, then blamed AI because it sounds better than “we need to tighten our pants.”
The exact same patterns are observed in 1990, 2000, and 2008.
The Pattern You’ve Seen Before (And Here We Are Again in 2025)
Here’s what actually worries me: We keep doing this.
Insanity is doing the same thing over and over again and expecting different results.
This should give a clue to any researcher’s work on this topic. It bothers me… people that smart, why do so many of them try so hard to find a different reason to justify 2025’s hiring pattern?
You likely heard of Occam’s razor. Named after this guy:
Essentially, the simplest explanation for a phenomenon is the most likely to be correct.
In this case, you do not need complex data collection and stats formulas to force the association between AI and the decline of junior roles on the market.
Every economic shock, every time, the same pattern. Junior hiring craters. It stays depressed. And when it comes back, if it comes back, it never returns to the old baseline.
1990-1991
This is a screenshot from the Congressional Budget Office report in 2005.
They saw that the 1990-1991 recession was characterized by a high rate of destruction of existing jobs, which persisted into the early stages of the recovery.
Of course, they saw a slightly different pattern in the 2001 employment recession trend, but it still also follows a cycle, surprise!
The Dot-Com Crash (2001)
In the late 1990s, tech companies hired aggressively with minimal qualification barriers, leading to the hiring of many “marginal developers” who might not have otherwise qualified for tech roles.
When the dot-com bubble burst in 2000, the speculative frenzy ended abruptly. Then it’s history, there was no Merry Christmas in the year 2000.
What followed was a prolonged and severe contraction in employment.
Entry-level positions disappeared as companies retained senior staff and eliminated junior roles.
New college graduates who entered the job market between 2001 and 2003 faced a brutal reality: 6–12 months of unemployment became standard, with many spending even longer searching for work.
Unemployment rises in the expected manner when the economy hits a recession. But in the current downturn, the increase for young college graduates was larger than in past cases, as unemployment grew from an historical low point of 1.7% in 2000 to 3.1% in 2002… implying that this group (grads) has been disproportionately hurt by the weak labor market.
When hiring resumed around 2004–2005, companies demanded “experienced” candidates. Read this report.
Effectively locked out graduates who hadn’t secured positions before the crash. Many from this cohort never recovered; they either left tech entirely or remained in lower-paying, less desirable roles, creating a “lost generation” of workers whose career trajectories and earnings never caught up.
The Great Recession (2008-2012)
Worse than the dot-com crash.
Youth unemployment in the U.S. hit 19.1% in July 2010, the highest rate on record since 1948.
In the UK, youth unemployment peaked at 22.5% in 2011. Entry-level opportunities collapsed across all sectors.
The lasting damage, however, wasn’t the unemployment rate itself but the duration of its effects.
When tracking recession-cohort graduates, multiple official reports and studies reveal profound long-term consequences.
In the UK, those starting careers during the downturn experienced approximately 6% reduction in real hourly pay even one year after leaving education, with wages not recovering for up to six years. U.S. research found that even 15 years after graduation, recession graduates still earned 2.5% less than those who entered during better economic times.
Beyond wages, the impacts extended into fundamental life outcomes.
Research tracking the 1982 recession cohort through midlife found that those who entered the labor market during downturns experienced measurably higher divorce rates and greater rates of childlessness. Most strikingly, they suffered increases in mortality that appeared in their late 30s…
To sum up,
Every Recession’s Playbook:
1990-1991: “It’s globalization” → Juniors cut 40%
2001-2003: “It’s offshoring” → Youth unemployment 2x
2008-2012: “It’s automation” → Entry-level -50%
2022-2025: Same patternAnyway, back to today.
The 2022-2025 Freeze
This is the fourth iteration (of what I have written) of the same movie.
Starting in 2022, tech hiring froze.
Tech job postings fell 36% below pre-pandemic levels as of mid-2025.
Entry-level postings dropped hardest, with roles for workers with less than one year of experience declining 50% between 2019 and 2024.
The proximate cause was the Federal Reserve’s aggressive interest rate campaign. The Fed raised rates from near-zero in March 2022 to 5.25-5.50% by July 2023, a 5.25 percentage point increase in just 16 months, ending the longest period of cheap money in modern history.
What companies do when money is cheap?
They borrow, they expand, they spend to earn.
And what would happen when the money is no longer cheap, and when they want to borrow, they need to pay actual interest, which would bruise their business?
They cut costs.
Where to cut? Again, I was mainly an operational product person, which meant my sole responsibility was to tell our leaders how much I’d saved for the company every quarter. The easiest is automation and postponing hiring.
Junior roles hiring is undoubtfully the easiest target.
Job openings and hires are dropping to their lowest level since February 2021. The information sector saw job openings fall by 2.3 percentage points (same as you’ve seen in the dot-com), the steepest decline of any industry.
Unemployment for recent graduates began exceeding the overall unemployment rate for the first time in decades.
In 2023 alone, 226,000 tech workers were laid off, 40% more than the 164,744 laid off in 2022.
What else also happened in 22’? Yes, ChatGPT was released in November 2022.
So many prestigious institutions flood in to justify how ChatGPT (LLM) caused the junior hiring numbers to drop.
This theory itself is as obscure as saying how Salesforce’s stock price is driven by how many people have the first name Sunny.
The Real Pattern
Every major economic shock since 1980 has resulted in:
Disproportionate cuts to entry-level hiring
Permanent shifts downward in youth labor force participation
Long-term wage scarring for affected cohorts
Failure to recover to pre-shock hiring levels even during expansions
We’ve done this four times now in 25 years.
Each time, we invent a new story about why this time is different.
Maybe it is different.
Maybe the future AI really is replacing these jobs permanently.
But we said that about offshoring in 2001 and automation in 2008, and the real culprit both times was just economic cycles and managerial cowardice.
You can’t separate AI effects from macro effects when:
The timing of both shocks overlaps nearly perfectly (2022-2023)
The macro shock (interest rates + inflation) always affects juniors more than seniors
Firms have strong incentives to blame AI (sound strategic) rather than admit they over-hired during easy money
The researchers analyzed a period where both were happening simultaneously, and their statistical methods cannot tell them apart.
The Concerns (And Why This Study Bothers Me)
Harvard isn’t alone.
WEF has a report. MIT has studies. Nearly every top-tier institution you can think of is racing to publish the “AI killed junior jobs” argument.
Concern #1: A Study from a prestigious university only makes a wrong claim more influential when it’s wrong.
The Harvard study isn’t wrong because the data is not sufficient or the methods are sloppy. It’s wrong because the fundamental identification assumptions don’t hold, and people look up to their theory.
This is the problem: when a Harvard study says something, 90% of people assume it’s gospel. Institutional names are like designer labels; people stop asking if the product quality matches the price tag.
To be fair, this team did lay out the caveats in the body of the paper.
They admit their definition of “adoption” is a blunt tool, their timing is fuzzy, and they can’t fully rule out other confounders, like big tech firms suddenly finding religion about costs after binge-hiring in 2021-22.
But the mass media, pundits, and LinkedIn influencers do not read the footnotes.
They don’t care about the caveats, or they can’t tell the difference between a hypothesis and a proven theory.
As soon as Harvard/Stanford/MIT says “AI killed junior jobs,” the nuance gets lost, the story goes viral, and the institutional brand makes the claim stick, whether or not the correlation holds up.
Concern #2: AI exposure ≠ AI replacement.
The researchers use an index to classify which jobs are “exposed” to AI, such as writing, data analysis, customer service, and admin work. High-exposure jobs, in theory, can be automated by LLMs.
But “can be” never meant “has been”.
Let me emphasize this: Seniors with experience can correct AI’s bullshit faster than juniors can.
They know when the output is wrong. Juniors don’t have the industry experience to make the judgment call, which makes AI less useful for them, not more.
But AI’s error rate is still too high to replace a human entirely. If AI can do a 10-minute task on its own, great. But most junior work isn’t 10-minute tasks, it’s contextual understanding, context-switching, and figuring out what the senior actually meant when they said “make this better”.
The truth about junior roles is that they are responsible for a lot of the work —just the stuff seniors don’t want to do. When the economy is good, you hire juniors to do it; they can learn from the job and become a senior in three years.
Unfortunately, when the fund dries out, you force seniors to do it themselves, maybe with AI making it 10-15% faster.
I’ve written about this before. 95% of enterprise AI pilots fail to show any measurable business impact. MIT report found that despite $30-40 billion in spending, only 5% of AI projects achieve any ROI.
And what do they actually use AI for?
Summarizing emails and taking meeting notes. Most people find it useless after an hour. The best-performing AI models in enterprise settings achieve 76% accuracy at best. Would you trust a bank clerk who’s wrong 24% of the time?
AI adoption stays at the surface level.
The adoption is not fundamental enough to replace entire roles.
At best, it helps seniors work slightly faster on tasks they don’t want to do, the grunt work they used to delegate to juniors.
Basically, it’s “the CFO cut headcount, and now the senior is working longer hours with a chatbot.”
Concern #3: This lets companies off the hook for a structural problem.
When you blame AI for something that’s actually a macro-economic, over-hiring, cost-cutting spiral, you spend resources solving the wrong problem.
What worries me the most is that we’re going to “fix” a non-existent AI problem while ignoring the real issue.
What Actually Matters
I’m not going to waste your time with “embrace lifelong learning” platitudes. For new grads or for those with children around this age.
Here’s what would actually help:
For Individuals (Especially If You’re Stuck)
Get any experience you can
I’m sorry that you were in a difficult situation like this. There's no other thing I can recommend aside from taking on anything —internships, freelance, volunteer work —anything that lets you point to “I did this thing in a real context.” The gap between “college grad” and “1 year of experience” is enormous right now.
Close it however you can.
The data shows recession-cohort grads suffer wage scarring for 15+ years.
The best protection is getting in before the door fully closes, or finding another door.
Target smaller firms and non-tech companies
Small firms, non-tech companies with tech needs, startups, they’re still hiring because they didn’t over-hire in the boom, and their money is tight, so there’s no sugarcoat in that they pay much less, sure.
But they’re a door in.
Stop blaming AI for your struggles
I know it’s easier said than done. What about “I’m competing with 500 other people for every entry role because the economy isn’t ideal?”
Blaming never helps, I know, but we all do because it gives us some certainty and makes us feel better.
At least blame the right things, so you direct your anger to the root cause and have a chance to fix the actual problem.
The reality is that most companies aren’t using AI nearly as much as they claim. They’re just not hiring because budgets are tight. That can change with time.
To end this…
You want to know what keeps me up at night? It’s not AI.
It’s this:
We’re in the middle of the fourth major “juniors get screwed” cycle in 25 years. Dot-com crash, Great Recession, and now this.
Each time, we invent a new story about why it’s different.
Offshoring, housing collapse, and AI. Each time, the real cause is the same: economic shock + companies invest less in developing talent when money is tight + use AI as a scapegoat because it sounds better than no budget.
And each time, we lose a chunk of a generation’s career trajectory. As I said earlier, the ramifications of unemployment for a few years will last a lifetime.
They start late, they earn less, they have worse health outcomes, and they don’t develop the skills or relationships they would have. And ten years later, we wonder why we have a “skills gap” and “leadership pipeline problem.”
Again, this Harvard study is well-executed research.
No matter how well-designed and constructed a house is, if it's on a shaky foundation, it's going to collapse and hurt the people who believed in the house.
The authors are smart, careful, and honest about limitations.
But the framing—” GenAI as Seniority-Biased Technological Change”—is going to get picked up by every tech publication and every board deck as proof that cutting junior hiring is inevitable and necessary.
It’s not. It’s a choice.
Companies chose to over-hire in the boom. They chose to cut juniors first in the bust. They’re choosing to keep hiring frozen even as the economy stabilizes. And they’re choosing to blame AI instead of admitting they’re just not willing to invest in training people.
AI might be making that choice easier to justify. But it’s not making the choice for you.
You are.
