The Borrowed Ladder: How AI Is Changing Everything for New Graduates and Interns in Tech

Entry-level tech hiring is collapsing. Junior engineers who use AI don't understand their own code. And the industry won't feel the full consequences for another decade. Here's the honest picture.

A Deceptively Promising Start

Imagine starting your first data engineering job in 2025. You open your laptop, face a task that would have taken a week to learn two years ago, fire up GitHub Copilot or Cursor, and have working code in an hour. Your senior colleagues are impressed. Your manager notes your output. You feel productive, capable, ahead of schedule.

Now imagine it's three years later. A production pipeline fails in a way nobody can immediately explain. Your team looks to you — a now mid-level engineer — to trace the problem through five interconnected systems. And you realize, with growing dread, that you've never actually had to do that before. Every hard problem was quietly solved by AI before it became a learning experience.

This is the paradox at the center of the new graduate and intern experience with AI tools — and it's one the industry is only beginning to talk about with honesty.

The Numbers That Tell the Story

The data on AI's impact on junior engineers and new graduates is striking, sometimes alarming, and almost uniformly underreported in the enthusiastic coverage of AI productivity gains.

Entry-level tech hiring decreased 25% year-over-year in 2024. Entry-level postings dropped 60% between 2022 and 2024. Google and Meta are hiring approximately 50% fewer new graduates compared to 2021. Salesforce announced it would halt junior hiring for 2025 entirely, citing AI productivity gains from existing senior staff. A Stanford University study found that employment among software developers aged 22 to 25 fell nearly 20% between 2022 and 2025, coinciding directly with the rise of AI-powered coding tools.

A 2024 survey of hiring managers found that 70% believe AI can perform the work of interns. 57% of those surveyed said they trust AI's work more than the work of interns or recent graduates. The logical, if brutal, conclusion from those two findings: why spend the time and budget training a student when AI can produce similar outputs immediately?

Internship postings on Handshake, a leading internship recruitment platform, dropped 30% since 2023 — while internship applications rose 7%. Supply of eager graduates increasing. Demand collapsing.

The computer science degree, once a near-guarantee of early career stability, has shifted dramatically. Computer science graduates now face a 6.1% unemployment rate, and computer engineers face 7.5% — among the highest across all university majors. Stories of graduates sending hundreds of applications, failing to receive interviews, and pivoting to industries they never intended to enter have moved from anecdotes to a documented pattern.

The Myth of the Great Equalizer

The most seductive story about AI coding tools and junior engineers is that AI levels the playing field — that a fresh graduate with GitHub Copilot can perform like a five-year veteran. Organizations circulate this narrative because it justifies both the cost of AI tool subscriptions and the reduction of hiring budgets. It feels democratic. It feels exciting.

The research says otherwise, consistently and clearly.

A Fastly survey of 791 professional developers found that senior developers ship AI-generated code at nearly two and a half times the rate of junior developers — 32% versus 13% of all shipped code is AI-generated among seniors versus juniors respectively. The reason is not that seniors use AI more aggressively. It's that seniors can evaluate the output. They have the experience to recognize when code "looks right" but isn't — when the logic is subtly broken, when the edge case isn't handled, when the security assumption is wrong.

One senior developer in the survey summarized it plainly: "AI will bench test code and find errors much faster than a human, repairing them seamlessly." A junior developer in the same survey described the flip side: "It's always hard when AI assumes what I'm doing and that's not the case, so I have to go back and redo it myself."

The gap extends into production behavior. Among junior developers, just 13% say over half of their shipped code is AI-generated. Among senior developers, 32% say the same — nearly two and a half times higher. Senior engineers are not only using AI more, they are trusting it more in high-stakes environments. That trust is earned by the very experience that junior engineers haven't yet had time to build.

The analogy that captures this best: senior engineers are using AI the same way a great chef uses a knife — faster, safer, more precise. Junior engineers are using a knife they haven't been taught to handle. The output can look identical until the moment it matters.

The 17-Point Comprehension Gap

There is a finding from Anthropic's own research that deserves to be at the center of every conversation about AI tools in engineering education and early career development. Research shows a statistically significant 17-point comprehension gap when junior developers learn with AI assistance — 50% code understanding versus 67% — with a Cohen's d of 0.738, indicating a substantial effect.

What this means in practice: junior engineers who rely on AI to generate their code understand what they've built significantly less than peers who write it themselves. And understanding your own code is not an academic nicety. It is the fundamental precondition for debugging it, extending it, explaining it to stakeholders, and taking professional responsibility for it.

The traditional learning pathway in engineering was, in important ways, built on productive struggle. Debugging a system that won't cooperate forces you to develop a mental model of how it works. Reading documentation until you understand an error message teaches problem-solving patterns that transfer to novel situations. Sitting with a failing test and working through why it fails builds the instinct that separates an experienced engineer from a fast typist.

AI interrupts this process by handing you the answer before the struggle can create understanding. The output is faster. The learning is less. When this happens at scale, across a generation of engineers entering the field, the downstream consequences are serious.

The Succession Planning Problem Nobody Is Talking About

Here is the structural problem that follows from the two findings above, and it operates on a timeline that makes it easy to ignore today and very difficult to fix later.

The cohort of engineers currently entering the industry with heavy AI assistance will become mid-level engineers in 2027–2029 and senior engineers in 2029–2032. If those engineers have never developed foundational debugging skills and architectural judgment — because AI handled those problems before they became learning moments — the industry will face a genuine quality and capability crisis at the exact moment when AI systems are expected to be most deeply embedded in critical infrastructure.

As one engineering leader stated directly: "The concern isn't that AI eliminates jobs, but that it eliminates learning pathways."

The mathematics of succession planning makes this concrete. Organizational capability depends on a continuous pipeline — juniors become mid-level engineers after three to four years, mid-level become seniors after five to seven years. Companies that dramatically cut junior hiring today are not saving money. They are borrowing from the future. In five to ten years, when those companies need senior engineers with architectural judgment, debugging instincts, and the ability to mentor the next generation, they will find the pipeline empty.

AWS CEO Matt Garman called the idea of replacing junior developers with AI "one of the dumbest things I've ever heard" — specifically because of this succession risk. Senior engineers retire. They move on. They burn out. If there is no junior pipeline being developed into the mid-level and senior engineers of the future, organizations will face a talent crisis with no path out.

The companies that have already understood this — the ones investing in junior hiring and deliberate mentorship even in the AI era — are making a bet that will compound positively. The companies eliminating junior roles for short-term efficiency are making a bet that feels smart in 2025 and will feel catastrophic by 2033.

Senior vs. Junior: The Real Performance Picture

The question of whether AI equalizes senior and junior performance is worth examining carefully, because the answer is more nuanced than either "yes" or "no."

In a large-scale 2024 field experiment by researchers from MIT, Harvard, and Microsoft, covering 4,867 professional developers across Microsoft, Accenture, and a Fortune 100 firm, AI-assisted developers completed 26% more tasks on average. Critically, the productivity boost was largest for junior and newer hires — the group that lacked prior context used AI to scaffold and accelerate their work most effectively. Senior developers, already familiar with the codebase, saw little or no measurable speed increase.

This seems to support the equalizer narrative. But the study measured task completion speed, not code quality, not long-term maintainability, not production incident rates.

A separate analysis by METR tested senior engineers in large, familiar codebases. In that environment, minutes saved on boilerplate were erased by time spent reviewing, fixing, and discarding AI output. Senior engineers who already knew the solution found AI added friction rather than acceleration.

A July 2025 analysis of over 10,000 developers across 1,255 teams by Faros AI found that teams with high AI adoption interacted with 47% more pull requests per day and 9% more tasks — but developers were juggling more parallel workstreams, a pattern historically correlated with cognitive overload and reduced focus. More volume, more context switching, not necessarily better outcomes.

And by early 2026, 35% of professional developers acknowledged that AI tools still struggle significantly with complex tasks — architecture decisions, security considerations, edge case logic, performance under real-world conditions. These are precisely the tasks that define senior engineering work and that junior engineers are supposed to be learning.

The honest summary: AI helps junior engineers produce more output. Senior engineers produce better output with it. In production environments, better consistently beats more.

The Vibe Coding Trap

There is a new archetype emerging in engineering teams that experienced engineers describe with a mixture of understanding and alarm: the engineer who can write a great prompt but cannot explain a stack trace.

This person has real skills — they are fluent in AI tools, comfortable with fast iteration, adept at generating working prototypes. What they cannot do is debug a system that AI didn't build, reason through a failure mode that wasn't in the training data, or explain to a stakeholder why a data pipeline produced incorrect results for six months without anyone noticing.

"Vibe coding" — accepting AI output based on surface plausibility without deep understanding — looks productive until it doesn't. What it builds is a slow-burning accumulation of technical debt and undocumented assumptions. The structural analogy is good: building quickly on a weak foundation. The faster you build, the higher the structure gets before the foundation becomes the problem.

Junior engineers who fall into the vibe coding pattern may plateau at mid-level permanently — able to produce code but unable to take ownership of systems, mentor colleagues, or solve the class of problems that require genuine architectural judgment. The AI that accelerated their early career output may have quietly capped their long-term growth.

What the Best Junior Engineers Are Doing Differently

The picture is not hopeless. It does, however, require intentionality that earlier generations of junior engineers didn't need to think about consciously, because the environment enforced the learning.

The junior engineers who will build genuine expertise in the AI era share a common set of practices that distinguish them from peers who are building faster toward a ceiling.

They treat AI output as a starting point, not a finished product. Every AI suggestion is a hypothesis to be evaluated, not an answer to be submitted. They read the code. They change something. They understand why the change matters. They break it deliberately and watch what happens.

They do hard things without AI intentionally and regularly. When a mentor asks them to implement a module without AI assistance first, they don't see it as a punishment. They understand that the struggle is the product — that debugging an unfamiliar system for two hours teaches more than shipping ten AI-assisted features ever will.

They invest in the skills that AI cannot replace. Systems thinking — understanding how data flows through an entire architecture, where it can fail, where regulatory obligations attach. Communication and stakeholder management — the ability to translate between business requirements and technical reality. Judgment under uncertainty — the ability to evaluate confident-sounding answers critically and know when something "looks right" but isn't.

They ask why, constantly. When AI generates a solution, they ask why that approach rather than another. When a senior engineer reviews their code, they ask what the review was looking for. They are building the mental model that AI skipped over.

What Companies and Mentors Need to Understand

The engineering leaders and senior engineers who will navigate this period well are the ones who understand that mentorship in the AI era requires different strategies than mentorship in the pre-AI era — not less, more.

If AI handles the foundational tasks, aspiring engineers will not automatically build the foundational understanding. Mentors now need to be proactive about filling gaps that the learning environment no longer creates automatically. When a junior engineer submits AI-generated code that uses dynamic programming, the mentor's job is to ask: do you understand that algorithm? If not, the conversation needs to happen before the PR is merged, not after.

Code review in the AI era needs to evaluate comprehension, not just output. "Does this code work?" is no longer sufficient. "Do you understand why this code works?" is the question that matters for the long-term development of the engineer.

Training and onboarding programs need deliberate "AI-off" exercises alongside AI-assisted work — not as punishment or nostalgia, but as structured opportunities to build the understanding that AI assistance tends to shortcut.

The companies that approach junior development with this seriousness will build stronger senior engineers in five years. The companies that let AI do the teaching — which, as the research shows, it does poorly — will find themselves unable to explain why their engineering capability has quietly degraded.

A Message to New Graduates and Interns

If you are starting your career in data engineering or AI engineering right now, the landscape is genuinely more difficult than it was for the generation two years ahead of you. The entry-level market has contracted sharply. The competition for the positions that remain is intense. And the tools that make you more productive in the short term carry risks for your long-term development that are not obvious until they become problems.

Here is the most useful reframe: your goal is not to maximize output in your first two years. Your goal is to build understanding that will still be valuable in your eighth year, when you are the person the team looks to during the production incident at 2 AM.

Use AI aggressively and intelligently. It is a genuine advantage for someone who knows how to use it. But treat it as a sparring partner, not a ghostwriter. Question every suggestion. Break the code and rebuild it. Ask your senior colleagues not just "does this work?" but "is this the right approach, and why?"

The engineers who will define the field in ten years are not the ones who produced the most AI-assisted output in their junior years. They are the ones who built real understanding alongside the output — who learned what the AI was doing and why, and developed the judgment to know when it was wrong.

The AI is a tool. Expertise is still the goal. Those two things are not the same, and the gap between them is something only you can close.

Conclusion: Who Builds the Next Generation?

The industry is at a genuine inflection point. AI is creating short-term productivity gains that are real, measurable, and compelling. It is simultaneously eliminating many of the traditional mechanisms — entry-level roles, grinding through hard problems, learning by breaking things — through which the engineering expertise of the future has always been built.

The question is not whether to use AI. That question is settled. The question is whether organizations, mentors, educators, and junior engineers themselves will take responsibility for ensuring that AI accelerates real learning rather than replacing it.

If the answer is yes — if the industry invests in mentorship programs that adapt to the AI era, maintains junior hiring pipelines with deliberate development embedded in them, and creates cultures where understanding is valued alongside output — then the AI era will produce a generation of engineers more capable than any before it.

If the answer is no — if companies cut junior roles for short-term efficiency and assume AI will handle the rest — the industry will spend the 2030s discovering that it optimized for output and sacrificed capability, and that there is no fast fix for that mistake.

The ladder that every senior engineer climbed was built by the people who came before them. The question now is whether we are building that ladder for the next generation, or whether we are borrowing it.

Search This Blog

Data and AI