How to Assess Junior Engineers Beyond Degrees and Resumes

A degree and resume provide context about a junior candidate. Complement them with a task that requires debugging unfamiliar code, extending a partial system, and explaining decisions. If AI is allowed, assess how the candidate verifies its output.

The university examples below concern particular courses and admissions policies in the United States. They do not measure the competence of all graduates or establish that every degree has lost its value.

The credential just stopped certifying competence

The clearest evidence that the degree is decaying comes from the institutions that issue it. Failure rates in introductory computer science have spiked, prerequisite skills are eroding, and faculty are publicly declaring that their own admission signals no longer work. These are not opinion pieces. They are grade distributions and formal petitions.

35% failure rates: what happened in UC Berkeley CS

In Spring 2026, the failure rate in UC Berkeley’s CS 10 course hit 35.3%, up from under 10% in both Spring 2024 and Spring 2025, according to grade data reported by the Daily Californian. CS 61A logged 10.6% F’s and EECS 127 logged 16.8%. The department guideline is that roughly 7% of lower-division and 5% of upper-division students receive D’s or F’s, so a 35.3% failure rate is about a fivefold deviation from normal.

The average grade in both CS 10 and CS 61A fell to a C+ (around a 2.3 GPA), against a department target range of 2.8 to 3.3. Teaching professor Dan Garcia attributed the collapse to a “vast increase in academic dishonesty” from large language model use, noting that nearly 30 students in CS 10 alone were caught cheating on take-home exams.

Higher failure rates do not demonstrate that students who passed lack competence. In hiring, directly assess the skills needed for the job.

800 professors say the admission signal broke

The erosion is not only at the output. The inputs to the degree are weakening too. A November 2025 University of California San Diego Senate-Administration Workgroup report found that the number of first-year students with math skills below a middle-school level rose nearly 30-fold since 2020, the year UC suspended standardized testing.

More than 800 UC faculty, including seven of the nine math department chairs, signed an open letter calling for reinstating the SAT and ACT math requirement for STEM applicants, as reported by Inside Higher Ed. When the people who grant the credential write a formal petition saying their admission filter stopped working, that is the strongest possible signal that you cannot outsource judgment to the school’s brand.

It is worth being precise about cause here. The timing of the math-preparation decline does not by itself establish that suspended testing caused it. These problems predate public access to ChatGPT. AI is a separate accelerant layered on top: it masks who cannot reason and inflates who looks productive. The two causes are distinct, and they compound.

How AI both masks and inflates junior ability

AI breaks junior screening in two directions at once. It masks weakness, because a candidate who cannot trace a bug or reason about a data structure can still ship plausible-looking code and pass a clean take-home. And it inflates the appearance of productivity, because output volume no longer correlates with understanding.

That is the trap for entry-level hiring specifically. A senior engineer who leans on AI is a strong engineer with a force multiplier; their judgment backstops the model’s mistakes. A junior who leans on AI without judgment is a liability, because they cannot tell when the model is wrong. The whole point of hiring a junior is to bet on someone you can develop. If your screen cannot separate apprentice-ready juniors from prompt-jockeys, you are betting blind.

Why the resume and GPA were always a weak signal

Here is the uncomfortable part: the credential was never a great predictor of job performance. The decades of personnel-selection research are clear that what someone has done on paper predicts far less than what they can demonstrate. The credential collapse did not create this problem. It just removed the last excuse for ignoring it.

Work samples vs. resumes: what the meta-analyses show

The foundational meta-analysis by Schmidt and Hunter (1998) ranked predictors of job performance by validity. Work sample tests landed near the top. The combination of general mental ability and a work sample test reached roughly 0.63 average validity, with the work sample adding about 24% predictive power over a general ability test alone, per the published summary. Structured interviews scored around 0.51. Unstructured interviews, resumes, and years of experience trailed well behind.

For honesty, later research revised the numbers down. Roth, Bobko, and McFarland (2005) re-estimated standalone work-sample validity at around 0.33, attributing the drop to broader use across service-sector roles. So work samples are strong but not magic. The point holds in either estimate: a demonstrated work sample beats a credential proxy, and it beats it by a wide margin. The relative ranking is the part that has survived 25 years of scrutiny.

The hiring science was telling you to weight demonstrated ability over credentials long before AI arrived. Now that the credential certifies less, the case is no longer debatable. It is operational.

The screening response: structured, AI-aware work samples

The right response to credential collapse is not to ban AI or to trust the school harder. It is to assess real ability with a structured, AI-aware work sample and apply the same rubric to every candidate. The major tech companies have already converged on this, and their redesigns are the clearest template for what entry-level screening should look like in 2026.

Debug unfamiliar code, don’t solve clean puzzles

Google introduced a “code comprehension” round in 2026 that it describes as human-led and AI-assisted, where candidates analyze an existing codebase with Gemini available and are evaluated on prompt engineering, output validation, and debugging skills, according to Exponent’s breakdown. For junior candidates, a traditional algorithmic round is replaced with an open-ended engineering problem session.

The shift is from “can you write this function from scratch” to “can you reason about a system you didn’t write.” Interviewers now prioritize whether a candidate can trace through unfamiliar code, form hypotheses about failures, and validate fixes systematically. AI can also analyze unfamiliar code and find planted bugs. The task format alone does not prevent a model from helping. Ask the candidate to explain hypotheses, verify a fix, and adapt to changed constraints.

Score the reasoning and the tradeoffs, not the output

The second shift is what you grade. In the new format, candidates who recognized algorithmic tradeoffs and explained them scored well even when they could not implement both solutions, per Exponent’s analysis of the format. Successful candidates prompt, read, critique, fix, and iterate. The ones who fail treat AI as an answer machine and paste whatever it returns.

For a junior, this is the single most important thing to measure. You are not hiring them for the lines of code they produce on day one; AI produces those. You are hiring them for the judgment that decides whether the code is right, whether the approach scales, and what to do when it breaks. Build your scoring around questions like: Did they validate the AI’s output? Did they catch the planted bug? Can they explain why they chose one approach over another? Did they iterate when the first attempt failed?

Standardize one rubric across the whole junior pipeline

The third shift is consistency. A structured assessment means every junior applicant faces the same task and is scored on the same rubric. This is what turns the “work sample beats resume” finding into an operational default instead of an ad-hoc gut call that varies by interviewer mood.

Standardization matters more at the entry level than anywhere else, because junior candidates have the least differentiated paper. Without a consistent work sample, you fall back on the noisiest signals available: the school name, the GPA, the polish of the resume. The exact proxies that just stopped working. A single rubric applied uniformly is also your defense when leadership asks why “good school, good GPA” is no longer a sufficient filter. You can point to the same scored work sample for every candidate.

Why this matters more now: fewer junior seats, higher stakes

Junior hiring is shrinking, which makes accurate screening more valuable, not less. Entry-level hiring at the 15 biggest tech firms fell roughly 25% from 2023 to 2024, and new grads are now about 7% of Big Tech hires, more than 50% below 2019 levels, according to SignalFire’s State of Tech Talent report.

At the same time, underemployment for recent college graduates sits near record highs, reported around 43% for grads aged 22 to 27 as of late 2025. So the applicant pool is enormous and the average signal per resume is lower than ever. More candidates per requisition, fewer junior seats, and a weaker credential all push in the same direction: every junior hire has to be someone you can actually develop, and your screen has to be good enough to find them in a flood of indistinguishable resumes.

Analysts have started describing early-career engineers as stranded between AI agents and senior workers, because a senior developer plus an AI assistant reduces the need for junior support. That framing is debatable, but the strategic implication is not. The companies that keep building junior pipelines will win the next decade of senior talent, and the ones who screen well will get the apprentice-ready juniors instead of the prompt-jockeys.

How to run this in Kit

In Kit, prepare a task using a GitHub template repo. You control the initial code and instructions, including any bug to investigate or design decision to explain. The format does not prevent AI assistance.

Set the duration, estimated effort, instructions, payment, and grace period. Configure scored criteria in a team-review stage. Comparable assessment still depends on task design and reviewer practice. The Junior Engineer Pipeline template is a starting point to inspect and adapt.

Nothing above makes a credential a good filter, and Kit does not turn it into one. Recommended credentials works the other way round: you name a credential you would value, everyone who applies afterwards gets one optional email with free courses, and anyone who already holds it can send the badge. Kit never scores it, ranks on it, filters by it, or checks it — and a candidate who ignores the whole thing is invisible to you.

If you want the surrounding context, these reads go deeper on adjacent pieces of the problem:

• How to Hire Engineers When Everyone Has the Same AI covers candidate differentiation once every applicant has the same tools.
• Technical assessment beyond LeetCode examines alternative task formats.
• How to Structure Code Assignments Candidates Don’t Hate gets into the mechanics of a good take-home.
• How to Hire a Backend Engineer in 2026 is the adjacent role-specific guide.
• What Is an AI-Native ATS? gives the product context for why this is built in, not bolted on.

FAQ

Should I ban AI in junior interviews?

Choose rules that match what you need to assess. If the team uses AI, include work with that tool. You can separately assess foundations without an assistant, provided the purpose and conditions are clear. Allowing AI does not remove the need to check the candidate’s understanding.

What is a good work sample for an entry-level engineer?

A small, realistic task built on an existing codebase rather than a from-scratch puzzle. Hand them a partial system with a planted bug, ask them to extend a feature, and require a short written explanation of their tradeoffs. Score on whether they validated their approach, caught the bug, and can reason about a system they did not write. Keep it to a few hours, make it paid if you can, and run the same task for every candidate.

Does a CS degree still matter at all?

Yes, as information about education. Selected UC Berkeley course results and a petition about admissions do not determine the value of all degrees. Complement the resume with a job-relevant task and assess both the result and how the candidate learns.

Design an AI-aware code assignment for your next junior requisition in Kit: a template-repo task scored on reasoning, debugging, and tradeoffs, standardized across every applicant. Start your free trial.