How does interviewing change when take-homes and LeetCode no longer signal anything AI can't do?

The shift

Producing a plausible, working solution to a bounded, well-specified coding problem — a take-home, a LeetCode puzzle — goes from a scarce signal of individual skill to an abundant output any candidate can generate near-instantly by routing the problem through a model. The interview stops observing capability and starts observing tool access, which every candidate already has.

The axioms

• Take-homes signal the ability to independently build something without supervision — rests on scarcity of unsupervised, correct output.
• LeetCode-style puzzles signal algorithmic fundamentals and problem-solving speed — rests on scarcity of fast, correct pattern-matching under time pressure.
• Take-homes filter for effort and genuine interest, because faking hours of work is costly — rests on scarcity of candidate time as an honest signal.
• Live/whiteboard coding proves the code produced in front of you is the candidate's own — rests on scarcity of a controlled, observable environment.
• A resume and portfolio of past projects signal real experience — rests on scarcity of being able to fabricate a convincing body of work.
• System design rounds probe judgment on ambiguous, open-ended tradeoffs — rests on scarcity of experience-based judgment that isn't just a lookup.
• Behavioral interviews reveal how someone actually operates on a team — rests on scarcity of a consistent, hard-to-fake self-narrative under follow-up.
• References and work-history checks confirm claims are true — rests on scarcity of independent, accountable corroboration.

Invalid axioms

1. Take-homes signal independent ability to build something. Any candidate can paste the spec into a model and get a working submission with minimal understanding of it. The scarcity — unsupervised production of correct code — is gone; take-homes now measure who's willing to use AI well, not who can code without help, and most candidates use it whether allowed or not. Habit-trap: teams still weight take-home quality heavily in hiring decisions, and still burn candidate and reviewer hours grading submissions that no longer distinguish anyone.
2. LeetCode puzzles signal algorithmic problem-solving. Current models solve the vast majority of standard LeetCode-tier problems, including many "hard" ones, on the first try. The puzzle no longer separates strong engineers from weak ones — it separates people who've memorized patterns (human or model-assisted) from people who haven't. Habit-trap: companies still gate onsite loops behind LeetCode rounds and calibrate levels/comp off them, as if solving them were still rare.
3. Take-homes filter for effort as an honest signal of interest. Effort used to be expensive to fake, so investing it implied genuine commitment. AI collapses the cost of producing a plausible artifact, so a polished take-home no longer implies hours spent or interest shown. Habit-trap: recruiting pipelines still treat take-home completion as a costly, meaningful filter step rather than nearly free to clear.

Unchanged axioms

1. Live problem-solving with a human watching reveals how someone actually thinks. Verification of reasoning in real time — can you explain the tradeoff, catch your own bug, respond to a changed constraint — stays scarce. AI can generate an answer, but a live interviewer probing "why did you choose that, what breaks it" is testing judgment under follow-up, not output quality, and that's still hard to fake for a sustained session.
2. System design and ambiguous-tradeoff conversations test judgment on novel stakes. There's no clean lookup for "given this org's constraints, this traffic pattern, this team's skill gaps, what do you actually build." AI can supply options; deciding which one fits a specific, unstated context and defending it under pushback stays a human skill.
3. References and accountable work history confirm someone can be trusted with real responsibility. A model can't vouch for someone or be liable if the vouching is wrong. Corroboration from people who worked with the candidate, and willingness to put a name behind a recommendation, stays scarce and load-bearing.
4. Someone still has to be accountable for what the hire actually does on the job. Interviewing exists to reduce the risk of a bad hire; that risk, and who owns it, hasn't gotten cheaper just because code generation did.

New axioms

1. Distinguishing "used AI well" from "used AI as a crutch" during the interview itself. If everyone has model access, the useful signal shifts to how someone directs, checks, and pushes back on the tool's output — but almost no interview format is built to observe that yet.
2. Grading at the volume AI-assisted applications create. When applying (and passing early screens) gets cheaper for candidates, application volume rises, and manual review of take-homes or even resumes stops scaling — someone has to solve for triage at a volume the old process never anticipated.
3. Verifying that a live-coding performance reflects durable skill, not rehearsed pattern recall inflated by AI-assisted practice. Candidates can now drill against AI-generated variations of every common interview question at scale, which was expensive before and is nearly free now — the interview has to solve for a rehearsal advantage that didn't previously exist at this scale.

Where it breaks

Companies still price and sequence hiring loops as if take-homes and LeetCode were the expensive, discriminating stage (INVALID) while giving the cheapest amount of time to live judgment and system-design rounds, the parts that actually still discriminate (STILL HOLDS) — the loop spends the most effort on the round that now tells you the least.

Take-home grading at rising application volume (NEW) collides directly with "take-homes filter for genuine interest" (INVALID): the filter that was supposed to reduce reviewer load by weeding out low-effort candidates now increases it, because AI makes it cheap for every candidate to clear a bar that used to require real investment.

ai-axioms.vercel.app/axioms/engineering/swe-interviewing-ai-axioms