AI Hiring Needs a Baseline and a Co-Pilot Round

Treat the new format as an experiment

Jane Street’s November 2, 2026 post describes a limited test, not a wholesale redesign: the firm began trying some campus onsite rounds with less code-writing and more software design, review of code quality and intent, and LLM-produced code. It says most rounds remain traditional coding interviews without LLM assistance, and that the new formats are still evolving S1. That caveat matters. This is evidence that one firm is exploring a split, not evidence that an AI-enabled interview is a better predictor of job performance.

A single blended score can hide what the panel has learned. Did the candidate decompose the problem and catch a bug, or did the model supply the hard part? Did the candidate steer the tool effectively, or did a good output happen to arrive? Those are different signals, and teams should decide which the role needs before giving them one combined grade.

Score judgment, not AI theater

Canva’s October 20, 2025 interviewer guide offers a practical, explicitly older example of an AI-assisted rubric. It describes candidates clarifying scope and planning before implementation, using AI for routine work, pausing to check unfamiliar or complex suggestions against constraints, and explaining their choices afterward S2. That is Canva’s stated approach, not comparative validation that the format predicts stronger hires.

The useful signal is not prompt volume or tool-brand fluency. It is whether the candidate can define the problem, direct assistance toward a bounded task, verify the result, notice when it misses a requirement, and own the trade-offs. A polished transcript may show activity; it cannot substitute for a working solution and a candidate who can explain it.

Make the loop legible

For a lean hiring loop, use two short, disclosed stations: a small task without generative AI, then a distinct but comparable task with it. Tell candidates the permitted tools and constraints in advance. Keep the rubrics separate: score decomposition, correctness, and debugging in the baseline; score scoping, tool direction, verification, and explanation in the assisted round. Do not let a strong result in one erase a weak signal in the other.

This split is a recommendation, not a settled standard. Its value is that it makes the decision auditable: a panel can say whether it is hiring for independent problem-solving, AI-mediated delivery, or both—and see which evidence supports each judgment. If interviewers cannot agree on what they observed, adding AI has only made the score harder to interpret.