Take-home coding assignments are dead. Candidates paste them into ChatGPT, polish the output for ten minutes, and submit work that looks indistinguishable from genuine expertise. The hiring industry knows this, but most companies have responded with one of two bad options: drop technical assessments entirely or force everyone into high-pressure on-site whiteboard sessions.
A detailed new framework from engineering blog Incremental Forgetting proposes a third path: stop fighting LLM usage in interviews and start evaluating how candidates use these tools.
Go Deep on Real Projects
The strongest tactic is deceptively simple. Ask candidates about something they actually built, then keep drilling. What constraints shaped the architecture? What broke in production? What would they do differently?
Someone who lived through a project can riff on these questions for thirty minutes without breaking a sweat. Someone parroting LLM-generated answers will start contradicting themselves or retreating to vague generalities by the third follow-up. LLMs are great at plausible surface-level explanations but terrible at maintaining coherent depth across a winding conversation about real trade-offs.
Make Candidates Read and Review Code - With AI Help
The other two tactics flip the typical coding interview on its head. Instead of asking candidates to write code, hand them a large, messy codebase with outdated documentation and multiple languages. Let them use whatever AI tools they want. The evaluation shifts from "can you produce code" to "can you understand code and spot where the AI is wrong."
The code review variant is even sneakier: present a pull request that was partially generated by LLMs, complete with plausible-sounding but inaccurate comments. Candidates who blindly trust the AI's explanations reveal exactly the kind of uncritical thinking you want to filter out.
What You're Actually Measuring
The framework boils down to four things worth evaluating:
- Technical judgment - Does the candidate verify and cross-check what the LLM tells them, or do they accept the first answer?
- Problem decomposition - Can they break a fuzzy problem into specific, answerable questions for the AI?
- Domain knowledge - Do they bring real expertise that supplements what the model produces?
- Communication - Can they explain their reasoning to another human, not just copy-paste a model's explanation?
This approach makes practical sense. The daily reality for most developers now involves working alongside AI tools. An interview that bans AI usage is testing for a job that no longer exists. An interview that embraces it and watches how candidates navigate AI's blind spots is testing for the job people actually do.
The catch: these interviews are harder to standardize and require interviewers who genuinely understand the technical domain. You cannot hand a junior recruiter a rubric and expect this to work. That is probably why most companies will keep doing what they are doing - but the ones that adopt something like this will have a real edge in identifying who can actually think versus who can prompt.