If you have taken or run a technical interview in the last few months, you have probably felt the tension. The interviewer looks into the webcam and tries to tell whether your eyes are reading hints from Claude on a second monitor. Meanwhile, the candidate writes code on a virtual whiteboard that they would generate in three seconds at a real job.
Tech hiring is going through its biggest change in years.
Hiring developers, especially for classic Frontend and Backend roles, has changed for good. What worked for the last 15 years has stopped working because of AI tools. Below, we look at why the old hiring process broke, how Meta, Google and Coinbase rebuilt their interviews, and how to prepare for the new AI round.
A take-home task for a frontend developer ("build a small dashboard with state management") or a backend developer ("write a small REST service with authentication") used to be the gold standard: low stress and close to real work.
This stage has lost its value. In 2025, Gergely Orosz described the crisis of trust in remote interviews in his article “Tech hiring: is this an inflection point?” (The Pragmatic Engineer). Companies are flooded with AI-generated résumés, and in one case he described, 100% of candidates did the take-home task with an LLM.
How did the company find out? The candidates did not even remove typical model "hallucinations": fake endpoints and extra abstractions that looked neat but made no sense for the task. And when they were asked on the next call to explain a specific block in their own code, they froze.
The classic round where you write a single function from scratch also stopped giving an honest signal. Desktop AI assistants can now read the screen or listen to the interviewer in real time and show a ready answer on top of the browser window. A normal screening turned into a game of cat and mouse.
For Frontend and Backend engineers, this made the old tasks pointless:
The industry had a choice: bring everyone back to in-person interviews at the office, or change the rules. Leading product companies chose the second option. Their reasoning was simple: if generating code costs almost nothing, a developer's main value becomes verification, debugging and architecture.
The numbers support this. According to a Karat survey of more than 300 CTOs and SVPs (January 2026), 73% believe a strong engineer is now worth at least three times their salary, and 59% say a weak engineer in the AI era brings zero or even negative value. AI does not make developers equal. It makes the gap between them bigger. A classic interview does not see this gap.
Companies no longer look for people who can type syntax quickly from memory. They look for engineers who can read complex code written by others, spot hidden bugs (and AI creates plenty of them) and make technical decisions.
The solution was simple: allow AI in the interview, but change the tasks completely. In the new round, you no longer write code from scratch in an empty file. You work in an existing codebase, often broken or slow, and you are allowed to use AI.
In October 2025, Meta began replacing one of its two classic coding rounds with a 60-minute hands-on session with AI in CoderPad (according to the Aced guide). The other classic round, without AI, stays. The candidate works in an environment with three panels:
What's the catch? The AI cannot edit your files. It only answers in the chat. In 60 minutes, the candidate goes through three phases: find and fix a bug in unfamiliar code, implement the core logic and optimize it.
Google is still testing the new format as a pilot for junior and mid-level roles in some US teams. The candidate can use Gemini during the interview, but instead of writing code from scratch, they read, debug and improve existing code. Google evaluates "AI fluency": how the candidate writes prompts, checks the model's answers and debugs.
If you paste the task into the chat without asking the interviewer any clarifying questions, the model can easily lead you the wrong way. The focus has moved from writing code to reading someone else's logic quickly and judging AI suggestions critically.
In the article “Interviewing Engineers in the AI Era: Lessons from a Year of Rebuilding”, Coinbase writes that almost all new code at the company is now generated by AI and reviewed by people. So they rebuilt the interview to check how a candidate guides AI, evaluates the result and makes decisions where the model is wrong. The main live signal now is coding and debugging in a real repository. System Design with AI is still in early testing.
This round tests the areas where AI makes the most mistakes: the boundaries between components and what happens under the hood.
The most common mistake in this round is blind vibe coding in front of the interviewer. To show strong engineering skills, follow four rules:
null comes in?" Never use generated code that you can't explain in detail.The new format is good news for experienced developers. Algorithm puzzles matter less, and real engineering work matters more: reading code, debugging and making decisions, with AI as one of the tools.
How do you practice for this round? We explain it in "Tech Interview Prep in 2026: Why P2P Mock Interviews Work".