AI has made shallow engineering look convincing!
Scenario: Two engineers, with same briefs, same tools, and same available time.
The first opens their IDE, types "build me a payment service with a REST API and a database," accepts what comes back, tidies the formatting, and submits.
The second thinks and then writes: “I need to build a payment service that handles idempotent requests to prevent duplicate charges, communicates with a separate ledger service via an event stream rather than direct calls, and fails safely if the ledger is unavailable. What are the failure modes I should design for before writing any code?”
The prompt is the signal, and almost no interview process is currently designed to capture it.
Truth hiring teams are not saying out loud
AI is compressing the visible gap between shallow engineering competence and deep engineering competence. A junior engineer with good prompting habits can now produce output that looks, on the surface, indistinguishable from the work of someone with ten years of production experience. The code compiles with passed tests and valid API responses, but mostly happy-path.
What you cannot see from the output alone: whether they understood what they were building, designed for the failure modes that might surface in production, or whether the system they produced will hold up six months from now when someone needs to extend it.
This is the anxiety every hiring manager in a technical organisation is sitting with right now, even if they are not saying it directly. The interview process that worked in 2019, leetcode style end-to-end code generation for a very specific problem, is not sufficient. Because writing code is no longer the problem.
The question is not whether to allow AI in interviews, it is whether the interview can tell the difference between an engineer who directs AI with judgment and the one who follows it blindly.
The distinction that changes everything
There is a meaningful difference between using AI to generate code faster and utilising AI to think through systems more thoroughly, and it matters more than most hiring conversations acknowledge.
Generating a function with Copilot is a productivity tool. Most can do it these days, but it doesn’t answer if (or what) the engineer understands about the system they are building.
Constraint-driven prompting is different. It is knowing, before even a line of code exists, what the system needs to handle, what it needs to survive, and what it needs to tell you when it fails.
The best signal in an interview is identifying who is asking the right questions and reasoning well with AI before asking to build anything.
When companies say they want engineers who can work with AI, what most of them mean is engineers who can code faster, while they actually need is engineers who can think better with AI, especially in regulated, high-stake environments.
Interview strategy that CAN work well
1. Build live for a product requirement
Give the candidate a small product/user requirement, two hours of pair-session, & access and encouragement to company-approved AI tools.
The output matters, but there is something else to judge for. The prompts. Watch what they ask AI, do they immediately move to design, or discuss constraints first? Or if AI produced something, how do they verify to catch flaws and fix it.
This constraint-driven prompting cannot be faked. Candidate either understands the system/product before AI generates it, or not. The prompt is the first draft of the design and it tells you more about their mental model than the code that follows.
2. Take-home assessment + extension session
A take-home assessment: build a 2-service system from a product requirement. Provide reasonable scope, AI-tools access, and it needs to be submitted before the interview.
Then a 60-minute live session with one purpose: extend it.
Introduce a new requirement, a new service that needs to integrate, a constraint that changes the existing design, or a failure scenario the current system cannot handle, and ask them to extend what they built.
This would reveal whether they designed for change or just for today. You need an engineer who were already thinking about the system’s future, not just its present state.
The extension session also reveals how they reason under pressure with their own work. Do they defend original decisions or adapt? Do they understand why their design makes the extension hard, or are they surprised by it?
3. Code review of an AI-generated system
Give them a system, two services, a REST API, a data model, an event stream, all generated with AI and contains deliberate, realistic failure modes. Nothing that would fail a basic test, but everything that would cause a production incident in a regulated system.
Ask them to review it.
What they catch, what they miss, and how they articulate the risk tells you more than any coding test. A strong engineer will not just identify the problems, they will explain why each one matters in the specific context of the system, what the production consequence would be, and what they would do instead, even if it is just prompting to AI.
4. AI-assisted debugging and recovery
Give them a running system that is behaving incorrectly, and ask them to diagnose it with available AI tools.
What they ask as clarifying questions on the system, and then ask the AI to investigate, in what order, and how they interpret what comes back reveals whether they understand the system or are following suggestions. A shallow engineer accepts the AI’s first diagnosis and acts on it. A deeper one interrogates it, checks it against what they know about the system or the domain, asks the AI to consider alternative causes, and knows when it is pointing in the wrong direction.
Debugging reveals something the build formats cannot: how they reason when the system is not behaving as expected, and whether their first instinct is to trust the AI or verify it.
What deep engineering looks like when AI is in the room
1. Prompt is the mental-model
Not what they build, but what they ask before they build. An engineer who opens a session with constraints, failure modes, and explicit requirements has already done the design work. The AI is executing their thinking, not replacing it. An engineer who opens with a broad request is outsourcing the thinking. The difference is visible in the first thirty seconds and almost impossible to hide over two hours.
2. Domain judgment is something AI can never really close.
In regulated environments like fintech, AI might not know the whole context. The engineer who brings that context, who knows what to constrain and what to verify, is demonstrating something no amount of prompting skill can manufacture. It comes from having been in the room when these systems failed. AI cannot simulate that experience, it can only reflect what it has seen in code that someone else already wrote.
3. AI’s first answer is not the right or full answer
Strong engineers treat AI output as a draft, not a solution. They interrogate it, extend it, and push back when it misses something. The engineer who accepts the first response without verification is the engineer whose system will have the missing dead-letter queue, the mutable audit log, the duplicate charge that nobody designed against. The habit of verification is not something you can teach in an interview, but you can observe whether it is already there.
4. Operational thinking beyond building
Does the system they build tell you anything when it breaks? Logging, monitoring, alerting, graceful degradation, an AI will generate a working system that is completely opaque in production. Whether the engineer adds or asks AI to add the observability layer is one of the clearest signals of how they think about the difference between code that runs and systems that operate well with live customers.
The close
Interview processes have always been designed around the output entirely created by the candidate. That world is gone!
The engineers who should be hardest to hire, the ones who direct AI with judgment, who design for failure modes before writing code, who catch what the model missed before it ships, are not being reliably selected for by any process most companies are currently running.
That is not a candidate problem, but a process problem. And the companies that solve it first will not just hire better, but will build better.
