Tech Interviews in the AI Era: The Rules Are Being Rewritten

I recently read Coinbase Engineering’s Interviewing Engineers in the AI Era, an account of the year the company spent rebuilding its technical interview process.¹ That led me to review the latest practices and industry data from Canva, Meta, Google, Datadog, and others. One conclusion stood out: now that AI-generated code accounts for more than half of everything Coinbase merged in Q4 2025, the traditional whiteboard-and-LeetCode interview is rapidly losing its effectiveness. I previously explored the anxiety AI creates for knowledge workers in AI Anxiety for Knowledge Workers: Burnout & Mental Health. This article zooms in on one of the places where that anxiety becomes most concrete: the technical interview.

At its core, the interview dilemma in the AI era is a signal problem. The old signal—whether someone can write correct code from memory—is becoming distorted, while the new signal—whether someone can direct, review, and overrule AI—is still being calibrated. This article examines three major fault lines in that transition: the changing nature of the job, the cheating arms race, and the arrival of AI interviewers. It also identifies the skills candidates should focus on building.

TL;DR: Four Core Takeaways

• Interviews have not kept pace with the job: AI-generated code now accounts for more than half of Coinbase’s merged code, yet many interviews still test whether candidates can write a solution from scratch and from memory in 45 minutes. That raises the risk of both false positives and false negatives.
• Suspected cheating has reached scale: detection vendor Fabric’s proprietary model flagged 38.5% of interviews for suspected cheating. A new generation of cheating tools claims to keep shared screens completely clean, creating an escalating contest between detection and evasion.
• Companies are taking two paths: one returns to in-person interviews to verify fundamentals; the other permits AI and redesigns the assessment. The approaches are not mutually exclusive, and a growing number of companies use both.
• LeetCode and system design are not dead: LeetCode-style questions remain common, although how they are assessed is changing. System design remains an important dividing line for senior roles, while AI collaboration is a new requirement that candidates at every level should start building now.

1. Engineering Work Has Changed, but Interviews Have Not Kept Up

Coinbase’s July 2026 article, Interviewing Engineers in the AI Era, revealed a striking set of figures. AI-generated code represented only 5.7% of everything merged into its codebase in Q1 2025. By Q4 2025, it had crossed 50%, while humans still reviewed virtually 100% of the code.¹ Engineers are shifting their time away from writing code from scratch and toward authoring specifications, directing AI, reviewing AI-generated pull requests, and catching architectural mistakes that models produce with confidence.

“As building costs approach zero, choosing what to build, validating it, and shipping it safely become the bottlenecks.” — Coinbase engineering leader

Coinbase is not an isolated case. Google DORA’s 2025 survey of nearly 5,000 technology professionals found that 90% already use AI at work, up 14% from the previous year. Common uses include writing and modifying code, testing, and documentation. Karat’s 2026 survey of 400 engineering leaders in the United States, India, and China found that AI increased engineering productivity by 34% on average. The gains, however, are distributed very unevenly: AI is widening the gap between strong and weak engineers rather than leveling it.²

That raises an uncomfortable question: if nearly all the code a company ships is “AI-generated and human-reviewed,” why do interviews still ask candidates to write code from scratch, from memory, under a 45-minute deadline?³

2. Why Traditional Interviews Are Failing: The Signal Is Breaking Down

Coinbase’s internal analysis of its old interview process identified two failure modes pulling in opposite directions:

• False positives: candidates who have memorized standard solutions for URL shorteners and distributed queues can pass, even though what they have demonstrated is test preparation rather than the judgment the job requires.
• False negatives: candidates with strong architectural judgment but less deliberate pattern memorization can fail, even though they are often exactly the engineers best suited to an AI-native environment.

More surprisingly, Coinbase found an 84% correlation between two different interview rounds. In other words, two costly rounds produced almost the same signal—and neither produced the signal the company needed most.¹

Karat’s data supports the same conclusion: take-home projects and automated coding tests lose their ability to distinguish candidate skill fastest in the presence of AI because they evaluate only the final output, which AI can produce in seconds. Live interviews, by contrast, are becoming more valuable because they reveal how candidates break down a problem, make decisions, and collaborate with AI. Live technical interviews are used by 79% of companies in the United States and 87% in China.²

3. The Cheating Arms Race: Detection and Evasion

The other side of the deteriorating signal is suspected cheating at scale. The nature of the following data matters: Fabric’s figures come from interviews conducted on its own platform and were classified by its own detection model, using a threshold above a 40% predicted probability of cheating. Fabric also sells cheating-detection products. These numbers should therefore be treated as vendor-reported statistics, not as an independently confirmed cheating rate.^{4 5}

Cheating tools such as Cluely, Interview Coder, and Leetcode Wizard claim to use GPU overlays at the DirectX or Metal layer, placing answers beneath the screen-capture layer so that the screen shared with an interviewer appears clean. Conventional tab-change detection, browser lockdown, and ordinary screen sharing may not detect these low-level overlays. Most of these technical descriptions come from cheating-detection vendors and have not yet undergone comprehensive, independent technical audits. Companies are responding with behavioral analysis, in-person verification, and redesigned assessments. The latest tools even use an earpiece connected to an off-screen LLM, leaving no visible trace on the display.⁴

Consequences are beginning to surface. Media reports describe candidates who used AI to deceive interviewers but were exposed after starting work. These reports are mostly anonymous cases and observations from hiring managers, so they cannot establish how often this occurs. What is clear is that companies increasingly view the gap between interview performance and on-the-job ability as a cost to productivity and morale.⁷

4. Two Ways Companies Are Responding

Path 1: Return to In-Person Interviews and Strengthen Integrity Checks

• Google: Google CEO Sundar Pichai said the company would ensure “at least one round of in-person interviews” to verify fundamentals. Nearly all software engineering openings have returned to at least one in-person round.⁸
• Amazon: candidates must formally acknowledge that they will not use unauthorized generative AI tools during interviews. The policy retains explicit exceptions where use is authorized, and violations may result in disqualification.
• Anthropic: candidates may use AI to polish application materials and prepare for interviews, but AI is prohibited by default in live interviews and standard take-home assessments unless a prompt explicitly permits it. Interestingly, some technical assessments intentionally allow AI but still require frequent updates as Claude’s capabilities improve.
• McKinsey, Deloitte, and Cisco: these companies have added in-person verification for certain roles and stages, rather than requiring every role to be entirely in person. Deloitte had already reinstated an in-person stage for UK graduate recruitment in 2024.⁹
• Other measures include SAT-style proctoring, cameras aimed at the candidate’s workspace, and keyboard- or eye-tracking software.

Path 2: Embrace AI and Redesign the Interview

• Coinbase: the company rebuilt its interviews in three phases: a frontend pilot in the second half of 2025, expansion to repo-based backend exercises in January 2026, and a company-wide “AI Fluency” signal in March 2026. It defines AI fluency across three dimensions: Usage, Application, and Understanding Limits. The process now focuses on three baseline signals: repo-based coding and debugging, AI-assisted system design—which remains in early testing—and leadership and behavioral interviews. Coinbase also insists on not adding interview rounds.¹
• Canva: the company states, “Yes, You Can Use AI in Our Interviews. In fact, we insist.” Backend, machine learning, and frontend candidates are expected to use tools such as Copilot, Cursor, and Claude. Evaluation shifts toward decomposing ambiguous requirements, using AI strategically, and critically reviewing and repairing AI-generated code.¹⁰
• Meta: in 2025, the company officially confirmed that it was testing coding interviews in which candidates could use AI assistants. Third-party candidate guidance from 2026 indicates that some interview environments now offer multiple models, but Meta has not published a complete policy that applies worldwide.¹¹
• Datadog: the company has published formal AI-use guidelines for candidates. Certain roles have an explicit “AI coding interview”; AI remains prohibited by default in other live coding interviews, making the boundary transparent.¹²
• Karat’s “human-led, AI-enabled” model: a live interviewer observes candidates in an AI-enabled environment, assessing how they collaborate with AI, validate its output, and handle hallucinations. 62% of companies still prohibit AI during interviews, but the trend is clearly moving toward permission: 38% in the United States and 68% in China already allow it.²

The two paths are not mutually exclusive. Many companies combine a return to in-person interviews with permission to use AI: the in-person component verifies identity and fundamentals, while AI access makes the assessment reflect real work.

5. AI Is Also Taking the Interviewer’s Seat

The transformation runs in both directions. Job seekers are using AI to submit applications at scale, while the number of applications on LinkedIn rose 45% in one year to an average of 11,000 per minute.¹³ Recruiting teams are being flooded with AI-generated resumes, so companies are responding with AI screening and interviews of their own:

• Greenhouse’s 2026 survey covered 2,950 active job seekers in the United States, the United Kingdom, Ireland, Germany, and Australia. Among US respondents, 63% said they had participated in an AI interview, an increase of approximately 12–13 percentage points in six months.¹⁴
• The experience was poor, again primarily among US respondents: 38% had abandoned a hiring process because it included an AI interview, while another 12% said they would do so if asked. Of those who completed an AI interview, 51% never received any outcome notification.
• Candidates do not necessarily oppose AI itself; they oppose its opaque use.

The result is a paradoxical arms race: AI writes resumes, AI screens resumes, and AI conducts interviews. Candidates use AI to respond, and companies deploy more AI to detect AI. Both sides use AI while accusing the other side of using it.

flowchart LR
    A[Candidates use AI to mass-produce resumes] --> B[Recruiters receive 11,000 applications per minute]
    B --> C[Companies use AI for screening and interviews]
    C --> D[Candidates use AI to answer or even cheat]
    D --> E[Companies use AI to detect AI]
    E --> A

6. What Interviews Measure Is Shifting from “Recall” to “Judgment”

Across these companies, the capabilities that interviews in the AI era are converging on can be summarized as follows:¹⁵

1. The ability to direct AI: prompt quality, task decomposition, and knowing when to use—or not use—AI.
2. Review and judgment: distinguishing a correct solution from one that merely looks correct, and catching architectural mistakes that AI produces confidently.
3. Working in an existing codebase: triaging issues, debugging code written by someone else—or by AI—and understanding dependencies and rollback risk, because real work almost never begins with a blank slate.
4. Fundamentals still matter: Canva explicitly says code fluency and technical depth remain absolute requirements; they are simply evaluated in a different setting. Without strong foundations, a candidate cannot review AI effectively.
5. Taste and judgment: knowing what is worth building, which trade-offs are real, and when to overrule a model. Coinbase sees these qualities as what every interview signal is ultimately trying to measure.

7. The Skills Candidates Should Focus on Building

Instead of worrying about every change to the rules, candidates can invest their time in the capabilities that still matter.

mindmap
  root((Candidate skill map))
    Coding fundamentals
      LeetCode-style questions remain common
      Shift from speed to depth of reasoning
      Practice aloud instead of silently
    System design
      Explain trade-offs clearly
      A dividing line for senior roles
      Use AI as an adviser but retain final judgment
    AI collaboration
      Prompt precisely
      Validate output line by line
      Catch AI mistakes
    Code reading and debugging
      Understand existing codebases
      Diagnose problems from error messages
    Verbal explanation and communication
      Explain the reason for every decision
      Show why you accepted or rejected AI advice

1. Does LeetCode Practice Still Matter? Yes, but the Assessment Is Changing

“LeetCode is dead” is a common claim, but perhaps that conclusion is premature. The quality of the available evidence should be stated honestly. Multiple 2026 interview-preparation sources and candidate reports indicate that LeetCode-style questions remain common at large technology companies. Techinterview.org, which tracks the evolution of technical interviews, and TechScreen both say that most coding stages still contain one or two LeetCode-style questions. TechScreen describes underpreparing on the assumption that “LeetCode is dead” as one of the costliest mistakes a candidate can make.^{16 17}

These sources do not publish transparent, reproducible cross-company methodologies—and TechScreen itself sells interview-assistance products—so they demonstrate that the format remains common but cannot establish that it dominates at every company. One stronger piece of evidence is Google’s pilot interview format, which confirms that solid data structures and algorithms fundamentals remain necessary even when AI is allowed.

What has changed is how candidates should practice:

• Shift from speed to depth of reasoning: a few years ago, speed was the differentiator. Interviewers now care more about whether candidates can articulate their reasoning, answer follow-up questions, and adjust a solution when constraints change.
• Shift from silent problem-solving to practicing aloud: a common recommendation in interview-preparation communities is to solve fewer problems silently and spend more time explaining a solution while developing it, because today’s interview is more like a technical conversation. This is experience-based advice, not an empirically validated quantitative finding.
• Make LeetCode part of the strategy, not the entire strategy: Educative co-founder Fahim ul Haq reaches a sensible conclusion: LeetCode practice is still worthwhile in 2026, but it should support a broader preparation strategy rather than become the only one.¹⁸

2. Does System Design Still Matter? It Remains a Critical Divider for Senior Roles

As AI takes on more hands-on implementation, decisions about architecture and trade-offs become more valuable. Many experienced interviewers argue that system design matters more than ever. That conclusion needs a caveat: its growing importance in the AI era is a reasonable inference from changes in engineering work, not yet a quantitative finding supported by longitudinal, cross-industry data. The evidence we can confirm is:

• Coinbase includes “system design with AI” among the three baseline signals in its new interview process and explicitly asks candidates to demonstrate judgment while working with AI during design.
• Former Facebook interviewer Arslan Ahmad recalls that senior candidates whose design skills fell short could receive a down-leveled offer despite passing the coding bar—for example, E4 instead of E5.¹⁹
• The Pragmatic Engineer notes that coding rounds apply at every level, while system design is a decisive stage for senior positions and carries more weight as seniority increases.²⁰

Instead of memorizing a standard architecture diagram, practice explaining trade-offs: Why this database? What fails first if traffic grows tenfold? What happens during an outage? Also practice using AI as an adviser during the design process while retaining final judgment yourself.

3. AI Collaboration: A New Core Skill

Google’s 2026 pilot for some US teams hiring early- and mid-career engineers offers especially clear evidence. Candidates may use a company-provided Gemini assistant in a new “code comprehension” round, while interviewers directly assess AI fluency, including prompt quality, output validation, and debugging. This is still a limited pilot and does not mean that every Google software engineering interview has changed, but the direction is clear.^{21 22} Candidates can prepare in several ways:

• Use AI for real tasks in everyday work and build the habit of checking its output line by line.
• Practice precise prompting: provide enough context, state constraints clearly, and define the expected result instead of vaguely asking AI to “take care of it.”
• Practice finding AI’s mistakes. A study covering 16 models and 576,000 code samples found an average “package hallucination” rate—references to nonexistent packages—of at least 5.2% among commercial models and 21.7% among open-source models. Actual rates vary substantially by model, language, and prompting method.²³ The ability to spot such errors is precisely what interviewers want to see.

4. Code Reading and Debugging: More Common Day-to-Day Than Writing from Scratch

Google’s new pilot stage asks candidates to understand, debug, and optimize existing code rather than write from scratch. Coinbase likewise moved to exercises in which candidates find and fix problems in a realistic project. Several 2026 interview guides describe debugging AI-generated code as a new core skill.²⁴ The preparation is practical: read code in open-source projects, practice diagnosing problems from error messages, and review AI-generated code section by section as critically as you would review a teammate’s pull request.

5. Verbal Explanation and Communication: Explaining Matters More Than Merely Solving

An increasingly common integrity check and assessment technique is simply asking, “Why did you design it this way?” Take-home exercises are often followed by live questioning. Practice explaining the rationale for every technical decision, and be ready to show why you accepted or rejected AI’s advice. That reasoning is itself a strong positive signal.

6. Fundamentals Without AI: They Will Still Be Verified in Person

Google’s return to in-person interviews and Amazon’s stricter interview-integrity requirements mean employers will still test what candidates can do without AI. The role of fundamentals has changed: their purpose is no longer to reproduce memorized answers, but to make effective AI review possible. Without a foundation, candidates have no reliable way to decide whether AI output is correct. AI should reinforce weaknesses, not conceal them.

Conclusion

As AI evolves, the interview process is no longer something companies can refresh every few years. It has become a product that requires continuous iteration. Coinbase’s willingness to publish the questions it has not yet answered—including how long interview questions remain effective and how interview performance correlates with job performance—shows how early the entire industry still is. For candidates, the best preparation strategy has never been more honest. The durable advantage they can build consists of three transferable abilities:

1. Break ambiguous problems into verifiable subproblems (specifications, edge cases, and trade-offs)
2. Make decisions under uncertainty and explain the reasoning (taste and judgment)
3. Use AI as an accelerator rather than a crutch

Further reading on this site: AI Anxiety for Knowledge Workers: Burnout & Mental Health

References