How to Detect and Prevent Cheating in Technical Interviews
YöneticiNovember 20, 202612 viewsblog
Contents8
Software engineering recruitment faces a new integrity problem. With large language models and AI coding assistants widely available, algorithm tests and coding interviews that were trusted for years can be answered with polished, AI-generated solutions in seconds. This guide explains the warning signs interviewers can watch for, how human-written and AI-generated code tend to differ and, most importantly, how to design technical interviews that still reveal a candidate’s real ability.
The silent partner
Many candidates no longer take coding tests alone. When an AI assistant works in the background, the interview stops showing how the candidate solves problems under pressure, which is exactly what the interview was meant to measure. A bad hire in an engineering role is costly: delayed projects, production incidents and wasted onboarding.
Why traditional technical interviews are failing
Technical confidence can be masked
The gap between a candidate’s communication skills and their real system design ability can now be concealed with AI coding tools. “Does this code work?” is no longer a useful interview question. The critical question is: “Does the candidate understand why this code works?”
The flawless syntax fallacy
Someone with only basic programming knowledge can pass senior-level algorithm questions by copying AI output. AI does not make the small typos and missing semicolons that usually reveal how a person actually works.
Warning signs of AI assistance in interviews
When a candidate relies on hidden help, the traces often show up not only on screen but also in gaze, voice and the way they explain their work. None of these signals is proof on its own; they are reasons to ask follow-up questions.
Linear reading patterns
A person thinking through a problem looks around, pauses and moves unevenly across the screen. A candidate reading text from a second screen or phone tends to move their eyes horizontally and rhythmically. Reflections in glasses can also be revealing.
Voice and keyboard mismatch
Reading prepared text often sounds flat compared with explaining one’s own ideas. Long, flawless code blocks appearing without typing sounds may indicate pasted code.
Hesitation on “why” questions
If a candidate falters when asked why they chose this algorithm or data structure, or how they would explain its time complexity, the code may not be their own.
Human-written vs. AI-generated code
AI tools tend to produce the most “average”, rule-following code; people under time pressure are pragmatic and iterative. Comparing writing style in code is known as code stylometry. Typical differences:
Aspect
Human developer under time pressure
AI-generated code
Naming
Short or contextless names such as usr_dt or count
Sterile, overly descriptive names such as calculateTotalUserCount
Comments
Only for important design decisions
Long docstrings even for trivial functions
Writing rhythm
Pauses, typos, deletions and step-by-step changes
Large blocks appearing at once, without edits
Edge cases
Considered based on experience with real systems
Code may run but miss concurrency, error handling or edge cases
Stylometry is a hint, not a verdict. The most reliable check is still a conversation in which the candidate explains and changes their own code.
The role of proctoring and human review
For online coding tests and remote interviews, a proctoring layer makes the session observable without turning the interview into an automatic judgement:
Identity verification: Confirms that the person taking the test is the candidate who applied.
Browser and focus events: Tab switches and leaving the exam window are recorded and can be raised with the candidate during the interview.
Behavioural flags: Looking away from the screen repeatedly, a second voice or a second person in the room are flagged for review.
Human decision: A reviewer looks at the evidence and decides; flags are not proof of cheating.
In Qulak, coding and scenario questions run in the exam platform, and identity verification, locked browser and AI flagging are added through AI exam proctoring with QProctor, Qulak’s proctoring partner. Video-based technical interviews can be structured with video interview management.
How to write AI-resistant technical questions
Multi-file questions: Instead of a single function, give candidates a small project of four or five files. Working within a broader structure requires understanding the context, which makes copying much harder.
Time-boxed, multi-stage tasks: Test how candidates prioritize when several tasks compete for limited time. Prioritization depends on judgement, not typing speed.
Critique deliberately flawed code: Give candidates code that is inefficient or incorrect and ask them to debug it and explain the logic out loud.
Measuring AI fluency instead of banning AI
Some technology companies now allow AI openly in parts of the interview, because developers use these tools in real work. The question then becomes how well the candidate collaborates with AI:
Prompt quality: Does the candidate give the tool context and edge cases, or just say “solve this”?
Delegation: Does the candidate hand routine code to AI while keeping security, architecture and design decisions for themselves?
Verification discipline: Does the candidate test and question AI output, or paste the first answer into the system?
Whichever policy you choose, state it clearly to candidates in advance and apply it consistently.
The cost of undetected cheating
Security vulnerabilities and technical debt
An engineer hired on the strength of answers they did not understand may ship code whose logic nobody on the team can explain. Months later, this can surface as production outages, security vulnerabilities and hidden technical debt.
Fair hiring and equal opportunity
Long take-home assignments place a heavy burden on candidates and are easy to outsource to AI. Shorter, live interactive sessions such as pair programming are harder to fake and fairer, because they take place within working hours and measure the same thing for everyone.
Design technical assessments you can trust
Share your engineering roles with the Qulak team; let’s plan coding questions, structured interviews and the right level of exam security for your hiring process.
How can you tell if a candidate is using AI in a coding interview?
Common warning signs are linear eye movements while reading from another screen, long flawless code blocks appearing without typing and hesitation when asked to explain design choices. Browser focus events recorded by a proctoring layer add context. None of these is proof on its own; follow-up questions and human review are essential.
How can AI-generated code be recognized?
Code stylometry looks at naming, comments and writing rhythm. People under time pressure use short names and comment sparingly; AI output often has long, sterile names and lengthy docstrings even for simple operations. The most reliable check is asking the candidate to explain and modify the code.
How can coding questions be made resistant to AI?
Use multi-file project structures instead of single functions, time-boxed tasks that require prioritization and deliberately flawed code that candidates must debug and explain verbally.