Demo Day: Prevent and Detect Cheating in Your Tech Recruitment Process

Executive summary (business focus)

This session is a product demo/panel discussion on how to prevent and detect cheating in technical recruiting using:

• Codepad Screen (async assessments)
• Codepad Interview (live interviews)

It argues that cheating risk is high (including AI-assisted behavior), but that strong process design—especially realistic, job-relevant questions and structured monitoring—improves hiring signal and fairness.

Core framing: poor questions → poor answers → poor hiring signal (“shitty questions get shitty answers”).

---

Key ideas & tactics (what to do differently)

1) Improve interview signal by changing question quality

• Avoid generic “leak code / toy problems” that lead to memorized / internet / AI-generated answers.
• Use relevant, realistic, on-the-job tasks so candidates demonstrate true working capability.
• Use multi-part questions that build complexity progressively and include “nudges” that make one-shot AI responses less likely.

2) Set clear expectations with candidates (reduce “gray area”)

• Communicate early what’s allowed (e.g., external IDE) and what’s not (e.g., using LLMs, Googling).
• Use customizable candidate-facing communication templates/emails included in the product settings.

3) Use a layered cheating detection “stack”

In both async (Screen) and live (Interview), the demo emphasizes multiple controls working together, rather than relying on a single feature.

---

Frameworks / playbooks mentioned (as process guidance)

• Cheating-prevention layered approach (defense in depth):

  • Expectations → environment controls → real-time alerts → post-interview audit trail → interviewer training + follow-ups

• Pair-programming-style interview goal (explicit target):

  • Bottom-of-funnel live interviews should approximate pair programming embedded in the team context

• Interviewer training playbook:

  • Train interviewers to ask in-depth follow-up questions and force candidates to explain their reasoning line-by-line / deeply

---

Product features & operational controls (as described)

Codepad Screen (async assessment)

Key controls/settings demonstrated:

• Timer controls

  • Per-question timers and/or a global time limit

• Unusual activity alerts

  • Flags anomalies like unusually fast completion or odd behavior patterns

• Copy/paste controls

  • Can block copy and paste within the assessment environment

• Follow-up question generation (real-time)

  • If unusual behavior is detected, the system generates chat-based follow-up questions that ask candidates to elaborate on their solution

• Webcam proctoring

  • Periodic snapshots and flags unusual events such as:
    • candidate leaving view
    • multiple people appearing

• Full-screen mode

  • Locks the candidate to the IDE/assessment environment

• AI analysis / AI-assisted suspicion detection

  • Uses screenshots and AI to detect “strange” behavior (as described)

Detection/audit outputs demonstrated:

• Candidate dashboard indicators

  • “red exclamation” type signals for unusual activity

• Detailed report

  • Shows question-by-question performance plus an event timeline (“scrub through events”)

• IP tracking

  • Example signal: candidate starts in one location and completes in another

• Code similarity check

  • Surfaces duplicated code across submissions (signals may include LLM usage, Stack Overflow copying, or cross-candidate sharing)

• Keystroke-by-keystroke playback

  • Shows external paste events, time spent outside the environment, etc.

---

Codepad Interview (live interview)

Live interview features emphasized as overlapping but not identical:

• Similar toggles for:
  • copy/paste events
  • left-IDE events
  • suspicious behavior alerts

Additional live-experience operational controls:

• Waiting room

  • Prevents candidates from seeing questions early

• Ending pads

  • Immediate termination + auto-ending in bulk to prevent later reuse/leak of live assessment content

• Transcription + AI notetaker + summary

  • Helps interviewers focus while enabling post-call review

• AI assist integration

  • Used both to support candidates and reduce incentives/ability to leave the IDE for external tools—while maintaining transparency via playback of prompts/queries

• Follow-up questions

  • Designed to let interviewers “dive deeper” when candidates can’t explain decisions

---

Concrete examples / demo scenarios (actionable signals)

1. Real-time detection cancels a cheating attempt

   • Anecdote: a candidate using “InterviewCoder” on Codepad triggered real-time cheating detection and the interview ended.

2. ChatGPT-assisted shortcut detected

   • In the Screen demo, the panelist:
     • ran with ChatGPT on the side
     • pasted code externally
     • left the environment most of the time
   • Resulting signals:
     • external paste events
     • unusually fast completion patterns
     • generated follow-up question demanding explanation
     • candidate couldn’t explain (used as an authenticity/quality signal)

3. Location-based cheating trace

   • Example: started in Valencia; later completed in New York
   • Signal: IP tracking and completion trail

4. Post-interview audit summaries

   • Example summary format:
     • “X left pad events total of Y minutes”
     • “one paste of N characters”
     • other notable events (waiting room, etc.)

---

Key metrics / KPIs mentioned

• Cheat behavior prevalence in interviews (over ~100,000 interviews / “past few weeks”)
  • 50% of pads had left pad events
  • 20% of pads had external paste events

Note: not all events mean cheating (some can be legitimate), but disabling visibility prevents teams from knowing what’s happening.

(No revenue/CAC/LTV/churn metrics were provided; this was product/process-focused.)

---

Actionable recommendations (directly stated)

• Turn on multiple integrity features

  • timers, unusual activity alerts, copy/paste controls, webcam proctoring (if appropriate), full-screen mode

• Customize candidate communications

  • clearly define allowed/disallowed tools (LLMs, external IDE, Googling)

• Use follow-up questions when anomalies occur

  • require candidates to explain to validate authenticity

• Ask project-based, realistic questions

  • ideally problems relevant to a candidate’s first-week experience

• Train interviewers to go deeper

  • follow-ups are a primary defense against AI-memorized answers

• Adopt waiting room and auto-ending for live links

  • reduces information leakage and reuse of assessment content

---

Handling AI use (positioning)

• The panel suggests AI cheating is real and tools can be misused today.
• They do not claim to catch “deepfake candidates” specifically; they argue:
  • deepfaked avatars are less common, and other unusual behaviors would likely trigger alerts first
• Proposed modern approach:
  • incorporate AI use as part of the test (via AI assist) and measure transparency and capability rather than only trying to “fight AI”

---

Presenter(s) / sources

• Caroline Matz — Marketing team (moderator)
• Frank Hen — VP of Product
• Nathan Sutter — VP of Engineering
• Katherine — supporting moderator in the chat (named as available in background)