InterviewWatch includes a built-in applicant tracking system: post jobs and track candidates. Explore our ATS →
AI tool detection

Catch AI Assistance That Screen Share Can Miss.

Candidates using ChatGPT, Claude, Gemini, coding copilots or a stealth interview assistant rarely share it on screen. InterviewWatch detects AI at the system level, desktop apps, browser tabs, local LLM runners, browser-extension assistants, and the timing fingerprint of an answer read off a second device.

chat.openai.com tabActive
Cursor / WindsurfReview
Ollama (local LLM)Present
AI assistant DNSWeak signal
What we detect

From mainstream chatbots to purpose-built interview assistants.

The goal is not to chase a single brand name. InterviewWatch pairs built-in signatures with a remotely updateable tool catalog, so new AI assistants are covered without waiting for an agent redeploy.

AI chat & search

Web assistants detected by window title and browser context during the session.

ChatGPTClaudeGeminiCopilotPerplexity

Coding assistants & local LLMs

Developer AI and offline model runners that leave no network trace but run as processes.

CursorWindsurfOllamaLM StudioJanGPT4All

Stealth interview helpers

Purpose-built answer tools, hidden overlays and browser-extension assistants that inject answers.

Overlay appsMonicaSiderMerlin
How it works

Four independent surfaces, one reviewable finding.

No single signal convicts. The agent watches four separate surfaces on a two-speed loop, a fast 2-second loop for window and title changes, a slow 5-second loop for processes and DNS, then correlates what fires together into a scored incident.

Process matchCursor · Ollama · LM Studio Window titlechat.openai.com · gemini… DNS resolver cacheMonica · Sider · Merlin Answer timinglong pause → fluent burst Correlateby time · severity· interview context AI_TOOL incidentSeverity: Warning → CriticalScore impact: −40 to −60signed · timestamped

A process hit alone is context; a process hit plus a matching window title plus an answer burst is a strong, correlated finding.

Detection surfaces

Each surface has a different strength.

The strongest evidence is a known AI process or window title. DNS is a supporting signal; timing adds context. InterviewWatch weights them accordingly instead of treating every hit the same.

Process & title matchA known AI desktop app, local LLM runner, or browser AI tab appears during the session.
Strong
DNS egress traceA browser-extension assistant domain shows up in the resolver cache after baseline.
Support
Answer-latency patternA long reading pause followed by a fluent, low-variance typing burst, the fingerprint of reading an answer.
Context
Why screen share is not enough

What the call shows vs. what the machine knows.

On the shared screen

A candidate can share a single clean window while an assistant runs on a second monitor, a phone, or a capture-excluded overlay.

What InterviewWatch sees

The AI process, the browser tab title, the extension's DNS lookup, and the timing gap, regardless of what is shared.

Recorded video review

Slow, subjective, and easy to defeat with off-camera help.

Correlated integrity signal

Objective, timestamped, and delivered as evidence in the final report.

Metadata, not content

Detection uses tool presence, window titles, DNS lookups, event timing, and clipboard size and source. It does not capture what the candidate types, sees, or says.

No recording, no surveillance

No screen, audio or video recording. No eye or body tracking. No automatic rejection, every finding goes to a human with context.

FAQ

AI detection questions, answered.

Can InterviewWatch detect ChatGPT during an interview?
Yes. It flags web AI tools by window title (for example chat.openai.com, gemini.google.com), detects desktop apps by process, and picks up browser-extension assistants like Monica or Sider through the Windows DNS resolver cache. The surfaces are combined into one finding.
Does it read what the candidate types?
No. Detection is metadata only, tool presence, window titles, DNS lookups, timing, and clipboard size and source. InterviewWatch never records keystroke content, screen, audio or video.
Can it detect local LLMs that run offline?
Yes. The process detector recognises local runners such as Ollama, LM Studio, Jan, GPT4All and llama.cpp. They produce no network traffic but still run as processes the agent can see.
How does it keep up with new AI tools?
The agent merges a remotely updateable signature catalog over its built-in catalog, so new assistants are covered without redeploying the agent.
What if our team allows AI tools?
Treat the finding as context, not a failure. The report shows whether AI was active during a round where your policy matters. Humans decide what it means.

See AI assistance the way the machine sees it.

Add objective, timestamped AI-tool evidence to your remote interviews, without recording a single pixel of the candidate's screen.

Try nowRead the cheating-tool guide
Related reading