AI Agent & Bot Detection

More and more respondents are handing entire surveys to AI agents such as ChatGPT, Claude, or Gemini and letting them answer automatically. kicue includes a dedicated detection system for this new failure mode, alongside traditional bot and duplicate detection.

AI agent detection

A specialized model evaluates whether a response is likely to have been generated by an LLM agent. Example signals:

• Free-text writing patterns: unusually long and polite text, suspicious lack of hesitation, stereotypical synonym choice
• Option preferences: repeated neutral or safe choices that LLMs are known to favor
• Interaction log anomalies: zero mouse movement or keyboard events throughout the entire session

Detected responses are marked as pending and can be excluded from tabulation after review.

Bot and duplicate detection

Separately from AI agents, kicue also detects pure automation:

• Headless browsers: Puppeteer / Playwright fingerprints are identified
• Automation anomalies: missing mouse events, suspiciously uniform timing
• VPN-based duplicate responses: combinations of IP address, cookies, and browser fingerprint are used to catch the same person retrying over multiple VPN endpoints

These checks run both when the public URL is first accessed and right after submission.

How detected responses are handled

• Appear in the Flags list on the monitoring screen, tagged with the detection category
• Flag states: pending / confirmed / dismissed
• confirmed responses can be excluded from tabulation (see Flag Management & Exclusion)

About detection accuracy

• Machine-learning detection is never 100% accurate. Borderline responses stay in pending state, leaving the final decision to a human reviewer
• The model is updated continuously as new AI agents and automation tools emerge
• Multiple signals are combined rather than relying on any single indicator, to reduce false positives on legitimate respondents