Survey AI Response Detection

Detect AI-generated survey responses using multiple heuristic and statistical methods.

Methods (applied in order)

1. Linguistic analysis — Vocabulary diversity (Type-Token Ratio), perplexity scoring, burstiness patterns, formality consistency
2. Response patterns — Timing analysis, straightlining detection, semantic similarity clustering between responses
3. Behavioral signals — Completion time vs expected, sequential patterns, copy-paste indicators
4. Content quality — Generic/surface-level answers, absence of personal anecdotes, overly structured responses
5. Statistical outliers — Responses deviating from expected distributions

References

Based on academic research:

• Wu et al. 2025 — "A Survey on LLM-Generated Text Detection" (ACL Anthology)
• Yang et al. 2024 — "A Survey on Detection of LLMs-Generated Content" (EMNLP Findings)
• Dhaini et al. 2023 — "Detecting ChatGPT: State of Detection" (RANLP)

See references/detection-methods.md for detailed algorithm descriptions.

Input

• Survey response data (CSV, JSON, or text)
• Expected completion time (optional)
• Response metadata (timestamps, device info — if available)

Output

• Per-response flag level (high/medium/low) with confidence score
• Detection method evidence for each flag
• Aggregate statistics (flag rate, clean dataset size)
• Recommendation per response: remove, review, or keep