Scam-AI/spot-the-ai-receipt

Spot the AI Receipt 🧾

An interactive 2AFC (two-alternative forced choice) game built by [Scam.AI](https://www.scam.ai).

Each round shows two receipts side by side:

• —One is an authentic receipt from the public CORD-v2 dataset
• —One is fully AI-synthesized (GPT-4o generates the text, GPT-Image-1 renders the image) from our GPT4o-Receipt benchmark

Pick the AI one. After 10 rounds you'll see your accuracy vs. the human and LLM baselines reported in our paper:

Zhang, Ren, et al. — "GPT4o-Receipt: A Dataset and Human Study for AI-Generated Document Forensics" (arXiv:2603.11442)

Key finding: humans rate AI receipts as visually distinct from real ones (1.87/5 gap) yet only achieve F1 = 0.852 binary detection — well below LLMs like Claude Sonnet 4 (F1 = 0.975). The forensic signal is in arithmetic incoherence that humans rarely audit but LLMs verify trivially.

Production-grade detection: scam.ai.