'Do not trust your eyes': AI generates surge in expense fraud (www.ft.com)

Companies and auditors are sounding the alarm after a surge in expense fraud enabled by AI that can produce photo‑realistic receipts, invoices, and supporting documents on demand. Modern generative models — diffusion models and advanced image/text generators tied to LLMs — can fabricate convincing line items, vendor logos, timestamps and even plausible OCR‑friendly layouts, while stripping metadata that might reveal manipulation. Attackers are exploiting these tools to submit fake travel, meals and vendor invoices at scale, often blending synthetic documents with minor real details to evade rule‑based checks and human scrutiny — hence the warning, “do not trust your eyes.” The technical significance is twofold: first, visual forensic cues that once helped flag fakes are increasingly unreliable; second, typical automated rule systems and manual review struggle with the volume and sophistication of synthetics. The response will need to be technical and procedural: multimodal detection models that combine image forensics, anomaly detection on expense patterns, cryptographic receipt provenance, secure vendor APIs, and stricter metadata and invoice validation. Firms should also tighten audit trails, adopt real‑time verification (e.g., vendor-confirmed e‑receipts), and retrain staff. For the AI/ML community this is a call to prioritize robust media provenance, watermarking standards and deployment of defensive ML to restore trust in digital documents.