New image-generating AIs are being used for fake expense reports (arstechnica.com)

Businesses are seeing a surge in AI-generated fake receipts after recent image-generation model rollouts from OpenAI, Google and others. Expense platforms report sharp increases: AppZen found AI-created receipts made up about 14% of fraudulent documents in September (up from zero last year), Ramp flagged over $1M in fraudulent invoices in 90 days, and roughly 30% of US/UK finance professionals surveyed by Medius said falsified receipts rose after OpenAI’s GPT-4o launch. SAP Concur — which runs 80 million compliance checks a month — warned these forgeries “have become so good, we tell our customers, ‘do not trust your eyes.’” Technically, the change stems from accessible image synthesis and multimodal chatbots that let users produce realistic-looking receipts with simple text prompts instead of photo-editing skills or paid services. Platforms are detecting some fakes via embedded metadata that signals ChatGPT creation and by AI-based forensic pattern detectors, but attackers can strip metadata or iterate prompts to evade checks. The implication for AI/ML and finance teams is clear: expense controls must move beyond visual inspection to provenance signals, automated cross-checks with transaction data, behavioral analytics, and robust generative-content detection. Vendors and providers must also combine model watermarking, metadata standards, and tighter policy enforcement to limit this new low-cost fraud vector.