New testing reveals that the low-resolution modes of advanced AI models, specifically GPT-5.5 and GPT-5.6 generations, fail to accurately extract information from documents. Across 70 attempts using synthetic invoices, these models consistently invented details like company names and invoice years, even while correctly calculating financial totals. This tendency to hallucinate coherent but incorrect data, rather than producing random errors, poses a significant risk for applications like accounts payable, where accurate arithmetic might mask fabricated counterparty information. AI
IMPACT AI models' tendency to hallucinate coherent but incorrect data poses risks for automated document processing and financial applications.
RANK_REASON Analysis of AI model performance on a specific task. [lever_c_demoted from research: ic=1 ai=1.0]
- Accora Manufacturing Inc.
- Aozora Manufacturing Inc.
- GPT-5.5
- GPT-5.6
- INV-2025-0731
- INV-2026-0731
- K Northwind Trading Ltd
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
