Despite rapid advances in AI reasoning and coding, the decades-old PDF format remains a stubborn technical challenge, exposing a critical gap between model intelligence and real-world document reliability. (Source: Image by RR)

Hallucinations Remain a Persistent Risk in Extraction Tasks

Artificial intelligence can write software, solve advanced math, and generate photorealistic video — yet one of the most common file formats in the world continues to trip it up: the PDF. Originally designed in the early 1990s to preserve visual formatting across devices, PDFs were built to look identical everywhere, not to be easily interpreted by machines. As a result, extracting structured data from them remains a stubborn challenge, even for state-of-the-art AI models.

The problem, according to an article in theverge.com, became glaringly obvious when massive document dumps — including millions of Justice Department files related to Jeffrey Epstein — were released as poorly indexed PDFs. Optical character recognition (OCR) made the documents technically searchable, but formatting quirks, multi-column layouts, redactions, tables, handwritten notes, and inconsistent scans rendered them nearly unusable. AI models often summarize instead of extract, confuse footnotes with body text, or hallucinate content entirely.

The core issue lies in how PDFs encode information. Unlike HTML, which stores text in logical order, PDFs store coordinates and drawing instructions for painting a page. That makes it difficult for AI systems to understand editorial structure — what’s a header, what’s a footnote, what belongs inside a table, and what doesn’t. Researchers are increasingly building specialized vision-language models trained specifically on PDFs, including efforts from the Allen Institute for AI and Hugging Face, which recently “liberated” trillions of tokens from over a billion web-scraped PDFs for training purposes.

Companies like Reducto are taking a multi-model approach, segmenting documents into components — tables, charts, headers — before passing them to specialized parsing systems. This layered strategy improves accuracy, but the long tail of bizarre formatting edge cases remains a persistent obstacle. Like self-driving cars navigating unpredictable roads, PDF parsing appears 98% solved — yet that final 2% continues to block full reliability. And with high-value content from governments, courts, and enterprises locked inside PDFs, AI companies are now racing to crack a problem they once ignored.

read more at theverge.com