When the House Oversight Committee dumped 20,000 pages of Epstein documents last November, followed by over 3 million Department of Justice files, it exposed an awkward truth about artificial intelligence: despite billions in investment and endless hype, AI still struggles with one of the most basic digital tasks—reading PDFs. Luke Igel and his team discovered this firsthand while trying to parse through garbled email threads and barely searchable documents, revealing a massive gap between AI's promised capabilities and its real-world performance on fundamental document processing.
The AI industry loves talking about reasoning models, multimodal understanding, and artificial general intelligence. But last November, when actual government documents hit the internet, the technology face-planted on something far more mundane: reading text from a PDF.
Luke Igel and his friends were clicking through the House Oversight Committee's massive Epstein document release, trying to piece together email conversations and follow investigative threads. The experience was, in his words, "gross." The Department of Justice had processed the files with optical character recognition software, but the results were abysmal. Emails appeared garbled. Text searches returned nothing. The interface was practically unusable.
Then came the real test. In the following months, DOJ released over 3 million additional files. All PDFs. All requiring the same broken OCR technology that had already failed.
This isn't some edge case or obscure technical challenge. We're talking about reading typed text from digital documents—something computers have supposedly been able to do for decades. Yet here we are in 2026, with OpenAI raising billions at a $300 billion valuation and every tech giant claiming their AI can understand images, video, and human reasoning, but government agencies still can't make documents searchable.
The PDF problem reveals something uncomfortable about the current AI boom. While companies pour resources into flashy demonstrations of AI writing poetry or generating videos, the basic infrastructure work—the stuff that actually matters for day-to-day business operations—remains frustratingly broken.
