For years, businesses, governments, and researchers have struggled with a persistent problem: How to extract usable data from Portable Document Format (PDF) files. These digital documents serve as containers for everything from scientific research to government records, but their rigid formats often trap the data inside, making it difficult for machines to read and analyze.
“Part of the problem is that PDFs are a creature of a time when print layout was a big influence on publishing software, and PDFs are more of a ‘print’ product than a digital one,” Derek Willis, a lecturer in Data and Computational Journalism at the University of Maryland, wrote in an email to Ars Technica. “The main issue is that many PDFs are simply pictures of information, which means you need Optical Character Recognition software to turn those pictures into data, especially when the original is old or includes handwriting.”
Computational journalism is a field where traditional reporting techniques merge with data analysis, coding, and algorithmic thinking to uncover stories that might otherwise remain hidden in large datasets, which makes unlocking that data a particular interest for Willis.
The PDF challenge also represents a significant bottleneck in the world of data analysis and machine learning at large. According to several studies, approximately 80–90 percent of the world’s organizational data is stored as unstructured data in documents, much of it locked away in formats that resist easy extraction. The problem worsens with two-column layouts, tables, charts, and scanned documents with poor image quality.
The inability to reliably extract data from PDFs affects numerous sectors but hits hardest in areas that rely heavily on documentation and legacy records, including digitizing scientific research, preserving historical documents, streamlining customer service, and making technical literature more accessible to AI systems.
“It is a very real problem for almost anything published more than 20 years ago and in particular for government records,” Willis says. “That impacts not just the operation of public agencies like the courts, police and social services but also journalists, who rely on those records for stories. It also forces some industries that depend on information, like insurance and banking, to invest time and resources in converting PDFs into data.”
A very brief history of OCR
Traditional optical character recognition (OCR) technology, which converts images of text into machine-readable text, has been around since the 1970s. Inventor Ray Kurzweil pioneered the commercial development of OCR systems, including the Kurzweil Reading Machine for the blind in 1976, which relied on pattern-matching algorithms to identify characters from pixel arrangements.
These traditional OCR systems typically work by identifying patterns of light and dark pixels in images, matching them to known character shapes, and outputting the recognized text. While effective for clear, straightforward documents, these pattern-matching systems, a form of AI themselves, often falter when faced with unusual fonts, multiple columns, tables, or poor-quality scans.
Traditional OCR persists in many workflows precisely because its limitations are well-understood—it makes predictable errors that can be identified and corrected, offering a reliability that sometimes outweighs the theoretical advantages of newer AI-based solutions. But now that transformer-based large language models (LLMs) are getting the lion’s share of funding dollars, companies are increasingly turning to them for a new approach in reading documents.
The rise of AI language models in OCR
Unlike traditional OCR methods that follow a rigid sequence of identifying characters based on pixel patterns, multimodal LLMs that can read documents are trained on text and images that have been translated into chunks of data called tokens and fed into large neural networks. Vision-capable LLMs from companies like OpenAI, Google, and Meta analyze documents by recognizing relationships between visual elements and understanding contextual cues.
The “visual” image-based method is how ChatGPT reads a PDF file, for example, if you upload it through the AI assistant interface. It’s a fundamentally different approach than standard OCR allows them to potentially process documents more holistically, considering both visual layouts and text content simultaneously.
And as it turns out, some LLMs from certain vendors are better at this task than others.
“The LLMs that do well on these tasks tend to behave in ways that are more consistent with how I would do it manually,” Willis said. He noted that some traditional OCR methods are quite good, particularly Amazon’s Textract, but that “they also are bound by the rules of their software and limitations on how much text they can refer to when attempting to recognize an unusual pattern.” Willis added, “with LLMs, I think you trade that for an expanded context that seems to help them make better predictions about whether a digit is a three or an eight, for example.”
This context-based approach enables these models to better handle complex layouts, interpret tables, and distinguish between document elements like headers, captions, and body text—all tasks that traditional OCR solutions struggle with.
“[LLMs] aren’t perfect, and sometimes require significant intervention to do the job well, but the fact that you can adjust them at all [with custom prompts] is a big advantage,” Willis said.
New attempts at LLM-based OCR
As the demand for better document processing solutions grows, new AI players are entering the market with specialized offerings. One such recent entrant has caught the attention of document processing specialists in particular.
Mistral, a French AI company known for its smaller LLMs, recently entered the LLM-powered optical reader space with Mistral OCR, a specialized API designed for document processing. According to Mistral’s materials, their system aims to extract text and images from documents with complex layouts by using its language model capabilities to process document elements.
However, these promotional claims don’t always match real-world performance, according to recent tests. “I’m typically a pretty big fan of the Mistral models, but the new OCR-specific one they released last week really performed poorly,” Willis noted.
“A colleague sent this PDF and asked if I could help him parse the table it contained,” says Willis. “It’s an old document with a table that has some complex layout elements. The new [Mistral] OCR-specific model really performed poorly, repeating the names of cities and botching a lot of the numbers.”
AI app developer Alexander Doria also recently pointed out a flaw with Mistral OCR’s ability to understand handwriting on X, writing, “Unfortunately Mistral-OCR has still the usual VLM curse: with challenging manuscripts, it hallucinates completely.”
According to Willis, Google currently leads the field in AI models that can read documents: “Right now, for me the clear leader is Google’s Gemini 2.0 Flash Pro Experimental. It handled the PDF that Mistral did not with a tiny number of mistakes, and I’ve run multiple messy PDFs through it with success, including those with handwritten content.”
Gemini’s performance stems largely from its ability to process expansive documents (in a type of short-term memory called a “context window”), which Willis specifically notes as a key advantage: “The size of its context window also helps, since I can upload large documents and work through them in parts.” This capability, combined with more robust handling of handwritten content, apparently gives Google’s model a practical edge over competitors in real-world document processing tasks for now.
The drawbacks of LLM-based OCR
Despite their promise, LLMs introduce several new problems to document processing. Among them, they can introduce confabulations or hallucinations (plausible-sounding but incorrect information), accidentally follow instructions in the text (thinking they are part of a user prompt), or just generally misinterpret the data.
“The biggest [drawback] is that they are probabilistic prediction machines, and will get it wrong in ways that aren’t just ‘that’s the wrong word’,” Willis explains. “LLMs will sometimes skip a line in larger documents where the layout repeats itself, I’ve found, where OCR isn’t likely to do that.”
AI researcher and data journalist Simon Willison identified several critical concerns of using LLMs for OCR in a conversation with Ars Technica. “I still think the biggest challenge is the risk of accidental instruction following,” he says, always wary of prompt injections (in this case accidental) that might feed nefarious or contradictory instructions to a LLM.
“That and the fact that table interpretation mistakes can be catastrophic,” Willison adds. “In the past I’ve had lots of cases where a vision LLM has matched up the wrong line of data with the wrong heading, which results in absolute junk that looks correct. Also that thing where sometimes if text is illegible a model might just invent the text.”
These issues become particularly troublesome when processing financial statements, legal documents, or medical records, where a mistake might put someone’s life in danger. The reliability problems mean these tools often require careful human oversight, limiting their value for fully automated data extraction.
The path forward
Even in our seemingly-advanced age of AI, there is still no perfect OCR solution. The race to unlock data from PDFs continues, with companies like Google now offering context-aware generative AI products. Some of the motivation for unlocking PDFs among AI companies, as Willis observes, doubtless involves potential training data acquisition: “I think Mistral’s announcement is pretty clear evidence that documents—not just PDFs—are a big part of their strategy, exactly because it will likely provide additional training data.”
Whether it benefits AI companies with training data or historians analyzing a historical census, as these technologies improve, they may unlock repositories of knowledge currently trapped in digital formats designed primarily for human consumption. That could lead to a new golden age of data analysis—or a field day for hard-to-spot mistakes, depending on the technology used and how blindly we trust it.