The State of OCR in 2023: A Comprehensive Review of Tools for Investigative Journalism

Editor’s note: This article is published in collaboration with MuckRock. It serves as a major update to a review of OCR tools we published in 2019, which has remained a vital resource for data journalists and researchers. The technology has evolved rapidly; here is the current landscape.

Optical Character Recognition (OCR)—the technology that breathes life into static scans and photographs by transforming them into searchable, machine-readable text—is the bedrock of modern investigative journalism. For organizations like the MuckRock Foundation and DocumentCloud, OCR is not merely a feature; it is the essential bridge between thousands of pages of raw, chaotic FOIA filings and the meaningful, data-driven stories that hold power to account.

In the four years since Source last evaluated the OCR market, the field has undergone a radical transformation. As we revisit these tools, we see a landscape defined by the maturation of cloud-based AI, the decline of older open-source projects, and the emergence of high-performance, machine-learning-driven alternatives.

The Evolution of OCR: A Chronology of Change

The journey of OCR technology in the newsroom has moved from local, resource-heavy processing to sleek, cloud-integrated workflows.

• Pre-2019: The era of "install and hope." Journalists relied heavily on legacy software or local Tesseract installations that required significant technical overhead.
• 2019: Source published its initial review, identifying a mix of aging open-source engines and nascent proprietary cloud services.
• 2020–2022: The rise of Transformer-based models and large-scale cloud AI integration. Proprietary providers began offering "Document Intelligence" rather than simple character recognition, promising to understand layout, tables, and handwritten annotations.
• 2023: The current state. We see a clear bifurcation: powerful, "black-box" cloud services that require subscription fees but offer immense accuracy, and a new generation of open-source libraries like docTR that are beginning to challenge the performance of those cloud giants.

Supporting Data: The Five Contenders

To evaluate the current efficacy of these tools, we tested them against a diverse array of real-world documents, including scanned court filings, redacted FOIA releases, and complex handwritten notes. We focused on five primary contenders: two open-source (Tesseract and docTR) and three proprietary (Amazon Textract, Google Cloud Vision, and Azure Document Intelligence).

1. Tesseract (Open Source)

Maintained by Google since 2006, Tesseract remains the "old reliable" of the OCR world. It is a command-line engine written in C/C++.

• Strengths: Unmatched language support and ease of integration into existing workflows. It remains the default for DocumentCloud when no other engine is selected.
• Weaknesses: Tesseract is increasingly struggling with "dirty" data. Scanned documents with heavy noise, complex layouts, or handwritten elements often cause it to falter compared to modern neural-network-based tools.

2. docTR (Open Source)

Developed by Mindee and powered by TensorFlow 2 and PyTorch, docTR represents the new wave of open-source OCR.

• Strengths: It is arguably the most beginner-friendly tool available. Its performance on scanned documents and screenshots far outstrips Tesseract. It provides high precision and recall, making it a formidable choice for those who want open-source control without the headaches of older, unmaintained projects.
• Weaknesses: It currently lacks native support for handwriting, which keeps it a step behind the premium cloud services.

3. Amazon Textract (Proprietary)

Textract was in its infancy during our 2019 review. Today, it is a robust, mature service integrated directly into DocumentCloud.

• Strengths: Exceptional handling of tables and forms. It excels at extracting data from structured documents.
• Weaknesses: Its language support is strictly limited to Latin scripts, which makes it less versatile for international reporting compared to Google or Azure.

4. Google Cloud Vision (Proprietary)

The GCV API is a powerhouse, capable of identifying not just text, but faces, landmarks, and labels within images.

• Strengths: Unrivaled accuracy in messy, real-world scenarios. In our tests, GCV was the only engine capable of resolving the blurry, embossed seal on a government acknowledgment letter.
• Weaknesses: Implementation friction. The setup requires managing Google Storage buckets and asynchronous requests, which adds significant technical debt to the user.

5. Microsoft Azure AI Document Intelligence (Proprietary)

Azure has positioned itself as the most "developer-friendly" cloud solution.

• Strengths: Ease of setup. Its documentation is exhaustive, and the integration process is significantly less clunky than the Google equivalent. It performs impressively on non-Latin scripts and complex multilingual documents.
• Weaknesses: A tendency toward "hallucinated" formatting. It often inserts unnecessary punctuation or struggles with specific column-based layouts, requiring users to write cleanup scripts to normalize the output.

Implications for the Newsroom

The shift toward these tools carries profound implications for journalistic practice.

First, the barrier to entry has lowered. Because these tools are now available as Add-Ons within platforms like DocumentCloud, a reporter in a small newsroom can perform the same sophisticated analysis as a data team at a national outlet. No coding is required to run a document through Google Cloud Vision or docTR.

Second, the financial model of journalism is shifting. While open-source tools remain free, the most effective engines are now pay-per-page. This creates a new line item in investigative budgets: "AI Credits." Organizations must now account for the cost of processing vast caches of documents, which, while efficient, introduces a dependency on Big Tech infrastructure.

Finally, the decline of legacy projects serves as a warning. Several once-thriving open-source projects have been abandoned or archived because they could not keep pace with the rapid advancements in machine learning. This highlights the fragility of community-driven initiatives in an era where high-performance computing power is increasingly concentrated in the hands of major corporations.

Official Responses and Ethical Considerations

In our review, we noted that some vendors, like ABBYY, have moved away from pricing transparency, making it difficult for public-interest organizations to budget for their tools. This shift suggests a move toward enterprise-only contracts that may exclude smaller investigative shops.

Furthermore, we must address the "black box" nature of proprietary OCR. When a proprietary engine incorrectly interprets a handwritten note or misidentifies a character, the internal logic remains opaque. While our tests showed that GCV and Azure perform well, they are not infallible. Journalists must maintain a healthy skepticism of automated transcription, particularly when dealing with evidence in legal or high-stakes contexts.

Conclusion: How to Choose?

Choosing the right OCR tool depends entirely on the nature of your documents and your team’s technical capacity:

• For clean, machine-generated PDFs: Tesseract remains the efficient, cost-effective choice.
• For high-quality, "messy" scans and screenshots: docTR is the best open-source option for those wanting to avoid cloud costs and proprietary lock-in.
• For high-stakes, multilingual, or complex handwritten documents: Google Cloud Vision is the accuracy leader, provided you have the technical resources to manage the API.
• For structured data (tables, forms) and ease of implementation: Microsoft Azure Document Intelligence provides the best balance of power and usability.

As we look toward the future, the integration of Large Language Models (LLMs) with OCR promises to further revolutionize how we analyze these documents. We are no longer just looking for text; we are looking for context. The ability to automatically summarize, categorize, and extract structured data from thousands of pages of records is not just a dream—it is the next phase of the digital newsroom.

We invite you to explore the results of these tests in our DocumentCloud project and see for yourself how these tools handle the complex realities of the documents that define our world.