If you’ve used ChatGPT or Claude, you know it types an answer back. You ask a question, and it gives you text. It’s like talking to a very smart, very fast intern.

A newer kind of AI, however, doesn’t just talk. It does things. Imagine an AI that could look at your sales data, find trends, and then draft a summary email to your team. Or one that could scan a contract, pull out specific clauses, and compare them to a template. This is where AI moves from being a chatbot to a true digital assistant.

But here’s the thing: these “doing” AIs often rely on reading and understanding tables of numbers, just like the ones in your spreadsheets. And honestly, this is where things can get a little messy.

Why AI sometimes gets the numbers wrong

When you ask an AI to summarize a table, most people assume if it understands the table’s structure, it’ll get the numbers right. It seems logical, doesn’t it? If the AI can identify rows and columns, surely it can just copy the right data.

That’s a common misconception. The weird part is, even the most advanced AI models, like OpenAI’s GPT-4 or Anthropic’s Opus, can struggle with what researchers call data referencing errors (DREs). This means the AI might incorrectly cite a number, misstate a value, or even completely leave out a piece of data from a table, even when it clearly understands what the table is about. It’s like a junior analyst who understands the report’s purpose but keeps transcribing the wrong figures.

Why does this happen? These large language models (LLMs) are built primarily for language fluency, for generating coherent text. They’re amazing at understanding context, tone, and complex ideas. But when it comes to the precise, almost mechanical task of accurately pulling a specific number from a grid of data, they can sometimes be… careless. It’s a subtle but critical distinction. They get the gist, but not always the exact digit.

100 66 33 0 83.1 AI without Critic 66.6 Traditional LLM 45.2 Early LLM Answer Accuracy (%) *The answer accuracy of AI models when processing table data can vary significantly.*

New research, published on Hugging Face on June 30, 2026, investigated this problem. They found that DREs occur across all tested models, ranging from smaller 1.7-billion-parameter models to much larger 20-billion-parameter ones. This isn’t just a flaw in one specific AI; it’s a general challenge.

The real answer: An AI critic

So, what do you do when your star AI assistant sometimes misreads the numbers? You give it a proofreader. That’s the core idea behind a critic model.

Think of it like this: your main AI is a brilliant but sometimes hasty writer. It can draft a whole report based on a table. The critic model is a separate, smaller, highly focused AI. Its only job is to review the main AI’s output and specifically check for those data referencing errors. Did the main AI correctly cite the revenue figure? Did it miss any key numbers?

This critic acts as a quality control layer. When the main AI generates an answer, the critic examines it. If it spots a DRE, it flags it. Then, the main AI can either filter out the bad answer or try again, a process called rejection sampling. It’s like having a second pair of eyes, specifically trained to catch numerical slips.

This approach makes a huge difference. Researchers demonstrated that incorporating this data referencing critic significantly improves the main AI’s answer accuracy. We’re talking an improvement of up to 12.0%. That’s not a small bump; that’s the difference between a potentially misleading report and a much more reliable one.

1 Main AI generates ou... 2 Critic AI checks for... 3 If error, main AI re... 4 Final, more accurate... *The critic AI helps the main AI refine its output by checking for data referencing errors, leading to a more accurate final result.*

Even better, they trained a lightweight 4-billion-parameter critic model. That’s a relatively small AI, especially compared to the giant LLMs often in the hundreds of billions of parameters. This lightweight critic achieved an average F1 score of 78.2% in detecting DREs, meaning it’s quite good at its specific job. It effectively assists larger models without being a resource hog. You can read the full paper by Yuqing Yang and others here on Hugging Face.

What this means for your business

Imagine your CFO asking an AI to summarize quarterly earnings from a complex Excel sheet. If the AI, without a critic, pulls a revenue figure as $1.2 billion instead of $1.3 billion, that’s a 7.7% error. A small mistake, but it could lead to incorrect investor communications or strategic missteps. With a critic, that error is far less likely.

Or consider a supply chain manager using AI to optimize inventory. If the AI misreads a stock level for a critical component by even 10 units, it could trigger an unnecessary rush order, costing the company thousands in expedited shipping. The critic helps prevent these kinds of costly inaccuracies.

This isn’t about getting rid of human decision-makers. It’s about making your AI tools more trustworthy. It’s about ensuring the underlying data, when processed by AI, retains its integrity. For investors, this means looking for companies that aren’t just adopting AI, but adopting it smartly, with robust quality checks built in.

The caveat: Still need human eyes

Look, no AI is perfect. Even with a sharp critic model, errors can still slip through. The critic assists the larger models; it doesn’t make them infallible. This research shows a huge step forward in making AI more reliable for numerical tasks, but it doesn’t eliminate the need for human oversight.

For any critical business decision, especially those involving financial reporting, legal compliance, or significant capital allocation, you’ll still need human experts to verify AI-generated insights. The AI, even with its critic, is a powerful co-pilot, not an autonomous driver for your most important data.