.png)
A new study from Stanford, published in Nature Machine Intelligence (2025), highlights a fundamental flaw in today’s language models: they cannot reliably distinguish facts from other types of information. As these models are increasingly used in high-stakes fields like law, medicine, and journalism, this inability to separate verifiable facts from mere knowledge claims or assumptions poses a serious risk.
The research team evaluated 24 leading language models using the KaBLE benchmark—a rigorous set of 13,000 questions designed to test how well models handle factual accuracy and epistemic reasoning. The findings are stark:
- Systematic confusion between fact and non-fact: All tested models, including state-of-the-art systems like GPT-4o, frequently failed to recognize when a statement was a verifiable fact versus a belief, assumption, or knowledge claim.
- Sharp drops in factual accuracy: When challenged with questions that required distinguishing facts from other information, model accuracy plummeted. For example, GPT-4o’s performance dropped from 98.2% to 64.4% on tasks requiring fact discrimination.
- Superficial pattern-matching: Even when models answered correctly, the study found they often relied on surface-level cues rather than a robust understanding of what constitutes a fact. This means their “knowledge” is often just plausible-sounding text, not grounded in verifiable truth.
- Lack of factive reasoning: Most models do not grasp that true knowledge must be based on facts. They treat all information—whether factual, assumed, or believed—as equally valid, leading to potential misinformation and unreliable outputs.
The takeaway: before language models can be trusted in critical domains, they must be able to reliably identify, reason from, and communicate facts—not just generate convincing language. Without this, the risk of error, bias, and misinformation remains unacceptably high.
The NOAN Solution: Fact Control for AI-Native Business
At NOAN, we saw this problem coming. That’s why we built our platform around fact control—the ability to define, manage, and verify the facts that power your AI workflows.
What is Fact Control?
Fact control means giving users—not just the AI—authority over what counts as a fact in their business. As we wrote in our recent article:
“Fact control is the process of explicitly defining, curating, and updating the core truths that underpin your business. It’s about making sure your AI isn’t just guessing, but is grounded in the reality of your company, your market, and your goals.”
How NOAN Puts You in Charge
- Smart Blocks: Every key piece of business knowledge—your mission, pricing, customer segments, workflows—is stored as a “fact block.” These are editable, reviewable, and version-controlled.
- AI That Knows What It Knows: When you ask NOAN’s AI to generate content, strategy, or analysis, it draws only from your curated facts. No hallucinations, no guesswork.
- Fact-Driven Collaboration: Share specific facts with advisors, team members, or clients. Control who can view or edit each fact, and track changes over time.
- Continuous Fact Auditing: Easily review, update, or challenge any fact. The system keeps a transparent history, so you always know the source and status of your business truths.
Why This Matters
Stanford’s research shows that even the most advanced LMs can’t be trusted to know what’s true—unless we give them a foundation of facts. NOAN’s fact control system is designed to do exactly that:
- Reduce risk: No more AI-generated errors based on outdated or incorrect information.
- Build trust: Advisors, investors, and customers can see the facts behind your strategy.
- Move faster: With a single source of truth, your team and your AI are always aligned.
The Future: AI That Knows What It Knows
The Stanford paper is a wake-up call for anyone deploying AI in business. Pattern-matching is not enough. We need systems that can reason from facts, not just beliefs.
NOAN’s fact control approach is our answer: a platform where you define the truth, and your AI works from it—every time.
Want to see how fact control can make your business smarter? Read our full article or try NOAN today.