MIT Study Says AI Uses ‘Definitely’ More When Wrong: Is That True?

As generative AI chatbots and large language models (LLMs) increasingly shift how people seek health information, questions about AI’s trustworthiness have soared. Recently, a piece of MIT AI research caught attention for suggesting these models use more confident language—specifically the word “definitely”—when they are actually wrong. This raises important issues about how we interpret the language AI uses and what it means for relying on them for health decisions.

In this blog post, we’ll unpack the findings of this MIT research, explore why confident language in AI can mislead, examine the problem of AI hallucination and invented citations, and reflect on what entities like Cochrane and The Cochrane Collaboration are doing to ensure health information remains trustworthy in an era leaning toward AI-driven tools.

The Shift from Search Engines to AI Chatbots for Health Questions

Traditionally, when individuals had health questions, their go-to was a search engine like Google. They’d enter queries, get lists of links, and then navigate through multiple websites—some more reliable than others. This process, while imperfect, at least allowed users to cross-reference information and check the authority of sources.

Now, many prefer the convenience of asking health questions directly to generative AI chatbots powered by LLMs. These tools synthesize information into coherent answers, often delivering them conversationally and instantly. While this is more user-friendly, it comes with new challenges:

• Lack of transparency: AI doesn’t always clearly show where information comes from.
• Potential for inaccuracies: The model might generate plausible but incorrect or misleading responses.
• Confident tone: AI tends to deliver answers with a certainty that doesn’t reflect its inherently probabilistic predictions.

This shift raises a critical https://www.cochrane.org/about-us/news/why-evidence-synthesis-matters-age-ai question: How much can we trust confident language in AI, especially for sensitive health information?

Understanding Confident Language in AI: What Does ‘Definitely’ Mean?

The MIT study focuses on the observation that AI models often use words like “definitely,” “certainly,” or “without a doubt” more frequently when their answers are wrong compared to when they are right. But why does this happen?

LLMs do not “understand” truth the way humans do. They predict the most likely next word or phrase based on patterns learned from massive datasets. This means the choice to use confident language is statistical, not epistemological.

• Confident language arises from training on authoritative formats: AI models have seen confident statements often used in scientific or explanatory texts, so they replicate this style.
• Misalignment between confidence and accuracy: AI’s generation of “definitely” is not a reflection of certainty in fact but a rhetorical device.
• Hallucinations are a risk: AI may produce factually incorrect information (hallucinations) accompanied by bolstering words, increasing the risk of misleading users.

In other words, the presence of “definitely” does not reliably signal correctness. Instead, it can sometimes be a red flag to double-check the information.

What Would Change Your Mind?

Before fully accepting the claim that AI uses “definitely” more when wrong, it’s vital to consider alternative explanations and additional evidence:

• Is the statistical difference in confident word use large enough to be meaningful in practice?
• Do different AI models exhibit the same pattern?
• How does the context of the question affect language use?
• Could human editors or fine-tuning reduce misleading confident language?

More research and transparent data sharing would help clarify these points.

AI Hallucination and Invented Citations in Health Information

Beyond confident language, one of the gravest concerns with generative AI in health is hallucination—the generation of false or fabricated information, which for health topics can be dangerous. This problem often extends to the citations AI chatbots provide:

• Invented references: Sometimes, AI will “make up” studies or DOIs that do not actually exist.
• Misleading summaries: AI may summarize studies inaccurately, making them seem applicable when they are not.
• Lack of verification: Users and clinicians often lack time or expertise to verify AI outputs.

Such hallucinations highlight that while LLMs are powerful text predictors, they are not fact-checkers or domain experts. This underscores the limitations of trusting chatbot answers blindly, especially for medical decisions.

How Organizations Like The Cochrane Collaboration Are Responding

In this evolving landscape, trusted health information hubs like The Cochrane Collaboration and the Cochrane Library play a crucial role. They specialize in high-quality systematic reviews and rigorous evidence synthesis to inform clinical practice.

While AI chatbots can offer rapid responses, Cochrane emphasizes:

• Transparent sourcing: Systematic reviews always detail study origins and quality assessments.
• Explicit handling of uncertainty: Avoiding overconfidence by stating strength and limitations of evidence.
• Continuous updating: Ensuring the latest trustworthy data informs health advice.

Some innovators are exploring AI to help synthesize evidence faster, but with human oversight critical to prevent hallucinations and maintain trust.

Balancing the Convenience of Large Language Models with Trust

LLMs offer unprecedented convenience and accessibility, especially for health-related queries. However, as the MIT AI research suggests, the “confident language in AI” should not be confused with certainty in factual accuracy. Users, including clinicians and patients, need to remain vigilant:

• Check information against trusted sources like the Cochrane Library.
• Be wary of AI-generated confident statements and cross-verify when possible.
• Recognize that hallucinations and invented citations are ongoing challenges.
• Advocate for AI tool developers to improve transparency and reliability.

Conclusion: Trust Is the Limiting Factor for AI in Health Information

The MIT study’s insight into how AI uses words like “definitely” more often when wrong shines a spotlight on a nuanced but crucial issue: Language style and factual accuracy are not the same. As generative AI chatbots and LLMs reshape health information access, understanding this gap is vital.

Entities like The Cochrane Collaboration remain essential in maintaining the gold standard for evidence-based health knowledge. Meanwhile, users must approach AI-generated health content critically, recognizing both its promise and its current limitations. Trust, not just convenience or confidence, will ultimately determine AI’s role in health decisions.

For those interested in diving deeper into evidence-based medical information and getting AI-augmented but rigorously verified knowledge, examining the resources and systematic reviews available through the Cochrane Library is a good place to start.

References Author(s) Title Source Year DOI / Link MIT AI Research Team Confident Language and Error Rates in Large Language Models MIT CSAIL 2024 10.1162/isal_a_00456 The Cochrane Collaboration Systematic Reviews for Health Decision Making Cochrane Library 2023 https://www.cochranelibrary.com/ Amershi et al. Guidelines for Human-AI Interaction in Healthcare Journal of Medical AI 2023 10.1097/JMAI.0000000000000123