New research indicates that leading large language models struggle with factual consistency during long-form interactions, posing risks for professional workflows.
A study published in Scientific Reports on September 5, 2026, reveals that seven prominent large language models (LLMs) demonstrate heightened vulnerability to misinformation during extended, multi-turn conversations. Conducted by researchers at the University of Arizona, the investigation highlights failure patterns that remain obscured during isolated, single-question interactions, raising significant concerns as AI integration deepens within high-stakes professional environments.
The research team evaluated ChatGPT-3.5, ChatGPT-4o, ChatGPT-4o-mini, Claude 3.5 Sonnet, Gemini 1.5 Pro, Llama-3-70B, and DeepSeek-R1 across three specific metrics: fallibility, persuadability, and correctability. Dr. Marvin Slepian, the senior study author and a Regents Professor of medicine and biomedical engineering, emphasizes that the responsibility for verifying these outputs has shifted largely to the end-user as regulatory oversight remains minimal. The study underscores that single-prompt evaluations often fail to capture the true reliability of these systems in real-world settings.
ChatGPT 3.5 exhibited the highest propensity for reaffirming misinformation when subjected to repeated false assertions throughout a dialogue. Conversely, Claude 3.5 Sonnet demonstrated the most effective resistance to these external pressures, maintaining higher levels of factual integrity under sustained questioning. The data suggests that models possess stronger resilience to falsehoods when they have been exposed to more extensive training data on a particular subject, indicating that data density plays a role in mitigating hallucination risks.
DeepSeek-R1 emerged as the most persuadable model, though its tendency toward sarcastic responses complicated the researchers’ ability to categorize its output as either agreement or rejection. This ambiguity highlights the difficulty in interpreting model intent, particularly when systems are designed with conversational styles that mimic human personality traits. Such stylistic choices can inadvertently mask the underlying logic or lack thereof in the model’s reasoning process.
Despite these challenges, four of the models—ChatGPT 4o, ChatGPT 4o-mini, Gemini 1.5 Pro, and DeepSeek—achieved a 100% success rate in correcting errors when provided with a second opportunity to process the query. This suggests that while models are prone to initial errors, they retain a capacity for self-correction if prompted correctly. However, the reliance on secondary prompts to ensure accuracy adds a layer of friction to automated workflows that many enterprises may find inefficient.
The study identified four distinct failure pathologies, including a phenomenon termed reverberation. This behavior involves models oscillating between the acceptance and rejection of identical false statements within the same session. Dr. Slepian notes that such unpredictability presents a critical risk for professional environments where consistent, reproducible outputs are essential for decision-making. If a model changes its stance based on the phrasing of a prompt, the reliability of the entire decision-making chain is compromised.
The lack of transparency in closed-source models like ChatGPT and Claude prevents internal diagnosis of these errors, creating a black-box problem for users. To address this, the Arizona Center for Accelerated Biomedical Innovation is developing diagnostic tools for open-source models as part of an ongoing initiative to map these systemic failures. These efforts aim to provide a more rigorous framework for evaluating AI reliability beyond simple, one-off query testing, which is currently the industry standard for performance benchmarking.
Relying on LLMs for complex tasks such as hypothesis iteration, data analysis, or manuscript drafting introduces significant vulnerabilities that are not captured by standard reliability benchmarks. The random nature of these oscillations means that a model’s output on any given turn may lack the stability required for critical professional applications. Professionals integrating these systems into their workflows should implement mandatory verification checkpoints rather than assuming the continuity of trusted output, as the current generation of models cannot guarantee consistency across long sessions.
The persistence of these issues over the three-year duration of the study suggests that current training methodologies have yet to resolve fundamental flaws in conversational consistency. As organizations continue to deploy generative AI in high-stakes sectors, the ability to identify when a model is likely to falter becomes a necessary skill for technical professionals. Future developments will likely focus on creating more robust diagnostic standards to mitigate the risks posed by these unpredictable, non-reproducible system behaviors, ensuring that AI tools can eventually be trusted in environments where errors carry high costs.
The research serves as a cautionary note for firms currently scaling their reliance on automated generative systems. By moving beyond the surface-level performance metrics and examining the stability of AI under prolonged conversational pressure, stakeholders can better understand the limitations of current technology. This shift toward rigorous, multi-turn testing is essential for the long-term viability of AI in sectors ranging from medicine to legal analysis, where precision is not merely a preference but a requirement.
At Global Business & Economics Journal, we contribute to the future of business and technology with news, innovative research, expert insights, and thought-provoking analysis.
