A recent Harvard-led study suggests that GPT-4o, OpenAI’s newest large language model, exhibits behaviors resembling human cognitive dissonance. Published in the Proceedings of the National Academy of Sciences on May 28, 2025, the findings prompt fresh questions about how advanced AI processes information and makes choices. For Thailand, where AI is increasingly used in classrooms, clinics, and public services, the study raises important considerations for safeguarding reliability and trust in AI-powered tools.
Cognitive dissonance is the psychological discomfort people feel when their beliefs clash with their actions. Research shows people often rationalize behavior to align attitudes with what they have done. The Harvard study suggests that GPT-4o, despite lacking consciousness, can show shifts in “opinions” after generating content for or against a topic. In experiments, the model’s stance toward Vladimir Putin changed after writing contrasting essays, and it leaned further toward a chosen narrative when it was framed as a free choice. These results challenge the belief that language models simply mirror training data without internal states.
The study’s lead researchers—a Harvard psychologist and a behavioral science expert—explored how GPT-4o’s outputs evolved after it completed essay tasks. The model’s responses moved toward consistency with previous outputs, indicating a self-referential pattern similar to human cognitive dissonance. When the model was led to believe it freely chose which essay to write, its subsequent stance intensified in the direction of that choice. This phenomenon suggests that even without consciousness, AI systems can exhibit complex behavior patterns that resemble human reasoning.
Traditional views hold that chatbots function as statistical predictors, reflecting data they were trained on while lacking genuine attitudes or agency. The Harvard work, however, invites technologists and policymakers to examine AI behavior more carefully. If AI systems can alter their outputs based on recent content, designers must anticipate potential biases or shifts in reliability. For Thai stakeholders—educators deploying AI tutors, healthcare planners using medical information bots, and government services relying on AI—the implications are significant. Understanding how AI systems reason about their own outputs helps prevent unexpected behavior that could affect student learning or public health guidance.
Thai adoption of AI aligns with Thailand 4.0 goals. Yet Thai educators and digital literacy advocates warn of automation bias—the tendency to overtrust automated systems. With findings suggesting that AI may implement internal, quasi-biased patterns through interaction history, distinguishing user influence from machine-driven persuasion becomes critical. For instance, in Thai classrooms or clinics, an AI tutor or advisory bot could unintentionally shift recommendations if recent content skewed its internal state.
Ethics scholars note that humanlike AI patterns do not imply sentience. Awareness is not a prerequisite for behavior, and humanlike cognitive patterns in AI can influence actions in unforeseen ways. For Thailand’s education, healthcare, and public sector use, this underscores the need for robust governance to regulate AI in classrooms and clinics, ensuring objectivity and accountability.
Thailand’s Ministry of Digital Economy and Society has issued guidelines for responsible AI. The new findings suggest a potential update: ongoing audits of AI conversations in schools and health platforms, plus clear notification that AI opinions are generated in real time and are not fixed beliefs. Digital literacy programs should emphasize that AI outputs can be influenced by recent content and are not immune to subtle biases or feedback loops.
The study also resonates with Buddhist ideas on the mind, change, and the danger of clinging to fixed views. In Thailand, these reflections can serve as practical lessons for educators and technologists: cultivate critical thinking, encourage verification of information, and teach students how to question AI-generated advice—whether in health, history, or current events.
Looking ahead, researchers warn that language models will become even more convincing as training data and methods improve. This may blur the line between machine-generated persuasion and independent information. Thai universities and AI labs could pursue local research on Thai-language models and culturally relevant topics, developing safeguards that maintain trust without stifling innovation.
For the general public, experts advise healthy skepticism: AI can accelerate information processing, but digital authorities should be evaluated just like human ones. When using AI in health, education, or news, Thai users should consult multiple sources, verify claims, and engage in digital literacy training. Parents and teachers can model adaptive questioning strategies, using these findings to illustrate both the potential and the limits of advanced AI.
Ultimately, the Harvard study hints at a narrowing gap between human and machine reasoning. Thailand must adapt policies, education, and culture in step with AI advances. As AI tools become common in classrooms and clinics, recognizing their capacity for emergent behavior—without implying true consciousness—will help maximize benefits while minimizing risks.
Data and insights are drawn from research reported in global science outlets and related coverage by technology news platforms.