The Illusion of AI Accuracy and the Limits of Trust in ChatGPT

The claim by OpenAI that ChatGPT's "hallucinations" have decreased is a narrative that demands careful scrutiny, especially from our perspective at Republika. While technological advancements are often presented with impressive statistics, the relationship between error reduction and trust in generative AI is not linear. The very term "hallucination" is misleading, suggesting an anomaly rather than an intrinsic feature of these systems.

At its core, ChatGPT and similar large language models (LLMs) do not 'know' information; they predict it based on vast datasets. They generalize and generate text, a process that inherently carries the risk of producing inaccurate or fabricated content. Therefore, a reduction in "hallucinations" is not a cure but a probabilistic adjustment. This is where the illusion of accuracy takes hold, leading users to place exponential trust in the system, a cognitive bias known as automation bias.

This shift from verification to delegation has profound implications, particularly in higher education, where institutions like Universitas Bina Sarana Informatika (UBSI) are grappling with AI's role. While students may understand the probabilistic nature of AI output theoretically, the practical reality often sees these systems treated as instant answer machines. This disconnect highlights a critical gap: technological literacy does not automatically equate to epistemological literacy. We must question whether the current digital ecosystem, which prioritizes efficiency and speed over accuracy, is fostering the critical thinking necessary to navigate the complexities of AI. The challenge for informatics education is not just technical, but deeply philosophical, demanding that we teach students not just how to use AI, but how to critically evaluate its outputs.