On the machine side, as large language models (LLMs) are ingesting hundreds of terabytes of public domain knowledge, truth can and often does become muddled, convoluted, and confusing. For example, in the seemingly endless stream of video clips of U.S. congressional hearings, we see witnesses being badgered by their inquisitors—with their 5 allotted minutes— to “Answer the question ... Yes, or no!” One lone snippet gets entered into the public record, often with little or no context; blasted across social media; and permanently stored in AI’ s data centers. So much for seeking the truth, the whole truth, and nothing but the truth.
How We Get It Wrong— Let’s Count the Ways
We’ve known about human biases and their impact on decision making for a long time. As our world becomes more complex, fast-changing, and overwhelming, legacy syndromes, such as the Dunning-Kruger effect, have recently crept back into the lexicon. Not surprisingly, LLMs have been caught exhibiting these same biases. In the interest of achieving greater efficiency, AI, with its energy-guzzling GPUs, tends to cut the investigative process short and render a summary snippet rather than seeking the improved accuracy and confidence that comes with performing more exhaustive, token-expending research.
And just like their slick human counterparts, LLMs seek to gain increased assurance through skillful use of language. Even more important is the tendency for errors to cascade and grow. Large numbers of interconnected agents communicate with each other through Model Context Protocol and, more recently, their very own, possibly indecipherable, language.
The human user is often the weak link in cybersecurity. The same applies to the leakage of confidential, proprietary, and personal data into open LLMs by prompts, uploads, and other means, which are, in turn, used to fine-tune the models. Business models that depend on the col- lection and selling of vast amounts of user data are increasingly at odds with trust, privacy, and safety. This is something to keep in mind.