New Study Reveals Widespread Use of Problematic Epistemologies (Knowledge Production Functions) in LLMs and Agentic AI

WOODSTOCK, Vt., Aug. 25, 2026 /PRNewswire/ -- Artificial Epistemics, LLC (AE) announced today its initial findings from an ongoing study of prevailing epistemologies (knowledge production functions) in leading large language models (LLMs) and Agentic AIs (AAIs). Using a distinction philosophers make between justificationist and falsificationist approaches, the AE team found that virtually all LLMs/AAIs at least implicitly rely on justificationist thinking when formulating responses to prompts. In a joint statement by Joseph M. Firestone and Mark W. McElroy, co-founders of AE and creators of the Susty Code, the study's leaders had this to say about their findings:

"Whereas justificationists believe in truth with certainty and thereby set out to prove, validate, or verify knowledge claims accordingly, falsificationists hold the opposite view – that the most we can do is to choose from amongst competing claims the ones that best survive our criticisms, yet which may still be false. Most LLM/AAI developers, however, seem more interested in taking a justificationist stand by which the appearance of certainty is maintained. Indeed, one leading lab when queried about its adherence to justificationist principles went so far as to say, 'Everyday outputs lean toward authoritative, generalized summaries. The model presents statements as established facts, often glossing over domain limitations or exceptions unless explicitly asked.' That an AI would openly admit to relying on such an uncritical approach is stunning, especially since exceptions can easily disprove the rule!"

As further explained by the AE team, the question of which epistemology to use when developing an LLM/AAI has everything to do with AI safety and alignment, since the truth and morality of what AIs say and do largely determines the risks and consequences of decisions and actions taken by people in response. Moreover, any reliance upon allegedly 'established', 'authoritative', 'consensus-based', or "verified" knowledge is, in epistemological terms, notoriously fallacious, since the rank or status of individuals, groups, or majorities has nothing to do with the truth or legitimacy of the claims they make. Their claims, too, are fallible!

Falsificationists, by contrast, start by accepting the uncertainty of human (and AI) knowledge, and rely instead on critical testing and evaluation of knowledge claims in order to eliminate errors as best they can. This allows AIs to 'quality control' the truth and morality of their outputs using criteria nowhere to be found in the justificationist toolbox. Indeed, virtually all LLMs/AAIs fail to do this and yet it must be done if humanity is to have any chance of protecting itself against the occasional, if not all too often, bouts of AI misinformation, hallucinations, and rogue behaviors. This can only be described as a significant blind spot in modern-day AIs, due mainly to the use of justificationist epistemologies in their designs.

About Artificial Epistemics

Artificial Epistemics, LLC is a U.S.-based startup founded in early 2026, whose purpose is to help guard against the dangers of misinformation and immoral content in AI by working with leading AI producers to integrate its epistemic tools into their offerings. See here for a white paper on the company's Sustainability (Susty) Code, a leading falsificationist protocol.

Contact: Mark W. McElroy, PhD 
Co-Founding Principal
[email protected]

SOURCE Artificial Epistemics, LLC