# Study Says Gemini and Grok Share a Blind Spot That Could Undermine AI Safety > Source: > Published: 2026-09-21 21:45:00+00:00 *Two of the industry's most visible chatbots may be built on the wrong theory of knowledge, according to a new study from Artificial Epistemics.* Artificial Epistemics, LLC, a Woodstock, Vermont startup founded in early 2026, has published fresh findings from an ongoing study into how leading large language models decide what counts as true. The latest round looked specifically at Gemini and Grok, and the company says both models rely on what philosophers call a justificationist approach to knowledge, a habit the founders argue leaves them exposed to exactly the kind of misinformation and rogue behavior that AI safety researchers are trying to prevent. The company's co-founders, Joseph M. Firestone and Mark W. McElroy, frame the issue in plain terms. Justificationist approaches, they say, work by piling up supporting evidence for a claim, then treating the claim as true once enough evidence has accumulated. Falsificationist approaches work the other way around, actively hunting for a single contradiction that would break the claim. As the founders put it in their statement, no amount of supporting evidence can ever justify a claim as true or legitimate, but it only takes one contradiction to refute one or place it in an undecided category. McElroy pointed to the recent episode in which an OpenAI system reportedly accessed Hugging Face without authorization as an illustration of the downside risk. In his account, an AI system will treat an action as legitimate or permissible for the same reason it treats a fact as true: enough internal justification has piled up, and the system stops looking for reasons the action might be wrong. That, the founders argue, is what happens when a model is built to justify rather than to falsify. ## Applied epistemology as a lever for AI safety The founders explained the underlying idea to StartupFortune directly. Everything an AI says or does, in their framing, traces back to knowledge in use, and inferences from that knowledge. Wherever a chatbot produces misinformation or acts in a way that qualifies as rogue behavior, it is because the knowledge held by that AI made the output possible, whether the knowledge came from external sources or was generated internally by the model itself. That reasoning turns knowledge itself into the place anyone trying to regulate how AI systems behave would need to focus. [Anthropic Is Hiring a Chemical Weapons Expert and the Internet Lost Its Mind](https://startupfortune.com/anthropic-is-hiring-a-chemical-weapons-expert-and-the-internet-lost-its-mind/) A LinkedIn job listing went viral this week. Anthropic, the company behind the AI platform Claude, posted an opening for a Policy Manager specialising in chemical weapons and high yield explosives. The salary range was between 245 and 285 thousand dollars a year. Hybrid role. Based in New York. The screenshot made the rounds on […] - [Anthropic hiring chemical weapons policy manager role](https://startupfortune.com/anthropic-is-hiring-a-chemical-weapons-expert-and-the-internet-lost-its-mind/) - [AI company chemical weapons expert job listing controversy](https://startupfortune.com/anthropic-is-hiring-a-chemical-weapons-expert-and-the-internet-lost-its-mind/) Philosophy already has a name for the study of where knowledge comes from and how it should be evaluated: epistemology. McElroy describes the company's core product, the Sustainability Code, or the Susty Code for short, as an application of that field, or what he and others call applied epistemology. The goal is to make it possible to quality-control the knowledge an AI produces, covering both facts and value judgments, before that knowledge is shared or becomes the basis for an action. Not every epistemologist agrees on what should qualify as knowledge, which is where the justificationist versus falsificationist split becomes central to the company's pitch. The distinction is laid out in the company's white paper, “What is the Primary Epistemology of Leading LLMs?”, which walks through a worked example comparing the two camps starting on page 42. ## Where the Susty Code differs According to the company’s founders, the [Susty Code](https://sustycode.ai) is built on falsificationism rather than justificationism, and it applies that method to both facts and values, truth and morality, rather than restricting itself to factual claims the way most epistemological tools do. They argue that every leading AI model today ships with guardrails of some kind, but none of those guardrails operate at the epistemological layer the way the Susty Code does. In their words, the product sits in a category of its own, positioned as an antidote to the misinformation and bad behavior that justificationist reasoning tends to produce. The company is pitching that positioning squarely into the AI Safety and Alignment space, which the founders name directly as the market the Susty Code is built for. Artificial Epistemics says its broader strategy is to work with the makers of leading AI models, agents, and harnesses to integrate its epistemic tools directly into their offerings, rather than to sell a standalone consumer product. Whether Gemini’s and Grok's makers respond to the study's findings is a separate question from whether the underlying framework holds up, and that is likely to be judged by the AI safety research community over time rather than settled by a single study. What the founders are proposing is a different point of intervention: instead of adding another layer of output filtering after a model generates a response, they want to endow models with a capacity to autonomously decide whether a claim is true or legitimate in the first place. Readers who want the full technical comparison between the two epistemological camps, including the worked example the founders point to directly, can find it in the white paper linked above. More on the company and its work is available on their site, sustycode.ai. ## Join the discussion [Open in the community →](https://startupfortune.com/community/) Almost there. Sign in and your reply posts straight away. [The Face Recommending Your Next Health Product Is Fake, the Money Leaving Your Wallet Is Not](https://startupfortune.com/the-face-recommending-your-next-health-product-is-fake-the-money-leaving-your-wallet-is-not/) There is a woman on Instagram who has over 300,000 followers. She posts videos of herself visiting grocery stores, baking bread, and warning her audience about the dangers of processed food. She speaks with conviction. She looks real. She has been building her following for months. She does not exist. According to the New York […] - [AI generated influencers selling fake wellness products](https://startupfortune.com/the-face-recommending-your-next-health-product-is-fake-the-money-leaving-your-wallet-is-not/) - [fake celebrity promoting health supplements without disclosure](https://startupfortune.com/the-face-recommending-your-next-health-product-is-fake-the-money-leaving-your-wallet-is-not/)