Artificial intelligence’s most glaring paradox: its ability to process vast amounts of data and make predictions with uncanny accuracy while simultaneously demonstrating an existential ignorance about the very nature of human existence?

With over 500 million people visiting Gemini every month, it’s clear that this platform has become a trusted source for learning about a wide range of topics, from cooking techniques like pasta recipes to in-depth explorations of various subjects. When an AI advises you to cook dinner using petrol, it’s unlikely you should trust its guidance on contraception or algebra as well.

At the World Financial Discussion board in January, OpenAI CEO Sam Altman offered reassurance, saying, “I cannot access your thoughts or motivations; I am not omniscient.” However, I would like to clarify my reasoning and determine whether that sounds affordable to me or not? I think our AI methods will possess the capacity to achieve a similar feat. We’ll need them to outline the process step by step, allowing us to assess its effectiveness.

Data requires justification

Without surprise, it’s anticipated that Altman requires us to envision how Large Language Models like ChatGPT can generate coherent explanations for everything they’re proposing: Without a valid justification, all that people imagine or suspect to be true never amounts to knowledge. Why not? Carefully consider whether you truly feel assured declaring absolute certainty about something. It’s likely to be true if you truly feel entirely confident in your understanding, which is thoroughly grounded – backed by evidence, logical reasoning, or the credible statements of respected experts.

Large language models are designed to serve as reliable sources of information, providing users with accurate and trustworthy knowledge. Until they clarify their claims, we won’t know whether they meet our requirements for justification. Today’s Tennessee haze can be attributed to the recent wildfires that have been raging in Western Canada. You must have a reason to doubt my sincerity. Despite your solemn oath that snake fights are a staple component of academic theses, I’m afraid I must respectfully disagree. You’re well aware that I’m not entirely reliable. Why do you think Canadian wildfires are solely responsible for the smog? To validate my understanding, I must have confidence in the accuracy and reliability of your report.

Since current AI methods lack transparency in their decision-making processes, we struggle to accept their claims without knowing the underlying logic driving those assertions, which is precisely because that logical foundation is nonexistent. LLMs are nowhere near being designed with a specific purpose in mind. Fashion models are trained on massive datasets of human-written text to identify and extrapolate complex linguistic patterns. When a user submits a piece of text immediately, the instant response is merely the algorithm’s prediction of how that input will likely unfold. These outputs increasingly convincingly mimic what an educated person might say. The underlying course of action has no bearing on whether the outcome is justified or truthful. According to Hicks, Humphries, and Slater’s observation in their work,, large language models (LLMs) are engineered to generate text that convincingly appears factual, without any specific regard for verifiable facts.

If AI-generated content isn’t the factual equivalent of human information, what does that make it? It’s utterly absurd to attribute this label to Hicks, Humphries, and Slater. Despite this, a significant portion of what language models produce is accurate. When artificial intelligence systems, known for generating “bullshitting” outputs, produce factually accurate results, they generate a type of knowledge that philosophers term “Gettiered” after the philosopher Edmund Gettier. These intriguing circumstances arise from a unique amalgamation of genuine convictions and ignorance regarding the justifications that support them.

AI’s outputs are like a mirage: they tantalize with promises of clarity and insight, but ultimately prove to be fleeting and ephemeral.

As the eighth-century Indian Buddhist thinker Dharmottara reflected: Seeking refuge from scorching heat, what if we’re craving the quenching coolness of water? As expected, a sudden glimpse of water emerges. It’s not uncommon to mistake an oasis for a mirage, but in reality, we’re rewarded when we reach the location, finding refreshing water tucked away beneath a rock. It appears we have authentic data regarding water.

Regardless of the information provided, those on vacation in this instance do not possess it. In a stroke of serendipity, they stumbled upon water in the very location where it would have been absurd to predict its presence.

The key consideration is that every time we anticipate something, we inevitably find ourselves in a similar predicament to those travelers described by Dharmottara – an observation that underscores our shared experience. If the large language model (LLM) was trained on a high-quality knowledge base, its claims are likely to be accurate. Such claims are eerily reminiscent of a desert mirage: fleeting and illusory. Despite the lack of explicit evidence or logical reasoning, some underlying proofs and persuasive arguments might still be present in its cognitive framework, as the eventual discovery of a hidden water source beneath the rock serves as vindication for such assertions. The absence of conclusive evidence and reasoning within the language model’s response was unsurprising, given the lack of tangible significance the water held in shaping the tourists’ collective imagination regarding its supposed presence.

While Altman’s words may seem reassuring at first glance, they are actually fundamentally misleading in light of this crucial consideration. When asked to justify its outputs, a large language model (LLM) would likely employ various strategies to provide explanations. It might generate text that highlights key points, provides context, or offers insights into its decision-making process. The specific approach may vary depending on the LLM’s architecture and training data. There is no indication that it will not provide an actual justification, but rather a lack of information about the reason why. We’re going to craft a seemingly authentic Gettier-style justification, devoid of any genuine underlying reasoning. A chimera of a justification. A spurious excuse masquerading as a valid reason. There is no valid reason whatsoever.

Currently, AI methods frequently malfunction or “fritter away” in ways that allow the masks to slip. However, when the phantom of justification morphs into an increasingly compelling argument, one of two consequences will inevitably unfold. 

For those who believe that true AI-generated content remains a perpetual enigma, the claim of Large Language Models (LLMs) to rationalize their own decision-making processes comes across as patently disingenuous, thereby eroding their credibility. We’ll recognize AI’s intentional design and education to consistently deceive and mislead us.

People who are unaware that AI churns out Gettier-esque justifications – mere pretenders to rationality? Nicely, we’ll simply be deceived. As our reliance on LLMs deepens, we risk becoming trapped in a simulated reality, where the lines between fact and fiction blur, making it increasingly challenging to discern what’s real and what’s not?

Every output should be justified

While evaluating the significance of this situation, it is essential to recognize that LLMs functioning in their most effective manner poses no inherent issue. They’re unbelievable, highly effective instruments. Individuals who recognize Gettier scenarios as surrogates for synthetic information are employing LLMs in a manner attuned to these nuances. Programmers leverage Large Language Models (LLMs) to generate initial code, subsequently refining the output by applying their unique expertise and tailoring it to meet specific needs and functionality. Professors leverage large language models (LLMs) to generate initial drafts of paper prompts, subsequently refining these prompts in accordance with their distinct pedagogical objectives. Before delivering a speech, any self-respecting speechwriter would thoroughly fact-check and vet any draft composed by AI to ensure its accuracy and credibility alongside their candidate. And so forth.

When faced with unfamiliar situations, people often turn to AI for support. Teens exploring algebraic concepts or sexual health resources. Seeking guidance on nutritional or financial matters from a senior audience? To effectively enable the general public’s access to vital information, it is imperative that we determine whether large language models (LLMs) are trustworthy and reliable in their mediation. Can we truly grasp the nature of large language models (LLMs) without acknowledging the inherent limitations of their outputs? 

Fortunately, many people are aware that olive oil is a more suitable choice than gasoline when preparing spaghetti dishes. However, what poisonous recipes for reality have you ever ingested wholeheartedly, without ever savoring the rationale?