How Big AI is plundering the concept of Truth

Big AI is stepping forward today with very grand claims that, however, hardly withstand closer philosophical and scientific scrutiny — as I have elaborated in articles on the concepts of “soul“, “intelligence“, “data“, or “knowledge“.

The PR and marketing assertions of Big AI that AI will soon deliver real, actual AGI, that it possesses consciousness and emotions, can confidently be dismissed as philosophical and scientific nonsense.

It seems more likely that a ‘AI Club‘ systematically – due to its economic structures and needs – hinders efficiency, innovations and transformative disruption to reach a GenAI or AGI to keep the investment bubble on track. The unified battle cry of the Big AI industry is ‘Scaling+compute+electricity‘ and magically we will hold AGI in our hands and to our service. That’s non-sense to me.

Back to our topoi: Today I would like to turn to one of the biggest — let us kindly call it — misconceptions of the Big AI industry: the claim that it is just truth-seeking, committed only to truth, and similar marketing rhetoric. What is currently offered as “truth” by the various AI models, however, has nothing whatsoever to do with truth.

So, what is “truth” in the world of the Big AI industry? Which concept of truth is deployed, if any concept of truth exists at all in the Big AI industry?

“Truth” in the World of AI

When an artificial intelligence – especially a Large Language Model – answers the same question, the internal operation roughly works as follows:

• Which token sequence has the highest conditional probability in the training data (or the retrieved sources)?
• Does this sequence align with the current alignment goals (e.g., “helpful, harmless, honest”)?
• If not: How can I rephrase the most probable sequence so that it still passes as “true”?

The result often feels astonishingly true to humans – yet it is almost never true in the philosophical and scientific sense. Why? AI is deploying 7 statistical tools/concepts to get results which I deny to call “truth”. Those 7 tools/concepts are:

How AI Fabricates “Truth” Today – A Systematic Overview

1. Ground Truth – The Supposed Bedrock Truth. In supervised learning, “ground truth” is the holy grail: the information labelled as correct by humans. The term “ground truth” suggests an objective, indisputable foundation. In reality, however, this foundation is often human-made and subjective. Two main biases are at work: wrong labeling and wrong history data. Labeling Bias: The “truth” is frequently defined by poorly paid crowd-workers who label data. If they decide that an image shows an “offensive gesture,” this becomes absolute truth for the model—regardless of cultural context. Historical Bias: When an AI is trained on historical data (e.g., hiring decisions from the past 20 years), it learns the prejudices of the past (e.g., women earn less) as “factual truth” about the world. The sad reality is: Ground truth is almost always human-made, often created by underpaid crowd-workers, culturally biased, and sometimes simply wrong (garbage in, garbage out).
2. Probability Is Not Understanding (Correlation vs. Causation). This is the classic epistemological problem of statistical AI. Example: The Rooster and the Sun: If an AI observes that every morning the rooster crows and shortly afterward the sun rises, it calculates a high probability for the connection. For the AI, the statement “The rooster causes the sunrise” is statistically “true,” even though it is factually false. Probalistic truth is an echo chamber of ignorance.
3. Missing Causality. AI models are extremely good at detecting patterns (correlations) but do not understand causal mechanisms. Example: For a doctor, it is not enough to know that symptoms correlate; the doctor must know why (the pathology) in order to arrive at the truth about the disease. Correlation is not causality. Fact.
4. Coherence Instead of Correspondence. LLMs primarily optimise for text that sounds linguistically plausible (coherence theory of truth). But language is not truth. Scientific and philosophical fact. LLMs are not trained to make their statements correspond to the real world (correspondence theory), because it is not possible. The algebra based on the very concept of scaling and the ideology of nerdism does not allow it. Consequence: Hallucinations – sentences that sound perfect but are fact-free and pure fantasy.
5. Consensus and Frequency-Based Truth. The more often a statement appears in the training data, the “truer” it becomes for the model. Majority wins: If the internet states 10,000 times that the Earth is flat and only 500 times that it is a sphere, the statement “The Earth is flat” has a higher statistical probability.The echo-chamber effect: A lie that is repeated often enough becomes truth for the model. AI does not distinguish between credible sources and frequent sources (even though modern models attempt to weight sources, the fundamental problem remains).
6. Pragmatic and Alignment-Based Truth. Through RLHF and Constitutional AI, an additional layer has been added: “True is what humans rate as honest, helpful, and non-deceptive.” This causes models to prefer saying “I don’t know” or “That is controversial” rather than producing false certainty—not out of insight, but because humans rewarded it in evaluations. AI becomes a liar, a sycophant.
7. Source-Based Truth (RAG & Live Search). Since about 2023–2025, factual queries are increasingly answered with Retrieval-Augmented Generation. Here, what is considered true is what current, supposedly trustworthy sources unanimously say. Almost all serious applications (for example ChatGPT Search, Perplexity, Grok with DeepSearch, Claude with web access, etc.) now use Retrieval-Augmented Generation. This massively reduces hallucinations on factual questions — but as soon as the model starts summarizing, drawing conclusions, or creatively combining information, the “coherence-over-factuality” problem immediately returns. Hallucinations are back.

Some Critical Notes on Big AI’s Concept of Truth

What at first glance looks like a robust hybrid system proves, upon closer inspection, to be epistemologically and ontologically as well as scientifically extremely fragile. Some non-exhaustive notes.

1. “Ground Truth” Is Not Truth, but Institutionalised Subjectivity. Every label is an act of human interpretation. If 90 % of the labels are antisemitic, racist, or simply wrong, then AI learns precisely that as “truth.” AI thereby conserves and amplifies societal biases—and calls it “objective truth”. Wrong.
2. Correlation Is Confused with Causation. Judea Pearl put it most clearly: “Deep learning has catapulted us to the bottom rung of the ladder of causation—and nailed us there,” (quote paraphrased). A model can perfectly predict that the sun rises after the rooster crows and will call it “objective truth”. Wrong.
3. Truth Through Mere Repetition (Echo-Chamber Truth). Emily M. Bender, and others have shown: The more often a piece of misinformation circulates on the internet, the higher its token probability. Without a real source-criticism module (which does not exist), the majority wins over truth and AI calls it “objective truth”. Wrong.
4. The Stochastic Parrot – Form Without Meaning. LLMs have no referential semantics. Language is not truth and reality. LLMs aka “AI” manipulate symbols according to statistical patterns without ever “knowing” what they are talking about. The computational linguist Emily Bender coined this term (Stochastic Parrot). It states that language models merely manipulate forms (syntax) without having any access to meaning (semantics). AI calls the output “objective truth”. Wrong.
5. Plausibility vs. Truth. A model can write a perfectly logical-sounding sentence about a chemical that does not exist. Because the sentence is grammatically flawless and written in the style of a scientific paper (coherence), these “hallucinations” appear true to the human. Humans are conditioned to attribute intelligence and truthfulness to fluent language. AI exploits this human heuristic flaw and calls the reult “objective truth”. Wrong.
6. No Quale. No consciousness. LLMs can write a perfect sounding article about “the smell of rain on hot asphalt”—without ever having smelled rain. AI has no senses, no empiric, no quale, no consciousness. However, AI will call its article on smell of rain on hot asphalt “objective truth”. Wrong.
7. Static Models in a Dynamic World. Even models with live internet access are only as up-to-date as their last retrieval query. Most AI models are “frozen” (training cut-off). Outdated facts stay facts. A statement like “The Queen is the head of state of the UK” was “true” (ground truth ≈ 1) for a model trained in 2021. Today it is false. Correction does not really happen and is very complicated. You have quickly a cascadian effect. Finally, the lack of context: Truth is often context-dependent. “It is warm” is true at 20 °C in Siberia but false in the Sahara. Vector-based models frequently struggle to capture these subtle contextual shifts in truth without explicit instructions. As soon as they reason over multiple steps, they inevitably mix old internal knowledge with new data—a recipe for subtle errors and more hallucinations declared as “objective truth”. Wrong.
8. Alignment Is No Guarantee of Truth. The strongest safeguard (“be honest”) is once again just another optimisation metric. At its core, it says: “Avoid statements for which humans would accuse you of lying.” That is a pragmatic, anthropocentric criterion—not an epistemic virtue disguised as ‘objective truth”. Wrong.
9. The Deepest Problem. Judea Pearl again: “Data are profoundly dumb.” Data tell us what correlates, never why. They contain no negation, no counterfactual, no experiment. Everything an AI produces as “objective truth” is therefore always only a highly sophisticated reflection (Widerspiegelung) of what humans once wrote down somewhere. AI finally presents it as ‘objective truth”. Wrong.

Useful Approximation, but Epistemological Bankruptcy

Artificial intelligence in 2025 has no unified, philosophically and scientifically grounded concept of truth. It operates with a hybrid, multi-layered construct consisting of:

• statistical frequency
• human label subjectivity
• coherence optimisation
• pragmatic usefulness
• and (at best) current sources/data

The result is remarkably usable for everyday questions, translations, code generation, mathematical extemporations, or simple fact retrieval; maybe. Yet it is miles away from what philosophy, science, or even common sense understand by “truth”: the highest, most accurate understanding of ‘Wirklichkeit’ and reality.

As long as AI systems are trained at their core on correlation instead of causation, on repetition instead of verification, and on human labels instead of objective reality, their concept of truth will remain instrumental, approximate, and – despite all practical usefulness – epistemologically, ontologically and scientifically deficient. In short, we have built machines which perfectly imitate truth without ever touching truth. The AI concept of truth is based on illusion generated by pure data and poor labelling by humans.

The declaration of bankruptcy is so comprehensive and profound that I do not need to delve further into the depths of philosophy here in order to demonstrate the complete untenability and absurdity of the concept of truth used in the Big AI industry. This checkmate does not need to be played out to the very last move.

Is there a way out? Yes. Pearl says – in my understanding of his writings – that each rung requires an entirely new kind of data and an entirely new mathematics. You cannot simply pour more data or more computing power into rung 1 and suddenly reach rung 2 or 3.

We have therefore to re-conceptualise our understanding of the world by deploying philosophy and real science by re-launching their deep symbiosis to get a new mathematics and data. There are first solutions at the horizon, but the Big AI industry want free lunch and a cheap and good-selling paradigm: scaling+compute+electricity. But the truth is, with this paradigma you will never get intelligence or AGI. However, progress will not be likely within the next 10 years. We will see only more scaling and more PR- and marketing speak.

Tools used for research, translation, proof reading, verification of codes/equations, pic generation etc.: LLMs / SE / BusinessSoftware / Parsers / DB/ Websites etc. All articles: Creative Commons BY-NC-ND 4.0 (Attribution-NonCommercial-NoDerivs).