If we do not have the capacity to distinguish what’s true from what’s false, then by definition the marketplace of ideas doesn’t work. And by definition our democracy doesn’t work. We are entering into an epistemological crisis. [1]
Thus Barack Obama in 2020.
The current standard bearer for the epistemological crisis is undoubtedly Barack Obama’s successor, Donald Trump. Mr. Trump often prioritizes convenient lies and demagoguery over truth and has in effect made lying one of his main political tools [5][6]. The consequence is a proliferation of false beliefs resulting in solidification of tribalism, division, and a weakening of democracy.
Obfuscation and perversion of the truth has been an important tool of dictators since time immemorial. Daodejing says:
Therefore the wise ruler [ … ] tries by keeping people in ignorance to keep them satisfied and those who have knowledge he restrains from evil.
The dismantling of epistemology is a standard tool in the playbook of established and aspiring dictators. One of the first things a dictator would do is to take control of the mass media so that he (are there any female dictators?) could spread useful disinformation. In the past a coup would often start with the takeover of the state TV broadcaster. It’s a bit more complicated today.
He will then take control of all the “truth checkers” such as independent media, government institutions and authorities, and the judicial system. Science may go last since it is useful for the development of military technology but it is best to keep the scientists locked in.
I will write more about the multitude of converging reasons for the epistemological crisis in future posts. Here I will instead focus on what truth is.
Epistemological frameworks
What is truth? And how do we know something is true? This is a question that goes back at least to the ancient Greeks. Aristotle defined truth thus in the Metaphysics:
To say of what is that it is not, or of what is not that it is, is false, while to say of what is that it is, and of what is not that it is not, is true.
Aristotle’s theory is close to what’s today called the correspondence theory that says that true statements correspond accurately to objective reality; a statement is true if it matches the actual state of affairs in the real world.
Correspondence theory is one of many definitions of truth. Others include:
- Coherence theory: Truth is determined by how well a belief fits with other established beliefs. A statement is true if it’s logically consistent with the broader system of accepted truths.
- Pragmatic theory: Truth is what works in practice – ideas are true if they lead to successful predictions and outcomes in real-world application.
- Consensus theory: Truth is what a relevant community agrees upon through discourse and investigation. Knowledge is socially constructed through collective agreement.
- Constructivist theory: Truth is constructed through social, historical, and cultural contexts rather than discovered. There is no single objective truth independent of human interpretation.
- Postmodernism: Questions grand universal truth claims while examining how knowledge and truth are embedded in language, power relations, and cultural narratives. Doesn’t reject empirical evidence but analyzes how scientific knowledge is socially constructed and historically contingent.
- Social relativism: Truth claims are relative to particular social groups or cultures, with different groups potentially having equally valid but incompatible truths. Unlike consensus theory, doesn’t seek broader agreement.
- Cognitive relativism: Basic ways of thinking and reasoning are relative to cultural or linguistic frameworks. Even standards of logical coherence might vary between frameworks.
- Power-based epistemology: Truth is determined by who has the power to define it. Knowledge claims are analyzed in terms of their relationship to social power structures and whose interests they serve.
There are variations of the above such as the strong programme, standpoint epistemology, feminist epistemology, anti-realism, anti-foundationalism, epistemic relativism, and the list goes on.
Generative models
According to the active inference framework (AIF) introduced in earlier posts, organisms maintain internal generative models of the world that they use for perception and to guide their actions so as to maximize the desirability of future observations and thus their well-being1 (see e.g., this post and [2]).
Organisms continuously interact with their environment by choosing policies (sequences of actions) that maximize the probability of survival and procreation, and, in the case of humans, many other more abstract objectives far from the basic imperatives of life. The generative model is used to predict what observations each policy would lead to if enacted. The predicted observations are compared with some desired observations. The policy that is predicted to produce the most desirable observations is chosen for execution2.
On a basic level we have a rather good understanding of the kinds of things we want to see, hear and smell around us (and what we could do without). On higher levels our desires, like an exaggerated desire to accumulate wealth, may actually be detrimental to our well-being. Our generative models may be accurate while our desires are seriously off the mark.
The discrepancy between what we desire and what we actually observe is in AIF called surprise or surprisal and is approximated by a computable quantity called free energy [2]. An organism always strives to infer a policy that minimizes the expected free energy.
The quality of the generative model is determined by its ability to produce accurate predictions of what the organism will observe during the execution of a certain policy, whether those observations are desired or not. A generative model that produces accurate predictions enables the organism to reliably choose good policies.
Simple organisms have simple generative models and few policy choices. A bacterium observes the concentration of nutrients outside its membrane [3]. The policy inference part of the bacterium’s model basically says if you move that way, you will find more food (higher nutrient concentration), if you go the other way, you will find less. The bacterium uses this prediction to – you guessed it – move towards more food.
Humans have, in addition to their sophisticated built-in generative model, an extended model of the world based on tools such as language, mathematics, logic, and computer code (e.g., simulations). Humans can also use a multitude of additional sensors in addition to the natural ones such as telescopes, microscopes, and spectrometers to make observations. We can say that humans have an extended generative model with an enhanced sensory system at their disposal enabling humans to make much more accurate predictions and observations than other animals and to construct and choose among complex policies to satisfy complex desires. They can do more than move in the nutrient gradient although that kind of behavior is not unheard of either.
Without built-in (and now also extended) generative models that produce accurate predictions we would not survive. This is why I suggest the following definition of truth:
Truth is the accuracy of the prediction of observations obtained with the help of a generative model.
Example 1: The conservation of energy (part of a model of physics) says that the sum of kinetic energy and potential energy remains constant. Mathematically: \(mgh = \frac {mv^2}{2} \Rightarrow v = \sqrt {2gh}\). The model predicts that if we jump from the height of one meter we will hit the ground with the speed of 4.42 m/s. If we observe a number of such jumps we can conclude this to be a very accurate prediction.
Example 2: The “model of my life” contains the information that I was born on June 27, 1959. The model predicts that I will get a senior discount on train tickets which turns out to be true, based on observing my latest ticket purchases.
Multiple truths
The prediction accuracy of a model depends on the experiment. Let’s say that the model says “you will die of any snakebite”. This is obviously a poor prediction of the outcome of a comprehensive scientific examination of a large number of snakes. Some will not be venomous. But if the experiment instead measures survival chances in the jungle, then the model will do a rather good job as it basically tells you to avoid all snakes to be on the safe side.
Truth thus depends on how much the models has to say about the particular situation and the kind of “experiment” that will be performed. Most people’s models don’t predict the outcomes of experiments in the large hadron collider. My own generative model does a rather poor job at predicting where my golf ball will end up as a consequence of my swing.
Even though the truth depends on the experiment, it can still be kept value-free. We don’t need to specify any particular purpose or utility for the experiment. It can be any purposeful or worthless experiment.
Verifying predictions
Predictions produced by generative models can be evaluated with respect to accuracy using different types of verification mechanisms, all of which attempt to verify the prediction against reality. Organisms including humans continuously observe their environment and compare the observations to the desired observations. If they don’t agree, then the organism takes action to better align the actual observation with the desired observation. Our behavior is in a sense tantamount to continuous error correction, actual versus desired. If the model generates consistently bad predictions, then the model is modified through learning.
The most prominent society level truth checking mechanisms are science, journalism, and the judicial systems.
Scientists build sophisticated models of the world and make predictions using those models. They then run experiments and make observations. If the observations from the experiment agree with the predictions of the model then their belief in the accuracy of the model is strengthened; the model is verified. If not, then the model (or the experimental setup, if inadequate) needs to be modified.
Journalists typically check the accuracy of “predictions” after the prediction has been generated by somebody’s generative model and released in the wild. A current example (January 4, 2025) is the rumor (prediction), fact-checked by snopes.com, that a well-known venture capitalist’s family had rented an apartment in Venice Beach, Los Angeles, to Hunter Biden in 2019 but was stiffed $300,000 in rent payments, and that Biden offered to pay the debt in “art made from his own feces”. The actual observations made by snopes.com didn’t support this “prediction”.
Courts attempt to establish whose “generative model” has produced the most accurate description of the circumstances around a crime by comparing the narratives (predictions) produced by each generative model. They do this mainly by observing physical evidence and witness statements.
Summary
Generative models hold information about the ontology and the dynamics of reality. They can be used to make predictions about the reality. The accuracy of such a prediction is the truth. Organisms, organizations and other entities with generative models that consistently produce inaccurate predictions where it matters will eventually be eliminated by the dispassionate arbiter of truth, reality.
Links
[1] Jeffrey Goldberg. The Atlantic, November 16, 2020. Interview with Barack Obama.
[2] Thomas Parr, Giovanni Pezzulo, Karl J. Friston. Active Inference. MIT Press Direct.
[3] Chantranupong L, Wolfson RL, Sabatini DM. Nutrient-sensing mechanisms across evolution. Cell. 2015 Mar 26;161(1):67-83. doi: 10.1016/j.cell.2015.02.041. PMID: 25815986; PMCID: PMC4384161.
[4] Why truth matters. Ophelia Benson, Jeremy Stangroom. Continuum.
[5] Trump’s false or misleading claims total 30,573 over 4 years. Washington Post. January 24, 2021.
[6] False or misleading statements by Donald Trump. Wikipedia.
Appendix: Comparison with existing epistemological frameworks
The following table summarizes how the definition of truth (called “Definition” below) suggested in this post compares to the frameworks for truth introduced above.
| Framework | Comparison with Definition |
|---|---|
| Correspondence theory: True statements correspond accurately to objective reality; a statement is true if it matches the actual state of affairs in the real world. | There is no way for us to know the objective reality, the “ding an sich”. We can only make observations. Definition is based on observations and inferred mental states, not the “actual state of affairs”. |
| Coherence theory: Truth is determined by how well a belief fits with other established beliefs. A statement is true if it’s logically consistent with the broader system of accepted truths. | Consistency is a necessary but not sufficient condition for truth. The “system of accepted truths” also needs to be anchored in reality at a sufficient large number of points. |
| Pragmatic theory: Truth is what works in practice – ideas are true if they lead to successful predictions and outcomes in real-world application. | Truth in the pragmatic theory is judged not just by whether predictions are accurate, but by their practical value for human life and experience. A belief might make accurate predictions but still be considered less “true” if it leads to harmful outcomes. Definition is value-free. |
| Constructivist theory: Truth is constructed through social, historical, and cultural contexts rather than discovered. There is no single objective truth independent of human interpretation. | This is just wrong, at least in natural sciences. Reality does its thing wether we interpret it or not. Gravity is not a construction, its a law of nature. |
| Postmodernism: Questions grand universal truth claims while examining how knowledge and truth are embedded in language, power relations, and cultural narratives. Doesn’t reject empirical evidence but analyzes how scientific knowledge is socially constructed and historically contingent. | Gravity and other laws of nature are not “embedded in language” or “power relations”. They are embedded in reality. While power relations may affect the construction of models, the quality of the model will remain defined by how accurately it predicts reality. |
| Social relativism: Truth claims are relative to particular social groups or cultures, with different groups potentially having equally valid but incompatible truths. Unlike consensus theory, doesn’t seek broader agreement. | Same comment as above applies. Reality is the arbiter of truth. Aspects of reality may vary from culture to culture. The generative model may include aspects from many different cultures but these would still be conditionally true. The laws of nature stay the same across cultures. |
| Cognitive relativism: Basic ways of thinking and reasoning are relative to cultural or linguistic frameworks. Even standards of logical coherence might vary between frameworks. | Same comment as above applies. |
| Power-based epistemology: Truth is determined by who has the power to define it. Knowledge claims are analyzed in terms of their relationship to social power structures and whose interests they serve. | This theory has very little with truth to do at all. The pope once vehemently defended an earth-centric view of the universe. His authority didn’t make it true. |
None of the frameworks correspond exactly with the definition of truth offered here. There may be other frameworks that I have missed. Let me know!
Footnotes
- Desires can be expressed both in terms of observations and the resulting mental states, that in turn for a well-calibrated generative model have some useful correspondence to real-world states. I will here assume that desires are expressed in terms of observations as this is perhaps the most concrete and intuitive way to think about outcomes. This is particularly true when we talk about extended generative models that predict outcomes of scientific experiments. Such experiments first and foremost result in observations. The Higgs boson may not evoke any relevant mental state in most people. ↩︎
- The “desirability expression” actually has two terms: The first term reflects the immediate desirability of the observation or mental state. The second term promotes visiting mental states that one can observe. It basically tries to keep the organism out of the dark or other places where observations are inaccurate. ↩︎