The research program “Credition” has provided a solid framework to understand the phenomenon of believing and belief formation.

“Credition, the processes of believing, are fundamental brain functions that enable a non-human animal or human being to trust his/her inner probabilistic representation(s). Credition is based on neural processes, including perception and valuation of objects and events in the physical and social environment in secular as well as religious contexts. By predictive coding, credition guides one’s actions and behaviors through reciprocating feedback involving learning (Angel et al., 2017).”

Artificial Intelligence (AI) can be understood as a sequence of algorithms (that is, the mechanical application of some predefined steps) that is applied to a set of data and that generates a probabilistic representation of these data with the aim of making predictions, inferring a consequence or selecting the best possible option. AI can be understood as a machine that supports the formation and valuation of beliefs in the human and can be understood metaphorically as a belief-machine itself.

Artificial intelligence encompasses techniques such as clustering or pattern recognition. However, this paper focuses primarily on Reinforcement Learning (RL), where intelligent agents take actions in an environment to maximize a reward and punish mistakes. In RL, there are steps that are metaphorically described as perception (collecting new data), decision (selecting the optimal action), valuation (evaluating the outcome of a decision), or learning (the successive improvements in valuation obtained through repeated cycles of decision and valuation). For instance, a RL system could learn to play chess through cycles of selecting a move and valuing the possible positions.

For this reason, it is extremely interesting to examine AI from the lens of the credition process: understanding how AI works can give us insights to inform our hypothesis about the workings of credition. In parallel, acknowledging that AI should support belief formation helps to design it better and make this support as effective as possible.

Figure 1, taken from Seitz et al. (in press), presents a schematic representation of how a RL process can be understood in the context of credition, displaying the different levels of memory that are at play (Rolls, 2000). The inner probabilistic representation (which we refer to as “belief”) is built around the data received (perception) and is also used to value future actions (valuation). This representation is used to select the preferred course of action and predict its outcome. When new data are collected about the outcome (the prediction error), the probabilistic representation is updated through a new valuation in the process of RL. It is important to note that, according to this model, the credition process happens in a spontaneous and automatic manner, below the level of awareness.

For instance, one particularly interesting insight from the Creditions model and RL is the interpretation of the balance of exploration (which, in this context, we will understand the examination of new decisions or beliefs not tested before) vs. exploitation (the use of the existing belief system to make a decision in an efficient manner). This balance, which is key to the performance of RL algorithms, can be seen as an essential feature of belief formation and update that depends greatly on the psychological characteristics of each individual. In addition, In RL, each new data point is integrated into the probabilistic representation of the world. The specific manner in which this is done can be represented as an error minimization strategy. This model could be tested in experimental settings to improve our current understanding of how beliefs are formed and updated.

In addition, there also are interesting insights outside RL that can help us improve our understanding of the mechanisms of credition.

As explained in the section “Introduction,” the synergies between understanding AI and understanding belief are bidirectional, so there are positive outcomes that can be expected if we introduce what we know about belief formation into the way we design and use AI.

There are two main issues that plague the applications of AI: overfitting and machine bias. We need to remember that AI extracts patterns from the data that it receives for training, so it deeply depends on these data.

In overfitting, the data provided to the machine is not enough to be able to generalize. Much like a student that, instead of understanding, learns examples by heart, the algorithm fails when a new situation is considered.

In the same way, it is possible that the data we present to the algorithm does not represent reality fairly. For instance, it has been well documented that some algorithms disfavor African Americans, for instance when calculating their probability of recidivism in crime with the objective of deciding whether to grant them parole (Hajian et al., 2016). This was due to the database that was used for training containing a higher proportion of African American recidivists.

The problem with overfitting and machine bias is that they are not easy to detect. Most applications of AI are designed as black boxes, so we only have access to their specific predictions for every case but not to any reasoning behind them. Without detailed analyses, for instance, it is not possible to determine that the prediction is based on race, age, sex, or any other discriminatory variable. This means that when black-box AI is used in a high-stake decision (such as granting parole), this can have disastrous consequences.

As I have presented, black-box AI only shares predictions, and this can have disastrous consequences in high-stake problems. However, there is an emerging field within AI, Interpretable ML, which does allow for the understanding of models and their dynamics. This makes it possible avoid the problems derived from overfitting and machine bias (Molnar, 2020). The basic idea behind Interpretable ML is that, for many problems, it is possible to create AI that is so simple that can be expressed in rules understood by a human, and yet result in accurate predictions. These models would be the opposite to black boxes, they are transparent decision rules that can be understood and discussed.

For instance, Rudin (2019) developed an alternative to the parole algorithm that was based only on prior violent crimes and age. This shows that it is possible to create AI that better fits the way we form beliefs and is a more efficient support for our decision making.

Understanding AI and understanding belief formation have interesting bidirectional synergies. From explaining the logical derivation of beliefs and their internal consistency, to giving a quantitative account of mightiness, AI still has plenty of metaphors to illuminate belief. Potentially, these can be used to simulate belief systems and arrive to testable predictions.

In addition, acknowledging what our belief processes have that AI lacks (mainly, the capacity to reflect, rationalize and communicate that is allowed by semantic coding) makes it possible for us to focus on creating AI that can better support decisions such as Interpretable ML.

The original contributions presented in the study are included in the article/supplementary material, further inquiries can be directed to the corresponding author.

The author confirms being the sole contributor of this work and has approved it for publication.

The author declares that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.