How Humans Form Beliefs About Someone or Something -A Psychology- and Biology-Grounded Perspective Toward Belief-Aware Artificial Intelligence

Introduction

Human belief formation is not random. Decades of research in psychology, neuroscience, and cognitive science show that beliefs emerge through systematic, measurable processes that integrate evidence, prior expectations, emotional context, and lived experience.

While human belief updating often resembles Bayesian reasoning, it is approximate, biased, and dynamically unstable. Beliefs strengthen, weaken, decay, and sometimes resist change altogether. Understanding these dynamics is essential, not only for explaining human cognition, but also for designing artificial systems capable of adaptive, self-consistent intelligence.

This paper explores human belief formation from psychological and biological perspectives, and investigates whether simplified mathematical relationships can meaningfully approximate belief dynamics. Our ultimate motivation is to inform the development of a belief-aware cognitive architecture for synthetic, self-conscious intelligence systems.

Objective

To examine how humans form, update, and stabilize beliefs using empirical insights from psychology and neuroscience, and to explore simplified mathematical abstractions that capture these dynamics without claiming strict Bayesian optimality.

Goal

The goal of this work is to extract actionable principles of belief dynamics that can be implemented in artificial systems, specifically, to support the design of a belief-aware architecture capable of:

• Maintaining internal beliefs as dynamic states
  
• Updating beliefs through evidence integration
  
• Accounting for bias, confidence, and uncertainty
  
• Allowing belief strengthening, decay, and revision over time
  

Such an architecture is a foundational requirement for any system aspiring toward self-modeling or self-conscious intelligence.

Topic 1: Evidence Integration in Human Belief Formation (With Bias)

Approximate Bayesian Reasoning in Humans

Humans do not perform exact Bayesian inference. Instead, they behave as approximate Bayesian reasoners, integrating new evidence with prior beliefs in a manner that is:

• Noisy
  
• Context-dependent
  
• Emotionally modulated
  
• Biased by trust, authority, and personal experience
  

Rather than strict probabilistic normalization, belief updating in humans is better described as a confidence-weighted adjustment process.

A simplified conceptual model can be expressed as:

B_t+1​=B_t​+α⋅E_t​−β⋅D_t​

Where:

• $B_t$ = current belief strength
• $E_t$ = strength of new evidence
• $\alpha$ = learning or trust rate (how much the evidence is believed)
• $D_t$ = decay or destabilization term (time, doubt, or counter-evidence)
• $\beta$ = decay sensitivity

Illustrative Example: Belief Formation Through Personal Experience

Scenario

An overweight individual is considering whether a newly proposed diet plan can effectively reduce body weight.

Date: 10 December 2025

Initial State

• Current belief:
  • There is a 20% chance that this new diet works.
• Evidence available:
  • None (no prior experience or trusted confirmation)

$$B_{t} = 0.20$$B_t​=0.20

At this stage, the belief is weak and unstable, easily influenced by future information.

Date: 11 December 2025

Exposure to Secondary Evidence

The individual watches a health show on television hosted by a renowned dietitian.

• Prior belief: 20%
• Evidence type: Authority-based, indirect
• Subjective trust weight: Moderate

Assuming:

• Evidence strength $E_t = 0.30$
• Learning rate $\alpha = 0.5$

$$B_{t+1} = 0.20 + (0.5 \times 0.30) = 0.35$$B_t+1​=0.20+(0.5×0.30)=0.35

The belief increases to 35%, reflecting improved confidence, but it remains tentative, as the evidence is indirect.

Date: 05 March 2026

Direct Personal Evidence

The individual follows the diet consistently and experiences a 15 kg reduction in body weight.

• Prior belief: 35%
• Evidence type: Direct, experiential
• Reliability: High
• Emotional reinforcement: Strong

Assuming:

• Evidence strength $E_t = 0.80$
• Learning rate $\alpha = 0.8$

$$B_{t+2} = 0.35 + (0.8 \times 0.80) = 0.99$$B_t+2​=0.35+(0.8×0.80)=0.99

The belief now enters a high-stability regime (≈99%), where it becomes:

• Strongly resistant to change
• Highly confident
• Yet still theoretically revisable under extreme counter-evidence

Importantly, this is not absolute certainty. Human beliefs rarely reach true 100% confidence and remain subject to decay or revision if outcomes change (e.g., weight regain).

Key Observations From the Example

1. Not all evidence is equal
   Personal experience carries far more weight than indirect or authority-based information.
2. Belief strength is dynamic, not static
   Beliefs strengthen with reinforcement and weaken without it.
3. Confidence saturates asymptotically
   Beliefs approach stability rather than reaching absolute certainty.
4. Time matters
   Without continued reinforcement, belief confidence can decay.
5. Beliefs are revisable states
   Even highly stable beliefs can change when confronted with strong contradictory evidence.

Here is A Formal Computational Model for Belief-Aware Artificial Intelligence: Belief Dynamics Module (BDM)

Implications for Belief-Aware Artificial Intelligence

This analysis suggests that artificial systems aiming for human-like cognition should not treat beliefs as static probabilities or permanent memory entries. Instead, beliefs should be modeled as:

• Dynamical internal states
• Sensitive to evidence reliability and context
• Subject to decay, reinforcement, and stabilization
• Resistant – but not immune – to revision

Incorporating these principles enables artificial agents to exhibit adaptive confidence, contextual reasoning, and self-consistent belief maintenance, which are prerequisites for higher-order cognition and self-modeling intelligence.