Cognitive Processes in Humans and AI

Updated 24 September 2025

Cognitive Processes in Humans and AI are the mechanisms by which information is acquired, encoded, and processed, integrating algorithmic compression and structured reasoning.
The study highlights the integration of information theory with causal and compositional models to explain both biological and artificial learning and decision-making.
Hybrid neuro-symbolic strategies and adaptive memory systems are essential for mitigating biases and enhancing the synergy between human cognition and AI.

Cognitive processes in humans and AI refer to the mechanisms by which information is acquired, encoded, processed, abstracted, retained, retrieved, and utilized to guide behavior, reasoning, and learning. In contemporary research, the paper of these processes is grounded at the intersection of information theory, algorithmic complexity, neuroscience, behavioral experiments, and machine learning. This article reviews how foundational principles, formal models, and comparative studies elucidate the similarities and contrasts in cognitive architectures between biological and artificial agents, with an emphasis on algorithmic, computational, and resource-constraint perspectives.

1. Algorithmic and Information-Theoretic Foundations

The computational nature of cognition in both humans and AI is best understood through information-theoretic and algorithmic lenses (Gauvrit et al., 2015). Traditional views of intelligence—such as passing the Turing test via exhaustive look-up tables—are rejected as implausible for biological systems due to insurmountable resource constraints. Instead, cognition is posited to work as a form of algorithmic compression: human and animal minds detect and represent regularities with minimal code length, converging on concepts akin to Kolmogorov complexity.

Central to this approach is the notion of algorithmic probability $m(s)$ and Kolmogorov-Chaitin complexity $K(s)$ :

$m(s) = \sum_{p: U(p) = s} 2^{-|p|}$
$K(s) = \min \{ |p| : U(p) = s \}$
$K(s) = -\log_2(m(s)) + O(1)$

Here, $U$ is a universal prefix-free Turing machine and $p$ a program halting on output $s$ . This framework explains empirical psychological phenomena: people and animals are biased towards compressed, lower-complexity representations (e.g., preferring “chunkable” lists in working memory, underestimating truly random sequences if they are algorithmically simple).

Adaptation of these measures to short strings used in experiments is enabled by the Coding Theorem Method (CTM) and its ACSS implementation, while the Block Decomposition Method allows quantitative estimation of complexity for larger or multi-dimensional data.

Significance: This information-theoretic perspective not only models cognitive biases but also furnishes a testable, predictive basis for comparing biological and artificial intelligences under universal resource constraints.

2. Structured Models: Causality, Composition, and Intuitive Theories

Human cognition exhibits a propensity for causal modeling, compositionality, and the use of domain-intuitive priors (Lake et al., 2016). Unlike black-box pattern recognition, human learning incorporates:

Causal Modeling: Humans and advanced AI systems benefit from explicit generative models that reconstruct how observed data are produced by underlying programs (e.g., Bayesian Program Learning):

$P(x) = \sum_{p} P(x|Program=p)P(Program=p)$

Compositionality: High-level concepts (letters, words, scenes) are composed hierarchically of primitives (strokes, phonemes, parts) and relations, enabling rapid generalization through re-use and rearrangement.
Intuitive Theories: Innate or rapidly acquired “intuitive physics” (object permanence, gravity) and “intuitive psychology” (agency, intention) serve as structured priors for perception, prediction, and planning.
Meta-Learning (Learning-to-Learn): The ability to adjust inductive biases through developmental experience allows one-shot learning and adaptability in sparse-data regimes.

Contrast to AI: While current deep network architectures excel in specific tasks when given vast training data, their inability to construct explicit causal explanations or meaningfully decompose input impedes rapid transfer and explanation. Hybrid methods—combining deep learning with structured probabilistic programs or physics-based simulation—are identified as necessary to bridge this gap.

3. Dual-Process Frameworks and Cognitive Bias

Human cognition is not monolithic; it is often modeled as a dual-process or multi-process system (Booch et al., 2020, Rastogi et al., 2020). Key elements include:

System 1: Fast, intuitive, heuristic-driven, operating with limited information and minimal computation—yet effective in routine or familiar tasks.
System 2: Slow, deliberative, sequential, capable of explicit logic, abstraction, introspection, and override of System 1 errors.

AI research draws on this division through neuro-symbolic hybrids, combining deep learning (pattern-recognition) components with symbolic reasoning and meta-cognitive modules. Notably, cognitive biases such as anchoring emerge when humans over-weight external prompts (e.g., AI suggestions); this is captured in biased Bayesian models:

$P(Y|D, f(M)) \propto P(D|Y)^\alpha \cdot P(f(M)|Y)^\beta \cdot P(Y)^\gamma$

where the exponent $\beta > 1$ encodes the degree of anchoring to AI predictions.

Empirical Results: Task manipulations (e.g., increased deliberation time) mitigate biases, whereas increased workload or forced-correction requirements exacerbate cognitive shortcuts, leading to over- or under-correction of automated suggestions (Rastogi et al., 2020, Beck et al., 10 Sep 2025).

4. Memory, Knowledge, and Adaptation

Human cognition relies on multi-layered memory systems—short-term/working, long-term, episodic, and procedural (Chen et al., 2023, Salas-Guerra, 6 Feb 2025, Oakley et al., 3 May 2025). In AI, analogs are realized as attention-based working memory, external context databases, and dynamic long-term knowledge graphs.

Short-term/Working Memory: Temporarily holds relevant context for ongoing reasoning and dialogue. In AI frameworks, this is called “conversation context” (Salas-Guerra, 6 Feb 2025).
Long-term Memory (Interaction Context): Accumulates experience, user data, and historical knowledge, supporting personalization and adaptive behavior.
Dynamic Knowledge Management: AI frameworks now synchronize and validate which short-term contexts should be elevated to persistent, long-term storage using relevance thresholds:

$P = \sum_{i} w_i \cdot x_i$

where $P$ exceeds a threshold $T$ for data to be stored, simulating the selectivity of human memory.

Implications for Human-AI Interaction: Effective collaboration requires robust internal models in humans (biological schemata and neural manifolds) for error-correction and judgment (Oakley et al., 3 May 2025). Overreliance on AI tools risks atrophy of internal memory and critical schema construction, necessitating policy interventions to ensure foundational knowledge is retained and proceduralization occurs.

Cognitive processes are shaped not only by computation but also by social and cultural factors (Dancy, 2022). Cognitive architectures that integrate world knowledge (e.g., ConceptNet) and account for affective/physiological states can model how biases (including antiblackness) are encoded and emerge in decision-making:

ACT-R/Φ models declarative and procedural memory with physiological inputs (e.g., stress, fatigue), affecting availability and retrieval.
Knowledge graph association metrics (e.g., ConceptNet relatedness) modulate the salience of certain semantic associations, causing context-sensitive retrieval that can perpetuate social biases.

The design and development of AI systems are therefore not isolated from sociocultural context; inclusion of such processes is indispensable for fairness and validity in modeling cognition.

6. Comparative Evaluation and Limitations of AI Cognition

Benchmark studies have quantitatively measured the performance of AI systems across a range of higher-order cognitive tasks, revealing both remarkable strengths and clear deficits (Latif et al., 7 Dec 2024, Zhang et al., 30 Mar 2025):

AI models may outperform humans in structured domains: critical thinking, systematic reasoning, data literacy, and even certain creative and logical tasks.
However, adaptive, ill-structured problem-solving and genuine creative transformation—requiring representational change, abstraction, and deep semantic association—remain significant challenges for AI (Zhang et al., 30 Mar 2025).
AI “creativity” is often derived from combinatorial retrieval and probabilistic pattern matching, not the flexible, transformation-based processes observed in human creativity.
Overuse of AI-generated suggestions by humans can undermine metacognitive vigilance and deepen specific cognitive biases, with outcomes contingent on individual attitudes toward automation and the structure of review workflows (Beck et al., 10 Sep 2025).

7. Integration and Future Directions

The convergence of cognitive science and AI is characterized by a reciprocal interplay:

AI research adopts cognitive science principles (e.g., hierarchical processing, attention, reinforcement learning, meta-cognition, embodiment), mapping them to technically precise architectures (Mao et al., 28 Aug 2025).
Cognitive science leverages computational models to explicate mechanisms underlying perception, memory, learning, and decision-making, offering mathematical rigor and predictive power.
Emerging frameworks emphasize hybrid neuro-symbolic architectures, dynamic knowledge adaptation, continuous learning, and metacognitive self-monitoring (Nirenburg et al., 22 Mar 2025).

Open challenges persist in bridging structured causal modeling and black-box learning, mitigating bias, ensuring explainability, and maintaining human capacity for proceduralized expertise. The ongoing imperative is to design AI systems that augment and align with foundational human cognitive processes rather than replace or atrophy them.

Table 1: Key Algorithmic Constructs in Cognitive Modeling

Concept	Definition / Formula	Role in Cognition
Algorithmic Probability	$m(s) = \sum_{p: U(p) = s} 2^{-\|p\|}$	Pattern inference, prediction
Kolmogorov Complexity	$K(s) = \min\{ \|p\| : U(p) = s \}$	Compression, simplicity bias
Coding Theorem	$K(s) = -\log_2(m(s)) + O(1)$	Bias-explanation, predictivity
Memory Relevance Threshold	$P = \sum_i w_i x_i$ (retain if $P > T$ )	Selective long-term storage

These algorithmic tools bridge theoretical concepts and empirical modeling in studies of human and artificial cognition.

Developments in this field continue to inform the construction of AI architectures, the interpretation of cross-species cognition, and the design of systems that are both operationally efficient and human-compatible. Ongoing research addresses continuous learning, multimodal processing, resource efficiency, and ethical considerations in order to further refine the synergy between human and machine cognition.