Artificial Human Intelligence: The role of Humans in the Development of Next Generation AI

TL;DR

This paper argues that next-generation AI should be guided by a symbiotic relationship with human intelligence, proposing a threefold taxonomy (HIns, HAss, HInd) to categorize AI systems by their reliance on human input and inspiration. It surveys current state-of-the-art architectures, contrasts brain-inspired and hybrid approaches, and identifies key missing elements—such as higher-level abstractions, compositionality, and counterfactual reasoning—that hinder true human-level AI. The authors advocate for brain- and biology-inspired design principles to improve interpretability, robustness, and alignment, while acknowledging practical trade-offs with scalability and data efficiency. The work emphasizes realistic neuroscience-informed training, modularity, and embodied, autonomous learning as promising directions for ethically aligned, human-centered AI systems with broad societal impact.

Abstract

Human intelligence, the most evident and accessible form of source of reasoning, hosted by biological hardware, has evolved and been refined over thousands of years, positioning itself today to create new artificial forms and preparing to self--design their evolutionary path forward. Beginning with the advent of foundation models, the rate at which human and artificial intelligence interact with each other has exceeded any anticipated quantitative figures. The close engagement led both bits of intelligence to be impacted in various ways, which naturally resulted in complex confluences that warrant close scrutiny. Recent advances, such as DeepSeek, exemplify this interplay: the novel contributions, we argue, draw indirect inspiration from biological principles like modular neural specialization and sparse episodic encoding, addressing computational bottlenecks while aligning with human-inspired scalability. In the sequel, using a novel taxonomy, we shall explore this interplay between human and machine intelligence, focusing on the crucial role humans play in developing ethical, responsible, and robust intelligent systems. We briefly delve into various aspects of implementation inspired by the mechanisms underlying neuroscience and human cognition. In addition, we propose future perspectives, capitalizing on the advantages of symbiotic designs to suggest a human-centered direction for next-generation developments, focusing on the augmentation role of AI. We finalize this evolving document with some thoughts and open questions yet to be addressed by the broader community.

Figures (8)

• Figure 1: The kinds of inspiration sources in search of modeling intelligence with their corresponding levels of complexity and explainability within the context of Artificial Human Intelligence, a concept that bridges (through inspiration) between brain-inspired and human-independent intelligences.
• Figure 2: Summary of the manuscript detailing various topics of interest and where they are covered with associated section and/or subsection numbers.
• Figure 3: (a) In HIns Intelligence, developments in machine intelligence are primarily driven by insights from biology and neuroscience, leading to overlapping advancements in both fields. Conversely, in HAss intelligence, the overlap tends to diminish over time, resulting in more specialized assistance from either humans or machines, fostering the development of hybrid systems. Finally, in HInd intelligence, the intrinsic features of machine intelligence evolve freely, naturally causing a divergence in the fundamental principles governing human and machine intelligence. (b) A measurement of human feature alignment with respect to machines (CNNs, Transformers, Self-Supervised) as a function of accuracy fel2022harmonizing.
• Figure 4: Categorization of Human-centered intelligences and associated subcatagories.
• Figure 5: (a) An example of biomimetic progression of image resolution and color cues in the early months of development. Color maturization happens based on the wavelength of light i.e., red first accompanied by orange, followed by yellow, green, blue and purple towards the end of the spectrum. (b) Impact of training paradigms on performance with test images degraded by varying blur levels (quantified as the Gaussian kernel’s $\sigma$ in pixels): the blurred-to-high-resolution paradigm yields the strongest performance and generalization vogelsang2018potential.