Evolutionary Tempo Mismatch: Biological Evolution Unfolds over Millennia, Technological Iteration within a Decade
Human biological evolution and advances in artificial intelligence unfold on markedly different timescales.
From the perspective of natural evolution, the human body and brain were shaped over immense spans of time. Archaeological and anthropological research indicates that modern human brain volume had already approached its present level around 300,000 years ago, while the brain’s basic structure and modes of information transmission have changed very little. The human brain operates on roughly 20 watts of metabolic power; neural signals propagate at about 1 to 100 metres per second; and individual neurons typically fire no faster than 100 to 200 hertz. Cranial volume, heat dissipation, and metabolic homeostasis together constrain cognitive bandwidth, working memory, and parallel processing speed. Neuroplasticity enables people to learn and adapt, but it relies mainly on slow, localised synaptic updates and cannot compensate for the pace of biological evolution over millennia.
By contrast, AI systems are now updated on timescales of years, months, or even weeks. Current training frameworks can coordinate tens of thousands of specialised processors and perform synchronised computation at gigahertz clock speeds. Frontier-model architectures and safety mitigations can be reconfigured within weeks or months and redeployed across global infrastructure. On the hardware side, memristive in-memory computing systems, two-dimensional materials, and advanced thermal-management technologies are being developed to overcome bottlenecks in conventional computing architectures. This speed and scale far exceed biological evolution, but they do not make AI inherently more energy-efficient: training and operating large models still require vast amounts of electricity, cooling, data-centre capacity, and complex global supply chains.
The paper describes the disparity between millennial-scale biological evolution and sub-decadal algorithmic iteration as an “evolutionary tempo mismatch”. This is not simply a comparison of human and machine speed. It points to a structural tension: technological capabilities are expanding rapidly, while human biology and social institutions cannot change at the same pace.
In response to this mismatch, the paper proposes “transcending natural evolution”: using technology to advance and augment human perceptual, cognitive, and motor capabilities at a pace beyond natural evolution.
Crossing the Embodiment Threshold: AI Moves from the Digital Realm to Embodied Interaction
The embodiment threshold describes the transition from largely digital inference to continuous participation in physical, bodily, or neural activity.
AI has long operated mainly in the digital realm, processing symbolic information such as text, images, and code. Its outputs can shape human judgement and decision-making, but the systems themselves have generally lacked a closed loop through which they continuously sense the environment, act, and receive feedback. In this sense, they remain forms of disembodied intelligence.
Recent advances in multimodal models, robotics, and bioelectronics are beginning to blur this boundary. An AI system crosses the embodiment threshold when it forms a task-relevant closed loop with a physical environment, the human body, or neural tissue, supported by persistent or continuously available integration and sensory-motor coupling. This is not a binary boundary but a graded continuum. Robots that navigate and manipulate objects, wearable devices that continuously monitor physiological signals, and brain-computer interfaces that directly read or modulate neural activity represent different degrees of embodiment.
Flexible bioelectronics provide a critical material interface for sustained interaction between AI systems and the human body. Rigid silicon devices and soft biological tissues differ by several orders of magnitude in mechanical stiffness, so long-term wear or implantation may cause microtrauma, localised inflammation, and signal degradation. Surface-wrinkling techniques, nanoscale-ribbon buckling, self-healing structures, and textile electronics are improving device conformability, stretchability, and stability. Examples reviewed in the paper include smart textiles that monitor glucose and cortisol, artificial throats that capture laryngeal myoelectric signals and mechanical vibrations, and endovascular stent-electrode arrays delivered through the jugular vein to vessels adjacent to the motor cortex. These technologies show different routes from body-surface sensing to neural interfaces, but challenges remain in scalable manufacturing, biocompatibility, long-term signal stability, and power delivery.
This trajectory spans three continuous operational domains: disembodied intelligence in the digital domain, embodied intelligence in the physical domain, and human-machine integration in the neural domain. As interaction bandwidth increases, connections become more persistent, and information flows become bidirectional, AI may increasingly serve as an extension of human perception, communication, and movement. Whether this progression leads to deeper human-machine symbiosis will still depend on safety, ethical boundaries, and social governance.
Biological-Technological Stratification: Human-Machine Integration and Governance under Resource Constraints
Human-machine integration does not automatically guarantee a better future. The expansion of computing capacity and the wider adoption of bioelectronic interfaces are constrained by energy, materials, supply chains, and institutional conditions. These constraints directly shape who controls the technology and who can use it.
The paper calls the durable divide that could emerge “biological-technological stratification”. If high-bandwidth cognitive assistance, restoration of neural function, and predictive health monitoring remain available mainly to the few who can afford them, social inequality could extend beyond income and resources into cognitive bandwidth, physical capability, and health security. At the same time, large models trained on existing corpora and optimised for majority preferences may intensify cognitive convergence. If neural signals are collected and traded as ordinary consumer data, the protections surrounding neural privacy and biological intent will also be put at risk.
To address these risks, the paper proposes an adaptive governance roadmap. Rather than a three-phase plan tied to fixed calendar dates, it calls for different levels of oversight to be triggered by measurable technological indicators and the depth of human-machine integration.
Phase I focuses on infrastructural legibility. At the point of frontier-model release, it calls for standardised disclosure of compute allocation, energy use, and training-data provenance, together with immutable incident logs, rigorous red-teaming, and independent evaluation. As wearable and non-invasive interfaces develop, raw neural signals should also be placed under stringent protections for medical or sensitive data.
Once AI crosses the embodiment threshold, Phase II shifts the focus to embodied safety and user autonomy. Priorities include harmonised engineering and clinical standards, continuous assessment of algorithmic drift, robust software and physical fail-safes, and a functional right to disconnect. Legal frameworks will also need to distinguish among biological intent, algorithmic recommendation, and device-level execution when assigning responsibility.
Phase III addresses transnational equity. Drawing organisational lessons from the European Organization for Nuclear Research (CERN) and the International Atomic Energy Agency (IAEA), the paper recommends exploring cross-border oversight and public-interest review mechanisms, broadening access to baseline computational infrastructure and predictive health monitoring, and reducing dependence on a small number of large cloud platforms and actors that control critical supply chains.
Conclusion and Outlook
As humanity moves from natural evolution to what the paper calls transcending natural evolution, it faces a choice about how its future will be shaped. The value of this Position Paper lies not in forecasting specific technological milestones, but in placing people back at the centre of the debate: How should technology expand human capabilities, and what structural changes might follow for society? The paper argues that the goal should not be the unconstrained expansion of machine capability. It should be to keep the enhancement of human perception, cognition, and action at the centre of technological development, and to guide intelligent technologies and their governance so that humanity retains the agency to shape its future.
Science China Information Sciences