The Fundamental Matchup: Man vs. Machine

Executive Summary

The Most Important Matchup

For nearly all of human history, competing intelligence to humans has not been a going concern. The reason humans have dominion over the earth is their supreme intellectual abilities — not their physical strength. No other species has come close. Until recently, this was simply assumed to be a permanent feature of the world.

Computational intelligence platforms have been embedded in our lives for decades, but it wasn't until just the last four years that they began to compete in the realm of human intelligence. In 2022, frontier AI models had an IQ equivalent of 55 — below the threshold of functional human acuity. In less than four years, those same systems accelerated to over IQ 135, surpassing 99% of all humans. This capability is continuing to compound rapidly. As such, we need to carefully consider the impacts of this new intelligence — and, more urgently, how it will be shaped.

"Computer IQ" here is a benchmark-mapped score, not a measurement of cognitive capacity in the human-developmental sense. See Appendix B for framing choices and limitations.

Section 1

A History of Intelligence

Human Intelligence

For all of human history, the human mind has been the dominant intelligence on earth. Our intellectual supremacy and theory of mind are what let us work, subdue, and domesticate the natural world. No other creature has come close.

Human intelligence as we know it emerged gradually over hundreds of thousands of years of evolution. Early hominids developed increasingly complex brains, theory of mind, and social cooperation — capabilities that allowed small groups of biologically unremarkable animals to become the dominant species on Earth. By the time of Homo sapiens, the brain had developed the capacity for language, abstract reasoning, long-term planning, and cumulative culture: each generation could build on the knowledge of the last.

The standard psychometric model captures human cognitive ability as a normal distribution centered at IQ 100, with a standard deviation of 15. Applied to the 2025 global population of approximately 8.1 billion, this produces the distribution below. The shape has been stable across recorded history — the Flynn effect (a gradual upward drift of ~3 points per decade through the twentieth century) is the main documented deviation, and is averaged into modern calibration.

What makes human intelligence distinctive is not its raw throughput — it is its integration. The human brain synthesizes sensory, emotional, social, and analytical signals continuously and in real time. Intuition, embodied experience, theory of mind, and the capacity for genuine creativity and meaning-making all arise from this integration. These are the dimensions that have remained most resistant to machine replication.

Machine Intelligence

Mechanistic intelligence has a much shorter history. The first programmable computers emerged in the 1930s and 1940s. Early systems were purely rule-based: they executed instructions precisely but could not generalize beyond their explicit programming. The history of machine intelligence is largely the history of progressively relaxing that constraint — moving from hard-coded rules to learned statistical patterns to, most recently, emergent general-purpose reasoning.

The era-by-era table below traces the key inflection points from mechanical calculators to the present frontier.

Section 2

IQ Distribution Across the Computing-Device Population

Two facts together explain the shape of this moment. First: as recently as 2022, the most advanced AI systems registered an IQ equivalent of roughly 55 — below the threshold of basic functional reasoning by human standards. By early 2026, that number had crossed 135, placing frontier AI above 99% of all humans. This is not gradual drift. It is an acceleration with no historical parallel in any measured dimension of intelligence.

Second: it is not just that one model is now very capable. The entire computing-device population has migrated rightward in IQ-equivalent terms at extraordinary speed. In 2015, the median computing device sat at IQ-equivalent ~38. By 2026, AI-capable devices number in the tens of billions, with the median well into the high-average human range. The chart below shows both dimensions at once — frontier trajectory and population distribution, year by year.

Comparison at the High End (IQ 140+)

The human population at IQ 140 or above (the highly-gifted threshold) is approximately 80 million people worldwide — about 1% of the global population. As of mid-2026, frontier AI systems are in the IQ 130–135 range. The IQ 140+ crossover is plausibly within the next 12–24 months at current trajectory but has not yet occurred.

Section 3

Machines and Humans

IQ is a single-number summary that obscures most of what is interesting about how humans and machines actually differ. The most useful comparison breaks the question into specific capability dimensions, where one or the other has measurable, structural advantages.

Summary — Advantage by Category

At-a-glance view across the nine dimension families, sorted by who holds the advantage. AI strong = AI wins nearly all rows. Mixed = no clear winner. Humans strong = humans win nearly all rows.

Across the nine families: AI holds clear advantage in four (knowledge/memory, computation, input/output, economic/physical), humans hold clear advantage in four (reasoning, social/emotional/embodiment, adaptation/learning, creativity), and reliability is mixed. The full dimension-by-dimension detail follows in Section 4.

Why This Comparison Matters

Machines have real, structural advantages — speed, precision, recall, and scale that no human can match. Humans have equally real advantages of their own: contextual judgment, embodied experience, moral reasoning, and the capacity for meaning. Honestly respecting the relative strengths of each is the only way to shape AI development from where it is actually going, not from where we wish it were going.

This may not be the future everyone would choose. But it is the future that is arriving, and it is unlikely to be reversed. The most productive response is not to resist the comparison, but to understand it clearly — and to use that understanding to shape what comes next.

Section 4

The Comparison Table

★ = clear advantage. Humans lead on reasoning, social/emotional, creativity, and continuous learning. Machines lead on capacity, speed, scale, cost, and consistency.

Section 5

The Future Relationship: A New Division of Labor

This technology is not slowing down, and the window to shape it is narrowing faster than most people realize. As documented in the Exponential Divergence study, AI capability is growing on a fundamentally different curve than any prior technology — one that does not naturally plateau, and does not give institutions time to catch up. As the Fruit of the Spirit study explores, it is already producing measurable effects on human moral formation, particularly in children. The economic incentives, the geopolitical competition, and the utility of these systems are too strong for any plausible intervention to halt development. The question is not whether machines will keep growing more capable. They will. The question is what we do about it.

What is likely to emerge is a new division of labor — not between people, but between humans and machines. Economically valuable tasks rooted in computation, memory, recall, throughput, and consistency will increasingly migrate to machines. That raises genuine questions about where the economic value of human beings actually lies in the future. The positive version is that humans become more deeply human. The capabilities that machines cannot replicate — reasoning under ambiguity, intuition, genuine relationship, creativity, meaning-making, embodied wisdom, emotional and moral formation — become the primary source of what humans contribute. Machines do the heavy lifting of computation. Humans do the irreplaceable work of being human.

The negative version is that the machines generate enormous economic value, but the human systems around them — institutions, policy, education, social contracts — do not evolve quickly enough or equitably enough to absorb the change. The gains may concentrate, or they may simply be distributed unevenly, leaving a large portion of the population finding their economically valuable skills obsolete faster than alternatives can reasonably emerge. The transition, in that case, could become very rough, very fast, for a very large number of people.

The exact direction is uncertain. But the changes ahead will be significant — more significant, and faster, than anything most institutions are currently prepared for. The appropriate response is not to wait and see. It is to understand what is actually happening and to start preparing now.

Machines are going to keep getting smarter. The question worth asking is not whether we can stop that. It is what we choose to remain irreplaceably human at.

Section 6

Conclusion and What We Should Do

For nearly all of human history, the question of whether something else on earth could outthink us did not arise. In less than four years, that assumption has been overturned. Frontier AI systems now surpass 99% of humans on standardized reasoning benchmarks — and the trajectory shows no sign of slowing.

This is not a reason for despair. It is a reason for clarity. The comparison in this article is not a ranking of worth — it is a map of structural reality. Machines are extraordinarily good at computation, recall, speed, and consistency. Humans are extraordinarily good at judgment under ambiguity, genuine relationship, moral formation, creativity, and meaning. The productive path forward is to lean into that complementarity honestly, protect and cultivate what is distinctly human, and shape these systems before the window to do so narrows further.

The machines are not coming to replace us. They are arriving as something we have never had before: an entity that can do some of what we do, faster and at scale, but cannot do what makes us human at all. The question of this century is not whether we are still the smartest thing on earth. The question is whether we will remember what only we can be.

Appendix A

Methodology

Frontier AI IQ Estimation

"Computer IQ" or "Frontier AI IQ-equivalent" refers to a benchmark-mapped score on a scale calibrated to human IQ (median 100, standard deviation 15). For 2022 onward, this report uses the aiiq.org methodology developed by Ryan Shea: 17 benchmarks across 5 reasoning dimensions (Fluid Abstraction, Mathematical Reasoning, Programmatic Reasoning, Critical Reasoning, Agentic Reasoning), each hand-calibrated to an implied IQ via anchor curves, with conservative imputation for missing benchmark coverage. Pre-2022 values are imputed retrospectively using the same methodological framing applied to era-defining capabilities; they are intended for trajectory illustration rather than precise measurement.

Concrete anchors:

Computing-Entity Counts

Computing-entity counts use a broad definition: physical devices (PCs, smartphones, tablets, IoT, embedded systems) PLUS virtual instances (VMs, containers, cloud compute) PLUS AI agents (autonomous AI instances). Sources: Gartner / IDC for physical devices; IDC / RightScale / Flexera State of Cloud reports for VMs and cloud instances; public usage data from major AI assistants for AI agents.

IQ Distribution by Computing Device

Each computing device is assigned to an IQ-equivalent bucket based on what it has access to in a typical user session, not its silicon's intrinsic capability. The framing is "intelligence accessible via the device" rather than "intrinsic device intelligence." A 2025 smartphone with the GPT-5.2 app has access to IQ-137 reasoning.

Human IQ Distribution

The human IQ distribution is the standard normal-distribution model centered at 100 with standard deviation 15, applied to the 2025 global human population of approximately 8.1 billion. The Flynn effect is averaged into the calibration.

Memory and Recall

Humans hold roughly seven items in working memory at once (Miller, 1956) or roughly four chunks under stricter measurement (Cowan, 2001). Long-term storage is estimated at one to 2.5 petabytes, with retrieval that is lossy and reconstructive. Modern AI models have context windows of 200,000 to two million tokens — all simultaneously accessible with exact recall — and trained weights compress information from trillions of training tokens. The gap is structural: human working memory evolved under tight metabolic constraints, and AI models have no equivalent.

Computation and Precision

A human performs roughly one non-trivial arithmetic operation per second. A frontier GPU cluster performs 10¹⁸. For any task that decomposes into formal computation, machines are not in the same category as humans. For any task that resists formalization — judgment under deep ambiguity, aesthetic decision-making — the computation-per-second gap is irrelevant.

Symbolic vs. Fuzzy Reasoning

Symbolic reasoning — logical chains, formal proofs, programming, mathematics — is where machines have structural advantages. They do not fatigue, do not make arithmetic errors, can search exhaustively. Fuzzy reasoning — pattern recognition under ambiguity, reading a room, knowing when an answer is technically correct but contextually wrong — is where humans retain structural advantages. Most consequential real-world reasoning is fuzzy, not symbolic.

Causal and Counterfactual Reasoning

AI systems learn statistical patterns from data; they do not maintain explicit causal models in the sense Judea Pearl describes (Pearl, 2009). They cannot reliably distinguish "X and Y happen together" from "X causes Y." Humans do causal and counterfactual reasoning intuitively from age four onward. Active AI research (causal AI, world models) is attempting to close this gap, but no current approach has produced a system that reasons about cause and counterfactual at human levels of reliability across domains.

Social, Emotional, and Embodiment

Humans have continuous integrated multimodal experience grounding all higher cognition. Current AI models lack continuous embodied experience and cannot sense fatigue, hunger, or fear. Humans build relationships through repeated interaction, shared history, and reciprocal vulnerability. AI systems can model these dynamics and produce appropriate-sounding responses, but do not experience the emotional states they reference. The difference shows up in long-running relationships, trust under genuine stress, and shared meaning across time.

Adaptation, Learning, and Creativity

Continuous learning during deployment, one-shot acquisition from a single example, intrinsic curiosity, paradigm-shifting creativity, aesthetic experience, and continuity of identity are all structural to the human substrate. None are obviously solvable by larger models or more compute. The current frontier-AI workflow papers over most of these gaps with external scaffolding, but these are useful workarounds — not the underlying capability.

Cost and Scaling

Expert human time costs $150,000 to $500,000 per year in advanced economies. Equivalent capability via API access to frontier AI is roughly $500 to $5,000 per year per user — two to three orders of magnitude cheaper. This economics-of-expertise inversion is the structural driver of most of the changes documented in the EconFaithAI project's other reports.

Appendix B

Limitations

On "Computer IQ" as a Concept

1. IQ is calibrated to humans, not machines. Applying these tests to AI systems is not what the tests were designed for. The "IQ-equivalent" mapping is a benchmark-percentile translation that captures comparable performance on knowledge-and-reasoning tasks but does not certify equivalent cognition.
2. AI capability is narrow and uneven. A frontier 2025 AI system can solve International Math Olympiad problems but cannot reliably drive a car in unfamiliar conditions, cannot persist memory across sessions without scaffolding, and cannot self-improve in the way humans do across years.
3. Benchmark contamination. Some AI benchmark scores are inflated by training-data leakage from the benchmarks themselves. The IQ-equivalent mapping inherits this bias.
4. Pre-2024 IQ values are retrospective imputations. aiiq.org provides direct benchmark coverage from approximately March 2024 (Claude 3 Opus) onward. Earlier values are intended for trajectory illustration rather than precise measurement.

On the Distribution Comparison

1. The bucket assignment of devices to IQ ranges is a strong modeling choice. A 2025 iPhone is assigned to a high-IQ bucket because it can access GPT-5-class reasoning. But the device cannot use that reasoning autonomously; a human user invokes it.
2. Pre-2022 distributions are reconstructive. Pre-LLM-era distinctions between IQ buckets are largely directional (older = lower bucket).
3. The Flynn effect is not modeled separately. Human IQ scores have drifted upward by ~3 points per decade; this is averaged into the calibration.

What the Report Does Not Claim

• Does not claim AI systems are conscious, sentient, or possessing of human-equivalent general intelligence.
• Does not claim narrow benchmark performance equates to general capability.
• Does not claim the device-to-human ratio means anything specific about agency, autonomy, or control.
• Does not predict the trajectory beyond 2035.

Appendix C

References

Source Files

• Computer_IQ_Timeseries.csv / .xlsx — Year-by-year frontier AI IQ-equivalent with era classification. Sources: OpenAI / Anthropic / Google DeepMind technical reports; aiiq.org methodology; Russell & Norvig AI: A Modern Approach.
• Human_IQ_Distribution.csv / .xlsx — Standard normal model (μ=100, σ=15) applied to ~8.1 B global population. Sources: Stanford-Binet / Wechsler test norms; UN World Population Prospects 2024.
• Computing_Population_Distribution.csv / .xlsx — Year-by-year installed-base counts and IQ-bucket distribution. Sources: Gartner / IDC device trackers; IoT Analytics; AI-assistant MAU figures.

AI Capability and Benchmarks

• Shea, R. (2025). AI IQ — Intelligently Measuring AI Intelligence. aiiq.org. Primary source for 2024+ frontier AI IQ values.
• OpenAI. (2023). GPT-4 Technical Report.
• OpenAI. (2024). OpenAI o1 System Card.
• Anthropic. (2024). Claude 3 Model Card.
• Google DeepMind. (2024). Gemini Technical Report.
• Rein, D., et al. (2023). GPQA: A Graduate-Level Google-Proof Q&A Benchmark. NeurIPS.
• Hendrycks, D., et al. (2021). Measuring Massive Multitask Language Understanding. ICLR.

AI History

• Russell, S., & Norvig, P. Artificial Intelligence: A Modern Approach (4th ed., 2020).
• Nilsson, N. J. (2009). The Quest for Artificial Intelligence. Cambridge University Press.
• Turing, A. M. (1950). Computing Machinery and Intelligence. Mind, 59.
• McCorduck, P. (2004). Machines Who Think (2nd ed.).

Human IQ Measurement and Distribution

• Flynn, J. R. (2012). Are We Getting Smarter? Rising IQ in the Twenty-First Century. Cambridge University Press.
• Wechsler, D. (2008). Wechsler Adult Intelligence Scale (WAIS-IV) test manual.
• Stanford-Binet Intelligence Scales (5th ed., 2003) test norm tables.
• Mensa International norms documentation.

Human Cognition and Reasoning

• Miller, G. A. (1956). The Magical Number Seven, Plus or Minus Two. Psychological Review, 63(2).
• Cowan, N. (2001). The magical number 4 in short-term memory. Behavioral and Brain Sciences, 24(1).
• Ericsson, K. A., Krampe, R. T., & Tesch-Römer, C. (1993). The role of deliberate practice in the acquisition of expert performance. Psychological Review, 100(3).
• Kahneman, D. (2011). Thinking, Fast and Slow. Farrar, Straus and Giroux.
• Pearl, J. (2009). Causality: Models, Reasoning, and Inference (2nd ed.). Cambridge University Press.
• Tomasello, M. (2014). A Natural History of Human Thinking. Harvard University Press.

Computing-Device Population

• Gartner. Worldwide Device Shipments, quarterly tracker.
• IDC. Worldwide Quarterly Mobile Phone Tracker.
• IoT Analytics. State of IoT, annual report.
• Flexera. State of the Cloud Report, annual series.
• United Nations Department of Economic and Social Affairs. World Population Prospects 2024.

Companion Visualization

Appendix D

Project Files