Humanity Has Always Been a Human–Technology System

The wrong unit of analysis

Stories about technology usually begin with an object: the hand axe, plough, printing press, steam engine, transistor, smartphone, or artificial-intelligence model. The object appears, society reacts, and a new age begins. This narrative is legible because it gives causality a visible protagonist. It is also analytically weak.

No artifact operates outside a population capable of reproducing its use, an energy and material system capable of sustaining it, and institutions capable of assigning authority and resolving failure. A printing press without literate readers, paper supply, distribution, legal conventions, and communities of interpretation is metal and wood. A foundation model without data pipelines, accelerators, electric power, network access, evaluation, and organizational permission is a parameter file.

The more defensible unit of analysis is a human–technology system: a coupled arrangement of biological organisms, socially learned practices, artifacts, infrastructure, and rules. In such a system, causation is reciprocal. Humans design tools, tools restructure tasks, tasks alter skill formation, institutions distribute gains and risks, and those distributions shape which technologies are funded and retained.

This is not a rhetorical compromise between technological determinism and human exceptionalism. It follows from research on cultural learning, cumulative culture, niche construction, and gene–culture coevolution ^{[1, 2]}. Artificial intelligence is novel within this history, but novelty must be specified rather than presumed.

Two inheritance systems—and more

Biological evolution transmits genetic information with mutation and recombination. Human populations also transmit acquired information through imitation, teaching, language, ritual, apprenticeship, writing, and engineered media. Tomasello and colleagues distinguished imitative, instructed, and collaborative learning, linking high-fidelity social transmission to perspective-taking and shared intentionality ^[1]. Later cultural-evolution research has shown that human adaptation often depends on knowledge no individual could derive independently within one lifetime ^[3].

A minimal cultural-selection model can represent the frequency $x_i$ of a variant $i$ as

where $f_i$ is context-dependent cultural fitness and $M_{ij}$ represents transformation or transmission from variant $i$ to $j$. The equation is illustrative, not a claim that beliefs have fixed genes or that cultural change is reducible to one replicator law. Unlike alleles, cultural variants can be intentionally modified, horizontally transmitted, blended, externally stored, and adopted because of prestige, conformity, coercion, or institutional mandate.

Material artifacts add another inheritance channel. A novice does not receive only verbal instructions for a craft; the partially completed object, jig, notation, interface, or codebase constrains the next action. Institutions add still another: standards, property rights, scientific review, credentials, safety procedures, and bureaucratic records preserve some behaviors while extinguishing others. Human evolution after the emergence of cumulative culture is therefore neither purely genetic nor disembodied “memetic” change. It is multi-channel inheritance with unequal fidelity and power.

Gene–culture coevolution makes the coupling literal. Cultural practices can alter selection pressures, while genetically influenced capacities can alter which practices spread. Reviews in human evolutionary genetics document cases in which subsistence, settlement, and socially transmitted behavior changed the environments in which genetic selection operated ^[2]. The point is not that every modern technology has already rewritten the genome. It is that culture is part of the environment to which humans adapt.

The ratchet is retention plus reconstruction

Cumulative culture is frequently described through a ratchet metaphor: improvements are retained so later generations begin above the prior baseline. Mesoudi and Thornton sharpen the concept by distinguishing a mere sequence of changes from cultural evolution that produces a measurable improvement or elaboration beyond what individuals could readily invent alone ^[5]. Experimental comparisons have examined the social and cognitive mechanisms—teaching, communication, imitation, and prosociality—that enable human groups to accumulate solutions ^[4].

An abstract recurrence makes the retention problem visible:

where $C_t$ is usable cultural complexity, $\rho_t$ is effective transmission fidelity, $I_t$ is validated innovation, and $L_t$ is loss from population disruption, inaccessible records, incompatible standards, or failed teaching. Innovation receives most historical attention, but cumulative growth requires $\rho_t$ to remain high enough that new work does not merely reconstruct forgotten prerequisites.

Writing, diagrams, mathematical notation, libraries, version control, and machine-readable standards increase storage outside individual brains. They do not guarantee understanding. External records can preserve symbols whose tacit operating knowledge has vanished. Conversely, apprenticeship can preserve embodied skill without a formal specification. Robust cultures use overlapping representations: text, demonstration, artifact, test, and community memory.

The cultural-brain hypothesis models a long-run feedback in which larger and more connected populations can sustain more adaptive cultural repertoires, increasing the returns to learning capacities and sociality ^[6]. Such models do not establish one inevitable path toward larger brains or more complex technology. They clarify a positive feedback: social learning can create an environment in which further capacity for social learning is valuable.

Tools transform cognition by redistributing it

A tool does not simply amplify a fixed human capability. It reallocates cognitive operations across person, artifact, and group. Numeral systems move quantity into manipulable symbols. Algebraic notation makes structural operations inspectable. Maps stabilize spatial relations outside working memory. Scientific instruments transform inaccessible phenomena into traces that communities learn to interpret. Software compilers enforce formal constraints while concealing machine details.

This redistribution creates deskilling and reskilling simultaneously. When a navigation system performs route computation, unaided wayfinding may atrophy while coordination across a larger logistics system becomes possible. When a compiler automates register allocation, fewer programmers reason routinely at that level, while more can build higher abstractions. Whether the net change is emancipation, dependence, or both depends on failure modes and access to lower layers.

The historical pattern is therefore not “machines replace cognition.” It is cognitive decomposition. Some operations become automatic, others become newly valuable, and institutions redefine competence around the resulting stack. The dangerous point occurs when a society loses the ability to inspect a critical layer while continuing to depend on it.

Institutions are civilization’s error-correction machinery

High-fidelity transmission can preserve mistakes as efficiently as truths. Cumulative culture requires variation and retention, but reliable knowledge also requires selection mechanisms that are responsive to evidence. Science, engineering, courts, standards bodies, markets, archives, and democratic institutions are imperfect error-correction systems. They specify who may challenge a claim, what counts as evidence, how conflicts are recorded, and when decisions can be reversed.

Their performance cannot be inferred from information volume. A society may produce more documents while making provenance harder to trace, or communicate faster while shortening the time available for verification. Search engines increased access to documents; they also changed incentives for publication and attention. Generative systems increase the supply of plausible synthesis; without corresponding provenance and review, they can reduce the ratio of checked claims to total claims.

This gives an epistemic version of the cumulative-culture recurrence. Let $K_t$ be decision-relevant knowledge rather than stored content:

where $V_t$ is validated contribution, $E_t$ is undetected error, and $O_t$ is obsolescence. Increasing generation raises $V_t$ only if validation capacity scales; otherwise it can raise $E_t$ faster. AI policy focused exclusively on model output neglects the institutions that transform output into warranted action.

What is genuinely discontinuous about AI

Placing AI within cultural evolution should not flatten its novelty. At least four discontinuities deserve attention.

First, generative models operate over unusually broad representational domains. One system can transform prose, code, equations, images, and structured data, reducing translation costs between specialties. Second, inference is replicable at machine timescale: once infrastructure exists, a procedure can be executed across many contexts without training a human practitioner for each copy. Third, tool-using systems can join representation to action, making decisions that alter software, transactions, or physical processes. Fourth, the capital requirements of frontier training and datacenter deployment can concentrate control even while low marginal inference costs broaden access.

The Stanford AI Index documents rapid benchmark gains, falling inference costs for fixed capability levels, increasing business use, and continuing concentration of notable frontier models in industry ^[9]. These observations support neither inevitable abundance nor inevitable monopoly. Falling unit cost and rising fixed cost can coexist: inference becomes cheaper while frontier production remains capital intensive.

The difference between AI and a static archive is adaptive recombination. A model can construct a response not explicitly stored in any one source and can select tools based on context. The difference between AI and a human institution is equally important. A model does not bear legal responsibility, possess democratic legitimacy, or maintain grounded organizational memory merely because its language is fluent. Granting it action authority is an institutional choice implemented through software.

Automation acts on tasks, not occupations in one stroke

Predictions of total occupational replacement compress heterogeneous work into job titles. A task-based framework separates displacement from reinstatement. Automation shifts some tasks from labor to capital; technological change can also create new tasks in which labor has comparative advantage. Acemoglu and Restrepo formalize these opposing effects and connect their balance to labor demand and labor’s share of value added ^[7].

For AI, task exposure depends on more than whether a model can produce an answer. Let the economically automatable share of a task family be represented schematically as

where $C$ is technical capability, $R$ is reliability at the required horizon, $I$ is integration into real workflows, and $G$ is governance permission. A high benchmark score with poor integration yields little automation. Strong capability and integration with unacceptable liability may be confined to decision support. Conversely, modest models can transform work when outputs are cheap to verify and errors are reversible.

This predicts uneven adoption. Code, advertising variants, document triage, and structured analysis have abundant digital feedback and divisible tasks. Childcare, field repair, diplomacy, laboratory work, and executive accountability combine tacit knowledge, physical interaction, trust, and consequences that are expensive to verify. Portions of each can be automated without the occupation disappearing.

New tasks will cluster around specification, evaluation, exception handling, data stewardship, security, model integration, and institutional accountability. That does not guarantee that displaced workers transition smoothly or that new tasks match old wages and geography. Historical creation of work is an aggregate mechanism, not an individual compensation program.

Four conditional futures

Forecasts should identify variables that could make them wrong. The following are scenarios, not prophecies.

1. Broad augmentation

If capable models become widely accessible, verification tools improve, and education teaches decomposition and source criticism, AI can lower the cost of participating in technical culture. Small teams gain analytical and software capacity previously available only to large organizations. The bottleneck shifts from producing a first draft to selecting worthwhile problems and validating consequences.

2. Concentrated cognitive infrastructure

If frontier capability depends on a small number of datacenters, proprietary datasets, and vertically integrated platforms, organizations may rent not only computation but portions of their decision process. Switching costs then arise from accumulated workflow state, evaluations, permissions, and institutional dependence, not merely model weights. Cultural production remains broad while the infrastructure that ranks and acts on it becomes narrow.

3. Brittle abundance

If generation scales faster than verification, institutions receive more code, claims, applications, media, and strategic proposals than they can evaluate. Apparent productivity rises while latent error accumulates. Failures arrive discontinuously when an unverified dependency reaches finance, health, infrastructure, or security. This future is not caused by hallucination alone; it is caused by coupling uncertain output to high-consequence transitions without proportional error correction.

4. Accountable coordination

If provenance, auditability, capability-scoped permissions, contestability, and public-interest access become infrastructure, AI can support institutions rather than merely accelerate transactions. UNESCO’s recommendation frames AI governance around human rights, transparency, responsibility, data governance, and environmental and social well-being ^[8]. The operational challenge is converting principles into procurement rules, technical interfaces, liability, and rights of appeal.

These futures can coexist by sector and jurisdiction. A hospital may impose accountable coordination while entertainment experiences broad augmentation; military intelligence may concentrate; low-quality information markets may become brittle. “The future of AI” is likely to be a distribution of institutional equilibria.

The next phase of cumulative culture

Generative AI changes both terms in the cultural ratchet. It can increase innovation $I_t$ by recombining representations, translating across domains, and lowering experimentation cost. It can increase retention $\rho_t$ by making archives queryable and procedures easier to reconstruct. It can also increase loss $L_t$ when generated summaries replace primary records, model-mediated access becomes unavailable, or communities cease maintaining the expertise needed to detect subtle errors.

The crucial variable is complementarity with error correction. Systems that attach claims to sources, actions to logs, code to tests, and decisions to accountable principals can raise the throughput of validated culture. Systems optimized for frictionless plausibility can raise cultural entropy in the ordinary, non-thermodynamic sense: more variants, weaker provenance, and higher selection cost.

Human distinctiveness in this account does not rest on winning a benchmark against machines. It rests on participation in normative and institutional practices: deciding which objectives deserve pursuit, whose losses count, what evidence warrants coercive action, and how authority can be challenged. Machines can contribute to those practices without acquiring automatic legitimacy.

The deepest continuity from stone tools to AI is not a march toward disembodied intelligence. It is the repeated construction of systems that move cognition outside the individual and then reorganize society around the new distribution. The deepest risk is similarly continuous: confusing technical capacity with justified authority. Humanity’s future will not be made by humans on one side and technology on the other. It will be made by the institutions that decide how inseparable they are allowed to become.