The history of information processing can be told as a sequence of major inventions: writing, printing, calculating machines, computers, the internet and artificial intelligence. It becomes more revealing when we ask the same question about each invention: What was scarce before it arrived, and what became scarce afterward?
Seen this way, progress follows a chain of shifting bottlenecks. A technology makes one step easier. More people can then use more information. That very success puts greater pressure on the next step.
AI creates a new bottleneck through the same dynamic. When possible solutions emerge faster than we can assess their consequences, the scarce resource becomes our capacity to make sound decisions under uncertainty.
First, information had to outlive memory
Before knowledge could be recorded durably, its transmission depended on people, memory and oral tradition. Preserving information across absence and beyond a person’s lifetime was itself a constraint.
Writing changed that condition. An early writing system developed in Mesopotamia around 3200 BCE, with roots closely connected to recording goods. Information could exist outside human memory and be read again later. Writing researcher Denise Schmandt-Besserat traces this development in The Evolution of Writing .
But a stored text was not yet a widely available text. Once recording became possible, another cost became more prominent: producing additional copies.
Then, knowledge had to be reproduced
In manuscript cultures, every additional copy required labor. Writing materials, trained scribes and time limited the availability of texts.
From the middle of the fifteenth century, Gutenberg’s printing process using movable metal type made reproduction considerably more scalable in Europe. Its significance lay in combining reusable type with a practical production process, as the Gutenberg Society explains. Printing and movable type already had a history outside Europe.
The transition illustrates the pattern. A text still had to be written and typeset. But its contents no longer had to be copied by hand for every additional reader. As the reproduction constraint eased, distribution, literacy, selection and the assessment of competing claims grew in importance.
More information required more computation
Trade, administration, navigation and science increased the demand for reliable quantitative work. Calculation had mattered since antiquity; in the early modern period, logarithms, slide rules and mechanical calculators expanded the available tools.
The task was moving beyond preservation and copying toward processing: existing inputs had to yield useful results.
Eventually, the volume of records became a problem in its own right. Final tabulations for the 1880 US census were not completed until 1887. Hollerith’s punched-card technology was used for the 1890 census. The US Census Bureau documents this transition from manual counting to machine tabulation.
The bottleneck was tangible. An administration could collect data faster than it could process it. The next technological shift therefore targeted the systematic handling of large volumes.
Computers automated procedures; networks expanded access
Programmable computers made the automatic execution of formalized procedures broadly useful. Different processes could be expressed as programs, reducing the need to build a separate machine for each task.
Much of the work shifted to specifying those procedures: understanding requirements, writing software, integrating systems and finding errors.
Networks, the internet and later the web made distribution and access easier. These developments were distinct: the web emerged on top of existing networks. Tim Berners-Lee developed it at CERN in 1989, originally to support information sharing among researchers, as CERN’s history of the web describes.
Information could become accessible without physically transporting local copies. Yet accessible knowledge did not build an application by itself. Software development, operations and integration remained demanding. Cloud infrastructure later made computing resources easier to provision; translating an intention into a working system remained a scarce capability.
The historical sequence at a glance
The following overview compresses that history. It describes overlapping developments and relative constraints. No period had just one information problem across the entire world.
| Historical transition | Bottleneck eased | Next pressure point |
|---|---|---|
| Writing and durable records | Preserving information | Producing and sharing copies |
| Manuscript culture, writing materials and libraries | Transmitting and collecting knowledge | Scaling reproduction |
| Printing | Reproducing texts in larger numbers | Distribution, access and selection |
| Calculation aids and mechanical calculators | Solving quantitative problems | Systematically processing large datasets |
| Punched cards and tabulating machines | Processing repetitive tasks across large datasets | Automating a wider range of procedures |
| Programmable computers | Executing formalized procedures | Developing and integrating software |
| Networks, the internet and the web | Sharing information and making it accessible | Judging relevance and applying knowledge |
| Software platforms and cloud infrastructure | Provisioning complex systems | Designing and implementing solutions |
| Generative AI and coding agents | Producing solution candidates and implementations | Evaluating viability, consequences and errors |
These transitions do not form an inevitable chain of inventions. They share a dynamic: making one step easier increases the demands on what must happen next.
AI now accelerates the generation of possible solutions
Generative AI extends automation into a consequential part of the process. It can produce drafts, arguments and code without every individual solution step being explicitly programmed in advance. How well it works depends on the task and context. A generated result begins as a candidate.
That can reverse the relationship between decision and implementation.
Consider a simplified thought experiment. A software architect makes ten technical decisions in a day. Implementing them then occupies a team for several weeks. In that workflow, implementation limits throughput.
A coding agent can propose and execute numerous changes in a short period. Questions then accumulate more quickly. Does this change fit the architecture? What assumption does it depend on? Which side effect remains untested? Has the defect been fixed, or has only its visible symptom disappeared?
The backlog moves from implementation to evaluation. Adding agents can enlarge that backlog further. More generation capacity increases the number of proposals without automatically increasing the capacity to understand their consequences.
The new bottleneck: sound decisions and timely error detection
Decision capacity cannot be understood simply as the number of decisions made per hour. Automated decisions can multiply errors just as automatically.
Three capabilities become important:
- Generation capacity: How many possible solutions can a system produce?
- Decision capacity: How many alternatives can it meaningfully select among using relevant criteria?
- Falsification capacity: How many false assumptions can it detect and reject in time?
The third capability is particularly important in determining how much autonomy is viable. A passing test answers only the questions that the test actually examines. Even an independent checking mechanism needs suitable observations and can have blind spots. How Much Uncertainty Can an AI Agent Act On? explores this boundary further.
AI can also support evaluation. What matters is whether the process adds evidence: observed consequences, counterexamples, independent measurements or contradictions with the known state. Another persuasively worded answer alone does not provide that evidence.
With AI, the bottleneck shifts again
The historical hypothesis is that successful automation moves relative scarcity toward a less scalable part of the process. This is an interpretive model, not a historical law. Its relevance today could be investigated through review times, undetected errors and the effort required per verified result.
For AI, the implication is concrete. As solutions and implementations become easier to produce, progress depends increasingly on deciding which ones to adopt and recognizing in time when a decision was wrong.
AI creates a new bottleneck: the capacity to make sound decisions under uncertainty and evaluate their consequences. The next decisive innovation may therefore be a system that can reliably discover which of its possible answers will not hold up.