Intelligence amplification: revision 1 to 2

Revision 1

Revision 2

categories: ["enhancement", "minds"]

tags: ["cognition", "enhancement", "tools", "collective intelligence", "human-computer interaction", "ai"]

summary: "The project of raising human problem-solving capacity by improving the tools, notations and organisations that thinking runs on rather than by altering the brain."

summary: "The project of raising human problem-solving capacity by improving the tools, notations and organisations that thinking runs on rather than by altering the brain."

updated: "2026-07-27"

updated: "2026-10-29"

issues: ["The Flynn effect paragraph states a large empirical claim with no citation."]

---

```infobox

{

"caption": "Concept in cognitive science and human–computer interaction",

"caption": "Concept in cognitive science and human-computer interaction",

"rows": [

{ "label": "Also called", "value": "Augmentation, IA" },

{ "label": "Term introduced", "value": "1956, W. Ross Ashby" },

}

```

**Intelligence amplification** is the strategy of increasing what people can figure out by improving the systems they think with — notations, instruments, software, institutions — rather than by modifying the brain itself. Its proponents argue that essentially all historical gains in human cognitive capability have come this way, and that the biological routes covered elsewhere on this wiki are both harder and less effective by comparison. The idea is old enough to have a settled literature and unsettled enough that nobody agrees on how to measure the result.

**Intelligence amplification** is the strategy of increasing what people can figure out by improving

the systems they think with, such as notations, instruments, software and institutions, rather than

by modifying the brain itself. Its proponents argue that essentially all historical gains in human

cognitive capability have come this way, and that the biological routes covered elsewhere on this

wiki are both harder and less effective by comparison. The idea is old enough to have a settled

literature, and unsettled enough that nobody agrees on how to measure the result.

## Origins

W. Ross Ashby used the phrase "intelligence amplifier" in the mid-1950s, arguing by analogy with power amplification that a machine could amplify a small amount of human selective judgement into a large amount of problem-solving, provided the human supplied the selection criterion.[^ashby1956] Vannevar Bush had already sketched the practical version a decade earlier in "As We May Think", proposing the memex, a desk-sized associative index that would let a researcher build and share trails through a personal library.[^bush1945]

W. Ross Ashby used the phrase "intelligence amplifier" in the mid-1950s, arguing by analogy with

power amplification that a machine could amplify a small amount of human selective judgement into a

large amount of problem-solving, provided the human supplied the selection criterion.[^ashby1956]

Vannevar Bush had already sketched the practical version a decade earlier in "As We May Think",

proposing the memex, a desk-sized associative index that would let a researcher build and share

trails through a personal library.[^bush1945]

J. C. R. Licklider gave the position its most-quoted formulation in "Man-Computer Symbiosis", predicting a period in which humans set goals and formulate hypotheses while machines do the routine work that prepares the way for insight, and observing that most of what he did as a researcher was clerical.[^licklider1960]

J. C. R. Licklider gave the position its most-quoted formulation in "Man-Computer Symbiosis",

predicting a period in which humans set goals and formulate hypotheses while machines do the routine

work that prepares the way for insight, and observing that most of what he did as a researcher was

clerical.[^licklider1960]

Douglas Engelbart turned this into a research programme. His 1962 framework treated the human as one component of a system he labelled H-LAM/T — a human using language, artifacts and methodology, in which he is trained — and argued that capability could be raised by improving any component, that improvements interact, and that the highest-leverage target was the process by which capability itself is improved.[^engelbart1962] His group at SRI built the mouse, hypertext, screen editing, outline processing and videoconferencing, demonstrated together in December 1968 in a session later called the Mother of All Demos. Engelbart's disappointment was that the artifacts were adopted and the methodology was not: the industry took the mouse and discarded the bootstrapping argument.

Douglas Engelbart turned this into a research programme. His 1962 framework treated the human as one

component of a system he labelled H-LAM/T, a human using language, artifacts and methodology, in

which he is trained. He argued that capability could be raised by improving any component, that

improvements interact, and that the target with the most leverage was the process by which

capability itself is improved.[^engelbart1962] His group at SRI built the mouse, hypertext, screen editing, outline

processing and videoconferencing, demonstrated together in December 1968 in a session later called

the Mother of All Demos. Engelbart's disappointment was that the artifacts were adopted and the

methodology was not. The industry took the mouse and discarded the bootstrapping argument.

## The augmentation argument

The core claim has three parts.

**Cognition is not confined to the skull.** Andy Clark and David Chalmers argued that when an external resource plays the functional role that a belief would play — reliably available, automatically endorsed, easily accessed — there is no principled reason to place the boundary of the mind at the skin.[^clark1998] Written arithmetic, musical notation and double-entry bookkeeping made possible operations that no unaided brain performs, and they did so without touching neurons.

**Cognition is not confined to the skull.** Andy Clark and David Chalmers argued that when an

external resource plays the functional role a belief would play, being reliably available,

automatically endorsed and easily accessed, there is no principled reason to place the boundary of

the mind at the skin.[^clark1998] Written arithmetic, musical notation and double-entry bookkeeping

made possible operations that no unaided brain performs, and they did so without touching neurons.

**The brain is expensive to modify and slow to change.** Pharmacological effects are small, as [[nootropics]] documents; genetic routes are constrained by polygenicity, as [[genetic-enhancement-of-intelligence]] sets out; stimulation devices have a replication problem set out in [[non-invasive-neuromodulation]]; and surgical routes carry risks that healthy people will not accept, which is why [[brain-computer-interface|neural interfaces]] remain clinical devices for paralysis rather than consumer products. Even the most direct internal approach, the hippocampal [[memory-prosthesis]], has produced small effects in a handful of patients after two decades of work. Tools, by contrast, can be revised weekly and distributed at near-zero marginal cost.

**The brain is expensive to modify and slow to change.** Pharmacological effects are small, as

[[nootropics]] documents. Genetic routes are constrained by polygenicity, as

[[genetic-enhancement-of-intelligence]] sets out, and stimulation devices have the replication

problem described in [[non-invasive-neuromodulation]]. Surgical routes carry risks that healthy

people will not accept, which is why [[brain-computer-interface|neural interfaces]] remain clinical

devices for paralysis rather than consumer products. Even the most direct internal approach, the

hippocampal [[memory-prosthesis]], has produced small effects in a handful of patients after two

decades of work. Tools, by contrast, can be revised weekly and distributed at near-zero marginal

cost.

**Improvements compound.** A better tool for building tools raises the rate of improvement rather than the level, which is why Engelbart insisted on working on the improvement process itself.

**Improvements compound.** A better tool for building tools raises the rate of improvement rather

than the level, which is why Engelbart insisted on working on the improvement process itself.

```compare

{

## Routes

**Instrumental.** Search, version control, spreadsheets, simulation and statistical software changed which questions are askable; [[ai-protein-design|learned structure prediction]] did the same for questions about proteins. The gains are real and notoriously difficult to isolate, since they show up as changes in what work gets attempted rather than as faster completion of fixed tasks.

**Instrumental.** Search, version control, spreadsheets, simulation and statistical software changed

which questions are askable, and [[ai-protein-design|learned structure prediction]] did the same for

questions about proteins. The gains are real and difficult to isolate, since they show up as changes

in what work gets attempted rather than as faster completion of fixed tasks.

**Notational.** Positional numerals, algebraic notation, graphs and diagrams restructure problems so that fewer working-memory slots are needed. This is the oldest form of amplification and the one with the clearest historical record.

**Notational.** Positional numerals, algebraic notation, graphs and diagrams restructure problems so

that fewer working-memory slots are needed. This is the oldest form of amplification and the one

with the clearest historical record.

**Collective.** Groups can be organised to perform better than their members. Woolley and colleagues reported a general factor of collective intelligence in small groups that predicted performance across tasks and correlated with social sensitivity and equality of turn-taking more than with members' individual scores.[^woolley2010] A later meta-analysis by the same group supported the construct while narrowing the claims made for it. Markets, peer review and open-source development are the large-scale versions, each with characteristic failure modes.

**Collective.** Groups can be organised to perform better than their members. Woolley and colleagues

reported a general factor of collective intelligence in small groups that predicted performance

across tasks and correlated with social sensitivity and equality of turn-taking more than with

members' individual scores.[^woolley2010] A later meta-analysis by the same group supported the

construct while narrowing the claims made for it. Markets, peer review and open-source development

are the large-scale versions, each with characteristic failure modes.

**Environmental.** The Flynn effect — sustained gains in measured IQ across many countries through the twentieth century — is the largest documented change in population cognitive test performance, and it happened without any deliberate enhancement programme. Whether it reflects better nutrition, schooling, smaller families, or increasing familiarity with abstract test formats is unresolved, and several high-income countries have reported stagnation or reversal in recent birth cohorts.

**Environmental.** The Flynn effect, the sustained gains in measured IQ across many countries

through the twentieth century, is the largest documented change in population cognitive test

performance, and it happened without any deliberate enhancement programme. Whether it reflects

better nutrition, schooling, smaller families, or increasing familiarity with abstract test formats

is unresolved, and several high-income countries have reported stagnation or reversal in recent

birth cohorts.

**Machine.** The current version of the argument treats artificial intelligence as an amplifier rather than a replacement, and it is the route on which the ambitions of [[artificial-general-intelligence]] research and of augmentation have converged. [[human-ai-merger]] takes up the stronger form of this claim.

**Machine.** The current version of the argument treats artificial intelligence as an amplifier

rather than a replacement, and it is the route on which the ambitions of

[[artificial-general-intelligence]] research and of augmentation have converged. [[human-ai-merger]]

takes up the stronger form of this claim.

## Evidence from human–machine teams

## Evidence from human-machine teams

Chess supplied the first controlled case. After losing to Deep Blue, Garry Kasparov promoted "advanced chess", in which each player uses an engine. His summary of the freestyle tournaments that followed — that a weak player with a machine and a good process beat both strong players with weak processes and strong machines alone — became the standard citation for the centaur thesis. It has aged unevenly: as engines improved, the human contribution in top-level correspondence and freestyle play shrank toward selecting between engine lines.

Chess supplied the first controlled case. After losing to Deep Blue, Garry Kasparov promoted

"advanced chess", in which each player uses an engine. His summary of the freestyle tournaments that

followed, that a weak player with a machine and a good process beat both strong players with weak

processes and strong machines alone, became the standard citation for the centaur thesis. It has

aged unevenly. As engines improved, the human contribution in top-level correspondence and freestyle

play shrank toward selecting between engine lines.

Field experiments with large language models have produced a more textured picture. A randomised study of professional writing tasks found substantial reductions in time taken and modest gains in rated quality, with the largest benefit to initially weaker writers.[^noy2023] A field experiment with customer-support agents found average productivity gains concentrated among novices, with little effect on the most experienced staff.[^brynjolfsson2023] A large study of management consultants found improvement on tasks inside the model's competence and degraded accuracy on a task designed to sit just outside it, where participants accepted plausible but wrong output.[^dellacqua2023]

Field experiments with large language models have produced a more textured picture. A randomised

study of professional writing tasks found substantial reductions in time taken and modest gains in

rated quality, with the largest benefit to initially weaker writers.[^noy2023] A field experiment

with customer-support agents found average productivity gains concentrated among novices, with

little effect on the most experienced staff.[^brynjolfsson2023] A large study of management

consultants found improvement on tasks inside the model's competence and degraded accuracy on a task

designed to sit just outside it, where participants accepted plausible but wrong

output.[^dellacqua2023]

> [!key] The pattern across studies

> Machine assistance compresses the distribution: it raises the floor more than the ceiling, and it converts some skill differences into differences in the ability to judge when the tool is wrong. That is a different kind of amplification from the one Engelbart described, and it makes verification the scarce capability.

> Machine assistance compresses the distribution. It raises the floor more than the ceiling, and it

> converts some skill differences into differences in the ability to judge when the tool is wrong.

> That is a different kind of amplification from the one Engelbart described, and it makes

> verification the scarce capability.

## Criticism

**Attribution.** No one has shown how to measure amplification cleanly. Productivity statistics have not tracked the arrival of transformative information tools in the way naive extrapolation predicted, a puzzle economists have argued over since Robert Solow's remark about computers appearing everywhere but in the productivity statistics. The same measurement problem undermines the trend extrapolations discussed in [[accelerating-change]].

**Attribution.** No one has shown how to measure amplification cleanly. Productivity statistics have

not tracked the arrival of transformative information tools in the way naive extrapolation

predicted, a puzzle economists have argued over since Robert Solow's remark about computers

appearing everywhere but in the productivity statistics. The same measurement problem undermines the

trend extrapolations discussed in [[accelerating-change]].

**Offloading and dependency.** Sparrow and colleagues found that people are less likely to remember information they expect to be able to look up, and more likely to remember where to find it.[^sparrow2011] Whether that constitutes a loss or an efficient reallocation is disputed; the practical concern is that amplification which vanishes when the tool is removed has not changed the person at all.

**Offloading and dependency.** Sparrow and colleagues found that people are less likely to remember

information they expect to be able to look up, and more likely to remember where to find

it.[^sparrow2011] Whether that constitutes a loss or an efficient reallocation is disputed. The

practical concern is that amplification which vanishes when the tool is removed has not changed the

person at all.

**Homogenisation.** Tools that are used by everyone impose a common structure on thought. Shared search rankings and shared model outputs narrow the diversity of approaches that collective intelligence depends on, which is a mechanism by which amplification at the individual level could reduce it at the group level.

**Homogenisation.** Tools that are used by everyone impose a common structure on thought. Shared

search rankings and shared model outputs narrow the diversity of approaches that collective

intelligence depends on, which is a mechanism by which amplification at the individual level could

reduce it at the group level.

**Displacement of the goal.** Critics of the framing note that "intelligence" in this literature means whatever the benchmark measures. Amplifying performance on measurable tasks is not the same as amplifying judgement, and the historical cases most often cited — notation, printing, statistics — changed what counted as good reasoning rather than doing more of the old kind faster.

**Displacement of the goal.** Critics of the framing note that "intelligence" in this literature

means whatever the benchmark measures. Amplifying performance on measurable tasks is not the same as

amplifying judgement, and the historical cases most often cited (notation, printing, statistics)

changed what counted as good reasoning rather than doing more of the old kind faster.

## Relation to the rest of the field

Intelligence amplification is often presented as the safe alternative to building autonomous machine intelligence, on the grounds that a system with a human in the loop inherits human goals — a position with roots in [[transhumanism]] and a direct bearing on the arguments in [[existential-risk]]. [[nick-bostrom]] treats biological enhancement, neural interfaces and better institutions as alternative paths to greater collective capability, and notes that they are slow compared with the machine path, which is part of the argument in [[technological-singularity]] and in [[differential-technological-development]]. Critics of the safe-alternative framing observe that an amplified human is not obviously safer than an automated system if the amplification is doing the cognitive work and the human is supplying only approval.

Intelligence amplification is often presented as the safe alternative to building autonomous machine

intelligence, on the grounds that a system with a human in the loop inherits human goals. The

position has roots in [[transhumanism]] and bears directly on the arguments in [[existential-risk]].

[[nick-bostrom]] treats biological enhancement, neural interfaces and better institutions as

alternative paths to greater collective capability, and notes that they are slow compared with the

machine path, which is part of the argument in [[technological-singularity]] and in

[[differential-technological-development]]. Critics of the safe-alternative framing observe that an

amplified human is not obviously safer than an automated system if the amplification is doing the

cognitive work and the human is supplying only approval.

The unresolved empirical question is whether the compression effect seen in the assistance studies persists. If tools continue to raise the floor faster than the ceiling, the practical meaning of enhancement changes from making exceptional people more exceptional to making expertise less scarce — which is a different social outcome from the one that both proponents and critics of [[human-enhancement]] have generally argued about.

The unresolved empirical question is whether the compression effect seen in the assistance studies

persists. If tools continue to raise the floor faster than the ceiling, the practical meaning of

enhancement changes from making exceptional people more exceptional to making expertise less scarce.

That is a different social outcome from the one both proponents and critics of [[human-enhancement]]

have generally argued about.

## See also