Artificial
Intelligence

Artificial intelligence has stopped being a single technology and become the condition under which most other change now happens. In the space of a few years it has moved from research labs into the operating systems of work, science, media, finance and government — and the pace of release, adoption and capital formation has outrun the vocabularies and strategies most organisations use to make sense of it.

This report is a response to that gap. It is a strategic-foresight scan of the artificial-intelligence landscape, built to help leaders, educators, policymakers, investors and curious readers see the field whole: not as a stream of headlines, but as a system of forces that can be named, organised and reasoned about together.

It is not a prediction, a ranking of tools, or an investment thesis. Foresight of this kind does not forecast a single future; it widens the field of the plausible and strengthens the capacity to act under uncertainty. The aim is clarity, not certainty — a structured map that makes the next five years easier to navigate and to debate.

The argument the report makes, in one sentence, is that artificial intelligence in 2026 can no longer be understood as a capability story alone. Capability is still the engine, and it is still accelerating. But the questions that will decide what organisations can actually do between now and 2031 are increasingly about who is paying for intelligence and on what terms; what machines are permitted to do on our behalf, and who answers when they exceed it; and whether the public keeps consenting to the infrastructure underneath it all. Those are financing, authority and legitimacy questions, and they move faster than capability does.

The method follows the progressive-clarification logic developed by IF Insight & Foresight and The Long Game. A wide scan of raw signals is distilled into a smaller set of curated forces, which are then organised into a system of strategic spaces and the critical uncertainties that will shape how that system resolves.

Like the field it describes, this document is best treated as a living instrument rather than a finished verdict. It is designed to be revisited as the signals move, and to serve as the foundation for the next stage of work: building and stress-testing scenarios for the future of AI.

How to read this report

The report is organised in layers, from raw evidence to strategic structure. A reader in a hurry can take the executive summary, the system map in Figure 1 and the scenario frame in Section 9 and have the argument. A reader who wants to work with the material should read the twelve spaces in Section 8, which is where the forces stop being a list and start being a system.

• Signals — the non-curated raw material: articles, essays, filings and reports gathered continuously by the scanning system. They are the evidence, not the findings.
• Curated forces — five types built from those signals. Megatrends, Trends, Weak Signals and Wildcards are set out in Sections 3 to 6; Critical Uncertainties, the fifth type, have Section 7 to themselves because they carry the scenario work. Each is classified on the STEEP grid.
• Strategic spaces — twelve critical nodes of the AI system into which the forces cluster, each described by its function, its dynamics, the forces it holds and its tensions with the others.
• Toward scenarios — two Critical Uncertainties become the axes of a 2×2 matrix, producing four contrasting worlds. Each of the twelve spaces takes a different configuration in each.
• Signposts — eight observable indicators, each with a defined trip condition, so the map can be monitored rather than filed.

Artificial intelligence has crossed from a discrete technology into a general-purpose capability that is reshaping the economy, work, science, geopolitics and the human relationship with machines. This document applies the IF / The Long Game scanning method to the AI domain: a systematic immersion in signals and evidence, organised into Forces of Change, and then clustered into a system of strategic spaces.

The exercise runs over a five-year horizon, 2026 → 2031. It identifies 57 Forces of Change — 11 Megatrends, 30 Trends, 10 Weak Signals and 6 Wildcards — classified by STEEP dimension and typology, plus a fifth force type: 12 Critical Uncertainties. The forces cluster into 12 Strategic Spaces, each a critical node of the AI transformation system, defined by its function and by how it connects to and pulls against the other nodes.

Four throughlines run through the whole map.

The Strategic Spaces and Critical Uncertainties are designed as the input to scenario construction. Two uncertainties are recommended as the axes: CU06, whether frontier capability stays ownable, against CU11, whether the AI capital structure holds. They produce four contrasting worlds — Open Abundance, Salvage Intelligence, The Permission Economy and The Rationed Frontier — sketched in Section 9. Building them out is the next stage of the work.

Eight signposts, in Section 10, make the map monitorable: each is attached to a space, each names something observable, and each carries a trip condition.

The approach follows a progressive-clarification logic: begin by opening the map of change in all its complexity, then converge on a structured system of levers that can drive scenarios. Foresight outputs of this kind are not predictions — they are instruments to widen the field of the possible and strengthen the capacity to act under uncertainty.

2.1 Scanning deep dive

A systematic immersion in signals, evidence and dynamics with the potential to reconfigure the AI landscape. Forces of Change are organised on a STEEP grid — Social, Technological, Economic, Environmental, Political — and classified by typology. The STEEP discipline exists to guard against tunnel vision: it forces the scan to ask what is moving in five dimensions rather than following the loudest one.

2.2 From a list to a system: the twelve spaces

The forces are then clustered into 12 Strategic Spaces: critical nodes that represent the principal spaces of transformation of the AI system. This converts a dispersed list of factors into a map of systemic levers, where changes chain together and reinforce one another, or enter into tension. Each space is described by its function in the system and by its interconnections and tensions with the other nodes.

The twelve spaces fall into four families, and reading them that way makes the system legible. Three spaces form the engine room: a source of capability, a physical floor, and a price. Three convert that capability into activity in the world. Three decide who governs it and whether it can be believed. And three are where it meets knowledge, identity and meaning — which is where its legitimacy is finally settled.

One relation deserves naming in its own right. Energy, work and meaning — the substrate, livelihoods and culture — are increasingly contested by the same coalitions, in the same places, with the same argument. The report draws that as a single legitimacy circuit rather than as three separate spaces, because that is how it is being organised, and therefore how it will be contested.

2.3 Sources, and how each class is weighted

The scan draws on four evidence layers: an internal driving-forces intelligence base and continuous scanning system; institutional sources of record; the press of record; and a set of individually authored sources read closely for early detection and for vocabulary. Each class is weighted differently, and the report says which is which.

The fourth class earns its place and its discount in the same breath. Individually authored newsletters and podcasts are typically weeks faster than institutional coverage, and they are where much of the vocabulary of a new field is minted. They are also written by people with positions to advance. The rule applied throughout is to take the events and the language, and to second-source the numbers. Where a figure in this report rests on a single such source, it says so in the line that carries it.

2.4 The evidence base

The scan is instrumented, which means it can say where its own attention went. Between 27 June and 11 August 2026 the four scanning layers produced 65 dated force packs, from which 815 distinct AI-relevant forces were extracted after de-duplication. Figure 5 shows the distribution.

This matters for reading the report critically. Attention in a scan is not the same as importance in the world: a cluster can be large because a story is loud rather than because it is consequential. But publishing the distribution lets a reader see what the scan was looking at while it drew its conclusions, and check whether the structure of the report follows the evidence or the analyst.

2.5 The tempo of this domain, and how to read the horizon

One property of this domain deserves stating before the forces themselves, because it changes how the whole report should be read.

A weak signal is defined, in this method and in the field generally, as an early and ambiguous indicator worth monitoring precisely because it is not yet consequential. The implicit contract is that it has runway — years of it. That contract is what makes a five-year scan a five-year scan.

On this domain, at this moment, the contract does not hold. We can put a number on it, because we measured it against our own previous scan of the same subject: the interval between publication as a weak signal and confirmation as a trend ran to roughly six to seven weeks across three separate forces, in the summer of 2026.

The horizon in this report is therefore a consequence frame, not a schedule of arrivals. Several of the things described here as emerging have already emerged. What has not happened yet is their absorption — by firms, by regulators, by labour markets, by the public — and five years is the right window for that absorption to work through. Read the dates in the evidence, not the horizon on the cover, for timing.

The methodological consequence is worth stating plainly, because it applies to any organisation reading this. If signal latency in a domain is measured in weeks, a scan performed once a year is not a scanning capability. It is a snapshot with a decreasing half-life, which everyone will treat as current for months longer than it deserves. What is required instead is a sensing system: continuous, instrumented, and disciplined enough to hold contradictions rather than resolve them prematurely.

Each is stated below as a question with two contrasting poles, anchored to a strategic space, with a note on what would resolve it. The test each has to pass is strict: the poles must be mutually exclusive, each must be observable in the world, and the difference between them must change what an organisation would do.

Before the twelve spaces are described one by one, it is worth reading the map as a whole. A list of forces is not a system until the relations are drawn, and the relations are where most of the strategic content actually sits.

Three relations carry more weight than the rest. Each is a finding about a relation rather than about a force, which is precisely what a system map is for.

Two Critical Uncertainties become the axes of a 2×2 matrix, and each of the twelve Strategic Spaces takes a different configuration in each quadrant. Choosing the axes is the most consequential judgement in a scanning exercise, because it determines what the scenarios are about.

Two criteria are applied. Each axis must be genuinely unresolved on current evidence — an axis whose answer is already known produces four descriptions of the present. And each must change what an organisation would do next quarter, not merely what it would think.

The recommended pair is CU06 — does frontier capability stay ownable, or does it become a permissioned service? — against CU11 — does the AI capital structure hold, or does it unwind? Ownability decides whether AI strategy is a procurement question or a capability question. The capital structure decides whether the next three years are an expansion or a salvage operation. Between them they set the price, the availability and the terms of intelligence, which every other space inherits.

Where an organisation's exposure is operational rather than strategic, a second matrix is recommended instead: CU03, how much authority is delegated to agents, against CU05, whether governance holds. That pairing produces a sharper story about control and accountability and a weaker one about market position.

The four worlds below are sketches, not scenarios. Scenario construction — narrative, internally coherent, populated space by space, and stress-tested — is the next stage of the work, and the four names here are the handles it will be built on.

Scenarios are only useful if they are monitored. The eight signposts below are the indicators that would move this map: each is attached to a Strategic Space, each names something observable rather than something atmospheric, and each carries a trip condition — the specific event or threshold that should trigger a re-reading of the relevant forces.

They are deliberately few. A watchlist of forty indicators is a way of not watching anything.

A scan built during a period of enthusiasm tends to describe a monotonic trajectory. Two counter-currents run against the grain of this one, and both belong in the map. A map that only bends one way is not a map.

The first is enterprise disillusionment and cost discipline. Alongside continued adoption, the evidence base records the first credible reports of enterprise AI spending being pulled back rather than expanded, a crackdown on unbounded token consumption, routing layers cutting consumption by large margins, and architectures scrapped and rebuilt within months by companies that describe that churn as the fast path. The strategic content here is not scepticism about AI. It is a shift from betting on a model to building an architecture in which any model is replaceable — and it is the single most practical thing an operating executive can take from this report. It is carried in T10.

The second is the verification immune response. The obvious extrapolation from cheap synthetic content is a collapse of shared reality, and that possibility is retained here as a wildcard. But what the evidence actually shows through mid-2026 is a verification layer forming faster than expected: compulsory labelling regimes, platform-level anti-slop infrastructure, watermarking with acknowledged limits, and identity verification becoming an industry in its own right. This does not remove the risk. It changes its shape from collapse to cost — a permanent, unevenly distributed operating overhead on trust.

Two absences are worth recording as well, because a scan that only ever adds is not a scan but an accumulation. The environmental dimension of the STEEP grid remains the thinnest column in the force field, and that is a real property of the AI debate in 2026 rather than an artefact: the environmental argument is arriving through economics and politics — as electricity prices and land use — rather than as an environmental claim in its own right. And the interspecies and non-human-mind foundation-model signal that this scan has tracked previously produced no supporting evidence at all across ten weeks of continuous scanning, and now sits on the watchlist rather than in the report.

This report is a map, not a plan. Three ways of using it have proved most productive with the organisations we work with.

This is a desk-based foresight scan, not empirical research or technology due diligence. It synthesises secondary sources and an internal intelligence base into a structured interpretation of the AI landscape. A few limitations should be read alongside it.

The scan draws on: the ORION driving-forces intelligence database and four continuous scanning layers, which between 27 June and 11 August 2026 produced 65 dated force packs and 815 de-duplicated AI-relevant forces; a curated reading across the technology, business, science and policy press; the most relevant recent work of leading AI thinkers and three institutional sources — the Stanford AI Index, the International AI Safety Report and the State of AI Report; and fresh web research current to 11 August 2026.

Two individually authored sources — one a near-daily newsletter, one a podcast and newsletter pair — were read exhaustively across the window rather than sampled, because both run materially ahead of institutional coverage on this subject. Section 2.3 states how that class of source is weighted. Several concepts used in this report originate with them and are credited at the point of use.

The method builds on the foresight practice of IF Insight & Foresight and The Long Game. Full source detail accompanies this report in the supporting workbook.

Colophon

This is the 2026 edition, published 11 August 2026. It supersedes the June 2026 scan of the same subject. A companion document — What Ten Weeks Did to a Foresight Report — records every change between the two editions and the dated evidence behind each one, for readers who want the audit trail rather than the map.