Executive Summary

AI-assisted foresight has moved from speculation to practice, but most adoption still concentrates on discrete tasks. A 2025 World Economic Forum and OECD survey of 167 foresight practitioners in 55 countries found that two-thirds already used AI. Among users, the most common applications were trend analysis or clustering, scenario development and horizon scanning. End-to-end, customised integration across the foresight process remained rare. This is the gap that now matters.

The next transformation is not AI-generated foresight. It is the emergence of the living anticipatory system: machine-scaled, system-structured, human-governed, experientially inhabited and decision-coupled. Such a system can maintain evidence and strategic memory, structure relationships, execute methods, create participatory experiences, connect alternatives to decisions and learn from what follows.

The deepest discontinuity lies in systems and complexity. Foresight has long claimed to work with complex adaptive systems, emergence and interdependence, while many projects relied on lists, static maps and small scenario matrices because richer analysis was technically difficult and expensive. Semantic analysis, knowledge graphs, morphological fields, system dynamics, agent-based simulation and strategic twins are becoming feasible for far more teams. They can turn complexity into a manipulable strategic object. They can also produce complexity theatre: impressive representations without credible evidence, causal discipline or participatory legitimacy.

Greater machine capability increases the need for human and institutional responsibility. The critical human role is not a ceremonial approval at the end of an automated process. People and institutions must retain authority over purpose, boundaries, plural interpretation, values, legitimacy, commitment and consequences. Machines may open and challenge the possibility space. Consequential closure must remain named, contestable and accountable.

The practical implication is that organisations should stop asking only where AI can accelerate individual foresight tasks. They should ask what kind of anticipatory capability they are building, who governs it, how it connects to decisions, and whether it expands collective agency or merely automates the appearance of insight.

From the 2024 Foundation to the 2026 Inflection

The 2024 report remains the intellectual foundation because it did not define foresight as prediction. It began from multiple plausible futures, long-term and contextual analysis, systemic integration, complex adaptive systems, epistemological pluralism, participation, reflexivity and human agency. It also treated appropriation as the fragile bridge between anticipation and action.

Several capabilities described then are now visible in working practice. Frontier models can handle text, images, audio and code; agentic environments can work over files, datasets, browsers and software; persistent workspaces can preserve project context; and specialised agents can be combined with deterministic tools, knowledge bases and human review. The shift is not simply from weaker models to stronger models. It changes the unit of capability.

From generation to execution

In the first phase, a model produced a response. In the second, agentic environments began to perform bounded sequences of work: retrieve evidence, create intermediate artifacts, run analysis, verify results and revise. The object is increasingly a project rather than a prompt.

For foresight, this means that a method can become an executable protocol. Stages can have entry conditions, required evidence, tools, outputs, validation gates and hand-offs. The method is not merely stored in a manual or facilitator's memory. Parts of it can be encoded, inspected and improved.

From sessions to continuity

Persistent workspaces, governed knowledge layers and strategic memory make it possible to maintain an evolving intelligence base between workshops and reports. Evidence, classifications, relationships, assumptions, decisions and outcomes can accumulate instead of being reconstructed for every assignment.

Continuity does not mean permanent activity. A system that updates constantly without disciplined attention can become strategically reactive. It means that the organisation can recover why a conclusion was reached, detect when an assumption changes and reconnect new evidence to previous decisions.

From one model to an ecology of intelligence

The relevant architecture is no longer "the AI." It is a composition of frontier models, specialised agents, deterministic tools, structured databases, taxonomies, expert judgment, participants and governance. Different components have different rights and different failure modes.

The strategic question therefore changes from Which model is best? to How should different forms of intelligence be orchestrated, constrained, evaluated and made accountable?

The Field Is Moving, but It Has Not Yet Reached the System Level

The report's proposition sits inside a field that is already changing. The World Economic Forum and OECD's AI in Strategic Foresight: Reshaping Anticipatory Governance offers the most directly relevant empirical reference. Its mid-2025 survey covered 167 practitioners across government, business, academia and civil society in 55 countries. Two-thirds reported using AI in their work. Among those users, 69% applied it to trend analysis or clustering, 63% to scenario development and 60% to horizon scanning.

The pattern is important. AI is already entering the core method stack, but the report describes three levels of integration. Most practitioners remain at a first level of analysis augmentation: gathering, organising and synthesising information. A second group uses AI as a creative sparring partner for scenarios and wind-tunnelling. A third level, in which customised AI is integrated across the whole foresight workflow and includes agents, complexity mapping and continuous processes, remains rare.

This evidence supports the argument for a living anticipatory system, but it also corrects possible overstatement. The system-level future is not the current norm. It is an emerging practice frontier.

The same survey identifies the institutional weaknesses around that frontier. Respondents raised concerns about reliability, transparency, bias, limited inductive reasoning and the risk of deskilling. Among AI users, only 27% reported that their organisations had formal ethical guidelines in place. The obstacle is therefore not simply technical capability. It is the capacity to combine experimentation with governance, literacy and methodological integrity.

The 2026 OECD paper Exploring Possible AI Trajectories Through 2030 adds another necessary discipline. It presents four broad and plausible paths: Progress Stalls, Progress Slows, Progress Continues and Progress Accelerates. The evidence does not support confidently eliminating any of them. A foresight operating architecture should therefore not depend on one forecast of model progress. It should create value if AI capability plateaus near current levels and remain governable if capability and autonomy advance rapidly.

This report's contribution is not the claim that AI matters to foresight. The field has already established that. Its contribution is a proposed operational answer to the next question: what should foresight become when AI is no longer an occasional tool, but one component of a persistent socio-technical system?

What AI Changes, and What It Does Not

AI changes the economics, cadence and feasible complexity of foresight. It does not settle the normative questions that give foresight its purpose.

What changes

Scale and coverage. More sources, modalities, languages and domains can be monitored than a human team could review manually.

Speed and iteration. Classifications, system maps, scenario configurations and communications can be regenerated when evidence or assumptions change.

Continuity. Agentic pipelines can maintain watchlists, evidence chains, indicators and strategic memory between formal interventions.

Combinatorial reach. Machines can enumerate morphological spaces, test cross-consistency, compare configurations and explore heterogeneous actors.

Representation. The same evidence base can be rendered as a graph, scenario, simulation, artifact, film, conversation, spatial experience or decision interface.

Accessibility. Methods once limited by specialist software, coding skills or large production budgets can become available to smaller teams.

What does not change

Purpose. A model cannot legitimately decide which future is desirable for a community or which sacrifice is acceptable.

Framing. The focal question, system boundary, time horizon and stakeholder field are acts of power. Automating them can hide bias rather than remove it.

Meaning. Statistical, semantic or narrative coherence is not the same as strategic significance.

Legitimacy. Simulated participation cannot replace people whose rights, livelihoods or identities are implicated.

Commitment. A generated strategy is not an adopted strategy. Choice requires resources, opportunity costs and ownership.

Accountability. Someone must remain answerable for the decision, the exclusion, the risk and the consequences.

The practical division is not a simple sequence in which machines analyse and humans decide. Humans and machines interact throughout. The governing distinction is between opening and closing. Machines can open the field by expanding search, combination, challenge and representation. The right to close consequential questions must remain human, institutional and contestable.

The Machine That Scans: From Signal Scarcity to Attention Scarcity

Environmental scanning was historically constrained by analyst time. AI reverses the constraint. Collection, extraction and preliminary classification can become continuous. Trusted attention becomes scarce.

ORION's current reference model makes a critical distinction between a large signal layer and a smaller curated layer of driving forces. A signal is a sourced observation. A force is an interpretation that has survived judgment and acquired strategic meaning. The relationship between them is the value. A force without evidence is an opinion; evidence without curation is noise with a timestamp.

This produces five design requirements.

1. Provenance must be native. Every promoted signal, synthesis and scenario claim needs an inspectable path to sources, transformations and judgments.
2. Taxonomies must remain revisable. Classification improves retrieval, but a frozen ontology can make genuine novelty invisible.
3. Novelty must not be confused with importance. Viral or semantically unusual content is not automatically a weak signal with strategic consequence.
4. Promotion must be governed. Agents can recommend candidates. A defined process should decide what enters durable strategic memory.
5. Scanning must terminate in a question. More coverage creates value only when it changes what the organisation watches, debates, tests or does.

The WEF/OECD data show why this matters. Scanning and clustering are among the most common uses of AI in foresight. That means the first systemic risk is not a shortage of machine-generated material. It is an attention-allocation regime increasingly shaped by model rankings, source availability and embedded taxonomies.

The distinctive human capability is therefore not manual reading at heroic scale. It is curatorial judgment: defining the search space, noticing category failure, comparing sources, preserving dissent, recognising absence and allocating finite executive attention.

A mature system should also protect slow variables. Continuous feeds favour what is newly observable, measurable and publishable. Institutions need deliberate reviews of structural forces that may change slowly but determine the meaning of faster signals. Without this counterweight, real-time foresight becomes a sophisticated form of presentism.

The Systems Breakthrough: Complexity Becomes Workable

Foresight has spent decades asking practitioners to respect non-linearity, emergence, path dependence and interdependence. Yet many projects reduced that ambition to a PESTLE list, a trend radar and four scenario quadrants. The problem was not only methodological discipline. Richer system work was expensive to build, difficult to maintain and dependent on specialised technical skills.

The traditions behind this ambition are substantial. John Holland's work on complex adaptive systems established how non-linear interactions, adaptation and distributed agents can create emergent order. Fritz Zwicky's morphological approach, later developed into computer-aided General Morphological Analysis by Tom Ritchey, offers a way to structure multidimensional, non-quantified problem complexes and examine their possible configurations. These methods predate generative AI. What AI changes is their accessibility, speed and connection to broader evidence systems.

Semantic and graph intelligence

Embeddings, language models and graph methods can cluster heterogeneous evidence, propose relationships, expose bridge concepts, compare alternative taxonomies and visualise a corpus at multiple levels. The strategic move is from objects to edges: not only what forces exist, but what each enables, constrains, amplifies, legitimises or destabilises.

This makes system maps more dynamic, but not automatically more valid. A semantic association is not a causal relationship. A frequently co-occurring pair is not necessarily strategically important. Each edge needs a type, supporting evidence, confidence, provenance and the possibility of disagreement.

Morphological analysis and cross-consistency

Morphological work separates a problem into dimensions and possible conditions, then examines the configurations created by their combination. Ten binary dimensions already create 1,024 formal combinations; twelve create 4,096. A human team can define the dimensions and debate coherence, but software can enumerate the space, record cross-consistency judgments, identify coherent families and preserve the audit trail.

AI can assist in generating candidate dimensions, detecting redundancies, proposing consistency tests and explaining configurations. It must not silently choose the problem variables or treat linguistic plausibility as proof of consistency. The value is not automated scenario selection. It is a larger, more explicit and more contestable possibility space.

System dynamics and agent-based simulation

Code-capable agents can assist in specifying variables, feedback loops, actor rules and candidate equations; implement models; run sensitivity analyses; and render alternative behaviours. Synthetic agents can explore heterogeneous responses and help test assumptions about interaction.

The strongest use is not prediction. It is disciplined inquiry. A simulation forces causal claims into inspectable form, exposes missing variables, reveals counter-intuitive dynamics and identifies potential intervention points. In social systems, the limits remain serious: parameters are uncertain, actors are reflexive and the model can influence the reality it describes.

Strategic twins

A strategic twin should not be described as a perfect replica of reality. It is a governed, updateable representation of a strategic domain: evidence, forces, relationships, actors, assumptions, scenarios, indicators, options and decisions. Its value is rehearsal, traceability and learning.

A second meaning comes from the PAULO_TWIN — an experimental system that builds a governed twin of the author’s own foresight practice. A twin can preserve a decision-maker's frameworks, source priorities, workflows, permissions and strategic memory. This is not the automation of identity. It is an attempt to make the operating logic of judgment more explicit and reusable.

The danger: complexity theatre

AI can produce dense graphs, elegant simulations and plausible causal stories faster than institutions can evaluate them. Visual sophistication can create unearned confidence. Complexity theatre occurs when the representation becomes more impressive while its empirical, conceptual or participatory foundations remain weak.

A model becomes strategic only when its relationships are inspectable, its assumptions are challengeable, alternative models can coexist and the representation changes a real question or decision.

The Machine That Reasons and Imagines

Generative and reasoning models can contribute across the scenario process: widening the driver set, proposing uncertainty axes, testing independence, generating configurations, simulating stakeholders, identifying implications, stress-testing options and translating scenarios into multiple media.

The main failure mode is not crude nonsense. It is the plausible-but-empty future: a polished narrative that feels complete, reproduces familiar genre conventions and changes no decision. It often appears when a model moves directly from a topic prompt to four scenario stories without a governed evidence base, a system architecture or a decision context.

A rigorous process needs a reviewable analytical chain:

evidence -> signals -> forces -> relationships -> strategic spaces -> critical uncertainties -> configurations -> narratives -> implications -> options -> decisions -> indicators

The chain is not perfectly linear. New scenarios may reveal missing evidence; participation may contest the force set; a decision may change what is monitored. Its purpose is traceability. The narrative is one interface to the model, not the model itself.

The best division of labour is asymmetric. Machines are strong at search, comparison, combination, consistency checking, translation and rapid variation. Humans are strong at recognising absence, challenging the ontology, interpreting political meaning, judging surprise and deciding which tensions deserve commitment.

The objective is not to remove subjectivity. It is to make subjectivity visible, plural and accountable.

Futures Must Be Inhabited: From Representation to Appropriation

Analytical understanding does not guarantee organisational movement. A scenario can be intellectually accepted and still remain emotionally distant, socially unowned and institutionally irrelevant. The 2024 report described appropriation as the bridge between anticipation and action. The Long Game's practice turns that bridge into a design challenge.

This direction has a wider methodological foundation. UNESCO's Futures Literacy work treats the capacity to "use the future" as a way of changing what people see and do in the present. Experiential futures, as developed by Stuart Candy and Jake Dunagan, connect abstract future concepts to concrete manifestations in multiple media. Sheila Jasanoff and Sang-Hyun Kim's work on sociotechnical imaginaries reminds us that visions of technological progress always carry assumptions about public purpose, social order and the common good.

These perspectives matter because futures are not neutral content delivered to an audience. They are contested representations of who has agency, whose knowledge counts, what becomes desirable and what is treated as inevitable.

The Long Game combines Strategic Foresight with Speculative Design, Storytelling, artifacts and immersive experiences. AI expands each element:

• scenarios can become films, responsive worlds, spatial soundscapes and personalised narratives;
• artifacts can be generated, prototyped and varied rapidly;
• conversational characters can speak from different futures and respond to participants;
• digital environments can change according to collective choices;
• physical objects, mobile interfaces, projections, mixed reality and facilitation can operate as one experience;
• participant traces can update a living wall, strategic map, manifesto or decision record.

Horizontes da Educação 2050 offers a concrete glimpse. Its final event connected a strategic evidence base and four scenarios with virtual reality, interactive digital stations, a living physical wall, speculative tables and games, artifacts, collective commitments and an AI host speaking from 2050. Participants did not simply receive scenarios. They navigated, compared, interpreted and left traces inside a shared environment.

Experience can perform four strategic functions: reveal tacit assumptions, increase empathy, expose trade-offs and create shared memory strong enough to support action. It is an appropriation technology.

Yet immersive power creates its own ethical questions. Personalisation can isolate participants inside different narrative worlds. Emotional intensity can manipulate rather than illuminate. Cinematic coherence can make one future feel inevitable. Synthetic characters can simulate social legitimacy. The same capabilities that deepen understanding can also narrow agency.

Futurtainment Without Futures Theatre

Futurtainment is a promising design space because entertainment can widen participation, sustain attention and make abstract systems memorable. It is dangerous because a spectacular encounter can create the feeling of transformation without the organisational work of transformation.

The analogy with innovation theatre is precise. Futures Theatre occurs when leaders consume an exciting scenario experience, photograph it and return to the same assumptions, budgets and incentives.

A foresight experience should therefore pass five tests.

1. Decision. Which real decision could change because of the experience?
2. Evidence. Can participants trace what they encountered to forces, relationships and assumptions?
3. Agency. What can participants influence, and what is presented as a boundary condition?
4. Commitment. Does the experience end in an option, experiment, capability, allocation or public commitment?
5. Learning loop. What will be monitored, revisited or revised after the event?

If none of these changes, the experience is entertainment about the future. If it changes how a group understands a system, confronts trade-offs and commits to action, it becomes consequential rehearsal.

The opportunity is not to avoid entertainment. It is to master it while preserving methodological gravity: delight in service of judgment, narrative in service of systems and immersion in service of agency.

From Foresight Projects to an Anticipatory Operating System

The current body of work suggests an architecture broader than any single platform. Its five operational layers follow the flow shown in Figure 1.

1. Evidence and memory

Signals, sources, project materials, assumptions, decisions and outcomes are stored with provenance, permissions and retention rules. This is the durable institutional memory.

2. Sense-making and modelling

Taxonomies, semantic search, graphs, strategic constellations, system maps, morphological fields, causal models and simulations turn material into navigable representations.

3. Method and orchestration

Scanning, Sensing and Acting, preceded by Designing and Scoping and followed by Monitoring and Learning, becomes an executable sequence with entry conditions, tools, validation gates and output contracts. Agents operate inside the method; they do not silently redefine it.

4. Experience and participation

Narratives, artifacts, workshops, conversational agents, immersive media and physical environments allow people to encounter, contest and appropriate the models.

5. Decision, action and learning

Options are stress-tested, commitments are recorded, indicators are monitored, outcomes are compared with assumptions and new questions return to the evidence base.

Within the current ecosystem, distinct components demonstrate complementary roles.

ORION, positioned within The Long Game's evolving platform architecture, embodies the data, sense-making and method chain. Its current reference model separates Data and Information; Sense-making and Visualisation; and Embedding in Projects and Method. It also distinguishes sourced signals from curated driving forces and places human validation gates inside the method.

Futures.360 develops the literacy, concepts, tools and guided application required to use foresight platforms with judgment. Without this layer, organisations may gain sophisticated interfaces while losing the capacity to challenge them.

PAULO_TWIN and PAULO_AGENT_OS provide a control-plane model for governed retrieval, canonical-source priority, permissions, provenance, orchestration, workflows and approval gates. They show that agent capability and agent authority are different design questions.

The Long Game adds participatory imagination, speculative design, storytelling, artifacts and embodied experience.

These roles should not be collapsed into one undifferentiated product. Their strategic value lies in interoperability around a shared object model, explicit interfaces and common governance.

An Adoption Ladder for Organisations

The destination is not useful without a pathway. The following maturity ladder describes organisational capability, not model sophistication. An organisation may use an advanced model and still remain at Level 1 if its work is episodic, ungoverned and disconnected from decisions.

The ladder should not be read as a compulsory technology programme. Some organisations may need a strong Level 2 capability more than a Level 5 system. The correct level depends on the strategic focus, decision cadence, risk, regulatory context and capacity to govern what is built.

The progression also has two tracks. One is technical: data, tools, models and integrations. The other is institutional: literacy, rights, ownership, participation and accountability. Advancing on one while neglecting the other creates fragility.

Three Cases That Demonstrate Different Parts of the System

The report's argument is grounded in three forms of practice. They are not causal evaluations and should not be treated as proof that one architecture works in every setting. Together, however, they show that structural intelligence, agentic continuity and experiential appropriation are already more than abstract possibilities.

Case 1: The AI Deep Dive

The AI Scanning Deep Dive for 2026-2031 moved from 50 driving forces to 12 strategic spaces, 12 critical uncertainties and four configured scenarios. Its methodological contribution is the shift from a list of forces to an architecture of relations.

If twelve uncertainties are treated as binary for formal exploration, they produce 4,096 possible configurations before cross-consistency and strategic judgment reduce the space. The number is not the insight. It shows why a small number of hand-assembled scenarios can conceal the assumptions that excluded alternatives.

The configured scenarios connect capability, compute, energy, capital, infrastructure, agents, governance, geopolitics, work, trust and information. The radar remains useful, but insufficient. It shows what is moving; the strategic-space model shows how the system could change state.

The next step is to make that architecture live: connect signals and indicators to each space, update relationships as evidence changes, preserve alternative interpretations and allow strategy teams to run structured "what if" and wind-tunnelling exercises over the same object model.

Case 2: The agentic operating loop

ORION's July 2026 references describe a three-layer value chain, a large signal base beneath a curated force layer, purpose-scoped agents, semantic instruments, a codified foresight process and human-gated promotion. Futures.360 adds learning and guided application. PAULO_TWIN adds governed retrieval, workflow selection, permissions, provenance and approval gates.

Together they suggest a transition from projects to persistent capability. A scan becomes a living evidence base. A workshop becomes one intervention in a continuing loop. A report becomes a versioned view of a system. The practitioner becomes part architect, curator, facilitator, model critic and decision partner.

The unresolved frontier is also visible. If the methodological copilot reasons from general model knowledge while the proprietary evidence base remains operationally separate, the architecture is conceptually integrated but practically divided. Closing this retrieval-to-method-to-decision loop creates more value than adding another dashboard feature.

Case 3: Horizontes da Educação

Horizontes da Educação moved from an evidence base to strategic constellations and four scenarios for education in Portugal in 2050. The final event then distributed the model across space and media: virtual reality, interactive explorations, speculative games, artifacts, a living wall, collective input and an AI host situated in 2050.

The case demonstrates multimodal continuity. Evidence, system map, narrative, artifact and conversation refer to the same scenario architecture. It also reframes participation as a journey rather than a single workshop technique. Participants navigate, choose, make, compare and leave traces.

AI can reduce the cost of visual development, narrative variation, prototyping and responsive interaction. Coherent experience still requires creative direction, facilitation, spatial design, technical reliability and ethical care. Every sensory layer must deepen understanding of the same strategic system.

The Human Who Frames, Judges and Decides

The more capable the system becomes, the more explicit human responsibility must become.

Human value is often described vaguely as intuition, creativity or empathy. In an agentic foresight system, human and institutional responsibilities are more concrete.

• Framing authority: deciding the focal question, boundary, horizon and whose future counts.
• Epistemic judgment: distinguishing evidence from narrative convenience and recognising when categories are failing.
• Plural interpretation: preserving competing readings and ensuring that local, minority and dissenting knowledge are not averaged away.
• Normative judgment: deciding what is desirable, just, unacceptable or worth sacrificing.
• Political legitimacy: involving affected people and addressing unequal interests and power.
• Commitment: choosing options, allocating resources and accepting opportunity costs.
• Accountability: remaining answerable for consequences rather than attributing them to the model.
• Care: recognising lived realities that system representations inevitably compress.

This is consistent with wider AI governance frameworks. UNESCO's Recommendation on the Ethics of Artificial Intelligence places human dignity, rights, inclusion, transparency and monitoring across the AI lifecycle. NIST's AI Risk Management Framework similarly treats governance, mapping, measurement and management as continuing organisational functions, not a one-time technical test.

The objective is not "human in the loop" as a ceremonial approval click. It is human sovereignty over consequential closure. Decisions with significant social, ethical or irreversible effects require named owners, contestable reasoning and a record of how evidence and values were combined.

New Failure Modes of AI-Enabled Foresight

The risks extend well beyond hallucination.

Epistemic monoculture. Different teams using the same models, sources and prompt conventions may converge on the same supposedly diverse futures.

Synthetic participation. Simulated personas can support preparation, but they cannot replace people whose rights, livelihoods or identities are implicated.

Automation bias with better prose. Fluent reasoning can make weak causal assumptions harder to detect.

Complexity theatre. Dense maps and simulations can create unearned confidence.

Temporal myopia. Continuous updates can privilege what is recent over slow structural change.

Provenance erosion. Repeated summarisation can detach conclusions from sources and transformation history.

Power concentration. Those who own models, corpora, compute, taxonomies and interfaces may shape the imaginable future.

Capability atrophy. Practitioners who outsource framing, interpretation and writing too early may lose the ability to challenge the system.

Experience without action. Immersive futures can become a sophisticated form of organisational avoidance.

Permanent strategic reactivity. A continuous system can keep the organisation responding to signals while never sustaining a long-term commitment.

A mature practice should evaluate three levels: output quality, process integrity and decision consequence. Did the system produce a credible artifact? Was the process plural, traceable and governed? Did it improve a real choice, capability or learning loop?

Four Serious Objections, and What They Require

Objection 1: The value of foresight is the conversation, not the system

The strongest form of this objection is correct. Foresight creates value when people challenge assumptions, hear unfamiliar perspectives and build the capacity to act together. Turning that work into a platform can privilege the repository over the relationship and the workflow over the conversation.

The answer is not to diminish conversation. It is to build systems that make conversations better prepared, more plural and more consequential. Evidence and models should enter the room as contestable objects, not conclusions. Participation should be able to change the system, not merely comment on its outputs. The operating system is valuable only when it protects and extends collective sense-making.

Objection 2: Continuous AI-mediated scanning may degrade epistemics faster than it improves decisions

This can happen. More signals can amplify novelty bias, shorten attention cycles and create pressure to update strategies before structural evidence has changed. Automated synthesis can also hide source differences and reward the material most legible to models.

The response requires more than improved search. It requires an attention constitution: source diversity rules, promotion thresholds, protected reviews of slow variables, explicit stop conditions and a record of what the organisation deliberately chose not to follow. Continuous collection should not imply continuous strategic revision.

Objection 3: Model-generated corpora will converge and make futures less diverse

If organisations use the same foundation models over similar public sources, convergence is a plausible outcome. Asking one model for "diverse scenarios" does not create epistemological pluralism. It may produce stylistic variation around the same inherited assumptions.

Plurality must be designed upstream: different source communities, local knowledge, non-textual evidence, independent models, adversarial roles, minority reports and human participants with the power to reject the ontology. Diversity should be evaluated at the level of assumptions and system configurations, not the number of narratives generated.

Objection 4: Governed anticipatory systems may concentrate interpretive power in their owners

The owner of the corpus, taxonomy, model access and interface can decide what becomes visible, credible and actionable. A system may simulate participation while centralising the right to define reality.

This objection cannot be solved by a general promise of human oversight. Governance must specify rights: who can add evidence, challenge relationships, see provenance, propose alternative models, approve promotion, close decisions and appeal exclusions. In public or high-consequence settings, model governance is inseparable from institutional legitimacy.

These objections do not invalidate the living anticipatory system. They define the conditions under which it deserves to exist.

Strategic Recommendations for Leaders and Practitioners

1. Begin with a consequential decision, not an AI use case

Define the strategic focus, horizon, affected stakeholders and decision that the system should improve. A technically impressive workflow without a decision context will optimise output production.

2. Establish the governance boundary before increasing autonomy

Specify which actions agents may perform, which judgments require human validation, who owns consequential closure and how decisions can be contested. Capability and authority must be designed separately.

3. Build provenance before scale

Do not automate collection and synthesis faster than the organisation can preserve sources, transformations, assumptions and review history. Provenance is structural infrastructure, not an appendix.

4. Move from objects to relationships

Treat signals, forces, actors, strategic spaces, uncertainties, scenarios, options, decisions and indicators as related objects. A longer list of trends is not a more systemic understanding.

5. Encode one complete method with validation gates

Choose one foresight process and make its entry conditions, outputs, owners and stop rules explicit. Learn from a bounded end-to-end workflow before attempting an autonomous platform.

6. Invest in foresight literacy and AI literacy together

AI literacy without futures literacy creates fluent automation of weak methods. Futures literacy without practical AI experimentation leaves practitioners unable to understand the systems shaping their field.

7. Design plurality upstream

Diversify sources, models, participants and interpretive roles. Require alternative system maps and minority positions where consequences are significant. Do not treat narrative variety as evidence of epistemic diversity.

8. Use immersion as consequential rehearsal

Connect each experience to evidence, agency, a named decision, a commitment and a follow-up loop. Measure what changed after the experience, not how engaging it felt in the room.

9. Connect foresight to portfolios and learning

Record assumptions behind investments, policies, capabilities and experiments. Review indicators alongside decisions and compare outcomes with the futures that informed them. The final metric is not the amount of content generated, but the quality of preparedness and action.

10. Design for more than one AI trajectory

Build an architecture that creates value with current capabilities, degrades safely when models fail and remains governable if autonomy advances. Do not make institutional preparedness dependent on a confident technology forecast.

Questions for the Next Research Cycle

1. What is the minimum provenance chain required for an AI-assisted scenario claim to be trusted?
2. How can competing system maps coexist without forcing premature consensus?
3. Which foresight judgments should never be delegated, even if agents outperform people on consistency?
4. How should organisations use synthetic stakeholders without confusing simulation with participation?
5. Can morphological and simulation systems measure diversity at the level of assumptions and configurations?
6. What new professional roles emerge: anticipatory systems architect, corpus curator, scenario model auditor, experience dramaturg or agent governor?
7. How should public institutions govern strategic twins that contain contested assumptions about society?
8. What distinguishes useful continuous foresight from permanent strategic distraction?
9. How can immersive experiences preserve ambiguity rather than collapse futures into cinematic certainty?
10. What evidence would demonstrate that Futurtainment produced consequential rehearsal rather than Futures Theatre?

Scope, Caveats and Evidence Discipline

This is an analytical synthesis and design report, not an empirical forecast of AI capability. Demonstrated current practices are distinguished from proposed architecture and near-term possibilities.

The WEF/OECD survey is the strongest current field-level reference, but its 167 respondents came from established foresight networks and should not be treated as a statistically representative sample of every practitioner or organisation. Percentages describing specific AI uses refer to respondents already using AI.

The OECD AI trajectories paper was published in February 2026 and used capability evidence available up to October 2025. Its scenarios are plausible planning references, not predictions or assigned probabilities.

ORION architecture and corpus descriptions reflect internal July 2026 working references. Only public-safe methodological claims are used. ORION is framed within The Long Game's evolving platform architecture and ownership transition, not simply as an IF product.

Horizontes da Educação combined a long project journey with a final event. Public and internal participation figures refer to different stages and populations, so this report does not state one number as a definitive total.

The three cases are analytical demonstrations, not controlled evaluations. A future research edition should add interviews, decision-outcome evidence and longitudinal evaluation.

AI platforms and model capabilities change rapidly. The report's lasting claims concern architecture, method, governance, experience and institutional learning rather than model rankings or short-lived feature comparisons.

Conclusion

Foresight has spent decades asking people to think systemically with linear tools, embrace complexity with static diagrams, explore possibility with limited combinations, participate through compressed workshops and act on reports that quickly become obsolete.

AI does not automatically solve those contradictions. It makes them designable.

The field can now build continuous evidence systems, navigable knowledge graphs, morphological possibility spaces, executable methods, adaptive simulations and immersive environments at a cost and speed previously unavailable. The opportunity is not simply to produce more foresight. It is to build better conditions for collective attention, interpretation, rehearsal and action.

The responsibility is equally significant. Greater computational power can produce thinner imagination, hidden assumptions, synthetic legitimacy and concentrated interpretive power. Avoiding those outcomes requires more than a human approval step. It requires explicit rights, plural knowledge, inspectable models and accountable institutions.

The future of foresight is therefore a living anticipatory system: machine-scaled, system-structured, human-governed, experientially inhabited and decision-coupled.

The machine can scan, structure, compare, simulate and render. People and institutions must still decide what matters, whose knowledge counts, what is worth doing and who will answer for the consequences.

Primary foundation and practice corpus

• Soeiro de Carvalho, Paulo. How Generative AI Will Transform Strategic Foresight. 2024.
• Soeiro de Carvalho, Paulo. Substack publication and foresight essay corpus. Audited 18 July 2026.
• Soeiro de Carvalho, Paulo. Artificial Intelligence: A Scanning Deep Dive, 2026-2031. 2026.
• ORION. The Three Layers: A Reference Model. Internal working reference, July 2026.
• ORION. The Long Game Platform Technology Overview. Internal working reference, 9 July 2026.
• Futures.360. Public platform and foresight-core materials, 2026.
• The Long Game. Public method and positioning.
• Santander Portugal and project partners. Horizontes da Educação 2050 public reporting. 2026.
• PAULO_TWIN and PAULO_AGENT_OS. Governance and ecosystem architecture. Internal current state, July 2026.

External field and governance references

• World Economic Forum and OECD. AI in Strategic Foresight: Reshaping Anticipatory Governance. 19 November 2025.
• Hobbs, Hamish, Dexter Docherty, Luis Aranda, Karine Perset, Kasumi Sugimoto and Rafal Kierzenkowski. Exploring Possible AI Trajectories Through 2030. OECD Artificial Intelligence Papers, No. 55, 3 February 2026.
• UNESCO. Recommendation on the Ethics of Artificial Intelligence. Adopted 23 November 2021.
• National Institute of Standards and Technology. Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile. NIST AI 600-1, July 2024.

Intellectual and methodological anchors

• Miller, Riel, ed. Transforming the Future: Anticipation in the 21st Century. UNESCO and Routledge, 2018.
• Holland, John H. Adaptation in Natural and Artificial Systems. MIT Press edition, 1992.
• Ritchey, Tom. General Morphological Analysis: A General Method for Non-Quantified Modelling. Swedish Morphological Society, revised 2013.
• Jasanoff, Sheila and Sang-Hyun Kim, eds. Dreamscapes of Modernity: Sociotechnical Imaginaries and the Fabrication of Power. University of Chicago Press, 2015.
• Candy, Stuart and Jake Dunagan. "Designing an Experiential Scenario: The People Who Vanished". Futures 86, 2017, 136-153.
• Godet, Michel. From Anticipation to Action: A Handbook of Strategic Prospective. UNESCO Publishing, 1994.
• Inayatullah, Sohail. Questioning the Future: Methods and Tools for Organizational and Societal Transformation. Tamkang University Press, 2005.
• Schoemaker, Paul J. H. "Multiple Scenario Development: Its Conceptual and Behavioural Foundation." Strategic Management Journal 14, 1993, 193-213.
• Slaughter, Richard A. Futures Beyond Dystopia: Creating Social Foresight. Routledge, 2004.
• Weick, Karl E. Sensemaking in Organizations. Sage, 1995.

About the Author

Paulo Soeiro de Carvalho is a strategic foresight practitioner, researcher and entrepreneur working at the intersection of futures studies, strategy, innovation and emerging technologies. He is the founder and chief executive of IF Insight & Foresight and a co-founder of The Long Game, a strategic imagineering studio, and is the creator of the ORION and Futures.360 platforms for strategic intelligence and futures literacy.

He has taught foresight, strategy and innovation at ISEG – Lisbon School of Economics & Management, University of Lisbon, for close to two decades. There he was the Scientific Coordinator of the postgraduate programme in Foresight, Strategy and Innovation and the Executive Director of the ISEG MBA. He holds a PhD in Management Sciences from Université Jean Moulin Lyon 3, with earlier training in foresight and strategy under Michel Godet and Alain Charles Martinet.

He is a member of the World Economic Forum’s Strategic Foresight Network and of the Dubai Future Foundation’s Global Futures Society, and has contributed to European foresight work through the European Topic Centre on Sustainability Transitions and the European Commission. Across more than twenty-five years in the field, he has combined academic research, international advisory work and the hands-on design of anticipatory systems for public and private organisations.