The future with AGI: five scenarios and the evidence behind each
Most AGI futures are arguments about a bottleneck. Pick the bottleneck—compute, energy, experiments, institutions, or trust—and the future writes itself. Here is what the evidence says about each one.
A future with AGI is not one story, it is a family of scenarios that differ mainly in what they assume is the binding constraint. If cognition is the only bottleneck, gains arrive fast and broadly. If the constraints are physical—experiments, energy, supply chains, clinical trials, construction—then even a superhumanly capable system moves at the speed of the slowest real-world process it depends on. The evidence available in 2026 supports a slower, more uneven picture than the fastest public forecasts: measured productivity gains are real but inconsistent, labour exposure is high while displacement is patchy, electricity demand for data centres is growing fast enough to become a genuine constraint, and governance is fragmenting rather than converging. The useful planning posture is not to pick a timeline but to build institutions and systems that stay useful across several of them.
Every AGI forecast is a bet on one bottleneck
Strip the rhetoric from any AGI scenario and you find a single load-bearing assumption: what is the thing that currently limits progress. Optimistic scenarios assume the limit is cognitive — that we are short of good thinking, and a machine that thinks better removes the ceiling. Slower scenarios assume the limit is physical or institutional — that we are short of experiments, energy, trained people, regulatory throughput, and trust, none of which a smarter model manufactures.
Dario Amodei’s Machines of Loving Grace is unusually honest about this, which is why it is worth reading even if you disagree with it. It sketches a compressed-decades scenario for biology and health while stating repeatedly that the details are guesses, and it explicitly names the constraints that would slow everything down: the speed of physical experiments, human institutions, and the real world’s refusal to run at clock speed.[1]
So the productive way to read any AGI future — including this article — is to ask which bottleneck the author believes in, then check whether that belief is supported by anything measurable. The table below is the honest version of the disagreement, and it is worth pairing with the capability-profile framing that treats progress as a grid of level by domain rather than a single arrival date.[10]
| Assumed bottleneck | Implied future | What would confirm it | Current evidence strength |
|---|---|---|---|
| Cognition only | Fast, broad transformation within years | Sustained autonomous completion of expert work | Weak — partial, uneven capability |
| Compute and capital | Transformation gated by fab and datacentre buildout | Capability tracking compute spend closely | Moderate — historically consistent |
| Energy and grid | Physical infrastructure sets the pace | Datacentre demand colliding with grid capacity | Strengthening — demand growth is measurable |
| Physical experiments | Science accelerates unevenly, biology slowly | Lab throughput becoming the limiting factor | Moderate — argued by proponents themselves |
| Institutions and trust | Capability outruns permission to deploy | High exposure, low measured displacement | Strong — matches current labour data |
What we can actually measure right now
Before scenarios, the boring part: what is already documented. Three findings deserve to be in every AGI conversation because they consistently surprise people in both directions.
First, exposure is high but not uniform, and it is not the same thing as replacement. The ILO’s refined global index estimates a substantial share of employment in high-income countries sits in occupations with meaningful generative-AI exposure, with the effect concentrated in clerical work and falling heavily on roles disproportionately held by women — and exposure means tasks that could be affected, not jobs that will vanish.[2]
Second, measured productivity is messier than the demos. METR’s randomised trial with experienced open-source developers found that participants using AI tools took longer on real tasks in their own repositories, while believing they had been faster.[3] One study on one population is not a law, but it is a well-designed shot across the bow of every productivity claim built on self-report. The gap between perceived and actual speedup is the single most under-priced risk in enterprise AI planning.
Third, the macroeconomic modelling has not caught up with the marketing. Acemoglu’s analysis argues that the aggregate total-factor-productivity effect over a decade is modest under reasonable assumptions about how many tasks are genuinely automatable and how much cost saving each yields.[4] You do not have to accept his numbers to accept the structural point: task-level wins aggregate into macro effects far more slowly than a compounding-growth story implies.
| Question | What the evidence supports | What it does not support | Source |
|---|---|---|---|
| Are jobs exposed? | High, uneven exposure concentrated in clerical work | Uniform imminent replacement | ILO index |
| Do AI tools speed up experts? | Sometimes; context and familiarity dominate | Reliable across-the-board speedup | METR RCT |
| Will growth compound quickly? | Modest aggregate effect under standard assumptions | Automatic decade-scale growth explosion | Acemoglu |
| Is energy a real constraint? | Datacentre electricity demand is rising sharply | That capability is unconstrained by physics | IEA |
| Is governance converging? | Binding rules exist in some jurisdictions | A single global framework | EU AI Act |
Scenario one: compressed science
The most attractive AGI future is the one where discovery accelerates: disease mechanisms understood faster, materials designed rather than stumbled upon, climate and energy problems attacked with more shots on goal. This is the scenario proponents lead with, and it is not silly — several of its early forms have already happened in narrow domains where the search space is large and the verification is cheap.
The structure of that success is the key to forecasting the rest of it. Domains accelerate when three conditions hold: the hypothesis space is huge, evaluation is fast and automatable, and the ground truth is unambiguous. Protein structure, chip layout, and formal mathematics fit. Clinical medicine, ecology, and social policy do not — their bottleneck is the trial, the cohort, the decade of follow-up, and the ethics review, none of which a model shortens by being smarter.[1]
The honest expectation, then, is lumpy acceleration. Some fields go visibly faster within a few years. Others move barely at all until the surrounding infrastructure — automated labs, better instrumentation, faster regulatory throughput — catches up. Anyone promising uniform scientific speed-up is selling the cognition-only bottleneck.
- Accelerates fastest: large search space, cheap automated verification, unambiguous ground truth.
- Accelerates slowly: human subjects, long feedback loops, contested outcome measures.
- Does not accelerate: problems whose limit is consent, capital, or political agreement.
- Watch for: automated laboratories, because they attack the actual bottleneck rather than the imagined one.
Scenario two: the labour market reorganises rather than empties
The pop version of the AGI labour story is a cliff. The evidence so far describes something more like continental drift: task bundles rearranging inside job titles that keep their names. A role loses its drafting and its summarising, gains review and exception handling, and the headcount changes far less than the day-to-day content does.
Two forces make the transition harder than a simple substitution story predicts. The first is the entry-level problem: junior work is disproportionately the automatable work, and it is also how people become senior. Automating the apprenticeship without replacing it is a way to run out of experts in a decade. The second is the verification tax: when a system produces plausible output quickly, someone must check it, and checking is often slower and less pleasant than producing. Organisations that ignore this discover their productivity gain has moved, not appeared.[3]
Usage data from deployed assistants points in the same direction, showing patterns of augmentation alongside automation rather than a clean substitution of one for the other.[9] The scenario to plan for is not fewer people or the same people. It is different work, redistributed unevenly, on a timescale set by organisational change rather than model releases.
| Function | Most likely to be automated first | Most likely to grow | The trap |
|---|---|---|---|
| Software | Boilerplate, migrations, first drafts | Review, architecture, incident response | Review capacity becomes the bottleneck |
| Customer operations | Tier-1 responses, triage, summarisation | Escalation handling, policy design | Quality collapse hidden by response speed |
| Legal and compliance | Research, first-pass review, extraction | Judgement calls, negotiation, accountability | Nobody wants to sign for machine output |
| Finance | Reconciliation, reporting, variance drafts | Controls, scenario judgement, audit | Errors compound silently across periods |
| Research and analysis | Literature sweeps, drafting, formatting | Question design, source criticism | Fluent synthesis of shallow sources |
Scenario three: physics sends the invoice
The least glamorous AGI constraint is the one with the most documentation behind it. The IEA’s assessment of energy and AI is explicit that data centre electricity consumption is on a steep growth path, with AI the main driver, and that the local concentration of that demand — not the global total — is where grid stress actually shows up.[5]
This matters for AGI forecasting because it converts an abstract capability curve into a queue for transformers, substations, land, water and interconnection approvals. Those queues are measured in years and are not responsive to model quality. A superhumanly capable system that cannot be served at scale is a research result, not an economic transformation.
There is a pleasing irony here. The scenario in which AI solves energy is also the scenario in which AI needs energy first, and the second thing happens on a construction timetable while the first happens on a research timetable. Anyone modelling explosive AGI-driven growth should be asked, politely, where the electricity comes from and who approved the substation.
- Data centre electricity demand is growing fast enough to be a planning constraint, not a footnote.[5]
- Constraints bind locally — a national average hides a saturated regional grid.
- Interconnection, land and cooling water operate on multi-year permitting cycles.
- Chip supply, packaging and memory add their own physical queues upstream.
Scenario four: concentration and the governance gap
The risk that policy people worry about most is not the machine, it is the ownership. Frontier capability currently requires capital, compute and talent at a scale available to a small number of organisations. If AGI arrives inside that structure, the distributional question is not will there be gains but who is in a position to capture them, and what recourse does everyone else have.
Regulation is real but fragmented. The EU AI Act is binding law with a risk-tiered structure and phased obligations, which makes it the first substantial compliance surface for general-purpose models.[6] Elsewhere the picture ranges from voluntary commitments to sector rules to nothing. Companies operating across borders are therefore governing to the strictest applicable regime, which produces convergence in practice without convergence in law.
The international expert assessment on advanced AI is a good corrective to both the doom and the dismissal, because it sets out where evidence is strong, where it is contested, and where the honest answer is that nobody knows.[7] Independent scenario work such as the widely-read AI 2027 exercise is useful for the opposite reason: it makes a fast-takeoff assumption chain fully explicit so it can be argued with, and it should be read as a structured argument rather than a prediction.[8]
| Governance question | Status in 2026 | The unresolved part |
|---|---|---|
| Binding rules for general-purpose models | In force in the EU, phased | Enforcement capacity and technical standards |
| Pre-deployment safety evaluation | Voluntary commitments, some national institutes | No agreed threshold or independent authority |
| Compute and access equity | Concentrated among a few actors | Whether public compute changes the picture |
| Liability for autonomous action | Largely untested in courts | Who is accountable when an agent acts |
| Cross-border coordination | Dialogues and expert panels | No mechanism with binding force |
Planning for a future you cannot date
The good news for anyone running an organisation is that the robust actions are the same across most of these scenarios. That is unusual, and it is worth exploiting. You do not need to know when AGI arrives to know that portable data, measured workflows, explicit human authority, and the ability to change model providers are all good ideas in every branch of the tree.
The specific discipline worth adopting is scenario-conditional planning: write down what you would do differently if the fast branch were true, identify the earliest observable signal that would distinguish it, and then stop arguing until that signal appears. Signals beat timelines. A model that completes a week-long project unsupervised is a signal. A model that tops a leaderboard is not.
And keep the humility that the evidence justifies. The field has been wrong in both directions — spectacularly early on some capabilities, spectacularly late on others. The organisations that will do well are not the ones that guessed the date. They are the ones that could adopt a genuine advance in a quarter and could stop a bad deployment in a week.
- Define trigger signals, not dates: unsupervised multi-day task completion, verified domain expertise, reliable abstention.
- Keep human accountability attached to consequential decisions regardless of capability level.
- Protect the apprenticeship pipeline if you automate junior work, or plan to buy seniority later at a premium.
- Measure your own productivity effects; do not import a vendor’s benchmark or your team’s self-report.[3]
- Assume energy, compliance and trust gate deployment more than model quality does.[5][6]
Frequently asked questions
What would actually change if AGI arrived?
The most likely early changes are concentrated: faster discovery in domains with cheap automated verification, large shifts in the task content of knowledge work, and rising demand for review and accountability roles. Broad economic transformation depends on physical infrastructure, regulation and organisational change, all of which move on multi-year timescales that model capability does not shorten.
Will AGI cause mass unemployment?
Current evidence shows high task exposure but patchy displacement. Jobs are bundles of tasks, and automating some tasks usually reorganises a role before it eliminates it. The more concrete near-term risks are the loss of entry-level training paths and the hidden cost of verifying machine output at scale.
Could AGI cause an intelligence explosion?
The argument depends on AI research itself becoming the first superhuman domain and on that domain not being limited by compute, experiments or data. Those are open empirical questions. Expert assessments treat rapid recursive improvement as a possibility worth preparing for rather than an established trajectory.
Is energy really a limit on AI?
Yes, increasingly. Data centre electricity demand is growing quickly, and constraints bind locally through grid interconnection, land, cooling and permitting rather than through global generation totals. These queues run on multi-year timescales that are unaffected by model quality.
How should I plan when the timeline is unknowable?
Plan against signals instead of dates. Decide in advance what observable capability would change your strategy, such as a system completing a week-long project unsupervised, and build the things that pay off in every scenario: portable data, measured workflows, explicit human authority and provider independence.
Are AGI scenario documents worth reading?
Yes, if you read them as structured arguments rather than forecasts. Their value is that they make assumption chains explicit enough to challenge. Check which bottleneck the author assumes is binding, then ask what measurable evidence supports that assumption.
Sources and evidence
Primary and authoritative sources used for factual claims. Company research and executive forecasts are labeled as such in the article.
- 1Machines of Loving GraceDario Amodei · 2024-10
- 2Generative AI and Jobs: A Refined Global Index of Occupational ExposureInternational Labour Organization · 2025-05
- 3
- 4The Simple Macroeconomics of AIDaron Acemoglu, NBER · 2024-05
- 5Energy and AIInternational Energy Agency · 2025-04
- 6Regulation (EU) 2024/1689 (Artificial Intelligence Act)Official Journal of the European Union · 2024-07
- 7International AI Safety ReportUK Government and international expert panel · 2025-01
- 8AI 2027 scenarioAI Futures Project · 2025-04
- 9Anthropic Economic IndexAnthropic · 2025
- 10Levels of AGI: Operationalizing Progress on the Path to AGIGoogle DeepMind · 2023-11