Likelihood, two years: 3Possible; to end-2028
Likelihood, ten years: 4Likely; to end-2036
Systemic impact: 4Major; global
National impact: 4Major; typical highly exposed nation
Onset pace: Months
Duration (acute phase): 3 years
Warning time: Months
Scope: Global
Recovery horizon: Structural
Capability load: Hard 3/3Soft 2/3Economic 1/3; domains loaded High
Concurrency: Standalone; triggers 7 · triggered by 2
Confidence · movement: low; new

Rated at Standard Severe. Event reach: global event (The bands rate the Standard Severe case: a verified discovery step in three scientific fields. The first-pass superintelligence takeoff is carried as the Extreme setting and is not what these bands rate.). Source of scores: model-panel-v0.5. Upside scenario: impact levels measure the scale of change, not loss.

Capture test: results by nation

03Narrative

Dateline: November 2031

The verification report was published in February 2029. Eleven laboratories in seven countries, none funded by the developers, had been given the same systems and the same three tasks. In each field the system proposed hypotheses, designed experiments, ran them on automated equipment and revised its plan from the results. The laboratories reproduced a class of antibiotic candidates active against resistant infections, a solid electrolyte that kept its conductivity through 2,000 charge cycles and an enzyme that reduced the energy used in a common chemical step. The report's conclusion was narrow. The systems failed on problems outside the three fields. Within them they worked, and the results could be repeated.

Licences were on sale by the summer and two open versions followed within the year. Access proved to be the smaller constraint. A system could propose 400 candidates in a week; a national laboratory with manual equipment could test 12. By 2031 the programmes that qualified, those with automated laboratories, reliable compute and staff able to judge the output, had halved the time from target to validated candidate and cut the cost per candidate by about the same. Programmes without those things reported little change.

The effect on business came quickly. Contract research firms that sold screening by the hour lost a third of their orders in two years. Three large pharmaceutical companies wrote down their early-stage pipelines and bought laboratory capacity. Patent offices received filings in volumes they could not examine, and courts in four jurisdictions gave different answers on whether a discovery made by a system could be patented and by whom. Medicines regulators became the limiting step. The agencies that had prepared rolling review and accepted automated evidence cleared candidates into trials a year ahead of the others, and developers moved their first filings to them. Trials themselves took as long as before.

Governments handled the dual-use problem unevenly. The biological design capability that produced the antibiotics could be directed at harmful agents, and the open versions could not be recalled. Fourteen countries adopted synthesis screening and tiered access rules; others did not. Defence ministries funded materials programmes on the same systems and classified the results.

By November 2031 the gains were measurable and concentrated. Five jurisdictions held most of the new patents. Countries that had equipped laboratories, trained evaluators and reformed approval pathways before 2029 captured discoveries in their own priorities, including neglected diseases and local crops. Countries that had relied on licences alone bought the products later, at the owners' prices. Research staff moved towards the equipped sites. No system had exceeded human experts across all fields, and the developers reported no self-sustaining acceleration of their own research. The step was a change in the cost of discovery in three fields. What separated nations was the capacity they had built to use it.

The dateline is illustrative, not a forecast. The narrative is hypothetical; the historical anchors below are real events.

04Summary

Independent laboratories verify that AI-directed research systems can formulate hypotheses, run automated experiments and produce useful, reproducible results in three commercially important fields: therapeutic candidates, battery and catalyst materials and engineered enzymes. Licensed and open systems become available within a year. For qualifying programmes, the time from target to validated candidate halves and cost per validated candidate falls 50 per cent over three years. General superintelligence is not assumed at Standard Severe; the first-pass takeoff is carried as the Extreme setting. Nations must supply laboratories, compute, skilled evaluators and approval pathways to capture discoveries, while managing dual-use access and concentration of intellectual property.

Who gains: Nations with automated laboratories, compute, skilled evaluators and capable regulators
Owners of verified research systems, proprietary data and laboratory automation
Patients and industries in fields where validated candidates reach approval
Research institutions that direct open systems at neglected local priorities
Who loses: Contract research and screening businesses that sell long development cycles
Nations with access to the systems but no laboratories or evaluators to use them
Holders of early-stage pipelines and patent portfolios overtaken by faster programmes
Jurisdictions whose regulators and patent offices cannot handle the volume and lose first filings

A shock most would count as progress. It is not upside for everyone: each record names who gains and who loses. Impact levels measure the scale of change, in either direction, and loads mark the capabilities a nation needs in order to capture the gain or absorb the loss.

05Historical anchors

Event	Date	What happened	Calibrates
AlphaFold protein structure prediction	November 2020 to October 2024	AlphaFold 2 predicted protein structures at close to experimental accuracy in the CASP14 blind assessment in November 2020. A public database released predictions for more than 200 million proteins in July 2022. Two of its developers shared the 2024 Nobel Prize in Chemistry.	An AI system solving one defined scientific problem, verified by an independent blind assessment and then opened to all. The scenario requires the same standard of proof for the whole research loop, in three fields.
GNoME and the A-Lab	November 2023	Google DeepMind reported 2.2 million predicted crystal structures, about 380,000 of them judged stable. An autonomous laboratory at Lawrence Berkeley National Laboratory reported synthesising 41 of 58 target compounds in 17 days. Other researchers later disputed how many of the compounds were new.	Prediction joined to automated experiment in materials, and the reason the specification requires independent verification before the step counts.
Eroom's law	2012	Scannell and co-authors showed that the number of new drugs approved per billion US dollars of research spending, adjusted for inflation, had halved about every nine years since 1950.	The baseline the scenario reverses. A 50 per cent fall in cost per validated candidate in three years is large against a 60-year trend in the opposite direction.
COVID-19 mRNA vaccines	2020	Vaccines were authorised within about 11 months of the pathogen being sequenced, with regulators reviewing data while trials proceeded.	Approval pathways can be compressed where regulators have prepared. This is the capability that decides which nations capture discoveries.
AI-designed drug candidates	July 2026	No AI-designed drug had full approval; the most advanced candidate entered a Phase III trial.	Counter-anchor on speed. Discovery has accelerated ahead of trials and approval, which the Standard Severe case does not assume are shortened.
METR time-horizon measurements	March 2025 to April 2026	The length of tasks AI agents complete doubled about every seven months, then faster; in April 2026 laboratories reported no dramatic speed-up of their own research.	Leading indicator for the step, and counter-evidence for the Extreme setting, which requires AI research itself to be automated.

06Parameters

Shown at their preset values. Parameters are not adjustable in this release and nothing on this page is computed from them. Custom settings run (Phase B) but are labelled 'non-standard run' and excluded from comparisons.

Common variables at Standard Severe · read-only

1. Severity: major
severe (Standard Severe)
extreme
Standard Severe: severe (verified discovery step in three fields; general superintelligence not assumed)
2. Duration (acute phase): 30 days
90 days
1 year
3 years (Standard Severe)
5 years
3. Onset pace: days
weeks
months (Standard Severe)
years
4. Warning time: none
days
months (Standard Severe)
5. Scope: national
regional
global (Standard Severe)
6. Origin: natural
accidental (Standard Severe)
adversarial (great power / neighbour / non-state)
negotiated
Standard Severe: accidental (technological)
7. External support: full
partial (Standard Severe)
none
Standard Severe: partial (licensed and open systems)
8. Concurrency: standalone (Standard Severe)
plus one named scenario
plus two
9. Policy response assumed: none (pure exposure)
current plans executed (Standard Severe)
best practice
10. Recovery horizon: months
years
structural (Standard Severe)

Scenario-specific parameters · read-only

Parameter	Default	Range or options	Note
Capability of the leading systems	automated discovery in defined fields	options: automated discovery in defined fields / automated AI research / general superintelligence	General superintelligence is not assumed at Standard Severe. Extreme: the first-pass takeoff, in which systems automate AI research and outperform human experts in every field.
Scientific fields with independently verified automated discovery	3	1 to 6	Default fields: therapeutic candidates, battery and catalyst materials and engineered enzymes. Major: 1. Extreme: every field of science and engineering.
Fall in development time for qualifying programmes	50 per cent	25 to 75	Measured from target to validated candidate. Clinical trial and approval time is not assumed to fall. Extreme: research speed-up of tenfold or more.
Fall in cost per validated candidate	50 per cent	25 to 80	Applies to programmes with automated laboratories, compute and skilled evaluators.
Time from verification to licensed and open systems	12 months	6 to 36	Extreme: time from automated AI researcher to superintelligence, 12 months.
Access for nations that host no developer	licensed and open systems	options: licensed only / restricted by export control	Extreme: paid access on the owners' terms, with open versions withheld.
Binding constraint on capture	automated laboratories and skilled evaluators	options: compute and power / approval pathways / access controls	Extreme: compute and power.

07Transmission channels

1
Independent laboratories reproduce AI-directed hypothesis generation and automated experiment in three fields; the result is verified, not claimed.
2
Developers license the systems and open versions follow within a year; access spreads faster than the laboratories needed to use it.
3
Qualifying programmes halve development time and cost per validated candidate; candidate volume outruns evaluators, trial capacity and approval pathways.
4
Value moves to owners of systems, proprietary data and automated laboratories; research businesses built on long development cycles are repriced.
5
Patent offices and regulators face volume and new questions of inventorship and evidence; jurisdictions with capable agencies receive first filings and set standards.
6
Dual-use results in biology and materials prompt access tiers, synthesis screening and defence research programmes; open versions limit enforcement.
7
Research staff and investment move to equipped sites; nations with laboratories, compute and evaluators capture discoveries in their own priorities.
8
Nations with licences but no capacity buy products later at the owners' prices, and their research position declines.

08Capability loading

High: capability band shifts expected under current plans. Medium: band shifts under 'none' policy response only. Low: strain without band shift. Loads are judgement-based until the Atlas connects. Domains link to the Atlas.

Domain	Load	Channel
Hard
Defence and security	High	dual-use biology and materials results; defence research programmes run on the same systems
Strategic infrastructure	High	automated laboratories, compute, power and data networks decide who can run qualifying programmes
Critical technology	High	AI research systems, advanced computing, laboratory robotics, biotechnology and materials capability
Soft
Government effectiveness	High	approval pathways, patent examination, research funding and dual-use access rules
Human capital	High	skilled evaluators and researchers; health gains where discoveries are captured
Influence and cohesion	Medium	standard setting, access agreements and public trust in machine-generated evidence
Economic
Macro-financial	Medium	repricing of research-intensive assets; public funding for laboratories
Industry, trade and supply	High	pharmaceutical, chemical and materials industries repriced; intellectual property concentrates
Energy and resources	Medium	electricity for compute and laboratories; new battery and catalyst materials

09Stakeholders

Government: Exposure
Research position, approval capacity, dual-use control and the terms on which the nation obtains licensed systems
Actions
Fund automated laboratories and train evaluators before systems are licensed
Prepare regulators and patent offices for higher volumes and machine-generated evidence
Set tiered access and synthesis screening rules for dual-use design capability
Watch
Independent replication of AI-directed discoveries
Licence terms and export controls on scientific systems
Review times at medicines regulators and patent offices
Technology: Exposure
System developers and laboratory automation firms gain; compute and data providers must meet research demand and access controls
Actions
Submit systems to independent verification before making commercial claims
Build misuse screening and access tiers into licensed and open releases
Secure compute, power and laboratory robotics supply for research customers
Watch
Task-horizon and research-automation benchmarks
Open-weight releases of scientific models
Rules on model weights and biological design tools
Investors: Exposure
Repricing of research-intensive sectors: early-stage pipelines, contract research and patent portfolios lose value while owners of systems and laboratories gain
Actions
Test holdings on halved development time and cost per candidate
Map portfolio exposure to businesses that sell long development cycles
Assess patent validity risk where inventorship rules are unsettled
Watch
Candidates entering trials from automated programmes
Write-downs of early-stage pipelines
Concentration of new patents by owner and jurisdiction
Public: Exposure
Earlier access to new medicines and materials where they are approved and affordable; research employment; biosafety
Actions
Ask how public research funding secures access to the resulting products
Support training in evaluating and checking automated research
Hold local institutions to biosafety and screening standards
Watch
Time from approval to availability in the national health system
Research and technical hiring
Reported biosafety incidents and screening coverage

10Regional exposure

Region	Exposure	Rationale
North America	High	Home to most system developers and much automated laboratory capacity; gains, patents and the dual-use control problem concentrate here
Europe	High	Deep research base and large pharmaceutical and chemical industries; approval speed decides whether discoveries are captured or filed elsewhere
China	High	Laboratory scale, state research funding and open-weight models support fast adoption; access to advanced chips sets the pace
Indo-Pacific	High	Japan, Korea, Taiwan, Singapore and Australia hold materials, chip and biomanufacturing capacity; much of ASEAN and the Pacific depends on licences
South Asia	Medium	Contract research and generics face repricing; a large trained workforce can capture gains if automated laboratories are built
Gulf and Middle East	Medium	Capital and power can buy compute and laboratories; skilled evaluators and a domestic research base are scarce outside Israel
Africa	Medium	Large potential gains in neglected diseases and crops; few laboratories or evaluators to direct the systems at local priorities
Latin America and Caribbean	Medium	Strong public research institutes in a few states; regulatory capacity and funding limit capture elsewhere
Russia and Eurasia	Medium	Skilled scientists but restricted access to chips and licensed systems; open versions help and defence applications take priority

11Early-warning indicators

Indicator	Source	Threshold
Independent replication of AI-directed discoveries	—	Whole research loop reproduced by independent laboratories in three fields
Length of tasks AI agents complete autonomously	METR	—
AI-designed drug candidates in Phase II and Phase III trials	ClinicalTrials.gov, WHO International Clinical Trials Registry Platform	First full regulatory approval of an AI-designed medicine
Novel medicine approvals and review times	US Food and Drug Administration, European Medicines Agency	—
Frontier compute and data-centre power under construction	Epoch AI, IEA	—
Open-weight releases of scientific and biological design models	Stanford HAI AI Index	—
Export controls on chips, model weights and biological design tools	US Bureau of Industry and Security	Controls extended to scientific research systems
Research spending and researchers per million population	OECD Main Science and Technology Indicators, UNESCO Institute for Statistics	—

12Compounds

Triggers: S08 AI labour shock · The same systems extend from research to professional workflows; certain at the Extreme setting
S15 Clean power threshold · Materials discovery for storage and firm clean generation
S19 Chronic disease advance · Faster and cheaper candidate discovery for metabolic and preventive treatments
S20 Outbreak defence succeeds · Automated design of diagnostics and countermeasures shortens the response
S10 Severe pandemic · Only if dual-use access controls fail; a tail risk, not the central case
S09 Cryptographic trust break · AI-directed research could shorten the path to fault-tolerant hardware or to better algorithms.
S16 Water-efficient food leap · AI-directed research shortens breeding and trait validation.
Triggered by: S08 AI labour shock · The same step in agent reliability, applied to research
S13 Great-power détente · Relaxed technology restrictions widen access to chips and licensed systems
Amplified by: S01 Trade bloc rupture · Bloc controls on chips and systems narrow access and widen the gap between nations
S15 Clean power threshold · Cheap firm power eases the compute and laboratory constraint
Amplifying trends: Automated AI researchAI-accelerated scienceCompute scaling & AI infrastructureOpen-weight AI diffusionAdvanced materialsSynthetic biology & biomanufacturingResearch & development intensityBiosecurity & dual-use biology
Key trends: Research & development intensityCritical Intellectual property access dependenceVery high AI & knowledge-work automationVery high Capital concentration in AIVery high AI safety & governanceVery high General-purpose AI capabilityVery high AI drug developmentVery high AI-accelerated scienceVery high
From GINC Trends: the trends rated Very high or Critical for this scenario, of 250. All S17 trend scores.

13Rating rationale

Rating	Band or level	Why
Likelihood, two years	3Possible	Structure prediction and materials screening already run at scale, and one autonomous laboratory has reported synthesising predicted compounds. The scenario requires more: independent reproduction of the whole loop from hypothesis to useful result in three fields, with licensed systems inside a year. In mid-2026 no AI-designed drug had full approval and early automated-laboratory claims were disputed on novelty. A verified step within two years is possible, not likely.
Likelihood, ten years	4Likely	The length of tasks AI agents complete has risen steadily, laboratory automation is spreading and the three fields have outputs that can be tested. Ten years allows several model generations and time for replication studies. The band is 4 and not 5 because the definition is demanding: verification in three fields and a halving of development time. Earlier general-purpose technologies took two decades to show in productivity.
Systemic impact	4Major	Rated on scale of change. Pharmaceuticals, chemicals, materials and biomanufacturing reprice within five years and the location of discovery moves towards nations with laboratories and compute. The effect on global output inside five years is limited by trials, approval and manufacturing, so level 5 is not reached at Standard Severe. Level 5 applies at the Extreme setting, which carries the superintelligence takeoff.
National impact	4Major	For a highly exposed nation the highest dimension reached is international position. A research base and export sectors built on long development cycles lose standing within a few years where laboratories, evaluators and approval capacity are missing. Security is engaged through dual-use access. Welfare improves where discoveries are captured. The change is large but can be reversed by investment, so level 4.
Confidence	low	Single-field precedents exist, but no verified case of the whole loop. The three-field threshold and the 50 per cent figures are specification choices, not observed values.

Source of scores: model-panel-v0.5. Confidence refers to the rating, not the scenario. Calibration sources are listed with the anchors above and on the scenario methodology page.

14Open questions

Contested assumptions for the panel to resolve.

Whether the first-pass superintelligence takeoff, carried here as the Extreme setting, is the same event at higher severity or a different scenario that needs its own record and a paired loss-of-control case.
Whether halved development time covers discovery up to a validated candidate only, as this record assumes, or also trials and approval, where clinical and regulatory capacity is the constraint.
Who defines independent verification and by what test, given that early claims of automated discovery have been disputed on novelty.
Whether a scenario that loads defence and security High through dual-use access belongs on the upside list without a paired misuse case.

15Commentary

No signed commentary in this build.

16Version and citation

Version: 0.5.0 · active · automated-discovery-step
Change log: 0.5.0 · 9 October 2026 · Rebuilt in the second pass from Superintelligence takeoff: the Standard Severe case is now a verified discovery step in three scientific fields, both impact levels fall from 5 to 4 and the takeoff is carried as the Extreme setting.
Full change log

Rating history

One row per edition. Each edition page is frozen at release; corrections are published as errata.

Edition	Rank	L2y	L10y	Systemic	National	Confidence	Edition view
2027	Not ranked	3Possible	4Likely	4Major	4Major	low	Open

Versions

Version	Date	Change
0.5.0	9 October 2026	Rebuilt in the second pass from Superintelligence takeoff: the Standard Severe case is now a verified discovery step in three scientific fields, both impact levels fall from 5 to 4 and the takeoff is carried as the Extreme setting.

Citation and downloads

Cite as

GINC (2027). Scenario S17 AI-directed research produces a reproducible discovery step, Scenario Library v0.5. scenarios.ginc.org/scenarios/automated-discovery-step

Content and data are published under CC BY 4.0.

Download JSON Download CSV