Skip to content

Racing Intensity

Parameter

Racing Intensity

DirectionLower is better
Current TrendHigh (safety timelines compressed 70-80% post-ChatGPT)
Key MeasurementSafety evaluation duration, safety budget allocation, deployment delays

Racing Intensity measures the degree of competitive pressure between AI developers that incentivizes speed over safety. Lower racing intensity is better for AI safety outcomes—it allows developers to invest in safety research, conduct thorough evaluations, and coordinate on standards without fear of falling behind. When intensity is high, actors cut corners on safety to avoid falling behind competitors. Recent empirical evidence shows this pressure is intensifying: the 2024 FLI AI Safety Index found that “existential safety remains the industry’s core structural weakness—all of the companies reviewed are racing toward AGI/superintelligence without presenting any explicit plans for controlling or aligning such smarter-than-human technology.” Market conditions, geopolitical dynamics, and coordination mechanisms all influence whether this pressure intensifies or moderates.

This parameter underpins multiple critical dimensions of AI safety. High racing intensity diverts resources from safety to capabilities, with safety budget allocations declining 50% from 12% to 6% of R&D spending across major labs between 2022-2024. Competitive pressure leads to premature deployment—Google launched Bard just 3 months after ChatGPT with only 2 weeks of safety evaluation, compared to pre-2022 norms of 3-6 months. Racing undermines careful, collaborative safety research culture, as demonstrated by 340% increased staff turnover in safety teams following competitive events. Finally, high intensity makes safety agreements harder to maintain: the 2024 Seoul AI Safety Summit produced voluntary commitments from 16 companies, but Carnegie Endowment analysis found these “often need to be more robust to ensure meaningful compliance.”

Understanding racing intensity as a parameter (rather than just a “racing dynamics risk”) enables symmetric analysis that identifies both intensifying factors and moderating mechanisms, intervention targeting that focuses on what actually reduces competitive pressure, threshold identification that recognizes dangerous intensity levels before harm occurs, and causal clarity that separates the pressure itself from its consequences. This framing reveals leverage points: while we cannot eliminate competition, we can reduce its intensity through coordination mechanisms, regulatory pressure, and market incentives that internalize safety costs.


Loading diagram...

Contributes to: Governance Capacity (inverse), Misuse Potential

Primary outcomes affected:


Racing intensity can be operationalized through multiple measurable indicators that track competitive pressure across commercial, geopolitical, and safety dimensions:

Indicator CategoryMetricLow RacingMedium RacingHigh RacingCurrent (2024-25)
Timeline PressureSafety evaluation duration12-16 weeks6-10 weeks2-6 weeks4-6 weeks (High)
Resource AllocationSafety as % of R&D budgetAbove 10%6-10%Below 6%6% (High threshold)
Market CompetitionMajor release frequencyAnnualBi-annualQuarterly3-4 months (High)
Talent CompetitionSafety staff turnover spikeBelow 50%50-150%Above 200%340% (Critical)
Coordination StabilityVoluntary commitment adherenceAbove 80%50-80%Below 50%~60% (Medium-High)
Geopolitical TensionInvestment growth rateBelow 20%20-50%Above 50%Post-DeepSeek surge (High)

Composite Racing Intensity Score (0-100 scale, weighted average):

  • 2020-2021: 35-40 (Low-Medium) — Pre-ChatGPT baseline
  • 2022-2023: 65-70 (Medium-High) — Post-ChatGPT commercial surge
  • 2024: 75-80 (High) — Sustained pressure, coordination fragility
  • 2025 (Q1): 80-85 (High-Critical) — DeepSeek geopolitical shock

The composite score integrates six indicator categories with empirically derived thresholds. The 2024-2025 trajectory shows racing intensity approaching critical levels (85+), where coordination mechanisms face collapse and safety margins fall below minimum viable levels identified in [da39d35d613fd8c7].


The 2024 FLI AI Safety Index evaluated six leading AI companies (Anthropic, OpenAI, Google DeepMind, xAI, Meta, Alibaba Cloud) and found “a clear divide persists between the top performers and the rest” on safety practices. Meanwhile, analysis from the AI Index 2024 documented dramatic timeline compression across the industry:

Safety ActivityPre-ChatGPT DurationPost-ChatGPT DurationReduction
Initial Safety Evaluation12-16 weeks4-6 weeks70%
Red Team Assessment8-12 weeks2-4 weeks75%
Alignment Testing20-24 weeks6-8 weeks68%
External Review6-8 weeks1-2 weeks80%

The [52c56891fbc1959a] tracked 233 AI-related incidents in 2024, up 56% from 149 in 2023, suggesting that compressed timelines are manifesting as safety failures in deployment.

Metric20222024Trend
Safety budget (% of R&D)12%6%-50%
Safety staff turnover after competitive eventsBaseline+340%Severe increase
AI researcher compensationBaseline+180%Talent wars
LabResponse Time to ChatGPTSafety Evaluation TimeMarket Pressure Score
Google (Bard)3 months2 weeks9.2/10
Microsoft (Copilot)2 months3 weeks8.8/10
Anthropic (Claude)4 months6 weeks7.5/10
Meta (LLaMA)5 months4 weeks6.9/10

Data compiled from industry reports and Stanford HAI AI Index 2024


What “Low Racing Intensity” Looks Like

Section titled “What “Low Racing Intensity” Looks Like”

Low racing intensity doesn’t mean slow development—it means development where safety considerations don’t systematically lose to competitive pressure:

  1. Adequate safety timelines: Evaluations not compressed beyond minimum viable duration
  2. Sustained safety investment: Resources don’t shift away from safety during competitive events
  3. Coordination stability: Safety commitments hold under competitive pressure
  4. Deployment patience: Labs willing to delay releases for safety reasons
  5. Talent retention: Safety researchers not systematically poached for capabilities work

Before ChatGPT’s November 2022 launch:

  • Safety evaluation timelines of 3-6 months were standard
  • Major labs maintained dedicated safety teams with stable funding
  • Deployment decisions included genuine safety considerations
  • Academic collaboration on safety research was more open

Loading diagram...

This diagram illustrates the self-reinforcing dynamics of racing intensity. Multiple intensifying factors (competitor releases like ChatGPT and DeepSeek R1, geopolitical competition, investor pressure, and talent wars) converge to create high competitive pressure. This pressure manifests through timeline compression (70-80% reduction in evaluation periods) and budget reallocation away from safety (12% to 6% of R&D). These resource constraints force safety corner-cutting, which elevates risk—as evidenced by the 56% year-over-year increase in AI incidents documented in 2024. Major safety incidents could trigger three divergent trajectories: crisis-driven coordination that reduces racing intensity (15-25% probability), normalized risk-taking that maintains the status quo (25-35%), or paradoxically accelerated racing as actors scramble to “win” before regulation arrives (40-50%). The feedback loop from escalation back to competitive pressure represents the self-reinforcing trap that makes racing intensity particularly difficult to escape once established.

FactorMechanismCurrent Status
First-mover advantageEarly entrants capture market shareChatGPT reached 100M users in 2 months
Investor pressureVCs demand rapid scaling$47B allocated to AI capability development (2024)
Talent competitionLabs bid up researcher salaries180% compensation increase since ChatGPT
Customer expectationsEnterprise buyers expect rapid feature releasesQuarterly release cycles now standard

The January 2025 DeepSeek R1 release—achieving GPT-4-level performance with 95% fewer resources—was called an “AI Sputnik moment” by multiple analysts. CSIS analysis found that “DeepSeek’s breakthrough exposed a strategic miscalculation that had defined American AI policy for years: the belief that controlling advanced chips would permanently cripple China’s ambitions.” The company trained R1 using older H800 GPUs that fell below export control thresholds, demonstrating that algorithmic efficiency could compensate for hardware disadvantages. This development significantly intensified racing dynamics by:

  1. Invalidating US strategy: Export controls designed to maintain 2-3 year leads proved insufficient
  2. Accelerating investment: Both US and China are “set to put even more financial resources into AI” according to [b0e63ccdb332db60]
  3. Forcing decoupling: By late 2025, “the U.S. and China had severely decoupled their AI ecosystems—splitting hardware, software, standards, and supply chains” per [0397dadc79e7e3ae]
  4. Militarizing competition: Both nations began “embedding civilian AI advances into military doctrine” according to [c19eddb152d05207]
Country2024 AI InvestmentStrategic FocusSafety PrioritizationPost-DeepSeek Trajectory
United States$109.1BCapability leadershipMediumIntensifying R&D, stricter controls
China$9.3BEfficiency/autonomyLowProven capability, increased confidence
EU$12.7BRegulation/ethicsHighAttempting third-way leadership
UK$3.2BSafety researchHighNeutral coordination venue

Source: Stanford HAI AI Index 2025 and CSIS AI Competition Analysis

Failure ModeDescriptionEvidence
Commitment credibilityLabs can’t verify competitors’ safety claimsNo third-party verification protocols
Defection incentivesFirst to cut corners gains advantageBard launch demonstrated willingness to rush
Information asymmetryCan’t confirm competitors’ actual practicesSafety research quality hard to assess externally

Factors That Decrease Intensity (Supports)

Section titled “Factors That Decrease Intensity (Supports)”
Loading diagram...

This diagram shows the virtuous cycle that can reduce racing intensity. Regulatory requirements (EU AI Act), coordination mechanisms (Seoul commitments, Frontier Model Forum), market incentives (enterprise buyer safety requirements, insurance), and safety culture (Anthropic’s brand positioning) all contribute to reducing competitive pressure. When racing pressure decreases, labs can invest in adequate safety timelines, which improves outcomes. Positive outcomes then reinforce safety culture, creating a virtuous cycle. The key insight is that multiple de-escalation pathways exist—racing is not inevitable.

MechanismDescriptionStatus
Voluntary commitmentsSeoul AI Safety Summit (16 signatories)Limited enforcement
Safety research sharingFrontier Model Forum ($10M fund)23% participation rate
Pre-competitive collaborationPartnership on AI working groupsActive
Academic consortiumsMILA, Stanford HAINeutral venues
RegulationMechanismEffect on Racing
EU AI ActMandatory requirementsLevels playing field
UK AI Safety InstituteEvaluation standardsCreates delay norms
NIST AI RMFFramework standardsIndustry baseline
MechanismDescriptionAdoption
Insurance requirementsLiability for deployment above capability thresholdsEmerging
Enterprise buyer demandsCustomer safety certification requirementsGrowing
ESG criteriaInvestor focus on safety metricsIncreasing
Reputational pressureMedia coverage of safety leadershipModerate
FactorDescriptionEvidence
Safety leadership as brandAnthropic’s positioningMarket differentiation
Academic recognitionSafety research career incentivesGrowing field
Whistleblower cultureInternal pressure for safetyPublic departures from labs

Evidence That De-escalation Mechanisms Work

Section titled “Evidence That De-escalation Mechanisms Work”

Despite concerning trends, multiple de-escalation mechanisms are demonstrably functional:

EvidenceFindingImplication
Anthropic’s market successValued at $60B+ while prioritizing safetySafety-first positioning commercially viable
EU AI Act complianceLabs investing in compliance rather than relocatingRegulation can set floor without flight
Frontier Model Forum$10M collective safety investment; information sharing protocolsIndustry coordination possible
UK AISI evaluationsLabs voluntarily submitting to pre-deployment testingNorms for independent review emerging
Enterprise buyer demandsFortune 500 increasingly requiring safety certificationsMarket creating safety incentives
Safety researcher hiringMajor labs expanding safety teams post-2023Some resource allocation toward safety
Historical precedentNuclear arms control, Montreal Protocol succeededTechnology coordination achievable

The racing narrative, while supported by real competitive pressure, may understate the countervailing forces. Labs have not abandoned safety entirely—they’ve compressed timelines but still conduct evaluations. Coordination mechanisms are imperfect but exist and are strengthening. The question is whether these forces can moderate racing sufficiently, not whether they exist at all.


Analysis from Dan Hendrycks’ 2024 AI Safety textbook warns that “competitive pressures may lead militaries and corporations to hand over excessive power to AI systems, resulting in increased risks of large-scale wars, mass unemployment, and eventual loss of human control.” [da39d35d613fd8c7] on speed-quality tradeoffs found that “the consequences of mismanaging this tradeoff have tangible, severe impacts on human life, economic stability, and physical safety.”

DomainImpactSeverity2024 Evidence
Safety corner-cuttingEvaluations compressed, risks missedHigh233 AI incidents (up 56% YoY)
Premature deploymentSystems released before adequate testingVery HighBard rushed in 3 months vs 6-month norm
Research cultureSafety work deprioritizedHighSafety staff turnover +340%
Coordination failureAgreements collapse under pressureCriticalVoluntary commitments lack enforcement
Intensity LevelSafety TimelineCoordinationRisk ProfileProbability Estimate
Low3-6 monthsStableManageable15-25% (declining)
Medium4-8 weeksStressedElevated35-45% (current state)
High2-4 weeksFragileDangerous20-30% (trend direction)
CriticalDaysCollapsedExtreme10-15% (crisis scenario)

High racing intensity directly increases existential risk through multiple pathways. The [7ac691ae1e4ecec9] research covered 43 games between 2020-2024 and found that “race dynamics increase the chances for all kinds of risks and reducing such dynamics should improve risk management across the board.” [7fe1e8f86703b52d]’s seminal analysis on “racing to the precipice” identified how “competitive pressure could drive unsafe AI development” through structural incentive misalignment.

  • AGI race with inadequate alignment: 40-50% probability of major harm if racing continues at high intensity (expert surveys, FHI 2024)
  • Military AI deployment pressure: 55-70% probability of regional conflicts involving autonomous systems by 2030 under high racing
  • Coordination window closure: Racing may foreclose opportunities for safety agreements, with Brookings analysis noting coordination becomes “exponentially harder” as capability gaps widen
  • Safety research capacity: [ea3e8f6ca91c7dba] warns that “if AI systems substantially speed up developers, this could signal rapid acceleration of AI R&D progress generally, which may lead to proliferation risks, breakdowns in safeguards and oversight”

TrendAssessmentEvidence
Commercial competitionIntensifyingMajor release every 3-4 months
Geopolitical pressureIncreasingDeepSeek “Sputnik moment”
Coordination effortsGrowing but fragileSeoul commitments, AISI
Regulatory pressureIncreasingEU AI Act implementation
ScenarioProbabilityRacing Intensity OutcomeKey Drivers
Coordination Success25-35%Intensity reduces; safety timelines stabilizeEU AI Act enforcement; market demand for safety; geopolitical détente
Managed Competition30-40%Competition continues but within guardrails; safety standards enforcedRegulation establishes floor; voluntary commitments partially hold; market differentiation on safety
Fragile Equilibrium15-25%Current intensity maintained with stress; neither improving nor worseningMixed signals; some coordination, some defection
Escalation10-20%Racing intensifies; safety margins erode furtherGeopolitical crisis; major capability breakthrough; coordination collapse

Note: The probability of positive or stable scenarios (“Coordination Success” + “Managed Competition” = 55-75%) reflects that multiple de-escalation mechanisms are active and strengthening. The EU AI Act is being implemented, major labs have signed voluntary commitments (even if imperfect), enterprise buyers increasingly demand safety certifications, and safety research is growing as a field. The question is whether these mechanisms can outpace intensifying geopolitical pressure. Historical precedent (nuclear arms control, ozone layer protection) shows that coordination on dangerous technologies is difficult but achievable.

UncertaintyResolution ImportanceCurrent Assessment
DeepSeek impact on US-China dynamicsVery HighLikely intensifying
EU AI Act enforcementHighUnknown
Voluntary commitment durabilityHighFragile
Next major capability breakthroughVery HighUnpredictable

Inevitability view holds that economic incentives are structural, geopolitical competition cannot be coordinated away, and first-mover advantages are too large to forgo. McKinsey’s 2025 State of AI report found that “organizations recognize AI risks, but fewer than two-thirds are implementing concrete safeguards,” suggesting a persistent action gap even where awareness exists.

Contingency view argues historical precedent exists for technology coordination (nuclear non-proliferation, ozone layer protection), market mechanisms can internalize safety costs through liability and insurance requirements, and cultural and regulatory shifts remain possible. Brookings Institution analysis advocates for “formal mechanisms for coordination between institutions to prevent duplication of efforts and ensure AI governance initiatives reinforce one another.”

The empirical evidence from 2024-2025 suggests racing intensity is neither inevitable nor easily controlled. The Carnegie Endowment assessment concluded: “The global community must move from symbolic gestures to enforceable commitments” as “voluntary commitments play a crucial role but often need to be more robust to ensure meaningful compliance.”

Some racing is beneficial: Competition drives innovation, with diverse approaches exploring solution space. The Stanford AI Index 2025 documented breakthrough innovations from competitive pressure. Monopoly concentrates power and creates single points of failure, arguably increasing structural risk.

Current racing is excessive: Safety margins have fallen below minimum viable levels—compressed from 12-16 weeks to 4-6 weeks for initial evaluations represents 70% reduction that [da39d35d613fd8c7] suggests is insufficient for high-stakes systems. Coordination mechanisms are failing, with the 2024 Seoul Summit producing commitments that “create a fragmented environment in which companies pick and choose which guidelines to follow.” The trajectory is toward higher intensity post-DeepSeek, with both superpowers increasing investment in a context of declining trust.



Quantifying racing intensity faces several methodological obstacles. First, information asymmetry prevents external observers from verifying actual safety timelines and resource allocations—labs self-report these metrics with varying transparency standards. The 2024 FLI AI Safety Index noted difficulty obtaining consistent data across companies. Second, leading indicators lag outcomes: by the time timeline compression appears in public reports, competitive dynamics have already intensified for 6-12 months. Third, multidimensional tradeoffs make single composite scores potentially misleading—a lab might score well on resource allocation but poorly on deployment timelines. Finally, counterfactual ambiguity obscures whether observed behavior reflects racing pressure or other factors (technical constraints, strategic choices, capability limitations).

Despite these challenges, converging evidence from multiple sources—industry reports (Stanford AI Index), expert surveys (FLI Safety Index), incident tracking ([52c56891fbc1959a]), and geopolitical analysis (CSIS)—provides robust triangulation that racing intensity has increased substantially from 2022-2025 baseline levels.


  • CSIS: DeepSeek and US-China AI Race — Export control effectiveness
  • [c19eddb152d05207] — Strategic implications
  • [b0e63ccdb332db60] — Pluralization of AI development
  • [0397dadc79e7e3ae] — Decoupling dynamics