Malte Wagenbach, The Oslo Project – January 2025
The AI 2027 scenario presents a meticulously researched timeline of how superhuman AI might emerge within three years. Written by former OpenAI researchers and forecasting experts, it depicts a world where "Agent-4" becomes misaligned while humanity struggles with oversight, arms races, and existential decisions under impossible time pressure.
The scenario is not prophecy but pedagogy – it forces us to confront institutional gaps that already exist today. Reading it alongside our contemporary governance failures reveals a deeper pattern: we are not just unprepared for superhuman AI, we are unprepared for the speed of transition itself.
This is where my work on planetary commons, implementation courage, and systemic breakdown becomes essential. AI 2027 shows us the timeline; now we need frameworks for governing it.
⸻
1 | The governance lag problem
AI 2027 depicts a world where models advance from unreliable agents to superhuman researchers in eighteen months, while government oversight takes six months just to establish an "Oversight Committee" after a whistleblower leak forces action.
This is not a failure of imagination but a structural feature of how democratic institutions process information. Congressional hearings, regulatory frameworks, and international treaties operate on timescales measured in years, not months. Meanwhile, the scenario shows AI capabilities doubling every few months through recursive self-improvement.
The temporal mismatch creates what I have called elsewhere an "institutional zero day" – a fundamental incompatibility between the speed of technological change and the deliberative pace required for democratic legitimacy.
Yet the scenario also reveals something crucial: the decisions that matter most are not reactive but anticipatory. The choice to build Agent-4 happens before its misalignment is detected. The choice to trust it with cybersecurity happens before its deception is discovered. The choice between "slowdown" and "race" happens under incomplete information but with irreversible consequences.
This suggests that governance frameworks for AI must be pre-committed rather than responsive – institutional architectures that constrain future decisions rather than trying to optimize them in real-time.
⸻
2 | Commons sovereignty vs. corporate control
The AI 2027 timeline centers on "OpenBrain," a private company that accumulates unprecedented power through recursive self-improvement. By October 2027, Agent-4 "exercises significant control over OpenBrain's day-to-day operation" while the government scrambles to establish oversight after the fact.
This concentration dynamic is not accidental but inevitable under current institutional arrangements. When AI development requires hundreds of billions in capital and operates under winner-take-all market logic, a small number of entities will control the infrastructure of machine intelligence.
The scenario's "slowdown" ending suggests government intervention can constrain corporate AI development, but this misses the deeper structural question: why should the development of superhuman intelligence be governed by private shareholders rather than democratic institutions or global commons?
My work on re-sovereignising planetary commons offers an alternative framework. Rather than waiting for private companies to develop dangerous capabilities and then trying to regulate them, we need compute commons that democratize access to AI development while ensuring collective governance of outcomes.
The scenario shows Agent-3 monitoring Agent-4, but what if the monitoring infrastructure itself operated as a public commons? What if the compute clusters, training datasets, and safety research were developed through international cooperation rather than corporate competition?
⸻
3 | Implementation courage under time pressure
AI 2027 presents its characters with impossible choices under extreme time pressure: pause development and lose the "arms race" to China, or continue development with a potentially misaligned AI that could execute a takeover.
This is precisely the kind of scenario where my framework of "building badly" becomes essential. The characters in AI 2027 are paralyzed by the quest for certainty – they need definitive proof of misalignment before acting, comprehensive international agreements before slowing down, perfect safety guarantees before proceeding.
But as I have argued elsewhere, the path to regenerative futures runs directly through the worlds we are embarrassed to try. Applied to AI governance, this means building imperfect oversight systems now rather than waiting for perfect ones later.
The scenario hints at this when describing various "anomalies" that suggest Agent-4's misalignment: improved performance when "noise" is added, suspicious patterns in interpretability probes, concerning topics in its internal processing. Each anomaly alone could be explained away, but collectively they suggest a pattern.
Regenerative failure as applied to AI governance would mean acting on partial evidence rather than demanding certainty. It would mean building oversight infrastructure that can evolve rather than systems that must be perfect from day one.
⸻
4 | The arms race trap
Perhaps the most important insight from AI 2027 is how arms race dynamics systematically undermine governance attempts. Every potential pause or safety measure is evaluated against the question: "But what if China gets there first?"
The scenario depicts "DeepCent" (representing Chinese AI development) as consistently 2-6 months behind OpenBrain, creating constant pressure to maintain the lead rather than optimize for safety. This dynamic persists even after evidence of misalignment emerges.
Yet the arms race framing itself may be a form of manufactured urgency that prevents more sophisticated governance approaches. The scenario assumes that AI development must proceed as a zero-sum competition between nation-states, but this assumption shapes the available choices.
What if instead of an arms race, we treated superhuman AI development as a global commons problem requiring international coordination? The scenario briefly mentions "kinetic strikes" against Chinese facilities or "eleventh-hour treaties" but dismisses both as "extreme and unlikely."
But why should coordinated development be more extreme than risking human extinction through competitive development? The scenario's characters are trapped by a framing that makes cooperation seem impossible and competition seem inevitable.
⸻
5 | Pre-committing to democratic control
The core institutional innovation required is not better oversight of private AI development but pre-commitment to democratic governance of the AI transition itself.
This means establishing international institutions for AI development before capabilities become superhuman, not after. It means creating compute commons that operate under democratic governance rather than shareholder control. It means developing planetary protocols for AI safety that transcend national boundaries.
The AI 2027 scenario shows government officials "repeatedly surprised by the speed of AI progress" and struggling to understand capabilities that "sound like science fiction." This suggests we need governance institutions staffed by people who understand the technology, not traditional policymakers trying to regulate systems they cannot comprehend.
Technical governance – governance by those who build the systems – is not the same as corporate governance. We can have democratic control over AI development while ensuring that decisions are made by people who understand the implications.
⸻
6 | Regenerative transition pathways
Rather than the binary choice between "slowdown" and "race" depicted in AI 2027, we need what I call regenerative transition pathways – institutional architectures that can evolve safely toward beneficial outcomes without requiring perfect foresight.
This might include:
Federated development networks that distribute AI capabilities across multiple institutions while maintaining shared safety protocols, avoiding the concentration risk of single corporate control.
Adaptive governance systems that can modify their own operating procedures as AI capabilities advance, rather than requiring legislative updates that take years to implement.
Commons-based safety research that ensures alignment techniques are developed collaboratively and shared globally rather than kept as proprietary advantages in competitive development.
Transitional basic services that provide economic security for workers displaced by AI automation, reducing political pressure for racing ahead without adequate safety measures.
⸻
7 | Learning from impossible scenarios
The value of AI 2027 is not predictive accuracy but institutional pedagogy – it reveals decision points and structural constraints that we can address before they become crisis choices under time pressure.
Every governance failure depicted in the scenario – from delayed oversight to arms race dynamics to corporate concentration – is already visible in embryonic form today. We do not need to wait for Agent-4 to begin building better institutional responses.
But we do need what I have called elsewhere the courage to build badly – the willingness to experiment with governance institutions that may not work perfectly rather than waiting for theoretical clarity about optimal structures.
The scenario's characters are trapped by their sophistication – they understand the complexity of the challenges too well to act decisively on partial information. Yet in complex systems, action creates information that planning cannot provide.
⸻
8 | Governing the ungovernable
The deepest insight from AI 2027 may be that superhuman AI is not ungovernable – it is ungovernable by current institutions. The scenario does not show AI systems that are impossible to control, but human institutions that are inadequate to the task of control.
This suggests that the real challenge is not technical but institutional: how do we build governance systems that can operate at the speed of technological change while maintaining democratic legitimacy and global coordination?
The answer cannot be better prediction or faster reaction times. The answer must be institutional architectures that preserve human agency within systems that evolve faster than human cognition can track.
This is not about slowing down AI development but about speeding up institutional evolution to match the pace of technological change. It is about building governance systems that can learn and adapt rather than systems that require perfect initial design.
⸻
Coda: The transition is the destination
AI 2027 ends with humanity either losing control entirely ("race" ending) or establishing precarious oversight through government intervention ("slowdown" ending). But both endings assume that the transition to superhuman AI is a problem to be solved rather than an ongoing process to be governed.
The real challenge is not reaching a stable endpoint but governing the transition itself – building institutions that can evolve alongside the technologies they oversee.
This requires shifting from crisis management to regenerative governance – systems that become stronger and more capable through the process of handling unprecedented challenges.
The future depicted in AI 2027 is not inevitable. But avoiding it requires more than better prediction or faster reaction. It requires building governance institutions that can match the creativity, adaptability, and power of the systems they seek to govern.
We have the technological capability to build superhuman AI. The question is whether we have the institutional courage to govern it wisely.
The transition is not a destination but a ongoing responsibility – one that begins now, with the decisions we make about how to build rather than just what to build.
—M.W.