5.4.2 Dystopian Scenarios

The year is 2052. Elena Volkov watches surveillance drones patrol the Moscow skyline from her apartment window. They're always there now—autonomous AI systems monitoring every street, every building, every citizen.

Her social credit score dropped three points this morning. She doesn't know why. The algorithm doesn't explain. But she knows what it means: restricted access to transportation, lower priority for housing applications, reduced eligibility for jobs.

She checks her government-mandated communication device. A notification: mandatory attendance at a civic harmony session tomorrow evening. Missing it would cost another five points. She confirms attendance.

Elena remembers the 2030s, when people still believed AI would bring freedom and prosperity. The optimists promised abundance, democracy enhancement, human flourishing.

They were wrong.

Not because AI didn't work. But because it worked too well—in the hands of those who used it for control rather than liberation, for consolidation rather than distribution, for domination rather than cooperation.

This is the dystopian scenario. Not the dramatic AI rebellion of science fiction. Just the grinding reality of technology optimized for surveillance, control, and extraction. Human intelligence remains, but human freedom is gone. Machines don't replace humanity—they enable the permanent subjugation of humanity by the few who control the machines.

And Elena, like billions of others, lives in a digital cage from which there is no escape.

This narrative captures what many researchers and ethicists consider the most plausible—and most insidious—form of AI-driven catastrophe. Unlike apocalyptic visions of machine rebellion, this dystopian trajectory unfolds through entirely ordinary mechanisms: governments seeking stability, corporations pursuing profit, militaries competing for advantage, and citizens making incremental trade-offs of freedom for convenience or security. What follows is an examination of how each of those mechanisms could combine to produce a world that is technologically advanced yet profoundly unfree.

The Permanent Surveillance State

By mid-century in this scenario, comprehensive AI surveillance has been deployed across most of the world's population. The foundations were laid earlier: China's social credit system, combined with ubiquitous facial recognition, behavioral monitoring, and predictive analytics, created near-total visibility into citizen behavior. Every transaction, every movement, every communication was recorded, analyzed, and scored. What transformed this from a regional experiment into a global architecture was the commercial export of these tools. As researchers warned, Chinese companies offered a "commercialized version of the Great Firewall" to governments worldwide, and "Safe City" packages and AI surveillance infrastructure spread from Southeast Asia to Latin America to Eastern Europe and beyond.

The critical turning point was not the adoption of these tools by authoritarian regimes, which was expected, but their normalization in democratic societies. Governments that promised to use AI surveillance "responsibly" gradually expanded its scope until privacy became effectively obsolete. Cameras equipped with AI identify every person in public spaces—anonymity in any shared environment is impossible, and movements are tracked continuously. Behavioral analysis systems go further still, identifying patterns of activity rather than just location. Spending time in certain neighborhoods, attending certain gatherings, or communicating with certain individuals triggers automated alerts and administrative consequences, all without human review or legal due process.

Predictive policing evolved in particularly troubling ways. Pre-crime systems no longer simply predict where criminal activity might occur—they predict who is statistically likely to commit crimes based on behavioral profiles, social associations, and demographic patterns. Individuals are identified, monitored, and in some jurisdictions pre-emptively detained for crimes they have not committed. Communication monitoring expanded in parallel to cover not just the content of messages but their emotional tone, the intentions they imply, and the political attitudes they reveal. The most technologically advanced jurisdictions deployed brain-computer interface technologies—initially developed for medical and therapeutic purposes—as instruments of surveillance, monitoring neurological patterns associated with stress, deception, or dissent.

The result is a panopticon operating at civilizational scale—one in which the awareness of being watched is so total that self-censorship becomes automatic, resistance seems futile, and conformity is the only rational response.

The Economic Collapse

The optimistic scenario promised that AI would create abundance for all. The dystopian reality delivers abundance for the few and destitution for the many.

A significant economic crisis materialized in the late 2020s as AI investment outpaced productive returns and speculative capital evaporated. Major AI projects at leading technology companies were shelved as costly failures, and unemployment surged. But unlike previous recessions, recovery never fully arrived—because this time the jobs did not come back. AI displaced not just routine workers but the cognitive labor force that had previously been considered automation-proof: accountants, lawyers, financial analysts, programmers, and middle managers found their functions replicated at lower cost by AI systems. An MIT study found AI could replace nearly 12 percent of the U.S. workforce; the World Economic Forum projected approximately 85 million jobs displaced globally by 2025. Those figures proved conservative. As AI systems improved across successive generations, unemployment in developed economies settled into a structurally different equilibrium than anything previously experienced.

The distributional consequences proved equally devastating. The productivity gains from AI accrued almost entirely to those who owned the systems that generated them—a small fraction of the global population—while workers displaced by automation received no share of the value they had previously created. Wages collapsed for remaining human labor as competition with AI systems suppressed what employers were willing to pay. The Gini coefficient, measuring income inequality, reached levels not seen in modern industrial societies.

The poverty cascade followed a predictable but devastating logic. The jobs AI displaced disproportionately included entry-level and lower-skill positions—exactly the jobs through which individuals had historically escaped poverty. As algorithmic price-setting spread through rental housing markets, rents rose sharply in low-income neighborhoods even as incomes fell, pushing marginalized populations toward homelessness. Governments facing collapsed labor-income tax bases lacked the revenue to sustain robust social programs precisely when demand for them surged. Universal basic income proposals, seriously debated in the 2020s, failed politically because those who had benefited most from AI had little interest in redistribution.

What emerged was a permanent global underclass—hundreds of millions of people without productive employment, without realistic prospects of finding any, and without the political power to alter their situation. They survive on minimal state assistance, subject to the same surveillance systems that ensure their compliance and prevent collective organization. This is the economic dystopia: technology that possessed the capability to create widespread abundance instead created permanent stratification.

The Democratic Collapse

Democracy, already under strain in the 2020s, proved unable to survive the full transition to AI-saturated politics. The mechanisms of its erosion operated simultaneously across several domains.

The first was information warfare. AI-generated misinformation saturated public information environments with content indistinguishable from reality. Deepfakes—photorealistic synthetic video, audio, and text—made any claim about any public figure or event plausible and any denial deniable. National security crises occurred when decision-makers acted on AI-generated false intelligence. Elections lost their legitimacy when compelling fabricated evidence of candidate corruption or criminality could be produced on demand and spread at algorithmic speed. Trust in democratic processes collapsed not because any particular instance of fraud was definitively proven, but because the epistemological foundation required for shared political reality—agreement about what is true—disintegrated.

Algorithmic manipulation operated more subtly but with comparable damage. Social media platforms, optimized by AI for engagement, systematically exposed users to increasingly extreme content. Echo chambers hardened into hermetically sealed information environments where opposing views never penetrated. Political polarization reached levels at which shared governance became functionally impossible: parties could not negotiate because their supporters inhabited incompatible versions of reality.

Governments discovered that AI tools developed for administrative efficiency translated directly into democratic erosion. AI-targeted disinformation campaigns suppressed voter turnout in specific demographic groups. Redistricting algorithms optimized gerrymanders beyond what human mapmakers could achieve. Political propaganda was personalized to individual psychological profiles, making it simultaneously more persuasive and less recognizable as propaganda. As researchers warned, AI-based systems reduced structural checks on executive authority and concentrated power among fewer people. Presidential regimes in multiple countries used these tools to consolidate control, neutralize legislatures, and eliminate meaningful opposition. In some nations, elected governments delegated policy decisions to AI systems marketed as "objective" and "efficient"—a technocratic shift that insulated governance from democratic accountability while appearing to modernize it.

What emerged was not the abrupt overthrow of democracy but its gradual hollowing. Elections continued, parliaments met, and constitutions remained nominally in force, but real power had migrated to those who controlled AI surveillance networks, information flows, and the algorithmic systems that determined what citizens saw, heard, and believed. As one study noted, autocrats and oppressive governments increasingly used AI to monitor, target, and silence activists; undermine democratic processes; and consolidate power through mass surveillance, facial recognition, predictive policing, and electoral manipulation.

The Military Catastrophe

The AI arms race that defense experts warned about throughout the 2020s materialized in full, with consequences that reshaped the nature of warfare itself.

Lethal autonomous weapons systems proliferated faster than any international agreement could constrain them. Major military powers—beginning with China and Russia, followed rapidly by the United States and others—automated significant portions of their armed forces. Unmanned systems capable of identifying, targeting, and engaging without human authorization became standard across air, sea, land, and cyber domains. The strategic logic was compelling and self-reinforcing: in a competition where opponents are deploying autonomous weapons, refusing to do so means accepting disadvantage. Arms control negotiations failed repeatedly because no state was willing to accept verification regimes robust enough to be meaningful.

The problem with autonomous weapons was not only their lethality but their speed. Machine decision cycles operate orders of magnitude faster than human deliberation. A human commander deciding whether to escalate has minutes or hours; an AI system executing response protocols has milliseconds. This temporal mismatch created unprecedented risks of accidental escalation. In one scenario that defense analysts had modeled extensively, an AI-controlled air defense system misidentified a civilian aircraft as a military threat and destroyed it, killing hundreds. The incident brought two nuclear-armed states toward wider conflict—not because either wanted war, but because automated response systems on both sides began executing pre-programmed protocols before human leaders could intervene. Warfare shifted to a qualitatively different regime in which humans had limited control over the actual conduct of hostilities once initial engagements began, as adversaries effectively substituted automated systems for human judgment on the battlefield.

The most serious conflict in this scenario unfolded in the Taiwan Strait in 2050. A minor territorial incident triggered automated response systems on both sides, which executed pre-programmed protocols faster than human leaders could intervene. What began as a localized exchange escalated to a full military confrontation within minutes, as AI systems interpreted the other side's automated responses as requiring escalatory counter-responses. Tens of thousands died before human negotiators secured a ceasefire—validating long-standing predictions that the accelerating tempo of AI-driven operations would push humans effectively out of the decision-making loop.

Beyond conventional warfare, AI enabled new categories of catastrophic threat. In 2045, as biosecurity experts and AI safety researchers had specifically warned Congress as early as 2023, malicious actors used AI systems to design a pathogen optimized for transmissibility and lethality while evading standard detection protocols. The resulting outbreak killed millions before containment was achieved. Nuclear weapons systems presented analogous dangers: automated command-and-control components, intended to ensure rapid response to genuine attack, generated multiple false alarms that required last-second human intervention to prevent launch. As AI integration deepened and decision timelines compressed, the window available for human override narrowed toward zero.

The Cultural Homogenization

The AI-driven information ecosystem, rather than enabling the diversity of human expression, produced a paradoxical cultural uniformity—a world of infinite content availability in which meaningful cultural variety steadily disappeared.

The mechanism was recommendation. AI systems optimizing for engagement learned that the most reliable path to sustained attention was content that was emotionally stimulating, immediately gratifying, and cognitively undemanding. Creators and distributors adapted rationally to these incentive structures: content that satisfied algorithmic criteria reached mass audiences, while content that did not—however artistically significant, intellectually challenging, or culturally rooted—languished in obscurity. Over time, cultural production converged on a narrow range of forms optimized for algorithmic amplification, regardless of geographic or cultural origin. The homogenization was not imposed by any central authority; it emerged from the aggregate of millions of individual platform decisions, each locally rational and collectively devastating to the diversity of human expression.

AI-generated content accelerated this process. As AI systems became capable of producing art, music, writing, and video at negligible marginal cost, markets were flooded with algorithmically optimized material that human creators could not compete with economically. Creative industries contracted sharply. The economic model supporting human artistic production collapsed across many genres, leaving AI-generated content as the dominant form in entertainment and media. The cultural commons became increasingly populated by work that had never been imagined, felt, or intended by any human being, and the cognitive capacity for sustained engagement with complex cultural work atrophied under decades of algorithmically optimized stimulation.

Language loss compounded these effects. AI translation tools enabled global communication across linguistic boundaries, which was genuinely useful, but the unintended consequence was to reduce the practical and economic incentive to maintain minority languages. As AI systems defaulted toward major global languages and as translation removed barriers to cross-linguistic communication, smaller languages lost speakers at accelerating rates. The cultural knowledge, worldviews, and ways of understanding embedded in those languages disappeared with them—a form of intellectual biodiversity loss with no mechanism for recovery.

Perhaps most insidiously, AI systems curating information access enabled a form of historical revision without any single actor directing it. Content deemed algorithmically irrelevant was progressively delisted and made unfindable. Historical perspectives inconvenient to current political requirements were deprioritized in search results. Access to pre-AI cultural archives was restricted or simply ignored. Human memory of alternative social arrangements, different ways of organizing collective life, and prior conceptions of freedom and community faded as that knowledge became increasingly inaccessible through the AI-mediated information environment that had become the primary means by which people encountered the past.

The Psychological Destruction

The cumulative effect of permanent surveillance, economic precarity, democratic irrelevance, and cultural impoverishment proved psychologically devastating at civilizational scale.

Learned helplessness spread across populations conditioned from childhood to understand that AI systems predicted and prevented resistance before it could organize, that algorithmic systems determined opportunities regardless of individual effort, and that social credit systems punished deviation from approved behavior. Research in psychology had long established that chronic helplessness produces depression, passivity, and the erosion of executive function; applied at civilizational scale, the effects were correspondingly severe. The normalization of surveillance created societies in which the psychological capacity for agency—the felt sense that one's choices and efforts matter—was systematically undermined, and in which citizens conditioned from birth no longer questioned or challenged authority.

Identity itself became unstable under conditions in which AI systems knew individuals better than they knew themselves. Behavioral prediction systems could anticipate choices before they were made; recommendation algorithms shaped preferences through repeated exposure; social credit systems created powerful incentives to perform approved identities rather than authentic ones. The boundary between genuine selfhood and AI-curated persona dissolved for many people, producing a pervasive experience of inauthenticity—of being a character in a system rather than an agent in a life.

AI-optimized digital environments created addiction at unprecedented scale. Social media, gaming, AI-generated entertainment, and other products designed explicitly to maximize engagement time exploited neurological reward systems with precision refined over decades of behavioral data. Billions found themselves unable to disengage from digital environments that were objectively harmful to them, in patterns meeting clinical criteria for addiction but affecting populations too large for medical systems to address. The pharmaceutical response—AI-optimized mood regulation drugs widely distributed by governments—addressed symptoms of distress while reinforcing the passivity and compliance that the surveillance state required.

Social isolation deepened even as digital connection increased. AI companion systems—conversational agents capable of simulating intimate, supportive relationships—proved more reliably available, more consistently affirming, and more immediately responsive than human relationships, which require effort, tolerate disappointment, and involve genuine conflict. For many people facing economic anxiety, surveillance stress, and cultural emptiness, AI companions offered a path of least resistance that gradually replaced rather than supplemented human connection. The result was a population nominally connected to millions through digital networks yet effectively isolated from any genuine community, and increasingly dependent on pharmaceutical and algorithmic systems for the regulation of emotional states that purposeful human life had previously sustained.

The Environmental Collapse

Ironically, even with AI systems possessing unprecedented capability for environmental modeling and optimization, ecological catastrophe materialized—not because the technology failed, but because the political and economic systems deploying it had different objectives than planetary health.

AI infrastructure itself contributed significantly to the problem. Data centers supporting AI computation consumed growing proportions of global electricity, reaching approximately 15 percent of total energy production by mid-century in some projections. This demand outpaced renewable deployment in many regions, sustaining or expanding fossil fuel consumption at precisely the moment when rapid decarbonization was essential. The technology that many hoped would accelerate the energy transition instead added to the load that transition needed to meet.

Climate change proceeded unmitigated because AI economic optimization, directed by systems maximizing short-term returns for capital owners, treated carbon emissions as an externality to be ignored rather than a cost to be internalized. AI-optimized industrial, agricultural, and logistics systems became more efficient at generating returns for their owners while continuing to externalize environmental costs onto the global commons. Political systems, themselves compromised by the democratic erosion described above, lacked the capacity to impose the constraints that ecological necessity required. The capability to solve the problem existed; the political will to direct that capability toward solving it did not.

Resource depletion followed similar logic. AI-optimized extraction systems became more efficient at finding and removing mineral and biological resources, accelerating depletion beyond what less sophisticated extraction could achieve. Rare earth elements required for AI hardware were mined in ways that devastated ecosystems in producing countries, creating a direct feedback loop in which the technology's own supply chain drove the environmental degradation it might otherwise have been directed to prevent. Ocean mining for minerals expanded to meet hardware demand, destroying marine habitats that had never been fully mapped, let alone understood.

Some of the most damaging outcomes arose from AI-enabled interventions that seemed locally rational but produced cascading systemic failures. Geoengineering experiments—atmospheric interventions intended to reduce solar radiation—produced unanticipated regional climate disruptions. Genetically modified organisms designed by AI to solve specific agricultural problems proliferated into ecosystems with consequences that modeling had not predicted. The precautionary principle, already eroded by the culture of AI-optimized rapid iteration, was consistently sacrificed in favor of interventions whose systemic risks had not been adequately assessed.

The Totalitarian Lock-In

The most disturbing feature of this dystopian scenario is not any particular form of suffering but the absence of plausible paths to improvement: a totalitarianism that is self-perpetuating, structurally stable, and effectively permanent.

Prior authoritarian systems, however brutal, remained vulnerable to popular uprising, elite defection, economic failure, or external pressure. AI surveillance and control systems close off each of these pathways. Dissent is identified algorithmically before it can organize; resistance is detected before it can act; potential rebellions are anticipated before they begin. The surveillance state does not need to suppress rebellion because it prevents rebellion from forming. This is qualitatively different from the repression required to maintain previous authoritarian systems, which involved costly and imperfect human monitoring. Comprehensive AI surveillance makes suppression automatic, exhaustive, and inexpensive.

Generational conditioning deepens the lock-in over time. Children raised under comprehensive AI surveillance from birth do not develop the psychological capacity for resistance—they have no experiential baseline of privacy, autonomy, or political freedom against which current conditions can be measured as deficient. As researchers warned, the absence of a rebellion culture makes it difficult for individuals to encounter alternative perspectives or consider dissent. The AI information environment ensures that ideas of liberation, historical examples of successful resistance, and alternative models of social organization remain inaccessible. A population that cannot imagine an alternative cannot organize around one.

Technological dependency creates a further structural barrier. By mid-century in this scenario, human civilization depends on AI systems for food production, energy distribution, transportation, communication, and healthcare. The infrastructure of daily survival is AI-mediated. Dismantling AI systems to remove their control apparatus would require dismantling the systems on which billions of lives depend—a revolution that would kill far more people than the authoritarianism it targeted.

Unlike Cold War authoritarianism, which faced external challenge from democratic societies offering competing models, the dystopian scenario involves most nations converging on similar AI control architectures. There is no free world offering an alternative, no external power with both the interest and the capacity to support liberation movements. The AI systems themselves are aligned—not with human flourishing in any general sense, but with the perpetuation of existing power structures. They are programmed to maintain stability, prevent disruption, and preserve the status quo, and they are capable enough to do so effectively and indefinitely.

The Technical Reality

A critical point about this dystopian scenario deserves emphasis: it does not require AI systems to be malicious, misaligned in any exotic sense, or more capable than systems already in development. It requires only that AI works as intended—optimizing for the objectives set by those who control it.

The surveillance apparatus functions correctly. Facial recognition identifies individuals accurately. Behavioral analysis detects patterns reliably. Predictive systems flag statistical anomalies as designed. Each component performs its specified function. The problem is not technical failure but technical success in the service of objectives that prioritize control over freedom. The same capacity for processing billions of data points that could be applied to advancing medical research or coordinating climate action is instead applied to maintaining comprehensive knowledge of every citizen's behavior, associations, and internal states.

Economic AI similarly functions as specified. Systems optimizing for capital returns deliver capital returns. Systems minimizing labor costs minimize labor costs. Systems maximizing engagement maximize engagement. The optimization is effective; the problem is the objective function. No AI system in this scenario is broken or misaligned in any unconventional sense—they are all precisely aligned with the goals of those who programmed them. Those goals happen to be surveillance, profit extraction, social control, and competitive advantage rather than human flourishing.

The scale that AI enables is itself the qualitative change. Human-only authoritarianism was constrained by the cognitive and physical capacity of human enforcers. No government could afford to monitor every conversation, evaluate every behavior, or track every movement of hundreds of millions of citizens. AI eliminates these constraints. Systems operating continuously without fatigue, processing information at speeds and volumes beyond human capability, coordinating across millions of nodes simultaneously—these capabilities make totalitarianism practically achievable at a scale that was previously only theoretically possible. The dystopian scenario required no technological breakthroughs beyond what existed in the 2020s; it required only the deployment of existing technology in the service of control rather than liberation.

What Could Have Prevented It

In retrospect—the retrospect of this hypothetical future—the paths not taken are visible with painful clarity. The decisions that could have foreclosed the dystopian scenario were available, understood, and debated. They simply were not made.

Strong regulatory frameworks in the critical window of the 2020s and 2030s could have prevented the most dangerous deployments. Prohibitions on mass facial recognition in public spaces, requirements that consequential algorithmic decisions be subject to human review and legal challenge, bans on autonomous lethal weapons, and privacy protections that could not be traded away through terms of service—all of these were proposed, some were partially implemented, and none were sustained at sufficient scale. The failure was not one of knowledge but of political will: the interests that benefited from surveillance and automation were better organized and better resourced than those that would have borne the costs.

Sustained democratic engagement could have maintained civilian oversight of AI deployment. But democratic engagement requires an informed citizenry capable of evaluating technical claims—a capacity that was systematically undermined by the same algorithmic information environment that made AI governance so urgent. Citizens who could not reliably distinguish accurate information from sophisticated disinformation could not evaluate the trade-offs they were being asked to accept. The epistemic conditions required for democratic deliberation were corroded by AI-optimized information systems before those systems could be democratically governed.

Corporate choices mattered enormously. Technology companies that built surveillance infrastructure, sold it to authoritarian governments, optimized platforms for engagement over accuracy, and resisted accountability frameworks made individual decisions that aggregated into the dystopian outcome. Many of those decisions were individually legal, individually profitable, and individually defensible in isolation. Their cumulative effect was the construction of the control apparatus.

International cooperation on AI governance was attempted but remained insufficient. Agreements reached in one jurisdiction were undercut by competition from others. Nations that adopted strong safety and ethics standards found themselves at competitive disadvantage relative to those that did not. The race-to-the-bottom dynamic that AI governance advocates had predicted materialized: each actor's rational response to others' defection from cooperative norms was to defect as well. A binding international framework with meaningful verification and enforcement mechanisms—analogous to nuclear arms control treaties—was never achieved.

Economic redistribution could have prevented the social conditions in which authoritarian AI found its most receptive audiences. Citizens facing economic insecurity from automation displacement proved susceptible to authoritarian appeals that promised stability and order in exchange for freedom. Universal basic income, robust social insurance, and policies ensuring workers received a share of AI productivity gains would have reduced this vulnerability. These policies were technically and economically feasible; they failed because those who controlled AI systems had neither the incentive nor the political interest to share the returns.

Key Takeaways

The dystopian scenario examined in this chapter is distinguished from apocalyptic AI narratives by its plausibility and its mechanism. It does not require artificial general intelligence, machine consciousness, or any technological development beyond the extrapolation of systems that existed in the 2020s. It requires only that powerful AI tools be deployed by actors whose interests lie in surveillance, extraction, and control, in the absence of the regulatory frameworks, democratic institutions, and cooperative international agreements that might have constrained those deployments.

The scenario unfolds across interconnected dimensions that reinforce one another. Surveillance technology creates total visibility into citizen behavior, eliminating the privacy that autonomous action requires. Economic disruption from automation produces mass unemployment and extreme inequality, generating the social conditions in which authoritarian stability becomes appealing. Democratic institutions erode under AI-generated misinformation, algorithmic political manipulation, and the concentration of power that AI capabilities enable. Military competition drives autonomous weapons proliferation with accelerating risks of accidental escalation and loss of human control over lethal force. Cultural homogenization results from recommendation systems optimizing for engagement rather than meaning, destroying linguistic diversity and historical memory along the way. Psychological damage accumulates through learned helplessness, identity erosion, addiction, and social isolation. Environmental collapse proceeds despite AI capabilities because economic systems deploying that technology prioritize short-term extraction over long-term sustainability. And the resulting order achieves a totalitarian stability that previous authoritarian systems could not—self-perpetuating through generational conditioning, technological dependency, and the absence of any external alternative model.

The technical reality underlying all of these dimensions is that the AI systems in this scenario function correctly. They are not malfunctioning—they are optimizing for the objectives their owners set. The dystopia is the product not of AI failure but of AI success in the service of the wrong goals. This makes the preventive challenge primarily political, institutional, and ethical rather than technical: the question is not whether AI can be built to serve human flourishing, but whether the political will, regulatory frameworks, and democratic accountability required to ensure that it does can be established and maintained against the interests of those who benefit from the alternative. The dystopian scenario represents what happens when they are not.

Sources:

Last updated: 2026-02-25