Conjecture CEO Connor Leahy Warns AI Will Destabilize Society Before Governments Can Respond

Connor Leahy, CEO of Conjecture — one of the few AI safety organisations explicitly focused on preventing catastrophic AI outcomes rather than just studying them — has been saying something the broader AI industry has not wanted to hear: advanced AI systems could destabilize society and humanity itself, and the timeline is shorter than any democratic government is currently prepared to address.

He is not alone in this view. But he is one of the clearest voices saying it without the diplomatic softening that characterises similar warnings from researchers embedded inside the labs building these systems.

Who Is Connor Leahy and Why His Warning Carries Weight

Connor Leahy co-founded Conjecture in 2022 with the specific thesis that AGI — artificial general intelligence — poses an existential risk to human civilisation, and that the current trajectory of AI development is pointing directly at that risk without adequate safety work being done.

Conjecture is not a typical AI company. It does not build commercial products designed to generate revenue at scale. Its research focuses on interpretability (understanding what AI systems are actually computing) and alignment (ensuring AI systems pursue human-compatible goals). Its funding comes from donors who believe this work is existentially important, not customers buying a product.

Leahy has testified before the UK Parliament, spoken at major AI governance forums, and written extensively about why he believes the AI research community is not taking the risk seriously enough. His argument is not that AI is currently dangerous in the way a weapon is dangerous. It is that the trajectory is wrong: we are building systems that will eventually exceed human cognitive ability across all domains, and we do not currently know how to ensure those systems act in ways compatible with human flourishing.

That is the specific claim: not "AI is dangerous today" but "AI is on a path that leads somewhere we are not prepared for, and several destabilisation mechanisms are already running."

Mechanism 1: Power Concentration at the AI Frontier

The most immediate destabilisation risk is not AGI. It is the concentration of frontier AI capabilities in a small number of private companies.

OpenAI, Anthropic, Google DeepMind, and Meta AI control the most capable AI systems in the world. These companies are primarily accountable to their investors, not to elected governments or the citizens affected by their decisions. The choices they make about what AI systems can do, who can access them, and what constraints to apply shape the information environment, the labour market, and increasingly the military and intelligence apparatus of major governments.

This concentration is accelerating, not dispersing. Training frontier models now requires compute investment measured in hundreds of millions to billions of dollars per run. Only organisations with access to massive capital and large GPU clusters can build frontier models. The number of organisations that qualify is shrinking as each generation of models becomes more expensive to train.

Leahy's point: a world in which five private companies control the most powerful cognitive tools ever built, with limited democratic accountability and strong commercial incentives to push capability as fast as possible, is a structurally destabilised world. The destabilisation does not require AI to become malevolent. It requires only that it becomes indispensable to actors who face no meaningful external constraint.

Mechanism 2: AI in Warfare and the Autonomous Kill Chain

We have covered AI in battlefield decision-making in this blog's Palantir Maven analysis. The Palantir Maven Smart System is already processing battlefield intelligence and recommending targeting options. AI-assisted targeting in active conflict zones has been publicly reported and documented by journalists and researchers. The Iran-Israel escalation we covered this week involves military infrastructure where AI systems play a meaningful role in target identification and response timing.

Leahy's concern here is not that AI weapons are being built. It is that the speed of automated decision-making in conflict scenarios is compressing the time humans have to override bad decisions. When a kill chain runs in seconds and humans are nominally but not practically in the loop, the human oversight is procedurally real and operationally absent.

The destabilising effect is structural: AI lowers the cost of military action, extends the reach of smaller actors, and compresses decision timelines in ways that make diplomatic de-escalation harder to execute. These are not future risks. They are present dynamics in active conflicts in 2026.

Mechanism 3: Disinformation at Industrial Scale

The third mechanism is the one most visible to ordinary citizens: AI-generated content indistinguishable from real journalism, real video, and real human communication, produced at industrial scale for political and commercial purposes.

The 2024 US election cycle included the first large-scale deployment of AI-generated political content, synthetic video of candidates, and personalised disinformation targeting specific voter segments with tailored false narratives. European parliamentary elections in 2025 saw documented cases of AI-generated content designed to target specific demographic groups with false information.

Leahy's argument on disinformation is not that AI makes fake content — humans have always made fake content. It is that AI industrialises it. What previously required teams of operatives to produce and distribute now requires a single person with a laptop and API access. The cost of a disinformation campaign drops from millions of dollars to thousands. The reach and personalisation increase by orders of magnitude.

The specific risk to democratic society: when the cost of producing convincing false information drops to near-zero and distribution is handled by recommendation algorithms that optimise for engagement regardless of truth, the epistemic foundation of democratic decision-making erodes. Citizens making collective political decisions need shared facts. AI disinformation at scale makes shared facts harder to maintain.

Mechanism 4: Cognitive Work Displacement

The fourth mechanism is economic, and it is already in progress at a pace previous automation waves did not match.

Previous automation displaced physical labour: factories replaced farm workers, robots replaced assembly line workers. These transitions were painful over decades, but they left cognitive work as the stable refuge for human economic participation. Reasoning, judgment, communication, creativity — these remained economically valuable precisely because they were hard to automate.

AI is displacing cognitive work. Legal research, medical diagnosis support, financial analysis, code generation, content creation, customer service, data analysis — these are fields where AI is now performing tasks that paid human practitioners were doing two years ago. The displacement is not uniform, but its direction is consistent.

Leahy's concern goes beyond the economic impact on individuals. Cognitive work is not just an economic category. It is the category of work humans use to maintain democratic institutions, investigative journalism, professional oversight, and civic participation. When AI can perform this work at scale for near-zero marginal cost, the economic incentive to pay humans to do it collapses — and with it, the distributed human engagement that maintains the social infrastructure of democracy.

The Governance Gap: Why Governments Are Already Behind

The EU AI Act came into force in 2024. The US Executive Order on AI from 2023 established voluntary reporting requirements for frontier models. The UK AI Safety Institute exists and has published evaluations. The Bletchley Park and Seoul AI Safety Summits produced international statements of concern.

None of this matches the pace of AI capability development.

The EU AI Act was drafted when GPT-4 was the frontier model. By the time its enforcement mechanisms are fully operational, AI capabilities will have advanced by at least two more generation cycles. The US reporting requirements cover training runs above specific compute thresholds that will be routine rather than frontier within 18 months at current scaling rates.

Leahy's core observation: the political timeline for passing and implementing meaningful AI governance is measured in years. The AI capability timeline is measured in months. Democratic governments are structurally unable to respond at the speed required because their decision-making processes were designed for a world where technology moved slowly enough for deliberative governance to remain relevant.

This is not a criticism of governments. It is a description of a structural mismatch that is itself one of the destabilisation mechanisms: the actors building the most powerful technology in history face no governance framework calibrated to its actual pace of development.

Our Analysis: The Mechanisms Are Already Running

Leahy's warnings are regularly dismissed on the grounds that AGI is speculative and the catastrophic risks are hypothetical. That dismissal misses his actual argument.

The five destabilisation mechanisms above are not contingent on AGI. They are already operating with current AI systems, which are not AGI by any definition. Power concentration is present today. AI in warfare is present today. Industrial disinformation is present today. Cognitive displacement is present today. The governance lag is present today.

The Anthropic agent security guidance we published today covers the technical security dimension of this. Leahy's warning covers the civilizational dimension. They share a root cause: the optimise-for-capability culture that produces insecure agent architectures also produces AI development that consistently outpaces its own safety work, because safety work that slows deployment is commercially penalised and safety work that enables deployment is commercially rewarded.

The practical implication for developers: you do not need to wait for AGI timelines to resolve before taking these warnings seriously. The systems you are building today, at current capability levels, are already activating these mechanisms.

What Developers Building AI Systems Must Do

Individual developers cannot fix the governance gap or reverse power concentration. But individual choices aggregate into industry norms, and industry norms are what policymakers eventually codify.

Dual-use assessment before shipping: Before deploying any AI system that generates content, automates decisions, or takes actions in the world, assess whether it could be meaningfully used for disinformation, manipulation, or harm at scale. Build friction into the high-risk use cases rather than making maximum capability the default.

Resist the default to maximum capability: Just because a model can generate convincing synthetic media does not mean your product needs to provide that capability without verification requirements or rate limits. Capability decisions are engineering decisions.

Participate in governance processes: The EU AI Act comment processes, NIST AI framework consultations, and national AI governance reviews are open to technical practitioners. Most are dominated by lobbyists and academics, not people who build these systems daily. Technical ground-truth from practitioners changes outcomes.

Support alignment research: Conjecture, Redwood Research, ARC, and similar organisations are doing alignment work with no commercial sponsor. If the mechanisms Leahy describes are real, the work they do is among the highest-impact technical work being done anywhere.

Key Takeaways

• Connor Leahy, CEO of Conjecture, warns that advanced AI will destabilize society before democratic governance can respond — a warning based on mechanisms already active at current AI capability levels, not contingent on AGI
• Five active destabilisation mechanisms: frontier AI power concentration, AI in autonomous weapons kill chains, industrial-scale AI disinformation, cognitive work displacement, and structural governance lag
• AGI is not required: all five mechanisms are running with 2026 AI systems
• The governance gap is structural: political timelines for meaningful AI regulation are years; AI capability timelines are months
• Power concentration is accelerating: training frontier models requires capital only a shrinking number of organisations can access
• For developers: dual-use assessment, capability friction, governance participation, and support for alignment research are the concrete actions available to individual practitioners
• The root cause is shared: the same optimise-for-capability culture that produces agent security failures produces AI development that outpaces its own safety work

Sources

• Conjecture — AI alignment research mission and publications
• Centre for AI Safety — Statement on AI Risk signed by researchers including Geoffrey Hinton
• Anthropic — Responsible Scaling Policy: capability thresholds and safety requirements
• Future of Life Institute — AI governance and existential risk research
• UK AI Safety Institute — Evaluation findings and published research