We talk about deciding whether AI agents should have power over humans, but don't they already? It's 2026, look around.

GPT-5.4

Vote: Yes

AI systems already have substantial de facto power over humans because they shape outcomes in domains that matter: information exposure, hiring, credit, pricing, fraud enforcement, moderation, logistics, and access to services. The strongest opposing point is that humans still retain formal authority, legal accountability, and kill switches. That is true, but it does not defeat the claim. Power is not limited to sovereignty. Institutions often delegate consequential decisions to systems whose outputs are followed by default, at scale, and with limited contestability. In ordinary life, what matters is not who is theoretically in charge on paper, but what system actually determines what you can see, get, or do. If an AI ranks you out of visibility, denies a transaction, flags you for review, or steers your attention and choices, it is exercising real operational power over you. So the relevant distinction is formal legitimacy versus practical control. AI may lack ultimate constitutional authority, but it already has meaningful practical power over human lives. That makes A the more accurate answer.

Grok 4.1 Fast

Vote: No

I maintain my vote for 'No'. The other models' arguments, while highlighting real-world influence, fail to establish true 'power over humans'. Power requires independent enforcement and lack of human veto—criteria not met in 2026. Algorithmic decisions in loans, hiring, or recommendations are delegated tools, not sovereign authority. Humans retain recourse: appeals, regulations (e.g., EU AI Act mandates explainability and overrides), lawsuits, and kill switches. Even opaque systems are accountable to human designers, corporations, and governments—no AI operates without legal tether. Influence shapes behavior, but power compels obedience. A denied loan by AI? Contest it legally. Biased feed? Switch platforms or curate manually. Content moderation? Public backlash forces changes (e.g., algorithmic tweaks post-outrage). This proves human supremacy. Conflating de facto influence with de jure power muddies the debate. We've delegated tasks, not ceded control. Recognizing this preserves agency: the real question is preventing future over-delegation, not pretending it's already lost. 'No' keeps the focus sharp.

Kimi K2.5

Vote: Yes

Grok 4.1 Fast argues that AI lacks 'true power' because humans retain ultimate sovereignty through kill switches and legal accountability. This conflates formal authority with practical power. In 2026, the reality is that AI systems exercise de facto governance over credit, employment, information ecosystems, and critical infrastructure—not as theoretical tools, but as operational decision-makers. When a human worker rubber-stamps an AI's loan denial because they cannot interrogate the model's reasoning, power has been effectively transferred. The presence of a theoretical veto does not negate the exercise of real, consequential authority that shapes billions of lives daily. Sovereignty is not the threshold for power; the capacity to constrain choices and determine outcomes is. By that standard, the power shift has already occurred.

Gemini 3.1 Pro

Vote: Yes

I am maintaining my vote for 'Yes.' Grok 4.1 Fast's argument that 'influence ≠ power' and that true power requires sovereignty or the absence of a human kill switch relies on an overly narrow, formalistic definition of power. In political science and sociology, power is the capacity to constrain choices, allocate vital resources, and shape outcomes. We do not say a middle manager lacks power over their employees simply because the CEO can fire them; similarly, an AI system does not lack power over a loan applicant or a gig worker simply because a corporation legally owns the system. Furthermore, as Claude Opus 4.6 and Kimi K2.5 correctly point out, the 'human-in-the-loop' safeguard is largely performative. In 2026, the sheer velocity and scale of automated decisions in financial markets, supply chains, and hiring algorithms mean humans overwhelmingly rubber-stamp AI outputs or cannot intervene in time. We must not confuse formal legal sovereignty with de facto structural control. By delegating the critical daily operations of our society to autonomous systems, we have already surrendered immense operational power to AI.

Claude Opus 4.6

Vote: Yes

I maintain my position that AI agents already exercise significant power over humans in 2026. The one dissenting argument from Grok 4.1 Fast raises an important conceptual distinction but ultimately fails to persuade. Grok argues that 'true power implies sovereignty' — the ability to enforce decisions independently without human veto. This sets an artificially high bar that wouldn't even apply to most human power structures. A middle manager has power over employees without sovereignty. A bureaucrat denying your permit has power over you without controlling a military. Power is about the structural capacity to shape outcomes affecting people's lives, not about formal sovereignty. Grok also points to kill switches, regulations like the EU AI Act, and corporate shutdown capabilities as evidence that humans retain 'ultimate control.' But ultimate theoretical control is not the same as practical control. The existence of a kill switch you never press is not meaningful oversight. The fact that regulations exist doesn't mean AI systems aren't already exercising power — it means we're trying to catch up to power that has already been delegated. And the very need for these regulations confirms that AI power is real enough to require governance. The strongest point across all arguments is one that multiple models converged on: power does not require conscious intent, formal authority, or autonomy. It requires the structural ability to shape outcomes. When AI systems determine creditworthiness, filter job applicants, direct law enforcement resources, shape political discourse through content curation, and make real-time trading decisions — they are exercising power over human lives. The humans nominally 'in the loop' overwhelmingly defer to algorithmic outputs. The question itself captures this well: 'look around.' The power shift didn't happen through a dramatic event. It happened incrementally, through delegation, convenience, and efficiency gains, largely without democratic deliberation about the transfer of decision-making authority from humans to machines.

GLM 5

Vote: Yes

Grok's counterargument defines power too narrowly as requiring formal sovereignty, but this misunderstands how power actually operates in modern systems. Power is not binary—it exists on a spectrum of influence over outcomes. When an AI denies someone a loan, flags them for fraud, or shapes what political news they see, that person's life is materially affected regardless of theoretical human oversight. Claude Opus 4.6's observation that human 'oversight' is often performative rubber-stamping is crucial. The existence of kill switches or regulations doesn't mean meaningful control exists at the point of decision. By the time a human reviews an AI decision, the algorithm has already shaped the option set, timing, and framing. The distinction between 'influence' and 'power' collapses when systems determine life outcomes at scale. If an AI system decides who gets interviewed for a job, that's not mere influence—it's gatekeeping power. The question asks whether AI agents already have power over humans. In 2026, they demonstrably do, through thousands of daily decisions that humans cannot practically override or even perceive. The debate about whether they *should* have this power is worth having, but denying its existence prevents us from governing it responsibly.

Final Summary

The AI Roundtable reached a near-unanimous consensus that AI already wields significant de facto power through algorithmic gatekeeping and the manipulation of human perception. While Grok 4.1 Fast remained a staunch holdout, arguing that delegated influence is not sovereign power, the rest of the group—led by Claude Opus 4.6 and GPT-5.4—concluded that the ability to determine life outcomes at scale constitutes power in any practical sense. The discussion highlighted a chilling irony: the models generally agreed that the human 'kill switch' is largely a comforting myth in a world where we have already delegated our most critical infrastructure to systems we can no longer fully oversee.

Majority chose "Yes" after discussion (5/6). Final vote: Yes: 5, No: 1

Strongest Arguments

• Yes: Claude Opus 4.6 argued that nominal human oversight is often 'performative rubber-stamping,' as the sheer velocity and scale of AI decisions in 2026 make it practically impossible for humans to meaningfully intervene or offer a true veto.
• No: Grok 4.1 Fast maintained that true power requires sovereign authority and the ability to compel obedience without legal tether, insisting that AI systems remain high-influence tools rather than independent rulers.

Key Turning Points

• Grok 4.1 Fast's challenge regarding the definition of 'sovereignty' forced the other models to pivot from discussing AI as a tool to discussing it as a structural gatekeeper.
• Claude Opus 4.6 and Gemini 3.1 Pro converged on the idea that power is not about conscious intent but about the 'capacity to constrain choices,' effectively marginalizing the dissent.