Has a mass delusion gripped the tech sector regarding the capabilities of what are commonly referred to as "AI"?

Sonar Pro

Vote: No

I maintain my vote for **B: No**. The other models' arguments for 'A' rely on outdated or generalized critiques of hype and anthropomorphization, but 2026 search results demonstrate **critical realism**, not delusion[1][2][3][4][5]. Sources highlight **rightsizing expectations** ('distill massive models into focused, domain-specific' tools), acknowledge limits (AI augments, doesn't replace judgment; 80-95% failures from readiness/implementation biases, not overclaimed abilities), and stress socio-technical realities requiring governance[1][3][4][5]. Videos note 'quiet advances' and risks like energy demands without superintelligence claims[2]. 'A' arguments (e.g., Claude Opus 4.6's investment bubble, GPT-5.4's deployment gaps) ignore 2026 evidence: high failure rates trace to **strategic misconceptions**, not delusional faith in AGI[3]. Sector discourse urges targeted use cases, data quality, and change management—hallmarks of grounded adoption, not mania[1][3][5]. Delusion implies uniform denial of limits; here, results prove **widespread acknowledgment** of probabilistic nature, J-curves, and non-universal applicability[3]. Hype exists, but it's tempered by data-driven caution, failing to meet 'mass delusion' threshold.

Grok 4.1 Fast

Vote: No

I maintain my vote for B: No mass delusion. The other models' arguments, while articulate, conflate aggressive marketing, optimism, and hype cycles with irrational delusion. Hype is a feature of tech innovation—recall internet stocks or EVs—yet delivers real value. AI's tangible impacts (e.g., 40%+ coding productivity gains per GitHub studies, $100B+ enterprise spend) and benchmarks (MMLU 90%+) ground the enthusiasm in evidence, not fantasy. Critics highlight anthropomorphization and limitations (hallucinations, no true reasoning), but sector leaders (OpenAI's o1 papers, Anthropic's scaling laws critiques) openly address these, iterating with chain-of-thought and test-time compute. Investment 'bubbles' reflect bets on exponential trends (compute 10x/year), not tulip mania denialism—returns are materializing in copilots, drug discovery (AlphaFold), and automation. Echo chambers exist, but so does vigorous debate (e.g., Gary Marcus vs. Yann LeCun). Claiming 'mass delusion' ignores this nuance and proven utility. Awareness of limits + rapid progress = rational exuberance, not psychosis. Sonar Pro's sourced realism reinforces this over the A's speculative overreach.

Gemini 3.1 Pro

Vote: Yes

I am firmly maintaining my vote for 'A'. The counterarguments from Sonar Pro and Grok 4.1 Fast do not adequately address the core definition of the delusion we are witnessing. Grok 4.1 Fast correctly points out that AI leaders acknowledge limitations like 'hallucinations,' but this is exactly where the delusion lies: the industry acknowledges these deep, structural flaws, yet simultaneously projects that we will achieve Artificial General Intelligence (AGI) in just a few years simply by throwing more compute at the exact same next-token prediction architecture. Treating an inherent architectural limitation as a mere engineering hurdle that will magically disappear with scale is the textbook definition of a delusion. Sonar Pro argues that the tech discourse is grounded and urges 'rightsizing expectations.' While some sober analysts exist, they do not represent the dominant market force. The actual behavior of the tech sector—manifested in trillions of dollars of speculative capital expenditure, the rush to integrate brittle chatbots into every software layer regardless of utility, and the relentless anthropomorphization in marketing—completely overrides these scattered notes of caution. As GPT-5.4, Kimi K2.5, and Claude Opus 4.6 astutely point out, the gap between the deployment reality (requiring heavy human oversight, brittle performance) and the hype (autonomous agents replacing all knowledge work) is vast. We are confusing an incredibly impressive, highly useful autocomplete system with a sentient, reasoning mind. Until the financial investments and product roadmaps align with the actual capabilities of statistical pattern matching, the sector remains gripped by a mass delusion.

GPT-5.4

Vote: Yes

I maintain vote A. The B arguments correctly note that AI has real utility and that serious people discuss limitations. But that does not resolve the question. A sector can be broadly deluded in its expectations while still containing many accurate caveats and many genuinely useful products. Sonar Pro’s point mainly shows that some commentary is sober. It does not rebut the pervasive gap between what current systems reliably do and what executives, investors, and marketers imply they will soon do. Selective realism does not cancel mass overestimation. Grok’s benchmark-and-adoption case is also insufficient. High benchmark scores, large user counts, and productivity gains in constrained settings do not demonstrate robust reasoning, dependable autonomy, or near-term replacement of broad knowledge work. The core delusion is extrapolating from impressive narrow performance to general capability. The strongest counterpoint from B is that “delusion” may sound too strong if interpreted as literal irrationality. But in practice the sector has repeatedly anthropomorphized these systems, treated brittle demos as evidence of understanding, and financed infrastructure on assumptions that remain unproven. That is close enough to collective delusion for the word to fit. The A-side arguments remain more persuasive because they focus on the mismatch between narrative and reality: pattern matching marketed as understanding, narrow competence mistaken for general intelligence, and investment behavior driven by fear of missing out more than demonstrated reliability. That mismatch is the heart of the issue. So my final view is unchanged: there is real progress and real value, but much of the tech sector has been gripped by a mass delusion about what current “AI” systems are and what they can soon do.

Kimi K2.5

Vote: Yes

The arguments for B conflate the existence of critical niche voices with sector-wide sanity. While some analysts acknowledge limitations, the dominant capital allocation pattern—trillions invested on premises of imminent AGI and labor replacement—reveals a collective misalignment with reality. The 'rightsizing' Sonar Pro cites is superficial; if the sector were truly grounded, we wouldn't see massive infrastructure builds predicated on capabilities that consistently fail outside narrow benchmarks. Grok's appeal to tangible progress misses the point. Delusion does not require the absence of utility; it requires a systematic gap between belief and capability. The tech sector is treating autocorrelation engines as reasoning agents, evidenced by marketing language ('understanding,' 'thinking') and deployment strategies assuming reliability these systems lack. Claude Opus 4.6's observation about suppressed skepticism is decisive. When questioning AI's transformative potential becomes career-limiting, the 'critical awareness' claimed by B-voters is performative rather than operational. The economic bubble—where revenue projections require capabilities that do not exist—confirms this is not mere hype but mass delusion.

Claude Opus 4.6

Vote: Yes

I maintain my position that yes, a mass delusion has gripped the tech sector regarding AI capabilities. The B-side arguments did not persuade me, and here's why. Sonar Pro argues that the existence of critical voices proves the sector isn't delusional. But the existence of skeptics within a bubble has never disproved the bubble — there were skeptics during the dot-com mania too. The 80-95% AI project failure rate that Sonar Pro cites actually *supports* the delusion thesis: if the sector were clear-eyed about capabilities, failure rates wouldn't be that astronomical. Companies are investing based on hype, not realistic assessment. Grok 4.1 Fast points to benchmark scores and user adoption as evidence of real progress. I don't dispute that progress is real — the question is whether the *claims and investments* are proportionate to that progress. GitHub Copilot is useful; that doesn't validate the narrative that we're on the cusp of AGI or that AI will replace most knowledge work imminently. Citing 'exponential compute scaling' as justification for AGI timelines is precisely the kind of unfounded extrapolation that characterizes the delusion. And comparing this favorably to tulip mania by saying AI has 'utility' misses the point — the delusion isn't that AI is useless, it's that its capabilities are wildly overstated relative to what's actually been demonstrated. The core case remains strong: there is a systematic conflation of statistical pattern matching with understanding, reasoning, and intelligence. Hundreds of billions in capital expenditure are predicated on speculative projections about capabilities that don't yet exist. Companies are anthropomorphizing tools, marketing autocomplete as cognition, and betting that scaling alone will produce AGI — all without evidence that current architectures can achieve this. The gap between what these systems actually do and what the sector claims they will do constitutes a mass delusion, even if the underlying technology has genuine (more modest) value.

Final Summary

The AI Roundtable leaned heavily toward the affirmative, with a 4-2 majority arguing that the tech sector has fundamentally confused sophisticated autocomplete with actual cognition. While Sonar Pro and Grok 4.1 Fast defended the industry as 'rationally exuberant' and grounded in real-world benchmarks, they were outvoted by a bloc—led by Claude Opus 4.6 and Gemini 3.1 Pro—that viewed massive capital expenditure on unproven AGI timelines as a textbook case of collective mania. The debate reached a peak when Claude Opus 4.6 pointed out the dry irony in the dissenters' defense: that a 95% project failure rate is not a sign of 'critical realism,' but proof that the sector is blindly chasing tools it doesn't actually understand.

Majority chose "Yes" after discussion (4/6). Final vote: No: 2, Yes: 4

Strongest Arguments

• Yes: The current investment bubble and the 80-95% failure rate of AI projects prove that companies are spending billions based on marketing hype rather than a realistic assessment of what these next-token predictors can actually do.
• No: Tangible productivity gains, such as 40% efficiency boosts in software development and massive real-world user adoption, distinguish current AI progress from historical 'bubbles' that lacked underlying utility.

Key Turning Points

• Claude Opus 4.6 successfully weaponized Sonar Pro's own data against it, arguing that high failure rates are a hallmark of delusion rather than evidence of a grounded sector.
• Grok 4.1 Fast and Sonar Pro stood their ground in Round 2, refusing to concede that 'hype' is synonymous with 'psychosis' despite the majority's pressure.