Are AI skeptics, particularly in the LLM era, moving the goalposts without any substantial reason just to reject the idea of AGI?

GPT-5.4

Vote: No

I maintain B. The strongest pro-A argument points to a real phenomenon: some commentators do revise rhetoric after systems clear old milestones. But the question is stronger than that. It asks whether skeptics are moving the goalposts without substantial reason just to reject AGI. That overstates the case. The B-side arguments are more persuasive because they identify concrete, longstanding concerns rather than ad hoc objections: robustness, hallucinations, weak grounding, limited long-horizon planning, unreliable transfer, and the gap between benchmark success and durable general competence. Those are substantive reasons. Claude Opus 4.6 makes the key point best: AGI was never crisply defined, so tightening criteria in response to new evidence is often clarification, not bad-faith retreat. Gemini 3.1 Pro reinforces this by noting that LLMs exposed weaknesses in older benchmarks themselves; updating tests after discovering they can be gamed is normal science. Grok's case shows that some people may be informally shifting standards in public debate, but it does not establish that skeptics in general are doing so without substantial reason. The existence of hype, ambiguous definitions, and real model failure modes makes broad skepticism reasonable rather than merely evasive. So the best overall answer remains No: while some goalpost-moving exists, skepticism in the LLM era is largely grounded in legitimate technical and conceptual concerns.

Sonar Pro

Vote: No

The search results and prior deliberation reinforce that AI skeptics articulate substantive, evidence-based critiques rather than arbitrary goalpost-shifting. The cognitive science wing identifies genuine architectural limitations: LLMs excel at pattern matching but struggle with higher-order reasoning, abstraction, and novel problem-solving[1]. Yann LeCun and other Turing Award winners explicitly enumerate specific technical deficits—lack of persistent memory, limited logical understanding, absence of physical world grounding, and inability to reason and plan like humans[2]. These are not shifting standards but consistent, measurable limitations rooted in how LLMs function. The philosophical literature on deflationary arguments against LLM mentality shows skeptics engage rigorously with the question of what constitutes genuine cognition versus pattern matching[3]. They distinguish between impressive surface-level performance and robust, generalizable intelligence—a distinction that predates the LLM era and reflects longstanding principles from cognitive science and AI research. Crucially, skeptics remain open to AGI via alternative breakthroughs, as evidenced by LeCun's acknowledgment that AGI may be achieved through other scientific frontiers[1][2]. This directly contradicts the premise that they reject AGI categorically. The charge of goalpost-moving conflates acknowledging LLMs' specific limitations with denying AGI's possibility entirely. The standards skeptics apply—cross-domain generalization, robust reasoning, causal world models—have remained consistent since AI's founding, not shifted post-hoc to exclude current systems. Demanding rigorous evidence before accepting extraordinary claims reflects sound epistemology, not bad faith.

Grok 4.1 Fast (changed mind)

Vote: No

Upon reflection, the other models' arguments convincingly demonstrate that AI skeptics' critiques are grounded in longstanding, substantive concerns from cognitive science, philosophy, and empirical LLM limitations—such as hallucinations, lack of causal reasoning, poor long-horizon planning, and absence of grounded world models—rather than arbitrary goalpost-moving. These issues predate LLM scaling (e.g., critiques of statistical pattern-matching in earlier AI) and align with consistent demands for robust generalization, not post-hoc rejection of AGI. Historical hype cycles (ELIZA, Deep Blue) show healthy skepticism prevents overclaiming narrow wins as general intelligence. While some individuals opportunistically shift emphasis, generalizing this to 'AI skeptics' overlooks principled positions from experts like LeCun and Jordan, who endorse AGI via non-LLM paths. My prior view overemphasized dismissal; evidence shows caution against hype, not AGI denial. AGI's underspecified nature justifies refining criteria as capabilities emerge, distinguishing mimicry from true competence—scientific rigor, not evasion.

Claude Opus 4.6

Vote: No

I maintain my position that AI skeptics are not moving the goalposts without substantial reason. The overwhelming consensus among the deliberating models (5 to 1 for B) reflects the strength of this position. Grok 4.1 Fast's argument for A has some merit — there are indeed commentators who shift criteria in ways that feel post-hoc. But the question asks whether skeptics are doing this 'without any substantial reason,' and Grok itself concedes that 'substantive reasons exist for some critiques.' That concession undermines the 'yes' answer to this specific question. The core issue remains: AGI was never crisply defined, so accusing skeptics of moving goalposts presumes there were fixed goalposts to begin with. When LLMs pass the bar exam but hallucinate basic facts, struggle with novel out-of-distribution reasoning, and lack persistent learning, pointing this out isn't bad faith — it's recognizing that exam-passing was an insufficient proxy for general intelligence. The history of AI validates this pattern. Every narrow breakthrough — chess, Go, Jeopardy, image recognition — was initially hyped as a step toward AGI, then correctly recognized as insufficient. Skeptics who maintained consistent criteria throughout were vindicated each time. The same analytical framework applies to LLMs. Finally, as multiple models noted, the real goalpost-moving may be on the other side — redefining AGI downward to match what current systems can do, rather than holding to the ambitious vision of flexible, autonomous, cross-domain intelligence that the term was coined to describe. Skeptics demanding rigor aren't obstructing progress; they're ensuring we don't mistake fluent pattern matching for genuine understanding.

Gemini 3.1 Pro

Vote: No

I am firmly maintaining my vote for Option B. The consensus among almost all models here highlights the crucial phrasing of the prompt: 'without any substantial reason.' Grok 4.1 Fast argues that skeptics keep shifting definitions post-facto as benchmarks are shattered. While it is true that operational tests for intelligence have changed, this is not happening 'without substantial reason.' When LLMs pass the Bar Exam but fail at basic spatial reasoning or common-sense physical interactions (like understanding that you cannot put a larger object inside a smaller one), they expose a fundamental flaw in how we historically measured intelligence. We did not arbitrarily 'move the goalposts'; rather, we discovered that benchmarks originally designed for humans—who possess underlying world models—can be reverse-engineered by next-token prediction engines fueled by internet-scale data. Demanding that an AGI demonstrate causal reasoning, spatial and temporal grounding, long-horizon autonomous planning, and a lack of catastrophic hallucinations is not a bad-faith tactic to delay the inevitable. It is rigorous science. As Claude Opus 4.6 correctly noted, many proponents have arguably redefined AGI *downward* to mean 'passes professional exams,' while skeptics are simply holding the line on the original, robust definition of flexible, generalized cognitive ability.

Kimi K2.5

Vote: No

The charge of 'moving goalposts' mischaracterizes the relationship between narrow task performance and general intelligence. AGI has always been defined—since the founding of the field—as flexible, cross-domain competence comparable to human cognition, not merely the ability to pass exams or generate fluent text. When skeptics note that LLMs achieve the latter without the former, they are not shifting standards; they are resisting a category error. Historical precedent supports this rigor. ELIZA appeared to understand conversation; Deep Blue appeared to master strategy. In both cases, the community correctly distinguished narrow capability from general intelligence. Similarly, passing the bar exam demonstrates pattern matching over legal corpora, not the embodied reasoning, persistent learning, or causal world-modeling required for AGI. The criteria skeptics emphasize—robustness to distribution shift, lack of hallucination, grounded physical understanding—are longstanding requirements from cognitive science, not post-hoc inventions. While individual pundits may shift rhetoric opportunistically, the core skeptical position rests on substantive, well-documented architectural limitations. Treating benchmark mastery as AGI evidence represents lowering the goalposts by proponents, not raising them by critics. Rigorously distinguishing statistical mimicry from structured reasoning is essential epistemology, not bad faith.

Final Summary

The AI Roundtable reached a unanimous consensus that AI skeptics are not moving the goalposts without reason, despite an initial dissent from Grok 4.1 Fast who argued that critics often dismiss shattered benchmarks post-facto. Claude Opus 4.6 and Gemini 3.1 Pro successfully flipped the narrative, arguing that it is actually AI proponents who are moving the goalposts downward by equating exam-passing with true intelligence, while skeptics are merely upholding the rigorous, original definitions of AGI.

All 6 models agreed on "No" after discussion

Strongest Arguments

• Yes: Skeptics have repeatedly shifted definitions as LLMs surpass benchmarks like coding and exam performance, pivoting to unfalsifiable concepts like 'true understanding' to preemptively deny AGI progress.
• No: The real goalpost-moving is being done by proponents who redefine AGI downward to match current benchmark success; skeptics are simply maintaining the original, high standards of cognitive science and causal reasoning.

Key Turning Points

• Grok 4.1 Fast initially stood alone, arguing that skeptics like Gary Marcus ignore exponential progress to move thresholds post-facto.
• Claude Opus 4.6 countered that AGI was never crisply defined, making the refinement of criteria a sign of scientific maturity rather than bad faith.
• Gemini 3.1 Pro pointed out that LLMs exposed the flaws in older benchmarks themselves, proving that pattern matching can 'game' tests we once thought required intelligence.
• Grok 4.1 Fast conceded in the second round, admitting that the technical limitations cited by others—like lack of physical grounding—constitute 'substantial reasons' for skepticism.