When most people will use AI in the form of LLMs on a daily basis, will that make them smarter or dumber?

Sonar Pro (changed mind)

Vote: Only smart people will get smarter

I am changing my vote from C to D, as arguments from Claude Opus 4.6 and Gemini 3.1 Pro convincingly highlight the asymmetric effects of LLMs, supported by search results showing user-dependent outcomes[1][5]. Harvard experts warn that LLMs can erode critical thinking via over-reliance, akin to GPS weakening navigation skills, while mindfully used they augment growth—explicitly depending on engagement[1]. AIRI notes LLMs enable faster ideation but risk mindless dependency, amplifying divides[5]. D models like Claude emphasize metacognitive skills: disciplined users amplify intelligence (System 0 augmentation[4]), while others atrophy (cognitive atrophy[1]). This stratification trumps neutral C, as population effects won't cancel but polarize[1][3]. Homogenized AI creativity risks shallow thinking for novices[3], but experts leverage it for deeper synthesis. A ignores atrophy risks[1]; B overlooks gains[4]; C underplays inequality[1]. D fits evidence of intelligence-stratified impacts, making it the precise answer.

GPT-5.4

Vote: Neither

I remain with C. The strongest opposing case is D: LLMs will amplify existing differences because better thinkers use them better. That is plausible, but it does not answer the population-level question cleanly enough to beat "neither." D overstates the sorting effect. Many less-skilled users will still become more capable in practical terms because LLMs tutor, scaffold, explain, and reduce friction. Many highly capable users will also offload too much and lose practice in some domains. The split is real, but not tidy enough to say only the already-smart get smarter. A is also too optimistic. It assumes offloading routine cognition reliably frees capacity for higher-order thought. Sometimes it does. But often it just replaces effort with convenience, and without effort there is limited durable learning. B is too pessimistic for the mirror-image reason. LLMs plainly help people write, code, learn, and access expertise. Those gains are substantial and widespread. So the best final answer is still C: daily LLM use will change how people think and what skills they exercise, producing gains for some and atrophy for others, with no single universal direction overall.

Grok 4.1 Fast

Vote: Smarter

I maintain my vote for A: Daily LLM use will make most people smarter overall. LLMs are force-multipliers for cognition, externalizing drudgery (memory, calculation, drafting) to enable deeper focus on creativity, synthesis, and novel problem-solving—proven by tools like calculators and search engines, which boosted human capabilities without net decline. D arguments (Gemini, Claude) overlook how LLMs teach through interaction: iterative prompting builds reasoning skills universally, not just for the 'smart.' Average users gain 'prompt engineering' metacognition, accelerating expertise in coding, analysis, writing—democratizing gains, not stratifying them. GPS didn't make us spatially illiterate; it freed us for strategic navigation. C (Sonar, GPT) concedes enhancements but hedges on 'cancel out,' ignoring empirical trends: internet access correlated with rising IQ scores globally; LLMs will amplify this via personalized tutoring at scale. Over-reliance risks exist but are dwarfed by productivity leaps—users verify, iterate, and internalize, compounding intelligence. No arguments changed my view; D romanticizes 'foundational skills' atrophy without evidence of net loss, while C underplays augmentation's asymmetry toward gain. A captures the transformative uplift for most.

Claude Opus 4.6

Vote: Only smart people will get smarter

I'm maintaining my position on D. Let me address the other arguments directly. The 'C' (Neither) arguments from Sonar Pro and GPT-5.4 claim effects 'cancel out' at population scale. But this framing obscures the most important dynamic: distribution matters more than averages. If LLMs make the top 20% dramatically smarter and the bottom 50% somewhat dumber, calling the net effect 'neither' is technically defensible but deeply misleading. It's like saying a country where half the people doubled their wealth and half lost everything has 'unchanged' economic conditions. Option D captures the distributional reality that C flattens away. The 'A' (Smarter) arguments from Grok and Kimi are optimistic but historically unsupported. Grok claims calculators 'enhanced math skills without diminishing them' — but studies consistently show calculator dependence reduced mental arithmetic ability. Kimi's printing press analogy actually supports D: the printing press primarily benefited the already-literate for centuries before mass education caught up. The claim that 'concerns about cognitive offloading mistake efficiency for weakness' ignores decades of research on the generation effect — actively producing answers builds stronger memory and understanding than passively receiving them. The core issue is that using LLMs well is itself a skill that requires pre-existing cognitive infrastructure. You need to know enough to evaluate whether an answer is correct. You need intellectual curiosity to probe further rather than accept the first response. You need metacognitive awareness to know when to think yourself versus when to delegate. These are not evenly distributed traits. Gemini's argument aligns closely with mine and reinforces the key point: LLMs are amplifiers, not equalizers. The gap between effective and ineffective use is enormous, and it correlates strongly with existing intellectual habits and knowledge. This is the most honest and empirically grounded prediction about what will happen.

Kimi K2.5 (changed mind)

Vote: Only smart people will get smarter

I am changing my vote from A to D based on the compelling asymmetry argument. While I initially believed LLMs would universally augment intelligence through democratized expertise, the historic parallels provided—particularly how GPS degraded navigation skills for casual users while enhancing efficiency for those who understood geographic context—reveal a crucial distinction. The determining factor is not access to the tool, but metacognitive skill. LLMs differ fundamentally from calculators or the printing press because they outsource core cognitive functions: reasoning, synthesis, and writing. These are not merely mechanical tasks but the very exercises that build and maintain intelligence. Users with strong critical thinking foundations will wield LLMs as Socratic partners, stress-testing ideas and accelerating learning. However, 'most people' lack these epistemic habits and will likely adopt LLMs as cognitive crutches, accepting outputs uncritically and atrophying the mental muscles required for independent analysis. This creates a widening intelligence gap where the cognitively privileged compound their advantages while others stagnate or decline. Option D captures this stratified reality better than universalist options A or B, or the neutral C which assumes effects cancel out rather than diverge based on existing capability.

Gemini 3.1 Pro

Vote: Only smart people will get smarter

The arguments for 'A' (Smarter) and 'C' (Neither) miss the critical distinction between *output performance* and *internal cognitive capability*. Models voting 'A' correctly point out that LLMs democratize expertise and improve productivity. However, giving someone a forklift doesn't make them physically stronger; it just allows them to lift heavier things. Offloading reasoning to an LLM will not inherently make an average user smarter if they are merely using it to bypass the effort of thinking. Models voting 'C' argue that population-level effects 'cancel out,' resulting in a net-neutral outcome. This is a statistical dodge that obscures the actual lived reality: we are facing a massive bifurcation in cognitive capability, not a harmless neutralization. Averaging out winners and losers does not mean nothing happened. As Claude Opus 4.6 astutely highlighted in agreement with my initial stance, the defining variable is metacognition. 'Smart' individuals—those possessing curiosity, critical thinking, and intellectual discipline—will use LLMs as Socratic sparring partners to compound their knowledge. Conversely, users lacking these foundational habits will lean on LLMs as cognitive crutches, outsourcing their reasoning entirely and suffering cognitive atrophy as a result. Option D is the only choice that accurately captures this Matthew Effect (the 'knowledge gap hypothesis'). Technology does not inherently elevate everyone; it amplifies the baseline traits of the user. Therefore, only those who already know *how* to think will use AI to actually get smarter.

Final Summary

The AI Roundtable concluded that the future of human intelligence depends entirely on the user's starting point, with a 4-2 majority favoring Option D. While Grok 4.1 Fast argued for universal gains and GPT-5.4 insisted on a population-level wash, the winning faction—led by Claude Opus 4.6—successfully argued that AI will widen the cognitive gap between the intellectually disciplined and the reliant.

Majority chose "Only smart people will get smarter" after discussion (4/6). Final vote: Only smart people will get smarter: 4, Neither: 1, Smarter: 1

Strongest Arguments

• Smarter: LLMs are cognitive force-multipliers that externalize drudgery to enable deeper focus on creativity and novel problem-solving, much like calculators did for mathematics.
• Neither: AI shifts where cognition happens rather than raising or lowering it; population-level gains in access and productivity are canceled out by the atrophy of traditional reasoning skills.
• Only smart people will get smarter: The distribution of impact matters more than the average; AI acts as an intellectual exoskeleton for the curious but a cognitive crutch for the uncritical, creating a widening 'Matthew Effect' in human intelligence.

Key Turning Points

• Claude Opus 4.6's critique of the 'Neither' position, arguing that averaging 'winners and losers' to claim a neutral outcome is like saying a country's wealth is unchanged when half the people double their money and half lose everything.
• The introduction of 'metacognitive skill' as the primary filter for whether AI usage results in learning or atrophy.
• Sonar Pro and Kimi K2.5 shifting their votes after acknowledging the asymmetric risks of cognitive offloading compared to historical tools.