🔮 Does AI make you dumb? And why our forecasts suck #576

Exponential View · Azeem Azhar

Conventional forecasting and conventional thinking both fail during exponential regime changes like AI, because analysts anchor to linear models while users offload cognition to chatbots without anyone redesigning incentives for the new reality.

The mistake isn't bad numbers but the wrong curve: mean-reversion frameworks can't even see what an exponential trajectory looks like, so analyst price targets quietly encode a distribution that has nothing to do with reality. The same shape-blindness shows up in the debate over whether AI makes people dumber — the on-average answer matters less than the fact that we haven't articulated what thinking is now for, or what incentives, credentials, and workplaces would reward doing it ourselves. Both failures are downstream of refusing to notice that the basis has changed.

claim

Equity analysts are well-paid and diligent, but on average they are late when a technology hits a learning curve and demand goes exponential. They stay anchored to old linear assumptions while the curve runs away from them.

central 0.95 · novel 1.00

implication

If incentives produce the outcome of offloading important thinking to chatbots, the genuinely thorny question is what new system of incentives, credentials, and workplace design would motivate people to think for themselves again.

central 0.85 · novel 0.33

claim

The interesting question isn't whether AI degrades thinking on average. It's that technology has changed the basis on which we expect thinking to happen, and we haven't articulated the new expectations — an exponential gap.

central 0.90 · novel 0.19

implication

The real lesson isn't that analysts should revise their numbers higher; it's that mean-reversion frameworks can't even see what's coming during a regime change. Using them to evaluate the AI boom is using the wrong tool.

central 0.90 · novel 0.15

mechanism

An analyst's target price expresses a probability-weighted distribution over a company's possible futures. If you assume linear business-as-usual, you've not just picked the wrong point on the curve — you've picked the wrong curve entirely.

central 0.85 · novel 0.20

Open

· What concrete incentive or credential system would actually reward people for thinking rather than offloading to AI?
· What forecasting framework replaces mean-reversion during a regime change, and how would you know it's calibrated before the curve resolves?

Pipeline

source kind: url
generated by: anthropic+voyage
candidates: 24 (selected 5)
embeddings: voyage-3.5

Coverage

100% covered

Each block is one paragraph of the source. Darker means the decomposition captures it well; lighter means it was left out — the part of the document the summary doesn’t cover.

Considered candidates (19)

Below top-k · 15

evidenceMicron EPS forecasts tripled in five monthsc 0.80
In December 2025, consensus EPS for Micron's FY2026 was about $18.25. Five months later, with a bit of actual data in hand, the median view jumped to roughly $58 — raising the question of what the original forecast was even for.
mechanismWe still reward scarcity theatre from an era when cognition was scarcec 0.80
Tests, credentials, and career ladders were built when thinking ability was scarce and assumed to be the candidate's own. AI breaks that assumption, but the GPA-and-Ivy payoff structure remains — so people rationally offload the thinking.
mechanismMean reversion is a sociological tic, not just a modelc 0.75
Markets digest financial information well but stomach technological change poorly. Mean reversion is a powerful default heuristic that quietly fails on compounding processes — yet it's safer to keep using than to admit the model is broken.
evidenceEven Google's consensus jumped 40% in a yearc 0.70
Google is huge and only modestly exposed to AI, yet between May 2025 and May 2026 the median analyst view on it grew 40%. The mis-forecasting isn't a Micron-specific quirk.
evidenceConsensus expects an unexplained hyperscaler capex crash in 2027c 0.70
JPMorgan's chart of consensus forecasts shows hyperscaler capex and R&D dropping 25-40% as a share of revenue in 2027. The only rational driver consistent with analysts' actual job is a near-term AI revenue slump — which nothing in the data supports.
mechanismOutsourcing connection-making may weaken the musclec 0.65
Much of what we recognize as good human thinking is integrating ideas across experience, reading, and instruction. If we hand that connection-making to AI, the cognitive muscle plausibly weakens, the same way unused motor skills do.
exampleFinishing the marathon in a carc 0.65
Using ChatGPT to do thinking that matters is like driving a marathon: it only makes sense if you don't believe the race itself has value. The decision to offload reveals what people actually think is at stake.
contextThe FT's 'impossible maths of the AI boom' argumentc 0.60
An FT column argued that with AI capex growing 20% a year against revenues growing only 15%, the AI boom is heading toward one of the largest destructions of shareholder value in history. The argument leaned on consensus analyst estimates for 2025-2030.
mechanismNo one at major banks covers AI as a wholec 0.55
Coverage is sliced into chip, cloud, and software analysts, while the real demand action sits in private AI labs nobody formally tracks. The org chart of equity research can't see the system.
evidencePost-ChatGPT student essays got colorful but lost original ideasc 0.55
A study of 370,000 college personal statements found that after ChatGPT became available, essays grew more diverse and colorful in language but markedly less creative in their ideas.
exampleHeavy AI use can sharpen rather than dull thinkingc 0.55
At Exponential View, AI has absorbed the 'computer stuff' — typing, tedious emails, long-running coding and literature jobs — and the author reports reading more critically and reflecting more deliberately as a result, not less.
evidenceAlphaGo made human Go players better, not worsec 0.50
Henrik Karlsson shows that being trounced by AlphaGo actually raised the quality of human Go play — a counter-example to the simple 'AI atrophies our skills' story.
exampleWhiteboards, fountain pens, and two hours a week off-screenc 0.45
The team deliberately invests in non-AI thinking practices: whiteboard scoping sessions and a requirement that everyone spend at least two hours a week with pen and paper, no computer. Whiteboardmaxxing over tokenmaxxing.
claimIndividuals get more productive with AI but firms don'tc 0.40
Individual workers are seeing real productivity gains from AI tools, yet those gains don't translate proportionally to the organization. As one exec with a thousand Claude-Code-using engineers put it: 1+1+1+1=1.5.
exampleThe IEA spent two decades ignoring solar's falling costsc 0.40
The same pattern of consensus bodies under-forecasting exponential improvements played out with the IEA's solar projections — a comedic, decades-long sequence of upward revisions.

Redundant with selected · 2

evidenceAnalysts keep chasing the puck rather than getting ahead of itc 0.70 · sim 0.86
Median consensus targets for AI-exposed companies have been revised upward repeatedly over the past year and still sit well below the market price. Missing repeatedly is not bad luck — it's systematic blindness to exponential dynamics.
overlapped with: Consensus analyst forecasts are systematically late to regime shifts
caveat'Common sense' findings on AI and thinking are still thinc 0.55 · sim 0.84
The cognitive-atrophy studies cited are small-scale. Plenty of common-sense claims — like video games causing real-world violence — collapse once properly researched. Caution is warranted before declaring AI is making us dumb.
overlapped with: 'Is AI making us dumb?' is the wrong question

Low centrality · 2

contextAI engineering hiring concentrates at top US payersc 0.25
The latest state of the software engineering job market shows AI roles growing and concentrating in the highest-paying US companies, with Anthropic now the most-desired employer ahead of founding one's own company.
contextChina's AI optimism may be survival, not enthusiasmc 0.20
Zilan Qian reads Chinese AI optimism ambivalently: respondents may genuinely believe AI is good for them, or they may see it as necessary surgery — knowing flesh will be cut away.

Janitor

Non-content spans (acknowledgements, references, footnotes, headers, boilerplate) are dropped before the decomposition runs.

total spans: 88
kept: 78
dropped: 10
outliers: 6

content · 78
noise · 9
boilerplate · 1