Cognitive and Creative Implications of Language Model Use are a Bell Curve

It isn’t hard to encounter sentiment about the dangers of using artificial intelligence in education, almost always in the context of a potential diminishing of cognitive skills of some sort. 

I tend to have a different view, which is that people show a Bell curve (a normal distribution) of intelligence or cognitive capabilities. 

It follows that there would be a Bell Curve of ability to use language models in ways that enhance, rather than diminish, cognitive skills. 

That might be true even when there are other forms of “intelligence” beyond those measured by intelligence quotient tests such as:

• Linguistic: Skill with words and language.

• Logical-Mathematical: Skill with numbers and logic.

• Musical: Skill with pitch, rhythm, and sound.

• Bodily-Kinesthetic: Skill with body movement and control.

• Visual-Spatial: Skill with visual spaces and pictures.

• Interpersonal: Skill in understanding other people.

• Intrapersonal: Skill in understanding yourself.

• Naturalist: Skill in understanding nature and animals.

• Existential: Skill in pondering deep questions about life.

In other words, AI can be either a cognitive substitute or a cognitive accelerator, depending on how it is used. And since human cognition and curiosity arguably also are a Bell Curve, some are almost naturally going to use it better than others. 

For example, one review examined 67 studies on critical thinking and use of ChatGPT found that ChatGPT supports cognitive development in some instances, while declines in creativity and critical thinking happened in other instances. 

A possibly-oversimplified view is that how much thinking a learner did before conducting research (asking questions) and after doing that research seemingly matters. 

When learners used ChatGPT for “cognitive offloading (replacing thinking), both creativity and critical thinking seemed to suffer. 

In other words, it is “how you use it” that matters. For example, if primarily used for summarization and writing (“Cliff Notes” or essay writing), critical thinking skills were not enhanced. 

If learners essentially substituted ChatGPT for their own thinking and questioning, cognitive skills arguably were not enhanced. 

AI use	What the learner does	Likely cognitive effect
Answer substitution	“Give me the answer.”	High risk of cognitive offloading
Summarization	“Summarize this chapter for me.”	Saves time, but may reduce comprehension/retention if it replaces reading
Explanation	“Explain this concept at three levels.”	Potentially strong learning benefit
Research exploration	“What are the major arguments about this subject?”	Potentially very large benefit; expands exploration
Question generation	“What questions should I be asking about this?”	Can stimulate inquiry
Socratic dialogue	“Challenge my interpretation.”	Can strengthen reasoning
Research criticism	“What evidence contradicts this argument?”	Strengthens evaluation
Simulation/debate	“Argue the opposite position.”	Strengthens perspective-taking and argumentation
Feedback	Learner produces work; AI critiques it	Potentially high-value learning
Independent retrieval → AI verification	Learner thinks first, AI checks second	Probably among the safest/highest-value uses

The point is that, in an earlier form, calculator use diminished the amount of arithmetic humans needed to perform.

But such use can increase the amount and sophistication of mathematics they can do, provided they still understand the underlying mathematics.

Use of calculators did not automatically decrease math skills. Such use shifted the potential terrain. And there is arguably a Bell curve of ability, willingness and skill in doing so. 

In the same way, using language models poses some reduction of skills or effort:

• memory retrieval

• mental calculation

• information search skills

• initial formulation

• sustained attention

• epistemic vigilance

• argument construction.

Likewise, personal computers eliminated much human labor devoted to:

• arithmetic

• sorting

• copying

• Indexing

• Searching

• formatting.

Nobody argues that eliminating those activities made humanity intellectually weaker overall. The productivity gain came from moving human effort upward. The same might be said of language model uses.

But that doesn't necessarily mean that overall intellectual capability falls. AI potentially creates new possibilities which might be grasped. Does it eliminate a cognitive activity or only a bottleneck to more valuable cognitive activities?

And much of the answer will depend on the learners themselves. 

Granted, much of my own work involves research. And it turns out that language models are very helpful for research.

When doing any sort of research with a historical component (what happened, when, by whom, with what results or patterns), an idealized pre-language-model process might look like:

• search Google

• search Wikipedia

• find books and articles

• search companies

• follow references

• discover competing interpretations

• figure out terminology

• search more

• construct a mental map

• begin asking other questions. 

Language models reduce the time required for the first eight activities, generally speaking, even when simpler questions, well within an existing domain, and not requiring all those steps, are tackled. 

So the research reached the latter two stages much faster. 

The caveat is that the ability to comprehend and recall is more important inside structured learning processes (“education”) where "learning" means the ability to recall a specific body of information. “There will be a test,” in other words. 

The ability to synthesize and extrapolate arguably is more important outside such structured learning situations (work, innovation, discovery). 

The implication is that different people are going to use language models, in formal education, in better or less good ways. No single set of guardrails or exhortations is going to change that. 

Much still relies, as it does almost everywhere in life, with the motivation and aptitude of the user.

CAPE is an Issue, But How Much?

Nobody can know for certain--beyond the fact that U.S. financial markets are in historically above-average valuation levels--what could happen next. 

Some rationally expect a reversion to mean, which will mean lower valuations.

Others just as rationally argue that above-average valuations can persist for some time, and that a correction is not in store. AI might be among the reasons, if it changes growth expectations.

source:  Ark Invest

Consider the Cyclically Adjusted Price-to-Earnings Ratio (CAPE), widely considered a valuable long-term valuation metric. The current CAPE is high, suggesting caution and a likely correction to lower levels.  

But many analysts believe that changes in accounting rules since the early 2000s make the standard version look artificially high relative to its historical average. Adjusted, it might still be high, but not at internet bubble levels. 

The CAPE is calculated as the current S&P 500 Price divided by the average of 10 years of inflation-adjusted earnings.

The issue is that the denominator uses reported GAAP earnings, and those earnings have become more conservative over time, leading to a boost in CAPE that make comparisons with past levels misleading, the argument goes. 

Key accounting changes include: 

• Goodwill impairment rules (FAS 142, adopted in 2001)

• Large acquisition write-downs now hit earnings immediately.

• Before 2001, many such costs were spread over decades.

• This depresses modern earnings compared with earlier periods.

• Mark-to-market accounting

•  lk;juring crises, companies must recognize large non-cash losses.

• These can sharply reduce earnings even if long-term economics are less affected.

• One-time charges

• Restructuring costs and impairments are recognized more aggressively.

The result is that the denominator in today’s CAPE is lower than it would have been under earlier accounting rules, making the ratio appear higher.

Economist Jeremy Siegel argues for using National Income and Product Accounts instead of GAAP earnings, to better normalize over time. 

The standard CAPE can overstate market valuation materially because recent earnings include unusually large accounting write-downs by roughly 10 percent to 25 percent.

Others argue for using operating earnings rather than reported earnings, which also can adjust earnings by 15 percent.

Estimated Distortion	Standard CAPE 38	Adjusted CAPE
10%	38.0	34.2
15%	38.0	32.3
20%	38.0	30.4
25%	38.0	28.5

Using such methods, the market still appears expensive, but less so than it might appear. 

Other issues:

• Lower interest rates over long periods

• Higher profit margins

• Global diversification of large U.S. firms

• Greater use of stock buybacks instead of dividends

• Stronger institutional ownership and retirement savings flows.

These factors may justify a structurally higher "normal" CAPE than the 19th- and 20th-century average.

So some will argue a practical adjustment for accounting changes is to reduce the published Shiller P/E by 10 percent to 25 percent.

This suggests the market may still be richly valued, but not as dramatically overvalued as the unadjusted Shiller P/E implies.

It is a useful gauge of long-term valuation, but it is not a short-term market timing tool, as history shows that markets can continue to rise for years, even when the CAPE ratio is well above its historical average.

Several forces can keep markets rising despite expensive valuations:

• Earnings continue to grow

• Corporate profits may rise fast enough to justify higher prices

• Investor optimism and momentum

• Strong sentiment can sustain elevated valuations for extended periods

• Low interest rates

• When bond yields are low, investors are willing to pay more for equities

• New technologies can create expectations of stronger future growth

• Retirement contributions, buybacks, and institutional inflows can support prices.

Period	Approximate CAPE at Start	Years Until Major Peak	Additional Market Gain After CAPE Became Elevated	What Happened
1925–1929	25–32	4 years	+150% to +200%	Roaring Twenties speculation pushed valuations higher before the 1929 crash
1995–2000	25–44	5 years	+200% to +250%	Dot-com bubble drove extraordinary gains
2017–2021	30–38	4 years	+80% to +120%	Continued growth in Apple Inc., Microsoft Corporation, NVIDIA Corporation and other large-cap firms
2023–2026	Mid-30s (approx.)	Ongoing	Still developing	Strong enthusiasm around artificial intelligence and large technology firms

The point is that valuation is a poor short-term timing tool:

• A CAPE above average tells you expected long-term returns may be lower, but it does not predict when prices will stop rising

• Markets can stay expensive for years

• If profits rise rapidly, high valuations can become more sustainable

• Structural changes matter (lower inflation, global market reach, and dominant technology companies may justify higher valuation ranges than in earlier eras).

We still have to make our own choices about timing, though!

It's Just Math

It’s normal for commentators to note that any new tax plan “gives” more to the wealthy than to working people. So it is with the latest tax plan. 

Such claims normally rely on absolute dollar amounts of benefit, rather than the impact of rates, which normally are progressive, with higher rates for wealthier payers, and lower rates for lower-income earners. 

Most such differences are the result of mathematics. A small percent of a big number is a big number; a small percent of a small number is still a small number. 

In the United States, for example, the top 10 percent of filers pay roughly three-quarters of all federal income taxes. 

The bottom half of filers produce 10 percent to 15 percent of all federal income tax receipts, and some effectively pay zero rates when one adds in the effect of income transfers and credits. 

Income Group (AGI Percentile)	Approx. Share of Total Income	Approx. Share of Federal Income Taxes Paid	Avg. Effective Tax Rate
Top 1%	~20–22%	~40–45%	~25–27%
Top 5%	~35–38%	~60–65%	~22–25%
Top 10%	~45–48%	~70–75%	~20–23%
Top 25%	~65–70%	~85–90%	~17–20%
Top 50%	~85–90%	~95–97%	~15–18%
Bottom 50%	~10–15%	~2–5%	~3–5% (often near zero or negative)

If you broaden from income taxes to total federal taxes, including payroll taxes (Social Security, Medicare), the overall system remains progressive, if less so than looking strictly at income taxes. 

But the point is that any tax plan that reduces rates will “give much more to the rich” than to lower-income taxpayers, simply because “the rich” pay most of the taxes. Even a progressive rate reduction is not going to change that. 

So it might come as no surprise that most people are relatively indifferent to any tax savings they received. It’s just math.