Why NYC Private Schools Now Cost More Than Harvard - and What ChatGPT's Stock Picks Reveal About the Same Fallacy
Twelve Manhattan high schools now charge more in annual tuition than Harvard University.
Not more than a regional state school. More than Harvard. The cheapest school on the list runs $59,320 for 2025-2026. Harvard's annual tuition is the same number - actually a few dollars less. Parents writing those checks have presumably noticed. Most keep writing them anyway.
The same week this surfaced, a Harvard Business Review study published results from testing ChatGPT and DeepSeek against 4,978 Chinese companies. ChatGPT's forecasts ran 12.5% more optimistic and landed 13% farther from actual year-end stock prices. Expensive signal, systematically wrong.
Meanwhile, the CMO of Grillo's Pickles was asked at SXSW about ROI measurement for experiential marketing campaigns. His answer: "It's truly just the vibes."
These are not three unrelated stories. They are the same story.
When Prestige Becomes the Product
The private school paradox is not new, but the Harvard comparison makes it visceral in a way it hasn't been before.
Regression discontinuity studies - the gold standard for isolating school effects - compare students who scored just above and below the cutoff for selective school admission. The finding, replicated across multiple school systems, is that marginal admits to selective schools show "little effect of an exam school education on achievement." The students who barely got in performed about the same as the students who barely missed the cutoff. The school didn't change the trajectory.
After controlling for demographics, the private school advantage "disappeared and even reversed in most cases." The heritability of academic achievement is approximately 60%, and shared environment - the category that includes both family and school - accounts for only 20% of variance. School quality alone explains less than 2% of variance in test scores once you control for where a student started.
This connects directly to a mechanism well-documented in behavioral economics: price signals don't just inform perception, they alter experience. The brain preruns pleasure circuits based on price expectations before any actual evaluation takes place. A $60,000 tuition check triggers the same mechanism as a $45 bottle of wine served to someone who was told it cost $45. The experience of quality is partially constructed by the expectation of quality.
That's not a bug in how parents evaluate private schools. It's the mechanism private schools are selling.
The Network Asymmetry Worth Understanding
There is a genuine effect worth isolating: selective school networks provide real benefits, but they are concentrated almost entirely in lower-income students gaining access to high-status social capital for the first time. For students from already-connected families - which is to say, for most of the families writing $60,000 tuition checks - the marginal network gain is minimal.
The people paying most are precisely the people for whom it matters least. The Of Dollars and Data analysis puts the opportunity cost in concrete terms: $250,000+ in forgone tuition, compounded over roughly fifteen years, would grow to approximately $400,000 available for a down payment by the time a student turns 30. That is not a rounding error. That is a decade of compounding returns traded for a signal that research suggests is largely noise.
ChatGPT's 13% Error Rate Is a Prestige Problem, Not a Technical One
The HBR stock-picking study reads like a parable about confident tools operating outside their competence.
Researchers gave ChatGPT 4.1 and DeepSeek R1 identical financial inputs for 4,978 companies on Shanghai and Shenzhen exchanges - total assets, return on assets, leverage - and asked both to project December 31, 2024 stock prices. They ran this in June-July 2024, then validated against what actually happened.
ChatGPT issued "buy" recommendations 1.3 percentage points more frequently and projected prices 12.5% higher. When the researchers checked actual year-end prices, ChatGPT's absolute forecast errors were 13% larger than DeepSeek's on the same companies.
The mechanism is instructive: negative news about Chinese companies received coverage in Chinese-language media that never entered ChatGPT's training corpus. With no negative signal available, the model filled the gap with optimism. It wasn't malfunctioning. It was doing exactly what it was designed to do - pattern-matching toward the mean of available evidence. The problem is that the available evidence was systematically incomplete.
The study's authors recommend treating AI analysis as "a powerful but partial lens," supplementing outputs with local data, domain expertise, and cross-model validation rather than relying on single-source recommendations. That's careful and correct. It's also not how most organizations are using these tools.
This is the kind of pattern STI's research tracks systematically: how high-confidence signals diverge from actual outcomes across investment and decision contexts. The confidence of a recommendation is not evidence of its accuracy. The confidence trap is that the signals that feel most certain - consensus market forecasts, institutional endorsements, authoritative AI outputs - are often the ones most disconnected from ground truth. Going into 2025, the institutional consensus had U.S. stocks as the only game in town, rates heading down, inflation finished. All three were wrong.
Cross-Model Validation as Baseline Practice
The practical implication of the HBR finding isn't "don't use AI for financial analysis." It's "don't use a single AI model as your only lens on any domain with real stakes and uneven training data coverage."
The same foreign-bias logic applies anywhere one model's training corpus has systematic gaps: healthcare research in underrepresented populations, legal analysis in jurisdictions with sparse case law, market analysis in regions with different media ecosystems. The presence of a confident, articulate answer is not the same as the presence of a complete or accurate one.
If you're evaluating AI analysis tools against these criteria, our analysis tools can help surface where model confidence is and isn't matched by actual predictive validity.
Why Spectacle Stopped Working at SXSW
At SXSW 2026, brand marketers said something the industry usually keeps private: the era of the standalone pop-up is functionally over.
Winnifer Thomas-Cox of Accenture Song framed the shift as structural: experiential marketing has to become "an operating system," intentionally woven into every consumer touchpoint, not deployed once as an isolated activation. Jiayu Lin, CEO of PopSockets, was direct: "If you just pop up once at one place, it's still very fleeting in today's world." Adweek's coverage of the SXSW panel captured what most post-event brand summaries omit: the measurement gap behind the spectacle.
That gap is where the Grillo's Pickles CMO's answer lives. When a brand leader at a marketing conference describes campaign ROI measurement as "truly just the vibes," that is not humility. That is a description of how the industry has operated for years, said out loud for the first time. Single activations produce impressions, not data. Impressions produce cases for the next budget cycle. The cycle continues.
The resurgence of brand loyalty data from fast food chains tells the other side of the same story. McDonald's, Starbucks, and Chipotle are overhauling loyalty programs not because discounts stopped driving traffic - they still do - but because "deal loyalty is brand-detrimental." Brands that competed on deals trained their best customers to never pay full price. Starbucks built multi-billion-dollar loyalty infrastructure and discovered it had engineered discount dependency.
The brands recovering genuine loyalty are doing it through attitudinal commitment - emotional investment, occasion-specific relevance, identity alignment. That's harder to build than a viral pop-up. It's also the only thing that compounds. A one-time spectacle produces a one-time signal. Attitudinal loyalty produces recurring behavior without requiring the spectacle to recur.
The AI Bypass Problem Waiting to Land
There is a secondary implication from the SXSW conversation worth flagging: AI agents planning consumer itineraries may bypass brand touchpoints entirely. If your brand presence is a physical activation that an AI assistant decides isn't worth rerouting a consumer toward, the activation you built doesn't get seen. This isn't a future scenario - it's a planning constraint that should already be in the 2026 experiential strategy. Experiences that exist only for human discovery are becoming invisible to the layer that increasingly mediates discovery.
The Structure Underneath All Three Stories
Strip the context and three very different industries reveal the same architecture.
A premium signal - tuition, AI confidence, experiential spectacle - has been substituted for the underlying thing it was supposed to represent: educational outcomes, analytical accuracy, brand equity. The substitution worked for a long time because the signal and the reality were correlated often enough to feel reliable.
Private school graduates performed well; they just would have performed well anyway. AI recommended reasonable stocks; it just didn't know what it didn't know. Experiential activations generated buzz; the buzz just didn't translate into durable preference.
What's changing isn't that people are suddenly smarter about these substitutions. It's that the feedback loops are tightening. Private school ROI is getting quantified in ways it wasn't ten years ago. AI forecast accuracy can now be validated against actual price data at scale. Brand loyalty programs have longitudinal data deep enough to separate emotional commitment from deal-seeking frequency.
Signals that relied on opacity are becoming legible.
What Legibility Requires from Decision-Makers
The private school research points toward what actually predicts outcomes: prior student achievement and the socioeconomic conditions students bring through the door. The AI study points toward cross-model validation and local data supplementation. The brand loyalty data points toward attitudinal investment rather than transactional frequency.
In each case, the legible predictor is less comfortable than the premium signal.
"Your child is already high-performing and comes from a supportive home" is less reassuring than "the school we chose is exceptional." "Your AI model has systematic gaps that require validation" is less satisfying than "the AI said buy." "Brand loyalty requires years of attitudinal investment with no single measurable activation moment" is less fundable than "book the pop-up."
This is the decision intelligence gap - the distance between the signal that feels like certainty and the data that actually describes outcomes. It exists in education, in financial analysis, in brand strategy, and in most other domains where high-stakes decisions get made under information pressure. That gap is where expensive mistakes compound.
The question worth asking for any premium signal you're currently paying for: if someone stripped away the prestige layer and showed you only the outcome data, would you still sign the check?
For a systematic framework for approaching that question across investment and brand decisions, start with STI's research archive.