A 2026 study from Cornell and Carnegie Mellon analyzed more than 81,000 college applications and found a troubling pattern. The study found that lower-income students who leaned heavily on AI faced an admissions penalty nearly twice as large as higher-income peers using AI the same way. That held even after controlling for GPA and test scores. The gap was not about how much AI students used, but whether they had an expert checking the AI's work. That distinction applies to any high-stakes decision where AI is now part of the process.
The Data Point Most AI Coverage Is Missing
Most conversations about AI and inequality focus on cost and hardware, fixating on who can afford a paid subscription, who has a personal device, and who has reliable internet. The Cornell and Carnegie Mellon study points to a different and less discussed problem. Lower-income applicants in the study were 28% more likely to rely heavily on AI when applying to college. Their admission odds dropped by 83% compared to a 62% drop for wealthier peers, even with nearly identical academic profiles.
Also worth a look: Viridian Space awarded $225,000 NASA contract to advance satellite maneuverability in very low Earth orbit and New Roku App ‘Pevac Digital Signage’ Transforms Any TV into a Powerful Digital Signage Display.
The likely explanation is not their AI use itself. Instead, it points to the reliance on generic AI used by most students, which depend on generic web data to inform decisions that require specialized knowledge.
Why Most AI Tools Struggle With Specialized Decisions
An AI tool built to answer almost anything is only as good as what is publicly available online. College admissions is a compelling test case for specialized AI because it relies on information that is not actually public. Admissions committees do not publish their internal rubrics. The nuance that separates one strong application from another rarely shows up in the public datasets these AI models try to learn from.
That gap between what's public and what actually matters is exactly where Prepory has operated for over a decade. Since 2012, the global admissions coaching company has guided more than 20,000 students across 80 countries. Its coaches, many of them former admissions officers, represent more than 400 years of combined experience sitting on the other side of the table. That knowledge lived in the coaching relationships themselves, not in any public data set.
Recent developments in the space illustrate how domain-specific models attempt to bridge this divide. Earlier this summer, admissions coaching firm Prepory launched Rory, an AI application trained on historical student profiles, real outcomes, and institutional admissions insights rather than general web data. According to Prepory CEO Daniel Santos, closing the access gap requires democratizing genuine pattern recognition rather than relying on surface-level text generation. By training on specialized outcomes, such tools aim to preserve authentic applicant voice rather than smoothing out personal essays into generic AI prose.
What This Means for Teams Building AI in High-Stakes Domains
This approach points to a broader architecture pattern worth applying far beyond college admissions. When a decision is high stakes, largely irreversible, and dependent on nuance that is not public, a generic model trained on generic web data is insufficient. The people most exposed to that underperformance are usually the ones with the least access to an experienced human check.
Three questions are essential for any engineering or product team integrating AI into high-stakes workflows:
- Where does the model's knowledge actually come from? If the answer is the public internet, the model likely lacks the specialized judgment the decision requires.
- Who is checking the output before it matters? If the answer is no one, the product needs a substitute for that validation built into its design rather than relying on user prompting.
- Who is most likely to use this tool without a human backstop? That population is the exact demographic to design guardrails around first.
Addressing these questions highlights a structural problem that goes beyond a simple lack of access to human experts. The core challenge lies in the lack of access to what those experts actually know, insight that was never public to begin with. The AI industry has spent considerable energy helping users sift through raw public information. The larger opportunity is extracting the domain expertise that exists inside specialized practitioners and extending its reach to people who cannot access those experts directly.
