The Tutor and the Tripwire

I. The question of the week

The question that has organized the past year’s coverage of AI in higher education is not whether the technology will reach the classroom — it has — but who, after it arrives, will find themselves better off, and at whose expense. Through 2025 the conversation pivoted on a single claim with two faces: that generative AI lowers the cost of expertise, putting a competent tutor, drafter, and explainer in the hand of any student with a device; and that the same systems extend a regime of measurement, surveillance, and exclusion that has always borne down hardest on the students with the least margin for error. The two claims are not opposed. They describe the same deployments from different sides of the table.

The arc this column traces runs from late 2024, when the equity question still felt theoretical, through the spring of 2025, when optimistic rhetoric crested in the trade and consulting press, into the summer when critical framings overtook the boosterism for the first time, and on to a quieter fall in which the language shifted from promise to procurement. One inversion point is detectable in the data; the rhetoric flipped in 2025-Q2 and has not flipped back. What the inversion reflects is not a sudden discovery of harm — the harms were documented from the start — but the slow arrival of evidence into a discourse that had been running on projection.

This week’s piece is about the gap between those two registers — the promise and the procurement — and about what the institutions actually doing the procuring have learned in twelve months they did not know in the twelve before.

II. What we’ve been saying

The optimist case for AI in higher education was, through late 2024 and early 2025, articulated in a particular voice: the voice of the consultancy white paper and the keynote. It rested on a single empirical claim that lent it weight beyond rhetoric — the finding, surfaced in the AI Index Report 2024, that consultants given access to GPT-4 saw their performance improve across the board, with the largest gains accruing to those who had started below the median. AI, in this telling, was a leveler. The student who could not afford a tutor now had one; the writer whose first language was not English now had a copy editor; the worker without a credential could now perform tasks that had once required one. The narrowing-the-gap framing migrated quickly from labor economics into education and from education into a broader civic register.

By the first quarter of 2025 the trade press had absorbed this framing wholesale. Learning AI Is Not An Option Anymore, It Is Mandatory phrased the choice as an obligation to keep up, not a question of whether keeping up was desirable; Embracing AI: How Artificial Intelligence Is Shaping the Future of Work supplied the workplace vignettes. Both rested on the implicit equity claim — that the gains would be broadly shared — without examining it. Two Frameworks for Balancing AI Innovation and Risk admitted risk only in the form of organizational lag.

The second quarter was when the equity claim ceased to be subtext and became headline. AI Has the Potential To Be a Wealth Engine for Women put the framing in its title; Harnessing AI’s Potential: Building Pathways to Social Justice and Economic Equity supplied the academic-adjacent gloss. The argument in both was the same: AI’s lower marginal cost of expert assistance would benefit those who had been priced out of the existing market for it. Within higher education this argument did particular work. It allowed deans and provosts to describe expanded campus licenses for chatbots not as a productivity expenditure but as an access initiative — a tutor for every student, framed as a redistribution.

But the same quarter contained the inversion. AI in Hiring: Revolutionizing Talent Acquisition or Reinforcing Bias? sat in the trade press alongside the equity-promise pieces and asked whether the optimization at the other end of the credentialing pipeline — automated resume screening, video-interview scoring — would intercept exactly the students whom the in-classroom AI was supposedly serving. The contradiction was named explicitly: the same suite of tools that promised to lower the entry barrier was being deployed to raise the gate. From that quarter forward the critical framings outpaced the optimistic ones in the corpus, and they have not surrendered the lead.

By summer the critical voice had assembled its own canonical arguments. How AI Assessment Tools Affect Job Candidates’ Behavior documented that candidates change how they speak and present when they know they are being scored by a machine, with the changes least available to those without coaching. Beyond the bias: Designing AI for social fairness, notably published by a major consultancy, conceded what a year earlier would have been heresy in that register — that bias is not an edge case to be patched but a structural feature requiring design intervention from the start.

The library has had a voice on this from before the conversation began. Race After Technology (Benjamin, 2019) framed the question the trade press took five years to ask out loud: if the discrimination already in the workplace is widespread and well documented, whether outsourcing decisions to AI counts as a remedy depends entirely on what data the algorithm was trained on and what it was trained to optimize. The hopeful answer assumes the algorithm escapes the history; the empirical answer is that it inherits it. The Atlas of AI (Crawford, 2021) noted the structural reason the inheritance is hard to interrupt: public companies are pressured by shareholders to maximize return on investment, “commonly making ethics secondary to profits.” The equity-promise rhetoric of early 2025 ran on the premise that vendor and institution had aligned incentives with the student. The skeptical rhetoric of late 2025 had begun to test that premise.

What changed across the year, then, was not the underlying argument — the equity double-edge was visible from the start — but which edge the press was willing to put on its cover. The narrowing-the-gap headline did not disappear; it was joined, then surpassed, by the widening-the-gap one.

III. What’s been happening

While the rhetoric was negotiating its inversion, the ground was moving in a different rhythm. Three threads of actual deployment ran through 2025 in higher education and adjacent credentialing systems, each tracing the equity question more concretely than the discourse around it.

The first thread is the campus license. Beginning in late 2024 and accelerating through the first half of 2025, universities of all tiers signed institutional contracts with one of the major model vendors, often framed in the announcement language as a commitment to equitable access. The framing has a real referent — students paying out of pocket for the paid tiers were obviously not on a level field with those who could not — and a real cost: institutional licenses bind a campus’s pedagogy and its data to a single vendor whose pricing, terms, and model behavior the institution does not control. AI Market Trends 2026: Global Investment, Risks, and Buildout traces the financial logic from the vendor’s side: AI infrastructure spending is now indexed to institutional adoption at scale, and the multi-year contract is the unit of revenue projection. The equity framing supplied the moral cover; the procurement was a capital deal.

The second thread is the productivity evidence. The narrowing-the-gap finding that anchored the optimist case did not survive the year unchallenged. The ‘productivity paradox’ of AI adoption in manufacturing firms reported that AI introduction frequently produced short-term productivity declines, not gains, as workers and processes restructured around the new tool. AI adoption in the Enterprise 2026, a vendor-funded survey of 2,400 knowledge workers, found that the technology is working for individuals while “breaking organizations” — a phrasing that, even from an adoption-friendly source, concedes a gap between the personal upside and the institutional fallout. In higher education the analogous gap is between the individual student’s experience of a useful assistant and the institution’s experience of an assessment regime that no longer measures what it used to. Both ends of that gap have distributional consequences, and neither has been resolved.

The third thread, and the one most directly tied to the equity question in education, runs through hiring and credentialing. How do I maximize AI for hiring and talent management? catalogued, in the trade voice, the rapid normalization of AI in resume screening, video interviewing, and candidate scoring across the employers who will hire the students the universities are credentialing. Biases in Artificial Intelligence and Implications for AI Use in Public Administration: A Technical Perspective, published as the year closed, walked through the technical mechanisms by which large language models embed and amplify the demographic skew of their training data, with specific attention to the public-sector contexts in which the consequences are not appealable. The student who used the campus chatbot to draft a stronger cover letter is then evaluated by a vendor system whose error modes correlate with the same demographic axes the chatbot was supposed to help compensate for. This is not a hypothetical: the same year that saw Employees Won’t Trust AI If They Don’t Trust Their Leaders appear in HBR also saw widely reported incidents of AI-driven surveillance and identification systems producing false matches that ended in arrests — an end-of-pipeline analogue to the misclassification that begins in assessment software.

Within the educational institution itself the year produced a recurring pattern: a generative-AI policy issued by a provost’s office, often accompanied by a short faculty-development workshop and a vendor-supplied training module. The policy typically warned against academic dishonesty; the workshop typically demonstrated productive uses. Neither generally addressed the equity question directly — what happens to students with disabilities who rely on AI for access, what happens to multilingual students whose drafts are flagged by AI-detection software with known higher false-positive rates against non-native English, what happens to students whose financial circumstances make the free tier the only tier. These were the questions a campus had to answer in practice whether or not its policy named them.

AI for SMBs: Turning big tech advantages into small business strengths captured the parallel happening outside higher ed: the same equity-of-access framing applied to small businesses, with the same structural feature that the access on offer was access to a stack owned and priced by a small number of large vendors. The pattern repeats because the underlying procurement logic repeats. As an earlier essay in this publication’s social-aspects coverage observed in our briefing of 2025-06-29, the equity language and the vendor-consolidation reality have been moving together, not apart.

What was happening, then, was that institutions were buying AI faster than they were learning to govern it, and the equity claims were a function of the speed of procurement, not of any settled empirical case.

IV. Where they meet, where they miss

The rhetoric and the reality met, when they did, on the narrow finding that prompted the whole optimist case. The narrowing-the-gap result surfaced by the AI Index Report 2024 is real, replicable in the specific task domain in which it was measured, and consequential. People at the low end of a skill distribution do benefit disproportionately from access to a competent generative assistant on tasks the assistant performs competently. To deny this is to deny the most robust empirical finding the field has produced. The optimist case has a floor.

But the rhetoric and the reality miss in three respects, and each miss is structural rather than incidental.

The first miss is task selection. The narrowing-the-gap finding holds on tasks the model performs well; on tasks the model performs poorly or unevenly, the same access can widen gaps by giving more confident users — who tend to be the already-advantaged — the leverage to spot and correct model errors while less confident users adopt the errors as authoritative. AI Ethics, in the MIT Press Essential Knowledge series, framed this years before the chatbots arrived, asking who will be “empowered” and who “excluded” as AI takes over more activities — and the question is not about access to the tool but about access to the discernment needed to use it. The campus license addresses the first; it does not address the second.

The second miss is the pipeline. The student-facing benefit and the employer-facing harm are separated, in the discourse, into different conversations — one in the education press, one in the HR press — and treated as if they were unrelated. They are not. A student credentialed with AI assistance and then evaluated by AI screening is participating in a single process in which the institution has equipped both ends, but the equity arithmetic has been done only at the first. How AI Assessment Tools Affect Job Candidates’ Behavior is the rare piece that tries to look at both ends of the same pipeline; it remains an exception.

The third miss is governance. The equity argument was used to authorize procurement; procurement, once complete, became the constraint within which any subsequent equity work had to operate. The Atlas of AI (Crawford, 2021) is direct on the structural reason: vendors face shareholder pressure to make ethics secondary to profit, and institutions that have bound themselves to a vendor inherit that ordering. An equity rationale circulated in spring 2025 and a three-year renewal negotiated in spring 2026 leave the institution with markedly less leverage in the second meeting than it claimed in the first.

Where the two registers genuinely meet is in the quiet recognition, present in the late-2025 critical pieces and absent from the early-2025 optimist ones, that the equity question is not a question about the technology but about the procurement and the institution. AI does not widen or narrow access; deployments do, governed by contracts and shaped by incentives that the equity language has been used to disguise rather than examine. Race After Technology (Benjamin, 2019) put the test plainly: outsourcing a decision to an algorithm does not absolve the institution of the decision; it relocates the decision into a system that is harder to interrogate. The student arrested by a false facial-match, the candidate filtered out by a video-interview model, the multilingual essay flagged by a detector — all of these are decisions the institution made, executed through a vendor, in language that described the decision as access.

V. The longer view

The year that began with AI described as a tutor for every student is closing with AI described as a system for every transaction the student will subsequently be subjected to. Both descriptions are true. The longer view is that the equity double-edge is not a property of the technology to be optimized away but a property of the institutional choices that determine what the technology is used for, by whom, against whom, and on what terms. The optimist rhetoric of 2025-Q1 and the critical rhetoric of 2025-Q3 are not opposed positions to be balanced; they are two readings of the same procurement decisions, taken from two ends of the pipeline that those decisions configured.

The institutions that will look back on this year well are not the ones that bought the most AI or the ones that resisted it longest. They are the ones that wrote the equity claim into the contract — into the data terms, the audit access, the exit clauses, the disability accommodations, the assessment policy, the appeal process — rather than into the press release. The rest will discover, contract by contract, that the narrowing-the-gap finding was the smallest claim AI ever made about itself, and that the gap it widened in the meantime was the one between what an institution can say about a vendor and what it can do about one.