Will AI Become All-Too-Normal Edtech?
How to Prevent the New Tools from Deepening the Cognitive Equity Divide
Measuring Progress
There is a puzzle at the center of digital technology in the academy. Backed by billions of dollars in investment over the past thirty years, it is often heralded as a set of tools that improves learner outcomes by various means (retention, grades, test performance, equity of access, or student collaboration). Frequently there is evidence to support these claims. Yet despite the digital tools, six-year graduation rates at four-year institutions (perhaps the single best measure of student success) have improved only modestly, rising from 55% in the mid-1990s to 63% today, according to the National Conference of State Legislatures.
Even this modest gain may be illusory. An extensive study from 2022 in the American Economic Journal suggests that 95% of those graduation rate increases have not been achieved through improving learning and developing skills but by relaxing standards—that is, by simply raising grades (grade inflation and compression). There is little evidence, the study authors claim, that changes in institutional resources, which would include decades-long edtech investments, can explain the trend.
External corroboration of the loss of value in college grades is evident among the growing number of employers who prefer to do their own evaluation of recent graduates rather than trust the transcript. This is important context for the especially high unemployment rates for recent college grads and a significant loss of perceived value of a college education.
Edtech pessimists might point to these results as confirmation that the technology investments have been wasteful—missed opportunities to allocate resources, including educators' time, more productively.
But that isn't quite the story either. What makes this a puzzle rather than a sorry tale of costly distraction is that many of these technologies work as advertised. Developed by mission-driven and passionate edtech entrepreneurs, they are often good ideas and good products that, under the right conditions and training, can yield the intended benefits.
So why don't they, at least in aggregate?
Demographically Narrow Gains
Part of the answer is that while the benefits are often real and demonstrable, they are concentrated locally, say in introductory or "gateway" courses. These improvements are rarely sufficient to affect six-year graduation rates (they are, after all, a point along a much longer academic continuum), so the aggregate numbers change little despite some improvements under the surface. It could also be that graduation rates are a lagging indicator and that the boosts from at least the most recent tools haven't yet registered (the largest investments in US higher ed-focused edtech occurred during the past five years, during the height of the pandemic).
The deeper reason, however, likely lies in the gap between the tools themselves and their deployment in real-world usage. As is the case with technological integrations across market sectors, the benefits of these tools typically accrue to the organizations best positioned to make use of them. The best-prepared students at the best-resourced institutions will adopt these technologies with only marginal gains in retention and graduation rates, as they had already been doing well with these metrics; the less well-resourced will either not have access at all or struggle with a lack of preparedness within their communities.
What Makes AI “Normal Technology”
As Princeton computer scientists Arvind Narayanan and Sayash Kapoor argue in "AI as Normal Technology," published this past spring, AI adoption will not occur with the speed that many AI enthusiasts suggest because of the friction between the "methods" (LLMs) and real-world applications, and the back-and-forth iterations between app builders and users before they can become part of organizational workflows.
"Normal" in this sense is not a claim about the underlying tech itself but rather about its diffusion once the algorithms leave the frontier labs. (Their Substack series, just renamed with the title of that seminal article, is a worthwhile read.)
While AI adoption is rapidly accelerating among consumers and coders, its integration into standard business workflows is, as they predicted, taking a slower, more measured path. According to a recent S&P Global article, more than 40% of early corporate AI projects have been abandoned (other reports put the "failure" rate much higher). Businesses, of course, haven't given up on AI; they are iterating and applying new knowledge and tools in a fast-moving stream.
Edtech Distributional Challenges
The same dynamics are at work in higher ed (diffusion frictions and a need for complements), but there are also strong distributional effects that condition how or even whether such tools will be used effectively in the classroom. Making a general point, Narayanan and Kapoor note that the downsides of AI are likely to "mirror those of previous technologies that are deployed in capitalist societies, such as inequality."
As applied to higher ed, the risk is especially acute in the present context. During the past three decades, higher ed has experienced dramatic funding shifts, from chronic state disinvestment in public institutions (especially the already less-resourced community colleges and regional universities that serve most students) to tuition raises and simultaneous cuts in instructional and student support.
Over this same period, higher ed has become far more diverse, with more first-generation and older, working students than ever before, all of which by themselves drive greater demand for support and guidance. This makes the just-announced federal cuts to programs supporting first generation students especially challenging.
The institutional complements that make technology effective are scarcest precisely when the most powerful educational technology of our age demands them. The result could be a level of academic disparity that deepens in proportion to the power of the new technologies themselves. It is in this sense that AI could take its place as all-too-normal education technology, further widening the equity gap. If AI adoption maps to this same preparedness landscape, even the best technologies likely wouldn't change the overall academic performance trends.
When Tech is Well-Embedded: the Case of Georgia State
Crucially, elite resources are not required for success; all that might be needed is the right mixture of institutional creativity and resourcefulness. The earlier, pre-AI exemplary work of the Georgia State University (GSU) “GPS” project illustrates this well.
Faced with a graduation rate of just 32% in 2003, GSU pursued a model of change and human engagement at a system level, changing its processes to better serve its students. The university leveraged ten years of student data to build a predictive analytics platform that identifies at-risk students based on 800 different indicators. The predictive technology, however, has not been the intervention itself; it is a straightforward diagnostic tool that triggers high-touch, human-led support. The tech has helped uncover the needs, but the real success lies in the human interventions that followed.
Today the system prompts over 55,000 advisor meetings each year and helps target microgrants to prevent students from dropping out for minor financial reasons. The results have been transformative: By 2015, GSU had increased its six-year graduation rate by 23 percentage points while also eliminating achievement gaps based on race, ethnicity, or income. GSU's success is a pre-AI case demonstrating that intentionality matters and that progress is possible even in deeply resource-constrained contexts.
From Access to Cognitive Equity
Still, the risks today are not the same. They are at once more profound and centered less on access equity (free chatbots are ubiquitous) than on cognitive equity. To illustrate this, consider a peer-reviewed study from last year, "Beware of Metacognitive Laziness," which set out to determine the effect of different learning tools on "intrinsic" or "self-regulated learning" (SRL) and metacognition (thinking about thinking).
As the authors put it, "Metacognitive strategies in SRL, such as goal setting, self-monitoring, and self-evaluation, are essential for effective learning." Without the right external support, intervention, or scaffolding, however, "learners encounter diverse regulatory challenges, stemming from factors such as inadequate metacognitive strategies, low achievement motivation, and task complexity."
The study aimed to determine how the introduction of chatbots to support essay writing for non-native speakers of English as one of the four learning modalities affected SRL and, through this, retained learning. The researchers found that while ChatGPT could help students write better (better, in fact, than human tutors), there was little or no commensurate transfer of knowledge through the process.
Given this lack of knowledge transfer, the authors note, "we should be cautious about the integration of ChatGPT into teaching and learning. Although ChatGPT can quickly improve task performance, it does not significantly enhance intrinsic motivation and may trigger cognitive laziness." The worry is that students then become overly reliant on a tool that helps improve the quality of their work without a commensurate increase in motivation or knowledge.
Guided, And “Brain-First”
A recent preprint and non-peer-reviewed study from MIT, "Your Brain on ChatGPT," largely corroborates this by using EEGs to measure neural activities. While it's a weaker study (very small, unrepresentative cohort, questionable assumptions regarding EEG mapping to meaningful cognitive activity), it's nonetheless complementary and suggestive: the researchers found that in an essay-writing task, participants who relied on a chatbot from the start showed the weakest and narrowest neural couplings, consistent with outsourcing core semantic and memory operations.
Those who drafted on their own first and then used the chatbot for refinement, meanwhile, displayed broader re-engagement of memory and control networks and demonstrated significantly better recall of their own text. The preliminary implications are that "guided and brain-first, chatbot-second" work can lead to enhanced learning; reversing that order (or without guidance) might not only degrade learning but even, over time, weaken cognitive functions.
This situation fits into longstanding broader trends in academia, such as the use of "homework help" that enables students to perform better but doesn't improve learning or the motivation to learn. It’s a risk that’s also well understood at the leading frontier companies, and this is in part why they have all recently released educational versions of these chatbots (Anthropic’s Claude for Education, Google DeepMind’s Gemini for Education, and OpenAI’s ChatGPT Edu), that are institutionally-focused and intended to encourage student learning rather than simply acting as an answer source.
While these are encouraging developments, they are not by themselves likely to change the aggregate performance of higher education unless they too are coupled with sustained institutional commitments. The chatbot tutoring applications will be most effective for those whose SRL is already high and who have the extrinsic support to help ensure it remains so (the diffusion challenge reasserting itself).
An Infrastructure of Intentionality
Distinguishing among forms of cognitive load could provide an important principle for pedagogical design. Students, of course, are not waiting for such guidance, and most institutions haven't kept pace. While about 90% of college students use AI for their courses, fewer than 40% of higher ed institutions have comprehensive AI policies, much less implement classroom AI plans. Campus-wide policies around ethical use and academic integrity are required, but they will not be enough to avert the fate of AI becoming all-too-normal technology.
What's needed is a willingness to engage not only with the educational versions of the chatbots but also with the latest AI research, iterate quickly on the most promising new practices, and continually test new pedagogical strategies across the academy. It’s about charting one’s own path based on direct available data and evidence at one’s own institution. In a phrase, it's about establishing an infrastructure of intentionality.
Much of this will come down to sustained faculty development, moving beyond one-off workshops to campus-wide comprehensive programs that train educators in AI literacy, equitable teaching practices, and AI-resilient assignment design. This might include teaching faculty how to create learning pathways that use AI to reduce possible extraneous friction (however one defines this) while preserving the productive friction essential for learning.
Instruction could be oriented toward approaches less susceptible to the "metacognitively lazy" use of AI. GSU has helped demonstrate that a weaker technology paired with strong support can have outstanding results. Conversely, a technology orders of magnitude more powerful that is left unmediated could rapidly accelerate the cognitive equity divide.
Initial Suggestions
Focus on learning processes over submitted work (might include a "brain-first, chatbot-second" principle), requiring the submission of drafts, notes, and reflections to make student deliberative thinking more visible.
Make content personal and context-specific, tying assignments to specific in-class discussions, local issues, or personal experiences that are more AI-resistant.
Design for self-reflection and original thought, asking students to critique AI outputs and reflect on their own learning journey.
Use student collaboration techniques, such as the protégé effect (students presenting to each other), that could also use chatbots as intermediaries and (provisional) evaluators.
No doubt many others will continue to surface as new thinking and research findings emerge. See, for example, Dr Philippa Hardman’s post from last year on instruction and assessment in the age of AI. Jonathan Boymal’s recent post, “Making Thinking Visible, The Philosophy of Process-Based Assessment,” points to some of the promise and challenges with new forms of evaluation.
Regardless of the approach, agility is critical: test, succeed/fail fast, document and learn, iterate, and test again. The “jagged-edge” of development means that some tech may be nearing a plateau (with only incremental improvements ahead), while others will continue to climb for awhile. Text-based LLMs may be an example of the former; robotic and embodied AI are examples of the latter.
Averting an All-Too-Normal Outcome
Better-resourced institutions, their faculty, and students have significant advantages in the adoption of AI (we should openly recognize this), but just as the case of GPS at Georgia State has helped demonstrate, it's possible to achieve exceptional results when tech is deployed with a deeply intentional, human-centric framework. Educational versions of chatbots from frontier labs may be able to help but they’re not a substitute for institutional engagement.
In an era of higher ed disinvestment, growing student need, and likely the most powerful education technology the academy has ever grappled with, getting this wrong risks not only reproducing the uneven distribution of benefits that have long characterized edtech but also flattening the very cognitive work that higher education claims to value.
Building this infrastructure of intentionality is not only possible; it may also be our best hope of averting an all-too-normal outcome that would even more consequentially widen disparities in a society that may no longer be able to bear them.
Acknowledgements: I’d like to thank Julian Jacobs, John Katzman, and Henry Pickford for comments on earlier drafts of this essay.
