Big AI vs academia: for what comes next, follow the money

This blog was kindly authored by Michael Larsen, CEO and Managing Director at Studiosity, a HEPI Partner.

Since their public release nearly four years ago, no sector of society has been more profoundly impacted by increasingly powerful large language models (LLMs) than higher education. Student adoption of the technology exceeds 90%, while a growing body of research confirms that unconstrained chatbots diminish learning and lead to cognitive atrophy. At the same time, some universities are licensing the technology for their students to avoid being labelled AI luddites. Yet amidst all of this chaos, one question has been consistently overlooked: Why are the frontier models (OpenAI, Anthropic, Google etc.) so intent on embedding their technology in universities?

For clarity, there is no doubt AI has enormous potential in education, though that potential is conditional on a deep understanding of the increasingly well-researched risks associated with frontier models.

The $3 trillion driver

To answer this question, we need to understand the financial models and associated incentives that underpin the AI strategies of frontier model companies. When considering deployment of a generic LLM for their students, every university leader should be required to have a deep familiarity with the incentives that drive the LLM marketplace.

A few short years ago ChatGPT leapt over the first barrier of the Technology Acceptance Model – ease of use – by mimicking conversation. It has gone on to become among history’s fastest growing consumer applications. Overcoming the second barrier – usefulness – requires ‘smarter’ LLMs that don’t hallucinate. The solution has been scale: gigantic models trained using mind-bending volumes of data (obtained by means both fair and questionable) in huge data centres with insatiable appetites for electricity and water.

By 2028, approximately $2.9 trillion will have been invested globally in new data centres, with financial returns expected via new organisational purchasing on the fastest, most capable chatbots. Ostensibly, such purchases will be justified either by more productive human workers, or fewer of them.

Economically, Big AI is agnostic about how such productivity is achieved. The investment quantum is so large and so strategically existential that adoption prevails over impact and outcomes. The notion of constraining a powerful productivity chatbot when provided to learners – though educationally sound – simply introduces friction to mass student adoption, and risks the financial return. Releasing an easily disabled ‘learning mode’ by some LLMs is the visible extent of their commitment to authentic learning.

Cognitive growth versus token consumption

Higher education is bound by both duty and self-preservation to ensure graduates become value creators in the AI economy, rather than collateral damage. If today’s graduates are to flourish in an AI economy, they must first develop the very human capabilities of critical thinking, communication and collaboration that may then be amplified by the use of productivity-oriented AI. Mistakenly conflating ‘AI literacy’ with cognitive development ensures student technological dependence that undermines employability.

The current pressures universities face are immense and they necessitate strategic agility. Although most are redesigning learning & assessment, some are simultaneously deploying frontier/productivity chatbots that many staff and students report negatively impacting cognition. Further strain is introduced in some cases by the concurrent deployment of disconcerting digital surveillance tools that monitor students while they work.

While the education sector is in the midst of disruption and action is required, a ‘first principles’ response would prioritise the creation and dissemination of knowledge. This ethos is increasingly surfacing among faculty via the ‘Slow AI’ movement, which encourages a thoughtful, evidence-based and climate-sustainable approach. Far from a denial of AI’s importance, it is an acknowledgement of education’s uniqueness.

While the frontier models have provided a compelling view into the underlying capabilities of generative AI, they are neither the end state of AI in education, nor a product designed for learning. Such chatbots have certainly been optimised for adoption by someone. That someone is rarely a learner.

Reclaiming the purpose of higher education

Thequestion for academic leaders is not which frontier model to provide to students. It is: what is learning for?

Technology companies incentivised to reduce friction will remove the struggle upon which learning depends. Today, institutions have the autonomy to choose tools based on learning science, rather than token consumption. Which tools they choose matters far less than the principle underpinning the choice.

Enacting that principle could take the form of a sector procurement framework requiring a pedagogical value assessment of AI products and a minimum acceptable standard. Coupling authentic learning outcomes with the existing standards would provide both institutional clarity and public confidence.

Follow the money and the direction of travel is clear. As the capital sunk into model training and data centres compounds, so too will the pressure to convert it into usage, regardless of the impact upon human thinking and development. That pressure is not the responsibility of higher education. The sector’s distinction and obligation is to contribute to society and the economy through educated people, not users who consume tokens.

Studiosity is writing feedback and assessment assurance to support students and validate learning outcomes at hundreds of universities across five continents, with research-backed evidence of impact.