By Jim Shimabukuro (assisted by Perplexity)
Editor
Foundation models are large machine-learning systems trained on extremely broad, often multimodal datasets (text, images, code, scientific data and more), usually with self‑supervision at scale, and then adapted to many different downstream tasks such as question answering, prediction, and content generation.1,2 This makes them “foundations”: incomplete but general models that can be further specialized for particular domains, from science and medicine to law, education, the arts, and public policy.2
The University of Toronto Schmidt AI in Science Fellows describe a scientific foundation model as “think ChatGPT, but trained on massive datasets specific to fields like biology, astrophysics or chemistry,” emphasizing that such models become vital tools for scientists precisely where high‑quality, task‑specific data are scarce, because a single pre‑trained model can be fine‑tuned for specialized tasks or used to generate representations that improve downstream prediction models.3 In this sense, foundation models extend the transfer‑learning paradigm: they are trained once on broad, generic data and then reused across many applications, often via fine‑tuning or prompting, which is why they are central to the current AI “paradigm shift.”1,2
A widely cited technical and conceptual definition comes from the Stanford Center for Research on Foundation Models: these systems are models “trained on broad data (generally using self‑supervision at scale) that can be adapted to a wide range of downstream tasks,” with examples including language models such as GPT‑3, vision–language models such as CLIP, and other large pre‑trained architectures. The same report stresses that their significance can be summarized by “emergence” and “homogenization”: emergence refers to behaviors and capabilities, such as in‑context learning from prompts, that arise implicitly from scale rather than being hard‑coded, while homogenization refers to the consolidation of AI workflows around a small set of general models that are reused across tasks, domains, and even research communities. This combination yields extraordinary leverage—improvements to a few foundation models can propagate across entire ecosystems—but it also concentrates risks, because any biases or vulnerabilities in the shared foundation can be inherited by thousands of downstream systems.2
In the natural sciences and engineering, domain‑specific foundation models are already enabling new forms of discovery. The U of T “Foundation Models for Science” workshop, held at the Schwartz Reisman Innovation Campus, brought together about 70 participants from multiple continents to test scientific foundation models in real research workflows, with organizers arguing that such models are particularly valuable as “AI for science” enters an era where high‑quality training data, rather than model architectures, has become the bottleneck.3,4 Scientific foundation models trained on protein sequences, molecular structures, satellite imagery, or physical simulations can be adapted to accelerate tasks like drug discovery, materials design, climate modeling, and astrophysical signal detection, effectively lowering the cost of hypothesis generation and exploratory analysis.2,4 At the same time, initiatives like Schmidt AI in Science Fellowships at Toronto and elsewhere explicitly aim to diffuse this capacity across disciplines, so that researchers in veterinary science, environmental monitoring, and fall‑prevention in older adults can translate laboratory advances into real‑world applications and even commercial ventures.5
Beyond STEM, foundation models are increasingly treated as sociotechnical systems that reshape institutions, professions, and power relations, not just as neutral tools. A position paper on socially responsible foundation models argues that social science expertise is necessary “across all stages of foundation model research and development” to understand how these systems interact with existing power structures, how they might reproduce or disrupt inequities, and how technical interventions (for example, bias mitigation or content filters) actually play out in practice.6 This framing highlights that foundation models are embedded in an ecosystem: human‑generated data are collected and curated, models are trained on this data, adapted into applications, and then deployed into workplaces, classrooms, courts, and media infrastructures, with feedback loops that can reinforce or challenge existing social patterns.2,6 Because their capabilities are broad and often emergent, it is hard to anticipate all downstream uses, which complicates governance but also opens space for innovation in social science, public participation, and ethics.
In law and legal institutions, foundation models are beginning to transform both practice and regulation. The Stanford report surveyed early applications in law, including automated contract analysis, legal search, and assistance with drafting, all powered by language models adapted from a small set of general foundations.2 A 2024 paper on “Foundation Models and Fair Use” further notes that many such models are trained on copyrighted materials scraped from the internet, raising unsettled legal questions about whether training and generative use are protected under U.S. fair‑use doctrine, and what technical safeguards (such as reducing memorization or preventing near‑verbatim reproduction) might be necessary to reduce legal and ethical risk.7 At the policy level, recent work in international competition law characterizes “foundation models” as a critical economic layer in generative AI and proposes taxonomies and enforcement priorities—such as focusing antitrust scrutiny on “general public foundation models” that benefit from strong increasing returns—to prevent the market from “freezing” around a handful of dominant providers.8 These discussions underscore that foundation models are becoming infrastructure in legal and regulatory systems, not just productivity tools for individual lawyers.
Education and writing are emerging as especially important arenas for foundation‑model deployment and debate. A 2025 article in npj Science of Learning reviews “large multimodal foundation models” (LMFMs) in education and argues that they open “a new set of opportunities and challenges” by enabling highly interactive learning environments, richer formative feedback, and intelligent tutoring systems that can handle ill‑defined problems, while also risking new forms of bias, over‑reliance, and digital divide. The same work suggests that such models might even expand access for learners who cannot yet read or write by combining speech‑to‑text, text‑to‑speech, and visual interfaces, though this requires careful attention to infrastructure and equity.9 At the curricular and institutional level, writing‑across‑the‑curriculum practitioners are beginning to articulate AI‑aware pedagogy and policy, arguing that generative AI (powered by foundation models) should be governed by transparent, flexible, discipline‑specific policies that protect learning while recognizing that writing, in forms ranging from reflective journals to multimodal projects, is a central method of knowledge‑making rather than a peripheral skill.10 More generally, guidance for schools and higher‑education IT leaders emphasizes risk‑based approaches to AI policy that balance data protection, academic integrity, and age‑appropriate use, acknowledging that foundation models are becoming embedded in everyday educational tools.11,12
Social sciences and humanities are also engaging foundation models as both objects of study and instruments of research. Social scientists, for instance, are using large language and multimodal models to code qualitative data at scale, to simulate conversational agents for experiments, and to explore scenarios related to labor markets, misinformation, or governance, while simultaneously insisting on frameworks that treat these models as sociotechnical artifacts whose design choices, training data, and deployment contexts must be critically interrogated. Humanities scholars are beginning to use foundation‑model‑based tools for tasks such as corpus exploration, translation, and creative writing, even as debates about authorship, interpretation, and cultural representation intensify.2,6 As these disciplines adopt foundation models, they also contribute to broader conceptual vocabularies—such as thinking in terms of power, responsibility, and interpretive communities—that can inform the technical and policy development of the models themselves.6
Across all of these domains, the importance of foundation models stems not only from their immediate capabilities but from their role as infrastructural, cross‑cutting platforms. They centralize and fuse information from multiple modalities—text, images, structured data, protein sequences, and more—into shared representations that can then be adapted for specialized tasks across healthcare, law, education, robotics, and scientific discovery.2,3,4 At the same time, their scale and emergent behavior make them difficult to understand, evaluate, and control; their homogenizing role means that flaws and biases can propagate widely; and their development currently depends on concentrated computational resources, which raises questions about accessibility, openness, and global competition.2,8 These tensions are driving new research on transparency indices for foundation models, new tiered regulatory approaches, and proposals for public or academic infrastructure—such as national research clouds—to ensure that experimentation, oversight, and innovation are not confined to a small number of corporate actors.13,14,2 In this sense, foundation models are important not just as a technological milestone but as a focal point for rethinking how AI is built, governed, and integrated into scientific, educational, legal, and cultural institutions.
References
- Bommasani, R. et al., “On the Opportunities and Risks of Foundation Models,” Center for Research on Foundation Models (CRFM), Stanford University, 2021. https://crfm.stanford.edu/assets/report.pdf
- Bommasani, R. et al., “On the Opportunities and Risks of Foundation Models,” Introduction and §1.1–1.2. https://crfm.stanford.edu/assets/report.pdf
- University of Toronto, “U of T Schmidt AI Fellows bring foundation models to the forefront of scientific research,” 17 Feb 2026. https://www.artsci.utoronto.ca/news/u-t-schmidt-ai-fellows-bring-foundation-models-forefront-scientific-research
- Defy Gravity Campaign (U of T), “U of T Schmidt AI Fellows explore how artificial intelligence can accelerate scientific discovery,” 26 Feb 2026. https://defygravitycampaign.utoronto.ca/news-and-stories/schmidt-ai-fellows-explore-how-artificial-intelligence-can-accelerate-discovery/
- University of Toronto, “Schmidt AI in Science Fellows leverage accelerated research into real-world impact,” 15 Jun 2025. https://www.artsci.utoronto.ca/news/schmidt-ai-science-fellows-leverage-accelerated-research
- Davies, A. et al., “Social Science Is Necessary for Operationalizing Socially Responsible Foundation Models,” arXiv:2412.16355, 2024. https://arxiv.org/abs/2412.16355
- Henderson, P. et al., “Foundation Models and Fair Use,” Journal of Machine Learning Research, vol. 24, 2024. https://www.jmlr.org/papers/volume24/23-0569/23-0569.pdf
- Crémer, J. & de Montjoye, Y.‑A. et al., “Competition between AI foundation models: dynamics and policy implications,” International Journal of Industrial Organization / Industrial and Corporate Change, 2025. https://academic.oup.com/icc/article/34/5/1085/7942098
- Sinaulan, R. et al., “On opportunities and challenges of large multimodal foundation models in education,” npj Science of Learning, 2025. https://www.nature.com/articles/s41539-025-00301-w
- Writing Across the Curriculum (WAC) Clearinghouse & WAC Association, “Statement on AI and Writing Across the Curriculum,” 2025. https://wacassociation.org/ai-statement/
- Structural Learning, “Creating an AI Policy for Schools: A Practical Guide for 2025,” 18 Feb 2026. https://www.structural-learning.com/post/creating-ai-policy-schools-2025
- Microsoft Research, “Accelerate Foundation Models Research: Supporting a Global Academic Research Ecosystem,” 3 Oct 2023. https://www.microsoft.com/en-us/research/blog/accelerate-foundation-models-research-supporting-a-global-academic-research-ecosystem/
- Stanford CRFM, “The Foundation Model Transparency Index (October 2023),” 14 Sep 2023. https://crfm.stanford.edu/fmti/October-2023/index.html
- Stanford CRFM, “Drawing Lines: Tiers for Foundation Models,” 17 Nov 2023. https://crfm.stanford.edu/2023/11/18/tiers.html
Filed under: Uncategorized |
Educational Technology and Change Journal 
Leave a comment