In practice, the categories blur almost immediately. A group of colleagues begins experimenting with AI in their programme — bottom-up, clearly. Their institution funds a day to share what they learned. A dean sees one department has figured something out and tries to create conditions for it to spread. At what point does it tip? And does the direction even determine whether it works?
What seems to matter more is what happens when different levels are in motion at the same time: when institutional intent, departmental culture, and individual conviction overlap, reinforce, or work against each other. What does it look like when those levels connect productively? And what reliably gets in the way? These questions have direct consequences for how universities respond to disruption — generative AI being the most pressing current example — and for whether the people doing the actual work of teaching feel supported or simply managed.
How adoption actually works
Everett Rogers spent decades studying exactly this question across dozens of fields, and his findings are uncomfortable for anyone who manages institutions. Worth noting first: Rogers defined an innovation as any idea, practice, or object perceived as new — which means collective sense-making about what a change means, working through its implications for professional identity and purpose, counts just as much as adopting a new tool. Adoption of either does not move through broadcast; it moves through conversation. More precisely, it moves through interpersonal exchanges between people who trust each other, at the moment when someone is deciding whether to change what they do.
Rogers identified several distinct roles in this process (2003), and the distinctions matter. The most adventurous members of any system, the ones who experiment early, take leaps, absorb uncertainty willingly, are rarely the ones who bring the majority along. Rogers is direct about this: the most innovative member of a system is typically perceived as a deviant, accorded low credibility by average members. They prove that something is possible. They do not, by themselves, make it feel safe.
That work falls to opinion leaders. Rogers defined opinion leadership as the degree to which someone can “influence other individuals’ attitudes or overt behavior informally”, with the operative word being informally. Opinion leaders are not appointed from above or self-selected from below; they earn their position through technical competence, social accessibility, and — critically — conformity to the norms of their community. They are a step ahead of their peers, but recognisably of them: colleagues who have tried something in the same department, with the same student profile, under the same institutional constraints, and whose judgment others have learned to trust through repeated experience. When an opinion leader says a new approach worked, or is worth trying, it lands differently than any policy document or training slide could. Firm attitudes shift through interpersonal channels with near-peers, not through mass communication or top-down directives. A centrally issued AI policy can raise awareness. It cannot do the persuasion work.
There is a further wrinkle. Change agents (the educational developers, the learning designers, the academic support professionals) tend to have high enthusiasm and broad organisational reach, but they do not automatically carry the credibility of a near-peer. Rogers’ insight here is that the change agent’s most effective move is indirect: identify and cultivate the opinion leaders already present in local contexts, then work through those relationships to activate broader uptake. The change agent who bypasses this and goes straight to broadcast loses the social mechanism that actually moves people.
The micro level: pioneering groups and what they produce
In practice, this looks like a small team in a specific programme who begin experimenting together: redesigning an assessment with AI in the loop, rethinking how feedback works when students have access to language models, or simply trying to work out together what any of this means for what they are actually trying to do as educators. These groups tend to form around shared curiosity, shared frustration, shared stakes in a particular context. They are easy to call “bottom-up”, organic, unorchestrated, driven from the ground, but that framing can mislead. They need conditions to form, time and visibility to do their work, and something to connect to beyond themselves if what they learn is to matter at scale. In Rogers’ terms, they are where the early work happens that makes later adoption credible to others.
What these groups produce is not just practice knowledge. They produce shared meaning — ways of making sense of what is changing and why it matters — and from that, opinion leaders: people who can speak with authority about what worked, what failed, and what it actually felt like to navigate the uncertainty. This is the micro level doing its most important work.
But micro networks carry a structural limitation. Research on how instructors talk about their teaching (Roxå & Mårtensson, 2009) shows that these conversations happen inside small, trust-rich clusters, and Poole, Iqbal & Verwoord (2019) found that members of these networks are more likely to perceive high similarity of beliefs with each other, and that value within the network is tied to that perceived similarity. Homophily, the gravitational pull toward people who already share your assumptions, is what makes these groups feel safe enough for honest exchange. It is also what keeps their learning local. Structural holes open up between clusters. What a pioneering group in one department discovers rarely reaches another, even when it would be directly relevant.
This is where the institutional picture breaks down. Adaptation flourishes in one pocket of the university and stalls everywhere else: not because people elsewhere are resistant, but because they have never heard from a near-peer that the thing is worth trying.
The meso level: hubs and the flow of knowledge
Taylor, Kenny, Perrault & Mueller (2022) name the missing mechanism: hubs. A hub is an individual or group that sits across multiple local networks, energising cross-connections and improving knowledge flow between clusters that would otherwise remain isolated. Hubs do not replace the trust-based intimacy of micro networks; they bridge it. They carry the signal from one pioneering group to another — not just what worked in practice, but how people are making sense of the change, what questions they are sitting with, what the uncertainty feels like from the inside — preserving the near-peer quality of the communication.
This is also where the levels actually meet: not in a policy document, not in a mandate handed down, but in the work of someone who knows what is happening at the ground level and can create conditions for it to travel. The educational developer who knows which teams are experimenting, who the opinion leaders within those teams are, and how to bring those leaders into contact with each other across departmental lines — that person is doing the most consequential work available. Not broadcasting. Not mandating. Connecting.
Vaessen, Van den Beemt & de Laat (2014) put their finger on why this level is so persistently neglected: the informal learning happening inside these networks is, in their words, “largely invisible to the official framework of the organisation.” Institutions do not recognise it, do not fund it, and therefore do not protect it. The result is that a large proportion of the adaptive capacity already present in a university goes unregistered and unsupported, while resources flow toward the centralised instruments that have the least purchase on actual behaviour change.
Resilience as infrastructure
All of this — the micro networks, the opinion leaders, the hubs — describes how change travels. But there is a prior question worth pausing on: what allows people to keep adapting at all, under sustained pressure, without burning out or retreating?
I came across a piece by Lynn Berger in De Correspondent that reframed how I think about this. Her argument, drawing on seventy years of psychological research, is deceptively simple: it is primarily your relationships that make you resilient — not your training, your mindset, or your individual capacity to absorb stress. That finding led me to the work she builds on most directly: the Kauai Longitudinal Study.
Emmy Werner was an American developmental psychologist who spent four decades following the lives of 698 children born in 1955 on the Hawaiian island of Kauai. Her study tracked children born into serious adversity — poverty, family discord, parental illness — to understand what determined whether they flourished or struggled. Her findings, published in 1993, reframed how the field thought about resilience entirely, and they speak directly to what institutions facing AI disruption tend to get wrong.
One-third of the high-risk children thrived against every prediction. The consistent differentiator was not individual grit or personal resilience as a portable trait. It was relational: a close, stable bond with at least one trusted adult who functioned as a safe base under pressure. Berger traces how developmental psychology subsequently misread this finding, distilling it into the individualised resilience narrative that now saturates institutional language — the idea that resilience is something you build inside yourself, through the right habits and the right workshops.
What Werner’s data points to, underneath the developmental context, is a more general claim: people do not primarily sustain themselves under pressure by drawing on reserves within themselves. They draw on the quality of the structures around them — on whether there is someone to turn to, somewhere to process difficulty, a relationship stable enough to hold the weight. Remove that infrastructure and individual capacity, however substantial, tends to erode.
Resilience lives between us.
The parallel to universities under the sustained pressure of AI disruption is uncomfortable but important. Resilience across an institution does not accumulate through training programmes for individual adaptability: resilience workshops, AI literacy modules, prompt engineering guides. It accumulates through the relational architecture described above: the ongoing conversations between colleagues about what their work actually means, the spaces where it feels safe to admit uncertainty, to question whether a new approach is working, to talk about what the profession is becoming. It accumulates through micro-level pioneering groups generating local proof and local credibility; opinion leaders within those groups carrying the signal to the people most likely to act on it; meso-level hubs bridging across structural holes so that adaptation spreads rather than pools; and macro-level conditions (what the European University Association (2026) calls rightfully so ‘a culture of sharing’) that protect and fund all of it.
When institutions replace this architecture with compliance infrastructure, they crowd out the relational substrate that makes adaptation possible in the first place.
What this asks of leaders, developers, and those who join forces
There is a specific objection worth addressing before drawing practical conclusions. The network-based picture sketched here can look, from an institutional vantage point, like an argument for managed chaos: many small groups doing similar things, no clear direction, no shared framework. The coordination concern raised at the start of this piece resurfaces here. If six departments are each experimenting with AI in assessment, are we not duplicating effort? Would it not be more efficient to run one institutional pilot, document what works, and scale it?
The difficulty with this reasoning is that it treats contextual variation as waste. From a diffusion perspective, those six groups are not doing the same work. Each is producing local proof: evidence that a practice works in their context, with their students, in their disciplinary culture. That contextual specificity is what makes findings credible to near-peers in adjacent settings. A single institutional pilot, however well-documented, cannot generate the distributed trust that carries adoption across a university. What looks like duplication from the centre is often the process of adaptation working as it should.
The coordination problem is real, but it is a different problem than it appears. The issue is that teams cannot find each other, not that they are going in different directions. The solution is connection. Hubs do not tell groups what to do; they help different groups discover that their questions overlap, their experiments are adjacent, and their opinion leaders have something to say to each other. That is a fundamentally different kind of coordination than a shared roll-out.
This reframing has direct implications for the roles of people who do this connecting work. Educational developers, curriculum guidance, learning designers and the like: people who occupy positions that are neither fully central nor fully embedded in any single department are structurally well-placed to be hubs. They often already know which teams are experimenting. They attend the conversations where early adopters surface, and they have access to institutional channels that local groups do not. What is rarely asked of them explicitly is the network-brokering work: identifying opinion leaders within pioneering groups, creating low-stakes conditions for those leaders to encounter their counterparts elsewhere, and making the invisible visible — surfacing what is being learned in one corner of an institution so it can reach another.
This is different from delivering workshops or producing resources. It requires a different orientation to the role: less about service provision and more about circulation of people, stories, questions, and partial answers. It also requires institutional conditions that recognise and protect this kind of work, which tends to be slow, relational, and poorly captured by activity metrics.
For leaders, the ask is more structural. Find the pioneering groups. Protect their time. Make space, not just formally but culturally, for opinion leaders to speak in settings where their near-peers will hear them. Resist the pull toward premature standardisation, which tends to arrive just as local variation is producing the most useful learning. And where resources are genuinely scarce, invest them in connection rather than in coordination: not in another framework or another training module, but in the conditions under which different groups might find each other and compare notes.
In that light, the question of whether change should be top-down or bottom-up really does dissolve into something more useful. What seems to matter is whether the levels are in contact: whether what is being learned at the ground can reach people who might act on it, and whether the conditions set at the institutional level enable rather than crowd out the relational work that carries change forward.
A colleague who has genuinely worked through what AI means for their third-year seminar, and who sits down to talk it through with someone facing the same question in a different department, is doing more for institutional resilience than any policy on academic integrity.
The risk is not that these networks, these opinion leaders, these slow-built webs of distributed trust are too weak to carry institutional change. The evidence suggests otherwise. The risk is that the people in the best position to recognise and tend to this infrastructure keep looking straight through it — seeing inefficiency where there is adaptation, drift where there is sense-making, small talk where there is the actual work. The institutions that learn to look again discover something striking: that when you attend to how change actually moves, and build with it rather than around it, the adaptive capacity that unlocks is not incremental. It compounds.
Resources and further readings
Berger, L. (2026, June 1). Veerkracht zit niet alleen in jezelf, maar vooral in de mensen om je heen. De Correspondent. https://decorrespondent.nl/16961/veerkracht-zit-niet-alleen-in-jezelf-maar-vooral-in-de-mensen-om-je-heen/155dd925-85fe-04a3-25f3-f6a2e9a07898
Bolander Laksov, K., McGrath, C., & Serbati, A. (2026). Educational leadership of, for, with, and through academic development. International Journal for Academic Development, 31(2), 165–175. https://doi.org/10.1080/1360144X.2026.2665877
European University Association. (2026). For lasting impact, we must rethink staff development for teachers [Keynote address, EUA Annual Conference, Lisbon, February 2026]. EUA Publications.
Illingworth, S. (2026). What UK university AI policies actually do: A study of 96 institutions (HEPI Policy Note 71). Higher Education Policy Institute. https://www.hepi.ac.uk/reports/what-uk-university-ai-policies-actually-do-a-study-of-96-institutions/
Poole, G., Iqbal, I., & Verwoord, R. (2019). Small significant networks as birds of a feather. International Journal for Academic Development, 24(1), 61–72. https://doi.org/10.1080/1360144X.2018.1492924
Rogers, E. M. (2003). Diffusion of innovations (5th ed.). Free Press.
Roxå, T., & Mårtensson, K. (2009). Significant conversations and significant networks: Exploring the backstage of the teaching arena. Studies in Higher Education, 34(5), 547–559. https://doi.org/10.1080/03075070802597200
Taylor, K. L., Kenny, N. A., Perrault, E., & Mueller, R. A. (2022). Building integrated networks to develop teaching and learning: The critical role of hubs. International Journal for Academic Development, 27(3), 279–291. https://doi.org/10.1080/1360144X.2021.1899931
Vaessen, M., Van den Beemt, A., & de Laat, M. (2014). Networked professional learning: Relating the formal and the informal. Frontline Learning Research, 2(2), 56–71. https://doi.org/10.14786/flr.v2i2.92
Werner, E. E. (1993). Risk, resilience, and recovery: Perspectives from the Kauai Longitudinal Study. Development and Psychopathology, 5(4), 503–515.