Deep Learning: Which Groups Are Slow, Reluctant, and Refusing?

Deep learning is currently moving from policy into practice. At this stage, the important question is not only whether the concept is sound, but also how it spreads and is adopted by implementers. This is where Everett M. Rogers’ diffusion of innovations theory becomes relevant. Rogers explains that an innovation is never accepted by everyone at the same time. Adoption speeds differ because of perceptions of benefit, compatibility, complexity, the possibility of trial, and the observability of the innovation’s results (Rogers, 2003).

Rogers divides adopters into five groups. Innovators, around 2.5 per cent, are the group quickest to try new things. Early adopters, at 13.5 per cent, follow more selectively and often serve as social references for other groups. The early majority, 34 per cent, only move once they have seen evidence that the innovation makes sense and works. The late majority, also 34 per cent, tend to be more sceptical and wait until the innovation becomes increasingly common practice. Meanwhile, laggards, at roughly 16 per cent, are the slowest group; some only change when environmental pressure is strong, and some may even cling to old practices (Rogers, 2003).

These proportions are certainly not the result of measuring deep learning implementers in Indonesia. They are more appropriately used as an analytical lens to understand that policy adoption speeds can never be uniform. These categories are also not permanent identities or labels attached to individuals; they primarily describe a relative position within the adoption timeline of an innovation.

A DIFFERENT APPROACH

Deep learning deserves to be read as an innovation because it demands a transformation in the way of learning: mindful, meaningful and joyful learning; learning experiences that involve understanding, applying and reflecting; and support through pedagogical practice, learning environments, partnerships and digital utilisation (Kemendikdasmen, 2025). With demands for change that large, it is impossible for all implementers to move at the same speed.

Rogers also reminds us that an innovation is adopted more quickly if it has relative advantage, meaning it is perceived as better than previous practice; compatibility, meaning it fits the adopters’ values, experiences and needs; low complexity; trialability, meaning it can be tried on a limited basis; and observability, meaning its results can be seen. The term divisibility, often used to describe innovations that can be applied bit by bit, is essentially close to the notion of trialability. The easier it is to try, adapt and demonstrate the benefits of deep learning, the faster its adoption will be.

The implications for policy are highly strategic. Innovators should not be burdened with excessive procedures. They need room to experiment, build prototypes and produce examples of practice. Early adopters need access to evidence, networks and spaces for sharing, because they are the prospective opinion leaders who can give deep learning social legitimacy. The early majority cannot be won over with slogans alone; they need tested models, clear guidelines, implementation examples, mentoring, and evidence that deep learning can improve the quality of learning.

The late majority requires yet another approach. This group usually only moves once uncertainty has diminished. They need simple practices, low implementation risk, close technical support, and evidence from environments they trust. As for the slowest group, they should not be immediately branded as opponents of change. They may face resource constraints, bad experiences with previous changes, or simply have not yet seen the relative advantage of deep learning. A policy that merely issues commands can actually amplify resistance.

DIFFERENTIATED POLICY

For this reason, the implementation of deep learning should not adopt a ‘one size fits all’ philosophy. One size fits all may be administratively practical, but it is weak as a theory of change. Groups with different levels of readiness require different interventions. Some simply need room to act, some need to be convinced with evidence, some require intensive mentoring, and some must first be helped to reduce the complexity and risk of change.

The ministry can build a differentiated policy by mapping implementers’ readiness, rather than merely administrative compliance. Schools or regions that are already innovators and early adopters can serve as practice laboratories and demonstration centres. The early majority can be strengthened through the replication of proven practices. The late majority needs more systematic mentoring. Meanwhile, the slowest group needs a diagnosis of the underlying causes: whether the problem lies in perceived benefit, compatibility, complexity, resources, or trust.

Deep learning will fail as a transformation if its success is measured only by how quickly the term becomes known. The diffusion of innovation demands more than socialisation; it demands changes in belief, practice and experience. The ministry’s challenge is not to make all implementers move in unison, but to help each group move forward from its starting position. If policy can read these variations in readiness, deep learning will not remain a central policy, but will genuinely diffuse into a culture of learning.