Organisations across sectors are adopting AI without a map. Budgets are approved, vendors are chosen, tools are deployed, often before the people using them understand what they are actually adopting. Two MIT researchers, Fiona Hollands and Cynthia Breazeal, published a paper in 2024 that quietly challenges this order: before you deploy AI, the people who will live with it need to understand it first.
Literacy is the antidote to rushed adoption
Their argument is not new. It echoes what educators have always known: literacy precedes effective use. It becomes urgent when the subject is a technology that reshapes how decisions are made, how data is handled and where competitive advantage comes from.
What the research says
Hollands and Breazeal, working through MIT's RAISE initiative (Responsible AI for Social Empowerment and Education), describe a set of short, modular AI literacy curricula developed with the implementation partner i2Learning and released under a Creative Commons licence. The paper, published in The Science Teacher (vol. 91, no. 2), reports on the 2022-23 school year: an online questionnaire answered by 265 people worldwide, including 190 teachers who had used the curricula with roughly 12,000 students across 136 countries.
The findings are qualitative and self-reported, and the authors present them as such. Teachers reported three shifts:
- Knowledge of AI concepts increased. On the study's 0-to-10 scale, teachers' self-rated understanding rose from 3.8 to 6.0; their estimate of their students' understanding rose from 2.4 to 4.3. People moved from a vague idea of "machine learning" to knowing how models are trained, where bias enters and what a hallucination is.
- Optimism about AI's benefits grew. Teachers went from 5.6 to 7.1 on the same scale. It was, and this is the part that matters, informed optimism, formed after learning how the technology fails.
- Confidence rose. Most respondents said they now felt able to contribute to the future of AI literacy; fewer felt they could shape the future of AI itself. The gap is honest, and worth remembering.
Day of AI, the programme built on these curricula, has since grown well past the study: MIT RAISE's 2025 impact report counts more than two million students and 50,000 educators in 175 countries. Literacy has become a movement, at least in schools.
The IWH angle: autonomy before adoption
Here is where organisations tend to go wrong. The MIT framing is literacy for adoption: understand AI, then use it responsibly. That is a reasonable goal for a classroom. For an organisation the goal has to be one step wider, because understanding AI sometimes leads to a different conclusion.
Our position is that AI literacy should serve informed autonomy, not efficient adoption. Literacy that gives you the knowledge to say yes and the standing to say no.
The practical implication is stark. You need governance before tools. You need compliance review, data readiness and stakeholder alignment before deployment. And you need literacy so that people can ask the right questions:
- Does this tool belong in our workflow, or in a demo?
- What data are we feeding it, where does it go, and who has access?
- If it fails, what does the failure cost, and who notices?
- Are we solving a real problem or following a trend?
In the European compliance context (GDPR, NIS2, DORA, and the AI Act's obligations arriving through 2026 and 2027) none of this is optional. But even without regulatory pressure, organisations that skip literacy tend to deploy AI in ways that create long-term risk: shadow AI accounts, data leaving the building in prompts, models fed unvetted datasets, tools that amplify existing bias in hiring, underwriting or customer service.
The distinction that matters: the MIT study measures success as increased knowledge and readiness to adopt. We measure success as appropriate adoption, which sometimes means non-adoption.
The practical sequence
If you are considering AI deployment, the order matters more than the speed.
- Build literacy first. Short, modular, department-specific. Operations learns about process automation and its failure modes; finance learns what AI does and does not do in forecasting and compliance; HR learns about bias in predictive hiring. The MIT material shows this can be done in hours, not semesters.
- Establish governance before the first model runs. Data ownership, audit trails, an escalation path for anomalies, a review cadence. The same structures your security and compliance programme already uses; extend them rather than duplicating them.
- Ask the hard questions with authority. Literacy is what gives people the standing to be sceptical. Is this tool closing a gap or filling a budget line? Are we building organisational knowledge or outsourcing thinking?
- Plan for failure. Every model has edge cases and every vendor has an outage. Literacy includes knowing where the brittleness is and what happens when it breaks.
This approach is slower. It avoids the expensive reversal: deployed AI that creates a compliance finding, erodes customer trust or embeds a discriminatory pattern that takes a year to notice and longer to unwind.
Why it matters now
The window for literacy-first adoption is narrowing. As tools become easier to use, the temptation to skip the thinking and go straight to implementation grows, and the cost of that shortcut (regulatory penalties, reputational damage, teams disempowered by tools they cannot explain) keeps rising.
IWH's AI advisory and compliance work sits exactly in this space: building the literacy, the framework and the controls that make AI adoption a choice rather than a reflex.
The MIT research gives you the evidence. The decision is yours: adopt first and learn later, or learn first and then decide what to adopt.
Source: Hollands, F. & Breazeal, C. (2024). Establishing AI Literacy before Adopting AI. The Science Teacher, 91(2). Day of AI reach figures from the MIT RAISE 2025 Day of AI Impact Report.