AI adoption is visible. AI fluency is not.
That visibility gap is why twoprune is making a public research commitment. We are working to develop task-bounded AI-fluency assessment through internal research and external academic collaboration.
The work starts with construct boundaries, evidence-centred scenario design, integrated process-and-outcome evidence, opportunity-to-observe, explicit missingness and confidence handling, and a responsible validation programme. These are not supporting details around the product. They are the path from a promising product architecture to a defensible measurement method.
We are in early conversations with researchers and labs working across HCI, assessment, learning science and human–AI collaboration. These are early conversations, not announced partnerships. We are publishing the direction now because the construct should be challenged while it is being developed—not presented later as a finished answer.
The need is concrete. Organizations can see who completed an AI course. They can count logins, prompts and active users. They can ask people how confident they feel, test their knowledge of prompting, or inspect the finished work. All of that can be useful. None of it, by itself, reveals how a person actually handled the work.
The same output can conceal two different workflows
Imagine two people receive the same source pack and produce similarly polished recommendations.
One person frames the decision before prompting. They notice that two sources cover different periods, ask AI to surface contradictions, check the material claims against the original evidence and state what remains uncertain. They use AI extensively, but retain control over the conclusion.
The other accepts an elegant synthesis containing an unsupported assumption. The document reads just as well. A final-output review may make the two people look identical.
The reverse problem also occurs. A careful, selective user may appear less advanced than a high-volume user who rarely challenges an answer. Usage tells us that AI was present. It does not tell us whether the human used it well.
Existing measures capture useful pieces
The answer is not to dismiss existing approaches. AI-literacy scales can reveal perceived capability and attitudes. Knowledge tests can assess whether someone understands important concepts. A 2024 systematic review identified 16 validated AI-literacy scales across 22 studies, including both self-report and performance-based measures, while also finding that evidence for some measurement properties remained incomplete.[1]
