Beyond Better Prompts: The Case for Critical AI Literacy

For this week’s edition of the Lifelong Learning Club I interviewed Sam Illingworth from Slow AI. Sam is a Full Professor of Critical AI Literacy based in Edinburgh and the author of GenAI in Higher Education (Bloomsbury, 2026).

I appreciate how he shifts our focus from how to best use AI to the interiority of the human using it; reminding us that behind every algorithmic output are the hidden costs of human labor, historical biases, and a potential erosion of our own voice. Over to Sam.

AI literacy, as it is commonly taught, focuses on competence. Write a better prompt. Pick the right tool. Integrate AI into your workflow. The assumption is that AI is a tool, and the key question is how well you use it.

Critical AI literacy asks a different question. Should you be writing the prompt at all?

The difference between the two is the difference between technical skill and human judgment. AI literacy teaches you how to use AI as a tool. Critical AI literacy teaches you to notice how AI as a tool is changing you while you use it.

When teaching AI literacy stops at the level of competence, we might miss out on three things.

First, AI is never neutral. AI models reflect the biases in their training data and the choices made by the organizations that build them. Here is an example I can’t stop thinking about: working-class Victorian newspapers were printed on cheap paper with bad ink. Because of that, computer scanners struggle to read them. Consequently, these newspapers aren’t part of the datasets used to train modern AI models. Without us realizing it, the tight budgets of 19th-century printers are shaping what modern AI treats as history.

Second, AI changes you. Lean on AI often enough, and you start to outsource your judgment. You lose confidence in your own voice and adopt AI slop without noticing. It’s easy to celebrate getting an answer in three seconds. But when you stop wrestling with a paragraph, the thinking that used to happen in the background simply dies.

Third—and this can be the trickiest to grasp—is knowing when to leave AI alone. Knowing when not to use AI is harder, and matters more, than knowing how to use it. The AI industry is built on the premise of endless integration, but critical AI literacy gives you the permission and the framework to simply leave the machine alone.

So my working definition is this. Critical AI literacy is the ability to evaluate outputs for bias and error, to recognise how the tool shapes your thinking over time, to judge when it helps and when it harms, to understand the human and ecological costs behind it, and to make deliberate choices about when to refuse it.

This approach is anti-thoughtlessness. The aim is to use AI with your eyes open.

Before using AI,, a critically literate person pauses to ask:

1. What am I giving up by not doing this myself? (Will I lose the thinking or creating process?)

2. Whose perspective is missing from this output? (Knowing AI represents only a fraction of the world.)

3. What would I lose if this output were wrong and I did not notice? (Recognizing that AI hallucinations can sound confident and fluent.)

4. Who benefits from me using this tool right now? (Remembering the corporate incentives at play.)

5. Would I trust this output if I did not know it came from AI? (Checking if you trust the actual content, or just the technology.)

Because educators are being told to ‘integrate AI’ with no framework for working out what it helps and what it quietly undermines.

Take AI detection tools, where we can already measure the harm. AI detection tools are wrong up to 30% of the time, and they disproportionately flag non-native English speakers. Students who wrote every word themselves have been accused of cheating. The detection industry is worth millions and its products do not work.

An example for this is the word ‘delve’ which has become a popular shorthand for catching AI slop. But what most people don’t realize is that ‘delve’ is common in Nigerian business English. Because many of the data labelers who trained these AI models were based in Nigeria and Kenya—often paid less than two dollars an hour—underpaid workers in the Global South effectively shaped the model’s language. So, an AI detector that flags ‘delve’ isn’t necessarily catching a machine. It is penalizing the linguistic patterns of marginalized people.

This is where I think a lot of people get wrong with regards to AI slop. For me, AI slop is not really about vocabulary or em dashes. There was slop long before the advent of AI, and it is defined by people not taking any pride or agency in their work. There is nothing wrong with using AI to help interrogate an argument, rewrite a sentence, or provided additional context. The majority of students that I work with know this, and they are very aware of the need (and desire) to keep humans in the loop.

Yet, universities are writing AI policies that open with ‘How do we detect cheating?’ instead of asking, ‘How do we design assignments worth doing?’ This detection arms race is a complete distraction. Critical AI literacy asks the question that a detection tool never will: What are we actually trying to teach, and does AI help or hinder that?

Beyond the distraction of detection, there is an even bigger risk for students who are still forming their judgment. Researchers name this: ‘never-skilling.’ A trainee who leans on AI early and completely may never build the underlying judgement at all. The scaffolding goes up before anything is built behind it.

That is the stake in education. We are at risk of producing a generation fluent in operating tools, but entirely unpracticed in the thinking those tools were meant to support. Critical AI literacy is how we ensure we are designing learning that is worth doing—whether AI exists or not.

Deskilling is losing a capacity you used to have because a tool now does it for you. This is related to what I’ve mentioned in the previous question: never-skilling, where you lean on AI so early that the capacity never forms in the first place.

For example, in a 2025 study, experienced doctors who had grown used to AI help spotting pre-cancerous growths during colonoscopies were tested again without it. Their unassisted detection rate fell from 28% to 22%. Seasoned specialists, who had the skill, leaned on the tool, and quietly lost some of it. It took months, not decades.

However, this is not as simple as having the ‘willpower’ to resist AI. Indeed, blaming yourself for taking that path is like blaming yourself for breathing the air in a room someone else filled with smoke. We have seen this move before, with the personal carbon footprint, a framing popularised by an oil company. Guilt keeps you looking inward at your own character instead of up at who built the conditions.

So the one practical thing I would offer is this: before you turn to an AI tool, get a rough draft written down yourself. This means that you continue to develop your skills and it also results in far more creative solution, as AI by its very nature always regresses to the mean.

One, mostly. Stop reading confidence as a signal of correctness.

The newer and more capable these models get, the more assured they sound, and the assurance is climbing faster than the accuracy. As frontier models reason for longer, their answers can grow more incoherent while their confidence keeps rising. This means that on the difficult questions, the ones you most need help with, the model is most likely to be fluently and convincingly wrong.

So treat fluency as a warning to slow down. The formatting will be perfect. The tone will be certain. The citation will look real. None of that tells you whether the thing is true. The tone of certainty reads the same whether the model is right or whether it has invented the answer whole.

The second thing to always remember is who the model speaks for. I am, by most measures, the person these systems were built to reflect: white, Western, male at the comfortable centre of the data they trained on. When AI produces the ‘neutral’ professional voice, or the ‘average’ expert, it is producing a version of me. Before you trust an output, ask who the assumed default user is, and whose perspective never made it into the training data.

I am not advocating for the abolition of AI. I use them every day and I am not a doomer. It means meeting them with your attention switched on. The model is a fast, fluent, confident first draft of a thought. The judgement about whether that thought is any good has to stay with you.