The four levels of AI use

Four things people ask a model for, in ascending order of what they get back. Extract is "give me the answer". Explore is "help me understand this". Examine is "tell me where I am wrong". Extend is "help me build something I could not build alone". Almost nobody leaves the first, and the ladder is about the request rather than about the tool, so moving up costs nothing and takes about ten seconds.

The four levels#

1. Extract. "Give me the answer." Faster output, weaker learning. This is where almost everybody is.

2. Explore. "Help me understand this." You come away actually understanding it.

3. Examine. "Tell me where I am wrong." You get better at spotting what is true.

4. Extend. "Help me build something I could not build alone." You can do things you could not do before.

Two established ladders this is standing next to#

Ordering cognitive demand is not a new idea and the two dominant answers are both older than this one.

Bloom's revised taxonomy, restated by Anderson and Krathwohl in 2001, runs remember, understand, apply, analyse, evaluate, create. It is the standard vocabulary in education worldwide and it describes what a learner is asked to do. That is where it comes apart on contact with a model, because a model will perform every one of those six categories on request, including create. A taxonomy of what the learner produces stops discriminating once something else can produce it.

SAMR, from Ruben Puentedura, runs Substitution, Augmentation, Modification, Redefinition and describes what a technology does to a task. It is the closest structural analogue to this ladder and it is very widely taught.

It is also the cautionary case, and the caution applies here. Hamilton, Rosenberg and Akcaoglu, reviewing SAMR in TechTrends in 2016, named three problems: the absence of context, a rigid hierarchy implying that higher is always better, and an emphasis on product over process. They also recorded that SAMR is largely absent from the peer-reviewed literature despite heavy practitioner adoption, and that its theoretical and foundational evidence is thin.

Read that paragraph as being about this page too. A four-rung ladder with alliterative labels and an implied direction of travel is the same object, and it has less behind it than SAMR does.

What this ladder measures that the other two do not#

Bloom describes the learner. SAMR describes the technology. Neither describes the request, and the request is the only one of the three a person controls at the moment of typing.

That distinction is what makes the ladder usable rather than descriptive. "Where am I on Bloom's taxonomy" is a question a teacher answers about a curriculum. "Which of these four did I just ask for" is a question anyone can answer about the last thing they typed, and the answer is checkable in their own chat history.

The evidence relevant to the ordering is about the interaction rather than the label. Across nearly a thousand school students, an unrestricted group that took answers scored 17 per cent below students who never had the tool once it was removed, while a hints-only tutor group kept most of its gain. That is level one against level two, measured, in one subject. It supports the direction of the ladder and it does not validate the ladder.

Higher is not always better, and the hierarchy is the weak part#

Hamilton and colleagues' second objection lands squarely. Extract is the right level for a great deal of work: converting a file, finding a date, reformatting a table. Somebody operating at level four on a task that needed level one has wasted an afternoon.

The ladder is a diagnostic for a pattern rather than a target for a task. If every request you made last week was Extract, that is worth knowing. If a particular request was Extract, that is usually fine.

What this has not been shown to do#

Nothing has tested it. There is no study comparing people taught these four levels against people taught SAMR, Bloom or nothing, and no measurement of whether naming a level changes what anybody asks for. It sits in the position Hamilton and colleagues describe: adopted through teaching, absent from the literature, and plausible.

If you want an evidenced framework for ordering cognitive demand in a curriculum, use Bloom. If you want one for judging whether a technology changed a task, SAMR is better specified and more widely understood, with the criticism above attached. This one earns its place only on the narrow ground that it names the request.

Key sources

• Anderson, L. W. and Krathwohl, D. R. (eds.) (2001). A Taxonomy for Learning, Teaching, and Assessing. Longman.
• Hamilton, E. R., Rosenberg, J. M. and Akcaoglu, M. (2016). The SAMR Model: A Critical Review and Suggestions for Its Use. TechTrends, 60(5).
• Bastani, H. et al. (2025). Generative AI Without Guardrails Can Harm Learning. PNAS, 122(26).

Related SuperSkills research#

The sequence this sits inside, think, AI, think, and the prompt structure that moves you up it, goal, context, friction, standard. On the tooling ladder rather than the request ladder, the five rungs. On what level one costs, cognitive offloading and productive struggle.

About this framework#

Extract, Explore, Examine, Extend is used by Rahim Hirji in teaching and in the Mastering AI deck, most recently in September 2026. No claim of first use is made. No dated first publication exists for it and the Box of Amazing archive carries none, so it anchors to SuperSkills (Kogan Page, 2026) and to the deck. Bloom's taxonomy and its 2001 revision belong to Benjamin Bloom, Lorin Anderson and David Krathwohl, and SAMR to Ruben Puentedura. The reading offered here, that Bloom describes the learner, SAMR describes the technology and neither describes the request, is an interpretation by Rahim Hirji and is marked as an interpretation and not a finding.