AI frameworks compared

Six frameworks for using AI well, each set against the established alternatives rather than presented on its own. The comparison is the point: on most of these questions somebody has already published an answer with more testing behind it, and knowing which one to reach for matters more than adopting any single set. Where the outside framework is better, these pages say so.

Which one answers which question#

Think, AI, think is the spine. Write your own position first, use the model, then decide what to keep. Everything else sits underneath it. Nearest published relatives: Dell'Acqua's centaur and cyborg patterns, and Mollick's seven approaches for students.

The four levels, Extract, Explore, Examine, Extend, name what you are asking for. Nearest relatives: Bloom's revised taxonomy, which describes the learner, and SAMR, which describes the technology. Bloom is the evidenced one.

Goal, context, friction, standard is how to write the request. Nearest relatives: Lo's CLEAR, Google's TCREI, CO-STAR. CLEAR is tighter and TCREI is easier to teach; the only thing this adds is a slot for what the model should hand back to you.

Keep, share, hand over decides what gets delegated at all. Nearest relatives: Sheridan and Verplank's ten levels of automation, and Parasuraman, Sheridan and Wickens across four processing stages. Those are far better specified and are what a safety engineer should use.

The five rungs, Chat, Project, Skill, Automation, Agent, describe the tooling rather than the skill. Nearest relatives: Anthropic's workflow and agent distinction, and SAE's levels of driving automation.

The source rule handles a reference a model gave you. Nearest relatives: Caulfield's SIFT and the CRAAP test. SIFT is better and that page says so; the one amendment is that a model requires you to check a source exists before checking whether it is any good.

The standard every framework on this page fails#

In 2016 Hamilton, Rosenberg and Akcaoglu reviewed SAMR, at that point one of the most widely taught models in educational technology, and found it largely absent from the peer-reviewed literature despite heavy practitioner adoption, with thin theoretical and foundational evidence. They named three faults: the absence of context, a rigid hierarchy implying that higher is better, and an emphasis on product over process.

Every framework listed above is vulnerable to at least two of those, and none of the six has been tested against an alternative. What is evidenced is the mechanisms they are built on, which is a weaker claim and the accurate one. A memorable structure is a teaching aid, and the reason to prefer one is that it fits the decision in front of you rather than that it has been shown to work.

What the evidence underneath them does support#

Three findings recur across these pages and each is measured rather than argued.

Interface beats policy. Nearly a thousand school students split between unrestricted GPT-4, a hints-only tutor and nothing: with the tool removed, the unrestricted group scored 17 per cent below students who never had it, while the tutor group kept most of its gain. Same model, different interaction.

The gap is invisible until you measure unaided. Among 78 novice programmers, both AI groups produced working code and looked identical on every measure taken, until the tool was cut off and unrestricted users failed at 77 per cent against 39 for a scaffolded group.

It happens fast. In randomised trials with 1,222 people, the withdrawal effect appeared after roughly ten minutes and showed up as reduced persistence rather than lost knowledge.

These are older than the pages that describe them#

All six are taught material, carried forward through successive versions of a deck and most recently delivered in September 2026. None has a dated first publication, and each page says so in its own words rather than implying a coinage. They are also due a revision: the five rungs in particular track product features that did not exist as named things three years ago, and a ladder pinned to a vendor's menu will need rewriting.

They are offered as unique rather than as best. Where an established framework does the job better, the page for that framework names it and recommends it.

Key sources

• Hamilton, E. R., Rosenberg, J. M. and Akcaoglu, M. (2016). The SAMR Model: A Critical Review and Suggestions for Its Use. TechTrends, 60(5).
• Lo, L. S. (2023). The CLEAR path. The Journal of Academic Librarianship, 49(4).
• Wineburg, S. and McGrew, S. (2019). Lateral Reading and the Nature of Expertise. Teachers College Record, 121(11).
• Anderson, L. W. and Krathwohl, D. R. (eds.) (2001). A Taxonomy for Learning, Teaching, and Assessing. Longman.
• Bastani, H. et al. (2025). Generative AI Without Guardrails Can Harm Learning. PNAS, 122(26).

Related SuperSkills research#

The argument these operationalise, drift versus design. Applied to a student, how to use AI at university. Applied to a junior, how juniors become senior. On why a framework is not a substitute for measuring, the capability audit. On how this research grades what it cites, how this research works.

About these frameworks#

The six frameworks described here are used by Rahim Hirji in teaching and in the Mastering AI deck. No claim of first use is made for any of them except the source rule, which was written on 6 September 2026 and is dated from that page. Bloom's taxonomy, SAMR, CLEAR, TCREI, CO-STAR, SIFT, the CRAAP test and the levels of automation all belong to the authors named on the individual pages. The judgement that a framework should be chosen by which decision it fits, and that most of these have less behind them than the mechanisms they rest on, is an interpretation by Rahim Hirji, author of SuperSkills (Kogan Page, 2026), and is marked as an interpretation and not a finding.