How this research works

This page sets out how the research on this site is produced: where sources come from, how they are graded, how contradictory evidence is handled, where AI is used and where it is not, how corrections work, and how terms are attributed. It exists so that any claim here can be challenged against a stated standard rather than against a pile of links, and so that the standard can itself be criticised.

What this research is, and is not

It is practitioner research combined with an open synthesis of published evidence. It is systematic, longitudinal and cross-sector, conducted through direct engagement with organisations rather than through academic publication.

It is not peer-reviewed, and does not claim to be. It has not been through ethics review, it does not use randomised designs, and its sampling is not representative. Those are real limits and they are stated on the research pages rather than buried here. Where a claim rests on the practitioner base rather than on published studies, the page says so.

One further limit worth stating plainly: the practitioner research has not yet been published with full methodological detail. How many interviews, how many survey waves, who the respondents were, and what counts as an organisation being researched are not yet on this site. That is a gap, it is logged in the corrections ledger, and it is being written up.

How sources are selected

A source enters the evidence base if it bears directly on what increasingly capable AI does to human capability, and if its method can be described in a sentence. Studies are sought out for what they measure rather than for what they conclude, which is why the base contains findings that cut against the argument made elsewhere on this site.

Every external source is fetched and confirmed to resolve before it is published. No link is included on the strength of a citation seen elsewhere. Where a paper has a correction or retraction notice, it is linked alongside.

Institutional reports are included rather than excluded. They are often the best available data on scale, sentiment and demand. But several are published by organisations that sell the remedies they recommend, and where that is true the entry says so. No source is dismissed for its origin, and none is granted authority for it either.

How evidence is graded

Two things are recorded separately: what a source is, and how much weight it is allowed to carry.

Sources are typed as peer-reviewed, working paper, institutional survey, institutional modelling or compiled review. They are then placed in one of four tiers: A peer-reviewed research, randomised trials, systematic reviews, official statistics; B credible working papers, large field experiments, institutional research with a transparent method; C corporate and institutional surveys, commercial datasets, practitioner research; D expert interpretation, books, journalism, individual cases.

The rule that follows is the important part. Language tracks the tier. "The evidence shows" for A, "findings suggest" for B, "reported" or "self-reported" for C, "argued" or "observed" for D. Where a page rests on a single study it says so, and says what would strengthen the claim. Volume is not strength: five consultancy surveys do not outweigh one good field experiment.

Every entry also carries a field almost nobody publishes: what the study does not support. That field exists because most claims about AI and human thinking rest on a handful of papers, and most are quoted well beyond what their design can carry.

How contradictory evidence is handled

It is published, at the same prominence as supporting evidence. What we actually know about AI and human capability sorts nineteen claims into strong, emerging and unknown, and three of the unknowns undercut positions argued elsewhere on this site. Capability debt is listed as a framework without a validated instrument. Whether removing junior work impairs senior capability formation is listed as unproven, and it is a load-bearing assumption here.

Where evidence changes, the position changes and the change is dated. Where evidence is genuinely split, both sides are given rather than the convenient one. A page that only ever gains confidence is tracking commitment rather than evidence.

Where AI is used, and where it is not

This matters more here than on most sites, given the subject.

AI assists with: locating candidate sources, drafting and restructuring prose, checking internal consistency across a large estate, generating structured data, and mechanical quality assurance such as link checking and schema validation.

AI does not decide: what is true, what a study supports, how strong the evidence is, what position to take, or what to attribute to whom. Every source is opened and read. Every figure is checked against the original. Every claim of authorship is verified against a dated record before it is published, and where it cannot be, no claim is made.

This is Human at the Start applied to the research itself: the question, the constraints and the rejection criteria are set before anything is generated. The Delegation Boundary Map is the working version of the same idea, and it is used here rather than only recommended.

The honest caveat: this arrangement is only as good as the verification, and verification is only as good as the verifier's ability to detect the error. That is the same limit this research argues applies to everyone else, and it applies here too.

How terms are attributed

One rule. A term is credited to Rahim Hirji only where a dated first publication exists, and it is phrased as "has used the term since at least [date]", because earlier private or spoken use cannot be excluded and should not be implied. Terms without a dated first use are marked as used by SuperSkills, with no claim of priority. Established concepts from the research literature are attributed to their originators, and are never presented as coinages from this work.

Four terms currently carry dated claims. Four have had claims withdrawn for lack of a date. One, capability debt, has documented independent prior use by another author, which the page states. The full record is in the corrections ledger.

Provenance language is used precisely and not interchangeably. Predicted means forecast before the event with a dated source. Identified means recognised an emerging phenomenon others had not named. Developed means built an interpretation on existing ground. Coined means originated the terminology, with documentary evidence of first use.

How corrections work

Publicly, at /research/corrections, with the original claim quoted, what was found, and what changed. Nothing is removed from the log, including corrections that weaken an argument made here. Pages are reviewed on a rolling basis, with fast-moving subjects such as model capability on a 90-day cycle, and every page carries the date it was last reviewed.

Corrections from readers are welcome and are credited.

Conflicts and how they are disclosed

Rahim Hirji is the author of SuperSkills (Kogan Page, 2026) and earns income from speaking and advisory work grounded in the arguments made here. That is a straightforward commercial interest in these arguments being taken seriously, and it is the reason the evidence grading, the unknowns and this ledger exist rather than a claim that the interest does not exist.

No organisation funds this research. No entry in the evidence base is paid for, sponsored, or included at a third party's request. Where a cited institution is also a commercial partner or a client, the entry says so.

Reuse

The evidence base, the canon and the questions map are published as structured data under a Creative Commons Attribution licence, and individual entries carry permanent anchors so they can be cited on their own. Frameworks are free to use and adapt with attribution. For commercial licensing, training use or reproduction in publications, get in touch.

Related

The evidence base · what we actually know · corrections · the questions map · the canon.