In short: a multi-agent research system gathers, cross-checks and organizes evidence for us. People with doctoral-level expertise make every judgment, and we tell you openly which AI systems we used and for what.
Our position
AI lets a small, independent team review far more evidence, in more languages, and more often than traditional ranking processes allow. It also introduces risks: errors, invented sources, hidden bias and opaque reasoning. This policy explains how we capture the benefits while keeping humans fully accountable.
The core rule is simple: AI is an information aggregator, not a judge.
How the research desk works
Evidence is gathered by RankingLab's research desk, a multi-agent system that runs dozens of research agents in parallel across several leading AI research engines. For every candidate, one engine scans official sources field by field, a second engine checks each value independently, and where they disagree a third model reads the cited pages and decides from the source itself. Every value carries its source, an exact quote and a confidence level, and every quote is checked against the live page. All agents work to RankingLab's published rubric and research protocols. People decide: editors approve each step, and PhD assessors assign every score.
Specialist models are assigned to each task in the research protocol, and every agent is directed by RankingLab's own rubric, research protocols and checks. Doctoral-level experts are in the loop at every stage: they shape the research, review its results and make every judgment.
Permitted uses
Approved AI tools may be used to:
- Retrieve public information, such as institutional publications, curricula, accreditation reports and research outputs.
- Extract structured facts from documents, for example course titles, credit requirements or reported outcomes.
- Cross-check each value between independent engines, and check every quoted passage against the live source page.
- Classify evidence against our assessment rubric so that assessors can review it efficiently.
- Summarize long documents and research literature for assessors.
- Translate sources, so that evidence published in languages other than English is not overlooked.
- De-duplicate records and reconcile names and identifiers.
- Detect anomalies, such as inconsistent figures or implausible changes between editions, so that people can investigate.
Every claim extracted or summarized by AI must link to its original source. A claim without a verifiable source is not used.
Prohibited uses
The following are not allowed under any circumstances:
- AI final scoring. AI may not assign, change or approve any pillar score, overall score or rank. Scores come only from qualified human assessors, as described in the Methodology.
- Publishing AI-generated text without human editorial review. Every published text is reviewed, edited where needed, and approved by a responsible person.
- Exposing confidential material. Confidential institutional submissions, assessor notes and personal data may not be entered into any tool that uses inputs to train models or that does not offer adequate confidentiality terms.
- Judgments about individual people. We do not use AI, or any other method, to rank, profile or evaluate individual people.
- Generating fake evidence. AI may not be used to create sources, quotes, data or images that are then presented as real.
Human review and accountability
- Each institution or program is scored by at least two independent assessors who hold a PhD or equivalent doctorate and have research experience. They score blind to each other.
- If two assessors differ by more than 15 points on the 100-point scale for any pillar, a third senior assessor adjudicates.
- Assessors can and do reject AI-gathered evidence that is inaccurate, irrelevant or poorly sourced.
- Editorial responsibility for every publication rests with RankingLab, through its editorial panel, which reviews and approves everything before it is published. Contact: hi@rankinglab.org.
Note on the EU AI Act
Article 50(4) of the EU AI Act, which applies from 2 August 2026, requires disclosure when AI-generated text is published to inform the public on matters of public interest, unless the text has undergone genuine human review or editorial control and a person or organization holds editorial responsibility. Our published content is subject to that kind of review and responsibility. We disclose our AI use anyway, voluntarily, because transparency is part of our method. Ranking institutions is not one of the high-risk uses listed in Annex III of the Act.
Register of AI systems
For every edition we publish a register of the AI systems used, including:
| Field | What we record |
|---|---|
| Provider | The organization that supplies the system |
| Model family | The model or model family used |
| Purpose | The tasks it performed, such as retrieval, translation or summarization |
| Data handling | Whether inputs are retained or used for training |
The register is published with the edition notes on the Updates page.
Bias controls
AI systems can amplify existing imbalances. We actively manage this:
- Language bias. Much research and institutional information is published in English. We use multilingual retrieval and translation so that institutions publishing mainly in other languages are assessed on equal terms.
- Visibility bias. Large, well-known institutions produce more online content. Assessors are instructed to judge the quality of evidence, not its volume.
- Source checks. We compare AI outputs against original documents on a sample basis every edition, and we record error rates.
- Calibration. Assessors complete calibration exercises that include AI-assisted evidence packs, so that they know where tools tend to fail.
Logs and audit
We keep records of AI-assisted steps, including the tool used, the date, the task and the sources retrieved, for at least the life of the edition plus two years. For every value we record which engine reported it, with its source and quote, and how it was settled: by agreement between engines, by a model reading the source, or by an editor, whose decision is logged under their name. These logs support corrections, appeals and independent audit. They do not contain more personal data than necessary.
Questions and concerns
If you believe AI use has led to an error in a published result, please use the Review Requests & Corrections Policy. General questions about this policy can be sent to hi@rankinglab.org.