Evaluating AI search tools with confidence

A practical framework and browser extension for assessing search tools used in academic research, with special attention for AI-related aspects. Developed by information specialists from the University of Twente library. The TRUST framework is built up from five core principles: Transparent, Reliable, User-centric, Sound, and Traceable, inspired by the FAIR principles.

Install the extension

Install the extension via the Chrome Web Store by clicking here. For other installation options, read on.

Try in browser

Want to test the review tool without installing? A browser-based preview is available — use the Try button above.

View a report

See a finished, full TRUST review here in the form of a report exported by the review tool.

T
Transparent
How the tool works, where its data comes from, and what it cannot do.
R
Reliable
Results are accurate and verifiable, with real sources instead of fabricated ones.
U
User-centric
Works well for researchers and fits into everyday academic workflows.
S
Soundness
Search results are fair, unbiased, and ranked accurately.
T
Traceable
Every claim can be followed back to its original source.

Background

We’re showcasing the TRUST framework, the tooling built to support it, and our workflow at LIBER 2026. You can find our poster & abstract proposal here.

Poster

We're presenting the TRUST framework at LIBER 2026 (Trondheim, 1–3 July 2026).

Download the poster (PDF)

Zenodo DOI will be added after publication.

Conference abstract

We have proposed the TRUST framework for LIBER 2026 — a short paper making the case for a rigorous, principle-based way to evaluate AI-based search tools.

The extension

To make it easy to review a tool with the TRUST framework, we developed a free and open source browser extension. It’s available to install directly, but you can also find an online demo version below, as well as an example of a finished report generated by the extension.

How a review works

A TRUST review needs several forms of input: quality gates, scored questions, evidence, metadata, and final opinions. The tool turns this into an exportable & shareable report.

  • Quality gates: four yes/no checks covering privacy, intellectual property, and accessibility.
  • Scored questions: two per principle across five principles (ten in total), each scored 0 to 3.
  • Evidence: screenshots and notes captured in the browser, tagged to the question they answer.
  • Metadata: tool, version, and reviewer details recorded for every evaluation.
  • Final opinions: the reviewer’s conclusions and recommendations, captured alongside the scores.
  • Output: all of the above turned into a standalone, shareable HTML report.

Install the browser extension

Or try the live in-browser demo first, with no install.

Chrome and Edge

Add to Chrome from the Chrome Web Store; it auto-updates on new releases.

For manual installs or other Chromium browsers, see the GitHub Releases.

Firefox

Add to Firefox from Firefox Add-ons; it auto-updates on new releases.

View on Firefox Add-ons.

  • v0.9.6 Current release version
  • Apache-2.0 Open-source license — free to use and modify
  • Releases Release notes and downloadable builds on GitHub

Try it in your browser

Run the full extension in your browser — no installation required. It comes preloaded with a real review (ai2 Asta) so you can click through the evaluation, inspect the evidence, and read the generated report. New screenshots you capture in the demo are simulated, but everything else is live.

The TRUST web demo showing the evaluation tab: a split view with a simulated browser pane and the review sidebar scoring ai2 Asta against quality gates and rubric questions. The TRUST web demo showing the captures tab with a grid of evidence screenshots tagged to rubric questions.

Open the demo Works best on desktop.

See a complete example

A full, real TRUST review exported from the extension — the ai2 Asta evaluation referenced in the abstract.

View the example report Download the .zip

The report is a single HTML file that opens in any browser. The .zip holds the full review package, which you can load into the extension or drop into the Compare tool below.

Compare reviews

Upload multiple TRUST review ZIP files to compare scores, evidence, and verdicts side by side.

No review ZIPs handy? Download the ai2 Asta example to try it out.

Drop ZIP files here or

What’s next

While the extension and framework are ready to be used, there is still a lot to do. We are currently working on integrating TRUST into our daily work. Below you’ll find some examples of how we’re approaching that.

Central database

One of the immediate next steps we’re taking is creating a shared database that can hold all tools we know of: either because we already use them or because they are nominated, suggested, found, etc. This database will hold all relevant information about the tools, including the current TRUST review information, like progress, who reviewed, scores, conclusions, etcetera. Each column shows the average score for one principle, rated from 0 (does not meet it) to 3 (fully meets it): TR, RE, US, SE, TC. Nominate a tool.

Illustrative draft data. The table below is meant as an illustration of the idea. The actual database is being designed, and will store a lot more information, probably split over several joined tables/sheets.
Tool Category Verdict TR RE US SE TC Status
Loading…

Next steps

A single review is one step in a larger workflow — alongside contract management, the AI office, and the colleagues we support. This is what we are building next.

  1. 01

    Central database

    Build a central repository for every evaluation: tool submissions, review results, and procurement decisions. See the draft database →

  2. 02

    Monitoring & re-evaluation

    Establish a process to keep reviews current and trigger re-evaluation whenever a tool changes.

  3. 03

    Golden datasets

    Curate sets of queries with expected results, used during review for objective, comparable evaluations.

  4. 04

    Refine the workflow

    Implement the full process in collaboration with relevant departments — contract management and the AI office.

  5. 05

    Integrate into daily work

    Embed the workflow into outreach, education, and research support across the institution.

  6. 06

    Practice what we preach

    Apply TRUST to our own procurement decisions beyond this pilot.

Links and contact

Contact

Samuel Mok, s.mok@utwente.nl

Embedded Information Services, University of Twente

GitHub Repository
github.com/utsmok/review-extension
LIBER 2026 Conference
1 to 3 July, NTNU Trondheim, Norway
University of Twente, LISA-EIS
Embedded Information Services

Conference abstract · proposed for LIBER 2026

S. Mok, S.A. Kruit, G. de Jonge, J.M. van Eck — Embedded Information Services, Library, ICT Services & Archive, University of Twente.

Novel information search tools are changing how academics access knowledge. These AI-driven systems should augment, not replace, users’ critical thinking and information-literacy skills. As information specialists, our role is to act as informed intermediaries, guiding the scientific community toward tools that enhance research without compromising academic integrity [1]. Several principles and guidelines for designing and using trustworthy AI have been developed [2] alongside metrics for assessing these principles [3]. However, no current specific, rigorous methods to evaluate AI-based search tools exist. This abstract introduces the TRUST framework, a practical tool designed to evaluate AI-based search tools. It measures whether recommended tools are reliable, ethical, and effective, and match scholarly and governmental standards of rigor and transparency.

The TRUST framework is built on five core principles, inspired by the FAIR principles for data management [4]. Each tool review grades the subject on these aspects:

  • Transparent: Methodology, data sources, and limitations are clearly documented.
  • Reliable: Outputs are accurate, verifiable, and consistent. Sources exist and are not hallucinated.
  • User-centric: Fit for academic use and integrates with standard research workflows.
  • Soundness: Algorithmic fairness, bias mitigation, and accuracy of search results and rankings.
  • Traceable: All information generated and used can be traced back to a primary source.

Literature stresses that AI models operate as “black boxes,” making documentation alone insufficient for deep trust and transparency [5]. The TRUST evaluation protocol addresses challenges such as “consistent and reproducible results” from AI-based literature searches. However, this remains a critical point of failure in the current AI landscape due to the methods of retrieval and synthesis of information AI tools use [6]. Similarly, the framework checks if output is “factually accurate and verifiable” and “faithful to source material”, which especially generic large language models (LLMs) consistently fail to provide [7]. These evaluation protocol items are the main reason general LLMs are mostly categorized as “not recommended” for AI-based literature search. In order for these tools to pass these checks, they must evolve beyond keyword retrieval toward a nuanced understanding of academic context and hierarchy [8].

Our team of information specialists piloted the TRUST evaluation protocol. Examples of results include:

  • Recommended: ai2 Asta — restricts outputs to the Semantic Scholar corpus, exposes its reasoning chain, and is free and mostly open-sourced. It excels in Traceability and Reliability.
  • Needs review / provisional: Elicit — valuable for abstract summarization, but its ranking algorithm is not fully disclosed, limiting its Transparency. The free version is capped, affecting its User-centric score.
  • Not recommended: generic LLM interfaces (e.g. ChatGPT, Gemini) and general search-focused AI (e.g. Perplexity) — prone to citation errors and lacking rigorous, verifiable academic sources, failing the Reliability and Traceability criteria.

In the coming months, the TRUST framework will be refined into a standardized evaluation protocol with an interactive, integrated workflow — structured data entry coupled with automatic processing, ending in a database that can serve a public-facing list of approved tools accompanied by user guides. In the longer term, TRUST could be extended beyond search tools to assess AI features embedded in existing platforms, and integrate into institutional AI policies and initiatives.

References

Download sources as .bib
  1. L. A. Nogueira, S. T. Moltubakk, A. Fagervik, and I. B. Langfeldt, “Cutting through the noise: Assessing tools that employ artificial intelligence,” IFLA Journal, vol. 51, no. 3, pp. 735–749, 2025, doi: 10.1177/03400352241304121.
  2. European Commission, “Assessment list for trustworthy artificial intelligence (ALTAI) for self-assessment,” High-Level Expert Group on AI, 2020. [Online]. Available: https://digital.strategy.ec.europa.eu/en/library/assessment-list-trustworthy-artificial-intelligence-altai-self-assessment.
  3. L. McCormack and M. Bendechache, “A comprehensive survey and classification of evaluation criteria for trustworthy artificial intelligence,” AI and Ethics, vol. 5, no. 3, pp. 1973–1994, 2024, doi: 10.1007/s43681-024-00590-8.
  4. GO FAIR Foundation, “FAIR principles,” 2016. [Online]. Available: https://www.go-fair.org/fair-principles/.
  5. M. Mersha, K. Lam, J. Wood, A. K. AlShami, and J. Kalita, “Explainable artificial intelligence: A survey of needs, techniques, applications, and future direction,” Neurocomputing, vol. 599, art. no. 128111, 2024, doi: 10.1016/j.neucom.2024.128111.
  6. A. Tay, “A Librarian’s Guide to AI in Academic Search Tools,” Katina Magazine, 2025, doi: 10.1146/katina-052025-1.
  7. Á. Cabezas-Clavijo and P. Sidorenko-Bautista, “Assessing the performance of 8 AI chatbots in bibliographic reference retrieval: Grok and DeepSeek outperform ChatGPT, but none are fully accurate,” arXiv:2505.18059, 2025, doi: 10.48550/arXiv.2505.18059.
  8. I. Seth, B. Lim, Y. Xie, R. J. Ross, R. Cuomo, and W. M. Rozen, “Artificial intelligence versus human researcher performance for systematic literature searches: a study focusing on the surgical management of base of thumb arthritis,” Plastic and Aesthetic Research, vol. 12, 2025, doi: 10.20517/2347-9264.2024.99.