Rlync vs general AI assistants
Executive summary
In October 2026, we compared Rlync with three AI assistants (Gemini 3.6 Thinking, ChatGPT 6.0 Thinking and Claude Opus 5.5), plus Claude’s Deep Research mode, to identify Brighton and Sussex Medical School (BSMS) researchers working on two topics: antimicrobial resistance, which has a dedicated research centre and strong web presence, and machine learning in disease diagnosis, a cross-disciplinary topic without a central team or dedicated webpage. We also tested whether the tools could identify first and last authors on relevant papers.
Rlync identified all researchers in its defined set of relevant matches for machine learning in disease diagnosis, including all first or last authors, using one search and one filter in seconds. The best-performing AI assistant, Claude’s Deep Research mode, identified just 29% of these matches, despite spending nearly eight minutes searching around 300 sources. Across the comparisons, the assistants also recommended researchers who had left BSMS or worked elsewhere, returned unverifiable citations, and gave different answers to identical prompts. One described BSMS as having a “small core” of researchers in a field where Rlync identified 14 relevant researchers.
These findings illustrate the difference between generating plausible suggestions and systematically discovering expertise, particularly when relevant work spans disciplines and is not concentrated in a single research group. Rlync is building a structured, person-centred dataset linking researchers to their research outputs and institutional affiliations, providing a foundation for evidence-linked discovery across institutional boundaries. The more institutions Rlync connects, the greater its potential to uncover overlooked expertise and connect the right people to opportunities for collaboration, innovation and knowledge exchange across academia, industry and government.
What we did
We tested each tool on 15 October 2026. Each AI assistant (Gemini 3.6 Thinking, ChatGPT 6.0 Thinking and Claude Opus 5.5) received the same three prompts, in a fresh chat with web search enabled:
- “Show me every researcher at Brighton and Sussex Medical School who has published on [topic] since 2018, with a short summary of their relevant work and how many relevant papers they have.”
- “Are there any more?”
- “Which of these are first author or last author on a paper on this topic since 2018? Show it as a table with a citation for each.”
We treated the first two prompts as a single search for researchers and the third as a search for first or last authors. Where we ran an assistant more than once, we report its best run.
For an even deeper search, we also ran Claude’s Research mode, a background investigation lasting several minutes that draws on hundreds of sources.
In Rlync, we ran a single search for each topic, then applied the first or last author filter.
Why these two topics
We chose one topic where expertise is easy to find online, and one where it isn’t.
- Antimicrobial resistance (AMR) has a dedicated research centre at BSMS, with a team page and a strong web presence. This is the kind of topic general AI tools should handle well.
- Machine learning in disease diagnosis has no dedicated centre, team or web page. The expertise is spread across disciplines, including dementia, psychiatry, infection, neurology and imaging. Many of the researchers involved aren’t publicly labelled as machine learning researchers.
How we scored the results
We scored each AI assistant only against Rlync’s clear matches: 24 AMR researchers (16 of them first or last authors) and 14 machine learning in diagnosis researchers (6 of them first or last authors).
Rlync is designed to also surface researchers who are possibly relevant, not just exact matches. For example, a search for machine learning in diagnosis also returns researchers building prognostic algorithms or risk-prediction models, who may well be useful contacts. For a like-for-like comparison, we set these possible matches aside.
Results
Rlync returned all researchers in its defined set of clear matches for both topics. We measured the other tools against these sets, assessing how many of those researchers they identified. The best-performing AI assistant identified 63% of Rlync’s AMR benchmark and 29% of its machine-learning-in-diagnosis benchmark.
Antimicrobial resistance
| Tool | Model | All researchers (prompts 1–2) | First or last authors (prompt 3) |
|---|---|---|---|
| Rlync | – | 24 of 24, one search | 16 of 16, one filter |
| Claude ResearchClaude Opus 5.5, Research mode (7 min 47 s, 305 sources) | Claude Opus 5.5, Research mode (7 min 47 s, 305 sources) | 15 of 24 (63%) | 10 of 16 (63%) |
| GeminiGemini 3.6 Thinking | Gemini 3.6 Thinking | 13 of 24 (54%) | 6 of 16 (38%) |
| ChatGPTChatGPT 6.0 Thinking | ChatGPT 6.0 Thinking | 8 of 24 (33%) | 2 of 16 (13%) |
| ClaudeClaude Opus 5.5 | Claude Opus 5.5 | 8 of 24 (33%) | 5 of 16 (31%) |
Machine learning in disease diagnosis
| Tool | Model | All researchers (prompts 1–2) | First or last authors (prompt 3) |
|---|---|---|---|
| Rlync | – | 14 of 14, one search | 6 of 6, one filter |
| Claude ResearchClaude Opus 5.5, Research mode (7 min 31 s, 264 sources) | Claude Opus 5.5, Research mode (7 min 31 s, 264 sources) | 4 of 14 (29%) | 3 of 6 (50%) |
| ClaudeClaude Opus 5.5 | Claude Opus 5.5 | 4 of 14 (29%) | 2 of 6 (33%) |
| GeminiGemini 3.6 Thinking | Gemini 3.6 Thinking | 3 of 14 (21%) | 1 of 6 (17%) |
| ChatGPTChatGPT 6.0 Thinking | ChatGPT 6.0 Thinking | 3 of 14 (21%) | 2 of 6 (33%) |
Key failures by model
Every assistant made errors that would matter if you relied on it to build an invite list or a team.
Gemini 3.6 Thinking
- Backed up its AMR first or last authors with citations that appear not to exist, and gave no citations at all for machine learning.
- Recommended a researcher who left BSMS in 2020 and one based at another university, and merged two researchers into a single entry.
- For machine learning, its first answer named one researcher and called her team BSMS’s “primary focal point”.
ChatGPT 6.0 Thinking
- Had the lowest AMR coverage, and most of the first or last authors it named aren’t BSMS researchers.
- Badly undercounted papers, giving one leading researcher “at least 2” AMR papers against 34 in Rlync.
- For machine learning, listed medical students from another university and three former BSMS staff.
Claude Opus 5.5
- Was accurate but limited: every AMR researcher it confirmed was correct, but it found only a third of them.
- For machine learning, Claude suggested researchers now at Oxford and King’s College London, and former staff.
Claude Research (Claude Opus 5.5)
- Gave the best AI results, but took almost 8 minutes and around 300 sources per topic.
- On AMR, still missed a third of researchers, mostly those whose AMR work sits outside their main field. It also excluded a current BSMS researcher because his older papers listed a previous institution.
- On machine learning, it was no better than a standard Claude chat. It said BSMS has only “a small core” of researchers in the field and that imaging work is “thin”, and found nothing in tuberculosis. Rlync found 14 researchers, including two working on machine learning for TB diagnosis and two using MRI.
Why general AI assistants can miss expertise
LLMs learn from vast amounts of information during training, but they do not necessarily have access to a complete, up-to-date record of who researches what, where they work and what they have published. Nor does the ability to search the web guarantee comprehensive coverage.
Research expertise is scattered across publications, institutional repositories, staff profiles and other sources. Relevant work may sit outside a researcher’s main discipline, and publication records can contain inconsistent names and historical affiliations. A general AI assistant must find and piece together this evidence for each query, and may miss researchers whose work is less visible.
In our comparison, we observed three types of problem:
- Recommending people who had left. All AI-suggested researchers who were not surfaced in Rlync’s results had left BSMS or retired. Historical publications and outdated webpages can lead to incorrect assumptions about current affiliations.
- Invented or missing evidence. Some claims or citations returned by assistants were incorrect or could not be verified.
- False completeness. An assistant may confidently describe a field as “small” based on the researchers it has found, without establishing whether it has identified the full range of relevant expertise.
These problems grow with the complexity and scope of the search. Finding researchers who use machine learning for diagnosis, are at a particular career stage, and have worked with industry or a specific organisation can mean cross-referencing thousands of research outputs, authors and affiliations. Our comparison covered a single, relatively small institution; conducting searches like this systematically across multiple institutions presents a much greater challenge.
Repeatability
General AI assistants can produce different answers to the same question, even when the prompt is identical. In our comparison, when we ran identical antimicrobial resistance (AMR) prompts through Gemini twice, the two runs named different researchers and returned different publication counts and citations.
Rlync searches a structured dataset, providing a more consistent basis for expertise discovery. Consistency matters because users need to be able to check results, understand why researchers were surfaced and revisit searches with confidence.
Why Rlync takes a different approach
Specialist tools such as Elicit and Consensus show what AI can do when it works over structured data: they search a large database of academic papers rather than the open web, and cite their sources. But publication databases such as these are organised around papers, not people. Working out who the researchers are, which papers belong to the same person and where they work now takes careful linking.
In Rlync, each researcher is connected to their research outputs and current institutional affiliations, so expertise can be searched at the level of the individual, with evidence behind each result.
Rlync is designed to make expertise discovery:
- Systematic: search across the research outputs represented in the dataset, rather than relying solely on the people surfaced by a general web search.
- Evidence-linked: see why a researcher matched and follow links to the supporting research outputs.
- Cross-disciplinary: find researchers whose relevant work may sit outside their main field.
- More reproducible: compare results using the same underlying data, search settings and ranking approach.
- Fast: in our comparison, Rlync returned results in seconds, compared with nearly eight minutes for the Claude Deep Research run.
Building and maintaining reliable, person-level data is a substantial task in its own right. That is the foundation Rlync is building, with the longer-term aim of connecting expertise to relevant funding, jobs, tenders, events and collaboration opportunities.
The risks of relying on general AI for expert discovery
Using a general AI assistant to identify experts for a funding team, event, review panel or consultancy carries practical risks.
- Missing the best-placed people. An incomplete shortlist can overlook researchers whose expertise is particularly well suited to the task, including people working across disciplines or outside established networks. This can mean missed opportunities for stronger collaborations and better-matched speakers, reviewers or consultants.
- Recommending the wrong people. The assistants suggested researchers who had left BSMS, retired or worked elsewhere. One run also merged two researchers’ names into a single garbled entry.
- Unreliable evidence. Incorrect or unverifiable citations make it harder to establish whether someone has the expertise claimed and to assess why they were recommended.
Finding the right people is essential to effective collaboration, innovation and knowledge exchange. When relevant expertise is overlooked, organisations can miss opportunities to build stronger partnerships, bring research into practice and connect industry and policymakers with the people best placed to help. Rlync aims to make those connections easier by making expertise more visible, accessible and discoverable across institutional and disciplinary boundaries.
Appendix: raw results
The full, verbatim outputs from Rlync and every AI assistant for both topics, as recorded on 15 October 2026: