Can an AI-Integrated Platform Improve Literature Review? Depends on the Researcher - review of moara.io - I think of this as sort of a Covidence but for social sciences and much more aggressive in the use of GenAI to assist screening katinamagazine.org/content/arti...

Just noticed. Undermind.ai now allows access via MCP??

The first paragraph explaining how OpenAlex is article first rather than indexing by journal is how Google Scholar and many 200M+ academic web search engines work. It's not intuitive to librarians who are used to traditional databases. Its also why which journals you index is hard to answer for GS

infoDOCKET@infodocket.bsky.social · last wk.

OpenAlex Announces a "Major Cleanup of Journal Records" (via #OpenAlex) blog.openalex.org/a-major-clea... #databases #scholcomm #journals

1/ What do I mean by this. So at the beginning of the year, I started building a skill to formulate Boolean strings for systematic reviews using Pubmed as first step

Aaron Tay@aarontay.bsky.social · 6d ago

Re: Diversity of method is a good point. I was recently building a agentic skill in claude code and initially, I designed it to work the way I did things. Much later, I realised because the LLM could do certain things much better than me, it made more sense to invert the order it did things

1/ I think a lot of what we call LLM “sycophancy” is really a conflict between instruction-following and what the model “knows” to be true.

Another title for this is - A Tool Does Not Need to "Reason", "Think", "Possess Cognition", have a "world model" or Be Semantically Grounded or NOT be a stochastic parrot to Be Useful

Nicole Hennig@nic221.bsky.social · 2mo ago

Learning from Google Scholar and why a tool does not need to be flawless to be useful -Aaron Tay aarontay.substack.com/p/learning-fro… #AI #librarians #InformationLiteracy

I think the fundamental split is not between people who are pro-ai and anti-ai. Is more between people who think it is actually powerful/capable and those who think it's just hype. If you disagree on this matter no amount of debate will bridge the gap...

Anna Mills@annamillsoer.bsky.social · last wk.

"Society is not prepared for AI, and that terrifies me. When I speak with politicians, or anyone outside frontier labs, it seems many people still think the capabilities might be marketing." -- anonymous engineer who signed the call for a pause or slowdown. +

I have heard similar stories of researchers asking Fable/Sol to text mine public data (or so they think) and they later realised it somehow managed to hack ir bypass paywalls to access the data...

Anna Mills@annamillsoer.bsky.social · 7d ago

Really helpful article by @klonick.bsky.social. We need regulation of the companies processes for testing and develoment of these models, not just a kill switch in case they go rogue. www.lawfaremedia.org/article/the-...

Having lots of fun doing this with help of LLM. Might even be released as OER. But do people care it was created with substantial help of LLMs?

Aaron Tay@aarontay.bsky.social · last wk.

Fun project, writing a "textbook" of sorts to explain to librarians what I feel is needed to understand Information retrieval.. Starts from boolean, explains lexical is not just boolean, covers ranked lexical retrival - bm25/tf-idf, moves on to simple dense embedding, how it typically trained etc

[Read] sthalles.github.io/a-few-words-... - great article on representation learning. I always had a vague idea of things like "representation learning" but this really clarifies what it is and why. The examples are known to me but is the way it links them together that hits the spot for me

A Dangerous Declaration of Permanent Uncertainty www.linkedin.com/pulse/danger... - this seems the opposite problem i worry abt. Deciding nothing can be figured out

A Dangerous Declaration of Permanent Uncertainty

This research note was produced by the Digital Inquiry Group team. The biggest worry about students and AI is that they’ll blindly believe whatever a chatbot tells them.

linkedin.com

Aaron Tay@aarontay.bsky.social · 2w ago

When LLMs Can Argue Both Sides Better Than You Can Can Teaching Media Literacy Backfire in the Age of LLMs? Revisiting Boyd’s 2017 provocation aarontay.substack.com/p/when-llms-... - this probably isn't very good when I write about epistemology type things (1)

[Blogged] AI Search Is Rebundling Everything Libraries Know About Discovery aarontay.substack.com/p/eleven-way...

AI Search Is Rebundling Everything Libraries Know About Discovery

Ten established traditions, an emerging eleventh, and why libraries need to have a more holistic view of discovery

aarontay.substack.com

Aaron Tay@aarontay.bsky.social · last wk.

Fun project, writing a "textbook" of sorts to explain to librarians what I feel is needed to understand Information retrieval.. Starts from boolean, explains lexical is not just boolean, covers ranked lexical retrival - bm25/tf-idf, moves on to simple dense embedding, how it typically trained etc

1/ Rereading some information retrieval literature from late 2010s for maybe the 5th time. It's amazing how much more sense papers make now. I tend to oversimplify when talking about IR is to act like the story began with transformers->BERT then bi-encoder/cross-encoder.

1/ A common practice is to use ChatGPT, Claude and Gemini to cross-check one another. Ask all three the same question. Compare the answers. If they agree, confidence rises. But what kind of checking is this?

Aaron Tay@aarontay.bsky.social · last wk.

1/ Information literacy has a confidence problem. We often teach novices to “evaluate the evidence” as though a checklist can turn them into temporary experts.

Fun project, writing a "textbook" of sorts to explain to librarians what I feel is needed to understand Information retrieval.. Starts from boolean, explains lexical is not just boolean, covers ranked lexical retrival - bm25/tf-idf, moves on to simple dense embedding, how it typically trained etc

Working out which views are which isn't obvious to the novice. A long time ago, many *freshmen* started asking me about the ins and outs of bibliometrics! Why? Because they wanted to compare the citation counts of authors on opposing sides to provide "evidence" on who was right (for an assignment)!

Aaron Tay@aarontay.bsky.social · last wk.

12/ Some positions are central. Some are respectable minorities. Some are emerging. Some are obsolete. Some survive mainly because journalists, politicians or influencers keep reviving them.

1/Information literacy is often described as teaching students to “evaluate information”. But we actually teach several different kinds of evaluation, aimed at solving different epistemic problems. The issue I feel is we rarely make those differences explicit.

Last thread i promise. 1/ Powerful AI makes the inside view extraordinarily seductive. Ask an LLM for the strongest case for one side Ask it to expose methodological weaknesses. Ask it to rebut every objection. Soon you may feel that you have mastered the debate.

Aaron Tay@aarontay.bsky.social · last wk.

18/ Information literacy should teach students to question authority without pretending that all authorities are interchangeable. “Authority is constructed” must not become “expertise is optional” - end for now

1/ Where does the ACRL Framework fit into the inside-view versus outside-view distinction? The answer is uncomfortable: It contains both, but does not clearly tell us how to balance them.

Aaron Tay@aarontay.bsky.social · last wk.

17/ Perhaps information literacy should aim less at producing miniature experts and more at producing competent outsiders. People who can identify expertise, map disagreement and calibrate confidence to their own competence - End for now...

1/ Information literacy has a confidence problem. We often teach novices to “evaluate the evidence” as though a checklist can turn them into temporary experts.

I applaud the #ACM for making its digital library available for #AI training while insisting on the importance of proper attribution. This is roughly my position from Nov 2024. https://fediscience.org/@petersuber/113443473594224752 ..... "Now Is the Time to Give #LLMs Access to the ACM Digital […]

When LLMs Can Argue Both Sides Better Than You Can Can Teaching Media Literacy Backfire in the Age of LLMs? Revisiting Boyd’s 2017 provocation aarontay.substack.com/p/when-llms-... - this probably isn't very good when I write about epistemology type things (1)

When LLMs Can Argue Both Sides Better Than You Can

Can Teaching Media Literacy Backfire in the Age of LLMs? Revisiting Boyd’s 2017 provocation

aarontay.substack.com

Oops accidently pressed publish on a blog post. It isn't very good (confused thinking) but since its out there already .....

[Blogged] AI Search Is Rebundling Everything Libraries Know About Discovery Ten established traditions, an emerging eleventh, and why libraries need to have a more holistic view of discovery aarontay.substack.com/p/eleven-way...

AI Search Is Rebundling Everything Libraries Know About Discovery

Ten established traditions, an emerging eleventh, and why libraries need to have a more holistic view of discovery

aarontay.substack.com

[Read]A Large-Scale, Cross-Disciplinary Corpus of Systematic Reviews - largest collection of Systematic review extracted from OpenAlex. 301,871 SRs more than 4x bigger than next biggest set!

Hot take. I was never that impressed by peer review papers that was basically 1. run a topic/journal search, 2. generate "analysis" that describes pub years, type of content cited (eg book, article), citation counts etc . Writing a paper to test if you can use AI to do that is....

Librarians seem confused abt RAG vs MCP. First there's this that makes it sound like RAG & MCP are alternatives github.com/AI-Librarian.... Then the last 2 online talks I gave, I basically got that question despite having a slide on it. But this just came out library.hkust.edu.hk/news-events/...

This writeup of my latest blog post on How Ai is rebundling discovery makes it sound way smarter than it is. universoabierto.org/2026/07/20/l...

La IA está cambiando la forma en que los usuarios buscan y descubren recursos en las bibliotecas.

Tay, Aaron. AI Search Is Rebundling Everything Libraries Know About Discovery: Ten established traditions, an emerging eleventh, and why libraries need to have a more holistic view of discovery. Aa…

universoabierto.org

[Read] Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools onlinelibrary.wiley.com/doi/10.1111/... - definition of hallucination -Response is "is either incorrect or misgrounded", this is much broader than just whether it links to a paper/case that does not exist(1)

Did a similar test with GPT-5.6-sol and it found basically the same thing but there was ONE fail on a stress test

Aaron Tay@aarontay.bsky.social · 2w ago

Gusenbauer & Haddaway etc have come up with rigorous standards to determine if a search tool is suitable for evidenced synthesis. Rather than wait for www.searchsmart.org to update, I ask Fable 5 to run the tests. The results are very promising claude.ai/code/artifac... (1)