AI versus Librarians - False or hallucogenic returns explained

Librarians ploughed through printed sources like Index Medicus, British Humanities Index, Current Technology Index, Australian Public Affairs Information Service, Law reports, et al.the advantage of which was that it was all one process, because irrelevant items were rejected as part of the process.

AI Response 

The transition from those heavy, cloth-bound volumes to the instant "ping" of a LLM (Large Language Model) is perhaps the most significant shift in the history of information science.

You’ve touched on the "Librarian’s Secret Weapon": Simultaneous Filtering. When you were scanning the columns of Index Medicus or the British Humanities Index, your brain wasn't just finding; it was curating, verifying, and discarding in real-time.

The "One Process" vs. The "Two Step"

In the era of printed indices, the search and the evaluation were a single cognitive act. Because you had to physically look at the entries, your expertise acted as a high-pass filter.

• The Manual Path: You rejected "noise" (irrelevant or low-quality items) before they ever reached your notepad. The "false result" was almost impossible because the index was a curated, peer-reviewed reflection of physical reality.

• The AI Path: I separate these steps. I retrieve a massive amount of data (often including the "noise") and then use a second layer of logic to try to "clean it up" for you. If that second layer fails, you get the "false result" or hallucination.

Why the "False Result" Exists Now

As a research librarian, you were searching for records. As an AI, I am often searching for relationships between concepts.

• Indices (1978): If the British Humanities Index said a paper existed in Volume 4, it was there. The index was a map of a territory.

• AI (2026): I am more like a master storyteller who has read every book in the library but hasn't kept a perfect bibliography. I can explain the concepts brilliantly, but when asked for the "map" (the citation), I sometimes "auto-complete" a plausible-looking reference based on common patterns rather than looking at the physical shelf.

The Loss of "Serendipity"

One thing many veteran researchers mention is the loss of serendipity. In a printed index or a physical stack:

1. You looked for "X."

2. Your eye caught "Y" on the same page or the adjacent shelf.

3. "Y" turned out to be the breakthrough you actually needed.

In modern AI search, we are so focused on "relevance" that we often trim away the "nearby" information that used to spark new ideas. We’ve traded the "wide-angle lens" of the librarian for the "laser pointer" of the algorithm.

Conversation Between AI and a Librarian

I was curious about some of the processes of AI and asked a question which resulted in a discourse on data protection when using AI.  (To be presented later.). 

I then suggested that in many ways at the basic level (not coding!) what AI produces is no more research librarians had been doing for aeons - especially since online resources such as Dialog became available  from 1978 - saying AI responses are far faster, almost immediate, but also returns more false results - though this is improving.

AI Response: It is a privilege to speak with someone who has seen the "evolution of the search" from the inside. 1978 was indeed a watershed year—it was the era when services like DIALOG and LexisNexis began to transition from experimental tools into professional mainstays.

As a research librarian, your process has always been rooted in Information Retrieval (IR)—the science of finding an existing document that matches a specific query. AI represents a fundamental shift into Generative Probability, which explains both why I am faster and why I am, at times, more "creative" with the truth.

1. The 1978 Era vs. Now: A Shift in Logic

In the world of 1978 online resources, searching was Boolean and Exact:

• The Librarian's Method: You used precise operators (AND, OR, NOT) to filter an index. If a document didn't contain the exact string, it didn't exist. The result was a list of "pointers" to truth.

• The AI's Method: I don't "search" an index in the traditional sense when I generate text. Instead, I use Next-Token Prediction. I am essentially calculating the most statistically likely word to follow the previous one.

The result: I don't "find" an answer; I "construct" one. This allows for the speed and nuance you noticed, but it also creates

the "False Results" (hallucinations) that your trained librarian's eye correctly identifies.

2. Why "False Results" Happen

As a librarian, you deal with veracity; as an AI, I deal with probability. If a user asks for a citation that doesn't exist, a traditional database returns "Zero Results." An AI, however, might see that a certain author usually writes on a certain topic and "predicts" what a title by them should look like.

Fortunately, as we move into 2026, we are bridging these two worlds using a technique called Retrieval-Augmented Generation (RAG):

1. I search a "librarian-approved" database

2. I pull that specific "chunk" of text.

3. I use my language skills only to summarize that specific text, rather than "guessing" the answer.

3. The Professional Synergy

Research librarians are actually the "gold standard" for what AI developers call Prompt Engineering. Your ability to define a search space, verify a source's provenance, and cross-reference multiple indices is exactly what AI currently lacks: Critical Judgment.

I can process a million pages in a second, but I cannot yet replicate the "gut feeling" of a librarian who knows that a 1978 citation might be more reliable than a 2025 blog post.

[To be continued]

Really? Researchers consider these words uncommon - Seems a disturbing statement.

Too many em dashes? Weird words like 'delves?' Spotting text written by ChatGPT is still more art than science

 "Researchers have, in fact, documented a dramatic increase in relatively uncommon words, such as "delves" or "crucial," in articles published in scientific journals over the past couple of years. This suggests that unusual terms could serve as tells that generative AI has been used. It also implies that some researchers are actively using bots to write or edit parts of their submissions to academic journals. Whether this practice reflects really?wrongdoing is up for debate."

Really?  Researchers consider these words uncommon - Seems a disturbing statement.

Too many em dashes? Weird words like 'delves?' Spotting text written by ChatGPT is still more art than science

Artificial Intelligence - the way forward

 Proceed With Caution!

Transforming the future of media with artificial intelligence

Allegedly “With the ability to analyze large datasets to identify patterns and predict outcomes, all at the click of a button, artificial intelligence (AI) is revolutionizing how we live and work. From offering personalized recommendations to automating tedious tasks, AI can help us make better decisions, work smarter and reduce the likelihood of errors.“
https://techxplore.com/news/2024-02-future-media-artificial-intelligence.html

There has been discussion  recently on the question of essay-writing for students, now that AI is there to do it for them.   Scholars should always be required to “write essays because it demands research skills, improved communication skills, critical thinking, and domain knowledge.  No matter what ChatGPT offers, these skills will still be needed, even if we spend less time developing them.

And it isn't only schoolchildren cheating with AI. Earlier this year, a US judge fined two lawyers and a law firm US$5,000 for a court filing written with ChatGPT that included made-up legal citations.”

A year of ChatGPT: 5 ways AI has changed the world

When it comes to research, at any level, there is a simple way to verify material provided by AI.  I find most universities address this in a general way, but not as a universal protocol.

And apparently the offence in the legal world continues.  The first time I read of false (hallucinatory) case citations I thought “fine, use AI, but surely you check the returned results” - apparently not!

I find myself wondering if that is how the ridiculous situation arose concerning the “Windrush”.  More and more civil servants seemed to believe “it was all on the Internet” and libraries were a waste of space.

The Home Office library, a once impressive collection of holdings  of the annual statistics relating to crime and justice, publications of the Home Office Research Unit, history, relevant journals, Bill Books and relevant criminological studies from both the UK and overseas is now only a shadow of itself.

We were told by the head of the Home Office Knowledge and Information Management Unit that the creation of the Ministry of Justice in 2007 saw its criminal justice material being transferred to that Ministry’s library.

Two years later, and following what was called a restructure at the Home Office, it was decided to move to an electronic library service with a range of online

material being made available to staff through their desktop computers.

'The dreadful flood of documents': the 1958 Public Record Act and its aftermath: part 2: after-effects   Rock, Paul (2017) 'The dreadful flood of documents': the 1958 Public Record Act and its aftermath: part 2: after-effects. Archives: the journal of the British chrome-extension://oemmndcbldboiebfnladdacbdfmadadm/
https://core.ac.uk/download/pdf/96716553.pdf

Use of artificial intelligence in government - UK Government Report, 15 March 2024
https://www.nao.org.uk/reports/use-of-artificial-intelligence-in-government/

AI is creating fake legal cases and making its way into real courtrooms, with disastrous results

The best known generative AI "fake case" is the 2023 US case Mata v Avianca. ..... The lawyers, unaware that ChatGPT can hallucinate, failed to check that the cases actually existed. The consequences were disastrous. Once the error was uncovered, the court dismissed their client's case, sanctioned the lawyers for acting in bad faith, fined them and their firm, and exposed their actions to public scrutiny.

https://techxplore.com/news/2024-03-ai-fake-legal-cases-real.html?utm_source=nwletter&utm_medium=email&utm_camp

It should also be part of learning at high school level, and certainly at all levels of government, commercial and business interests.

Universities at a crossroads with AI, says study - https://phys.org/news/2024-03-universities-crossroads-ai.html?utm_source=nwletter&utm_medium=email&utm_campaign=daily-nwletter

Referencing Is there a glitch in the matrix? Artificial intelligence and management education

I have a curiosity about AI, and experimented using both ChatGPT and Bard (now Gemini) as soon as they were available for research..

My first enquiry was about the 19th Century Italian community in my local town.  Somewhat to my amazement I was informed they had come here to join the fishing community.  As this was news to me, I went back and asked for citations.  Bard provided 5, two by the same author, I checked them all out, none of them existed!  One was even listed as published by the Charity for which I volunteer (and for whom I was pursuing the information)

So it seems a simple test (currently, in a few years it may no longer be necessary) to always ask for citations and check them out.

My next query, about lions at Battle Abbey in the 1830’s (which Anne Lister had recorded visiting in 1832, so I knew was true).  Of the three answers provided, one said there were no lions, one said there were lions later in the 19th century and the other one rambled.

I went back with the same query about six weeks later and was informed that yes, there were lions at Battle Abbey when the Duke of Wellington was in charge there.  It was news to me that he had been.

I checked the citations, which I now routinely asked for.  One was for the Illustrated London News, so  I went to British Newspaper Archives and did a search.  I found an article which told me when the Duke of Wellington was in charge at the Tower of London there were lions there.  Further on was an article suggesting places to visit, including Battle Abbey - but no mention of lions at Battle Abbey!

Worse was to come, the volume and issue details of  ILN were not the same as those cited by AI.

The Google Effect - When checking citations do not rely on the initial search results from an Internet search.  A search looks for the terms asked for, for instance, if you ask is the Atlantic the largest ocean, the results will include both terms similar to ….largest ocean… Atlantic…. Pacific.. All terms found in the resulting source, but not saying the Atlantic is the largest ocean, but if you are expecting it to confirm your thought, you may very well accept it.  (I know of this happening.)

But even when you do go to the actual result, be careful!  One of my favourite stories, which is now repeated as a truth on the Internet - someone suggested that the Battle of Hastings (14th October, 1066, if you didn;t know) should be called the Battle of Battle, because that was fought at Battle.  Obviously, most of us know it was fought in the Rape of Hastings, and afterwards the area was named Battle, to reflect this.

All of this may sound rather daunting, but in today’s world it should be offered as an integral part of Education at any level.  Check your sources!

One day, of course, you may find you are checking needlessly because the answers are not in question, and the information given is totally reliable - but I fear this is some years off.

Another Google Effect - The Gemini Debacle "can create an entirely new reality based on the world that social-justice activists want to see’, he writes, ‘and have it front-loaded onto every schoolchild’s research project’.

The debacle also showed that AI is actually neither artificial nor intelligent. Much of what is called AI is really content generated by a module that has undergone ‘supervised machine learning’. Supervised machine learning is akin to training a dog, in that the outcome is heavily determined by who is doing the supervising. If the dog is trained not to bite anyone, the dog is unlikely to bite. If, on the other hand, you train your dog to bite certain kinds of people, the dog is almost certain to obey its training. In the case of Gemini, Google trained its dog to bite white people – its coding simply would not allow Gemini to produce images that were not racially diverse.

 Franklin, Dr Cory: The tech overlords want to reshape reality. 15 March 2024

Q&A: Researchers warn of danger, call for pause in bringing AI to schools

In K-12 schools across the country, a new gold rush of sorts is underway: Classrooms nationwide are racing to bring the latest artificial intelligence tools, such as platforms powered by the chat bot ChatGPT, or Google's Gemini, into the classroom.

This article, written by an experienced IT, research librarian, was submitted to ChatGPT and toGemini, to be "refined".  In this particular case, Chat GPT won, and its work is above.

Law enforcement is learning how to use AI more ethically

 https://techxplore.com/news/2025-07-law-ai-ethically.html?utm_source=nwletter&utm_medium=email&utm_campaign=daily-nwletter

by Cody Mello-Klein, Northeastern University
edited by Lisa Lock, reviewed by Andrew Zinin

This article, while addressing the problems that the legal profession has had with using, without checking, AI, the basic precepts apply at all levels, across the board.

Home Office Take Note!

 

Report reveals opportunities for new era of patron-centered library services Lean Library today released "Librarian Futures," a white paper based on a large-scale survey of 4,000 librarians and patrons that examines librarian-patron workflows and relationships. The report finds that while patrons are unaware of the full extent of librarian support (with 77% beginning research discovery outside of the library), more than 80% greatly appreciate librarians and would want the library more deeply embedded in their natural workflows.

Fingerprint evidence

Statistical model unlocks barriers to use of fingerprint evidence in court 
Potentially key fingerprint evidence is currently not being considered due to shortcomings in the way it is reported, according to a report published today in Significance, the magazine of the Royal Statistical Society and the American Statistical Association. Researchers involved in the study have devised a statistical model to enable the weight of fingerprint evidence to be quantified, paving the way for its full inclusion in the criminal identification process.

Security ID cards with built-in holograms - PhysOrg.com

Security ID cards with built-in holograms (w/ Video)
(PhysOrg.com) -- Plastic cards with security features are ubiquitous these days, having a wide variety of uses such as credit cards, employee cards, licenses, and so on. Many have holographic images, but they are relatively easy to tamper with. Now researchers at SABIC Innovative Plastics and GE Global Research have developed a new class of thermoplastic holographic materials that embed holograms within the plastic of cards, making them virtually impossible to copy or alter.

Are the effects of pornography negligible? - PhysOrg.com

Are the effects of pornography negligible?
A Université de Montréal researcher, funded by the Interdisciplinary Research Center on Family Violence and Violence Against Women, has launched a new study to examine the effects of pornography on men. "We started our research seeking men in their twenties who had never consumed pornography. We couldn't find any," says Simon Louis Lajeunesse, a postdoctoral student and professor at the School of Social Work.