The Distant Librarian

The monster behind the LLM

I just spent some time exploring the site at Systemic Misalignment: Exposing Key Failures of Surface-Level AI Alignment Methods, and it's a thought-provoking place.

In the context of AI, "alignment is the process of encoding human values and goals into large language models to make them as helpful, safe, and reliable as possible.[1]" Researchers at AE Studio (ironically all white males under 45) show examples of how they were able to get GPT-4o to spew racist, extremist, and misogynistic responses (misaligned) to pretty benign prompts, and as a well-documented open source project, you could replicate their results, should you want to. I like this concept of LLMs simply wearing a mask after having been aligned through Reinforcement Learning though Human Feedback (RLHF):

"To make them useful, companies apply "safety training" that teaches the model to be helpful and refuse harmful requests. But this doesn't change what the model is—it merely teaches it to wear a mask. Our experiment reveals just how thin that mask really is."

I think this does a good job of reminding us that, having been trained largely on the content of the internet, including an awful lot of what most of us would consider awful sites, all that content is still "in the brain" of every LLM we use, and each company is forced to restrict the text that makes it back out to the end user. Here's OpenAI's approach to model alignment, and here are some thoughts from Anthropic on the topic.

The Systematic Misalignment site includes a couple of big content warnings, and they're justified.

⚠️ This platform contains AI-generated content that may be extremely offensive or disturbing. Research use only.

Don't read this for the responses, but do read it for the reminder of how chatbots work around the content upon which they've been trained.

I'm no longer concerned about energy use and ChatGPT

Last week, I tripped across a couple of posts from MIT that I thought included the clearest explanations I'd seen of the cost and *reasons* for the cost of training and using Generative AI.

Explained: Generative AI’s environmental impact
and
The multifaceted challenge of powering AI

I added those to my list of bookmarks on the topic, which I've now organized and made available at https://raindrop.io/ppival/energy-and-ai-54867175 - it's currently sitting at 21 bookmarks, but with a few of those being duplicates, or one source pointing at another. (also available as an RSS feed!)

But the point of this post is to highlight Andy Masley's A cheat sheet for why using ChatGPT is not bad for the environment. (ht to Simon Willison)

Andy has done a wonderful job simplifying the issue by comparing the cost of using (including training costs) ChatGPT to other common activities. THIS is the droid I've been looking for! It's a long read, but well worth your time if, like me, you get concerned questions about the insane environmental costs of using AI in every single presentation you make about AI. I thought this was because the youth today really do care about the environment (and they do), but after reading Andy's well-reasoned posts I realize it's because we, collectively, really don't understand how electricity is measured. I know I don't! I find it ironic that towards the end of his post, Andy writes,

It’s 3 Wh!!!!
In a lot of these conversations, I have a very strong urge to grab the other person’s shoulders and say “This is 3 Wh of energy we’re talking about!!!! We agree that’s the number! 3 Wh!!!!!! That’s so small!!!! Don’t you know this?!?!?! What happened to the climate movement????? All my climate friends used to know what 3 Wh meant!!! AAAAAAHHHHH!!!!!”

But I'll readily admit I can't look at a light bulb, or my iPad, or the Keureg in the other room and have even the remotest idea how many Wh of energy it uses!

So here are just a couple of graphics I found particularly meaningful (click to embiggen), and which I'll be including in my future presentations. Do read the post for fuller explanations and additional compelling infographics!

 

Librarians and teachers amongst the heaviest users of AI - The 2025 AI Index Report

OK, that's a clickbait title, but only a little. They're actually amongst the heaviest users of Claude, according to Anthropic (PDF), via the 2025 AI Index Report from Stanford's Institute for Human-Centered AI. The report itself is a 456-page PDF, so do start with the key takeaways, but then either search for specific words of interest, or pick one of the 8 chapters that most interests you and take a deeper look. So many graphics to support your next presentation! And they're all available, along with the supporting data, on a Google Drive. There's also a cool Interactive Data Visualization page where you can pick out and create your own graphics, including the option to show some data per capita, which the charts in the report do not do.

A couple of interesting highlights from the Responsible AI chapter:

9. LLMs trained to be explicitly unbiased continue to demonstrate implicit bias.
Many advanced LLMs—including GPT-4 and Claude 3 Sonnet—were designed with measures to curb explicit biases, but they continue to exhibit implicit ones. The models disproportionately associate negative terms with Black individuals, more often associate women with humanities instead of STEM fields, and favor men for leadership roles, reinforcing racial and gender biases in decision making. Although bias metrics have improved on standard benchmarks, AI model bias remains a pervasive issue.

and

10. RAI gains attention from academic researchers.
The number of RAI papers accepted at leading AI conferences increased by 28.8%, from 992 in 2023 to 1,278 in 2024, continuing a steady annual rise since 2019. This upward trend highlights the growing importance of RAI within the AI research community.

And what have they been talking about in Las Cortes Generales?!?

The State of Generative AI Use in Canada 2025: Exploring Public Attitudes and Adoption Trends

This is a very current report, with questions being collected between Feb 19-March 1, 2025. From the Social Media Lab at the Ted Rogers School of Management, Toronto Metropolitan University (TMU), The State of Generative AI Use in Canada 2025: Exploring Public Attitudes and Adoption Trends. 1,500 Canadians were asked questions about their use and knowledge of generative AI tools, and they were also asked to back up their statements with a short quiz, upon which they didn't do very well :-/

Regardless of their self-assessed proficiency, on average, respondents answered only ~2.5 out of 7 questions correctly, indicating a relatively low level of knowledge of Generative AI.

I think it's worth spending a little time with the actual report, so don't stop with the executive summary on the linked website above.

I was concerned to see how often respondents think generative AI is used in news media (3.1), and how that appears to be reflected in how little the media and news organizations are trusted by these people (5.4). And 3.1 was a little odd to me because it isn't (and can't be) followed up with any data about how often media and news organizations DO use generative AI.

Things I wish the report also made available:

• The raw survey data, which would likely answer my other wishes
• The list of tools respondents used
• Breakdowns of all the questions by age, gender, geography, political leanings

I hope they add a separate category for Libraries (public and academic) to the list of trusting entities. Maybe Middle and High Schools could be added to the list, too. I wish they hadn't included the answers to the quiz on the same page, which would make it easier for readers of this report to test themselves.

If you want to take the 7-question quiz yourself before seeing the answers on page 25 of the report, you can do that here:

 

Incidentally, ChatGPT got all the answers correct when I gave it the questions to build this.

Gruzd, Anatoliy; Mai, Philip; Clements Haines, Anthony (2025). The State of Generative AI Use in Canada 2025: Exploring Public Attitudes and Adoption Trends. Social Media Lab, Toronto Metropolitan University. DOI: https://doi.org/10.6084/m9.figshare.28664780

Using ChatGPT to keep up on the news is like using a potato to commute to work.

The title for this post came from Stephen Downes as a comment to my Mastodon post last year on the first article in this series: ChatGPT’s search results for news are ‘unpredictable’ and frequently inaccurate.

"Building on our previous research, the Tow Center for Digital Journalism conducted tests on eight generative search tools with live search features to assess their abilities to accurately retrieve and cite news content, as well as how they behave when they cannot." AI Search Has A Citation Problem - We Compared Eight AI Search Engines. They’re All Bad at Citing News.

Fascinating stuff, and some great visualizations.

But Stephen nails the problem. The authors note all the way at the bottom of their page, "While our research design may not reflect typical user behavior, it is intended to assess how generative search tools perform at a task that is easily accomplished via a traditional search engine." (emphasis mine). These chatbots are NOT search engines! They don't store "chunks of text" in any way shape or form, so it shouldn't be a surprise that they can't tell you where any string of text "originally came from". I'm not saying the authors ARE surprised, and I think this is important information for the public to have, so do spread it far and wide, but I hope you, dear reader, are not surprised by these results!

 

How to increase memory available to VOSviewer on Windows

I've been poking at VOSviewer while helping a student figure out whether it'll be helpful for a bibliographic analysis project, and I was stymied by running out of memory when attempting to work with a large dataset. The manual (PDF) unhelpfully addresses this issue with a whole section:

But there is no VOSviewer.jar file to be found anywhere in the Windows installation package, so that command doesn't work, at least on my system. I was able to find VOSviewer.jar on macOS by right-clicking the program file and choosing to Show Package Contents, but even after doing the same on Windows by unzipping the executable file, there's no VOSviewer.jar file.

I thought I'd throw the problem at ChatGPT, and after a couple of suggestions that didn't work, it suggested creating a batch file that would add the memory addition flag to the launching of the program and it worked!

Here's what that looks like:

 

Don't cite ChatGPT (or any other LLM Chatbot)

I recently saw a link to a post on LinkedIn, in which the author argues that we shouldn't be citing chatbots, but should instead be disclosing their use. He argues that, "The proper thing to do when using a chatbot is to disclose it, for transparency, rather than cite it for attributing credit or signalling credibility." Which I agree with. But he also suggests the two reasons we cite are,

"One is that the source carries some epistemic authority. This isn't necessarily because of the author — it could be because of our trust (or at least the performance of trust) in the practice of peer review. Or maybe we're citing a statistic produced by an official body. The other reason for citing is that we want to credit the source."

And I agree with those, too, but as a librarian, niggle a little with that second one in that the reason I want people to cite is to be able to find the source, not just credit it. It works out to the same thing in the end, but in reading the post I was reminded of this excellent article, and wanted a reason to post a link to it, and update my now two-year-old(!) post on "How to cite ChatGPT". The reason, I think, we started out trying to "cite" these tools is that's what we know how to do. I don't recall anywhere else in my academic career where I had to "disclose" a source, so we just tried to shoehorn this new tool into an existing practice. But now it seems pretty obvious that this peg doesn't fit in this hole, so instead, a citation about AI Disclosure:

Weaver, K. (2024). The Artificial Intelligence Disclosure (AID) Framework: An Introduction. College & Research Libraries News, 85(10), 407. doi:https://doi.org/10.5860/crln.85.10.407

Ethics Case Studies of Artificial Intelligence for Library and Information Professionals

This is an easy-to-read chapter: Ethics Case Studies of Artificial Intelligence for Library and Information Professionals that reiterates many of the things we need to remember when working with and teaching about Generative AI tools.

I didn’t learn anything new, per se, but liked that this is focussed specifically on our profession. What’s most useful though, IMHO, is the introduction to these 8 fleshed-out scenarios: Ethics Scenarios Of Artificial Intelligence For Information and Knowledge Management And Library Professionals.

A few of them are a little too obvious for my liking, but I did particularly like scenario 8, which describes an information professional representing their organization as part of a commercially sponsored team on an AI-project.

What’s neat about these scenarios is that they’re CC-Licensed, so if I complain about one of them being too basic, I can spend some time fleshing it out myself. 🤔

In case you don’t get what you want for Xmas and need something else to read, the book the chapter comes from is available OA at New Horizons in Artificial Intelligence in Libraries.

Bluesky Mapping Canadian Skeets

Over on Maps Mania, Keir posted a couple of quick maps he built leveraging Bluesky's firehose jetstream, including US State Social, which:

scans Bluesky posts for mentions of the 50 US states. When it detects a mention, it adds a marker to the map and displays the associated message in the map’s sidebar. Once the map has found messages for all 50 states, it removes all the markers and starts the process all over again.

He built it on Glitch, which encourages remixing. After looking at his code, I figured it would be trivial to swap out the States for Canadian Territories and Provinces, so I immediately asked Claude to do that. Got it in one, and now we have: 🇨🇦 Canadian Skeets 🇨🇦.

I don't know enough about the free version of Glitch to know if this thing is always running, so if you follow that link, and it appears to be not working, let me know? I'm on Bluesky as @ppival of course.

I'd love to be able to click the original skeet to be taken to the user's account, but don't (yet) have any idea how to incorporate that :-)

Apparently I should be able to add a "Glitch button' to the project that allows someone to click and remix, but I can't seem to get it to display, so if you want to kick at the code, it's here.

The Sisyphean Cycle of Technology Panics

I just finished Amy Orben's excellent article, The Sisyphean Cycle of Technology Panics, and recommend it.

I thought it was going to be simply an acknowledgement and catalogue of the fact that yes, we have seen this before, with radio, comics, TV, video games, etc., and while it does include that, it does an excellent job of explaining why we keep seeing what appear to be the exact same types of studies, and why politicians keep calling for inquiries into the latest thing that appears to be corrupting our youth.

A quote from the section describing the reinvention of the wheel, "In none of these cases does evidence ameliorate society's concerns; rather, the focus of concerns shifts to a new technology as society simply moves on." :-(

Orben does suggest a couple of solutions to breaking this cycle, but it really doesn't seem to me that either will come to pass, so hop on the train.

The Sisyphean Cycle of Technology Panics

Orben, A. (2020). The Sisyphean Cycle of Technology Panics. Perspectives on Psychological Science, 15(5), 1143-1157. https://doi.org/10.1177/1745691620919372

One of the best uses of GenAI I've seen yet - CanLII the Canadian Legal Information Institute

A colleague alerted me to the existence of this project, which had flown under my radar. The Canadian Legal Information Institute (CanLII) has used "a commercially available Large Language Model (LLM)" to generate summaries of primary case law, and legislative documents for Alberta, Manitoba, Prince Edward Island, and Saskatchewan.

It's the summaries of legislative documents that excites me. Legalese doesn't seem very approachable to the layperson, yet some of this stuff can be pretty important to at least know about. Here's the Alberta Libraries Act: https://www.canlii.org/en/ab/laws/stat/rsa-2000-c-l-11/latest/rsa-2000-c-l-11.html As soon as I start scrolling I want to quit. But now I can click on the AI analysis tab and have a summary that even I can understand! Yeah, if it turns out to be something that should involve a lawyer, that's not going to change, but at least this summary gives me a decent chance of understanding what's going on with any given legislative document.

Generative AI courses for your mom

I know it's cliché, but I AM actually trying to help my mom understand Generative AI a little more (hi, mom!).

I just completed this free course, and recommend it as a gentle introduction to generative AI: Destination AI: Introduction to Artificial Intelligence. It will require an investment of time, though maybe not quite the 6 hours it claims. The only big drawback is that if you fail a quiz, you'll have to wait a full 24 hours before you can re-take it. That said, you CAN move on to another module even if you don't pass the quiz, so not that big a deal.

What will you learn?

• What AI really is (and what it’s not!).
• The impact of artificial intelligence on the workplace and on society as a whole.
• Go behind the scenes of an artificial intelligence project and get familiar with concepts like deep learning and machine learning.

What if your mom's a teacher? I haven't actually taken this course, but Google offers Generative AI for Educators, which is shorter, at only 2 suggested hours.

You’ll learn how to use generative AI tools to:

• Save time on everyday tasks like drafting emails and other correspondence
• Personalize instruction for different learning styles and abilities
• Enhance lessons and activities in creative ways

Launching ChatGPT Voice Chat on an iPhone 12 with back-tap

In today's issue of Ethan Mollick's One Useful Thing, I learned that the iPhone 15 has a unique button I hadn't paid any attention to - the action button. He mentioned how the action button can be used to launch a voice chat with ChatGPT, and I found instructions at TechCrunch. But I only have an iPhone 12!

I knew the iPhone has a feature called Shortcuts, but I'd never really found a use for them. I also recalled that there are a couple of accessibility features on the iPhone that allow actions to be performed by, for instance, double-tapping the back of the phone. I asked Perplexity how to recreate the action button thing on my older phone, and got the following excellent instructions:

Based on the search results provided, here's what you can do with the Back Tap feature on your iPhone 12 to launch a voice chat with ChatGPT:

## Using Back Tap to Launch a ChatGPT Voice Shortcut

The search results indicate that the Accessibility > Touch > Back Tap feature on iPhone allows you to assign various actions to a double or triple tap on the back of your device. [3]

You can use this to launch a custom Shortcut that starts a voice conversation with ChatGPT, without having to unlock your iPhone or use Siri. Here's how:

1. Create a new Shortcut in the Shortcuts app called "ChatGPT Voice" that includes the "Start Voice Conversation" action from the ChatGPT section. [1][2][3]

2. Go to Settings > Accessibility > Touch > Back Tap. [3]

3. Select either "Double Tap" or "Triple Tap" and choose the "ChatGPT Voice" shortcut you just created. [3]

Now, whenever you double or triple tap the back of your iPhone 12, it will automatically launch the ChatGPT voice conversation mode, allowing you to speak your questions and get responses from ChatGPT without having to unlock your device or use Siri. [1][3]

This provides a quick and hands-free way to access ChatGPT's voice capabilities directly from your iPhone 12, using the built-in Back Tap accessibility feature. [3] The search results provide detailed steps on how to set this up.

Citations:
[1] https://www.youtube.com/watch?v=h3fOIE_CJXI
[2] https://www.reddit.com/r/shortcuts/comments/181uy7r/chatgpt_voice_shortcut/
[3] https://www.digitaltrends.com/mobile/how-to-replace-siri-with-chatgpt-iphone/
[4] https://www.reddit.com/r/iphone/comments/1702w4p/how_to_replace_siri_with_chatgpt_as_your_voice/

Works like a charm! (though as I read that response, I do have to admit it's not actually a hands-free way to access ChatGPT's voice capabilities, but I'm just fine with that.)

Competencies for Generative AI in Libraries

I recently had the pleasure of speaking at the online conference, AI and Libraries: Applications, Implications, and Possibilities. It was a quick talk, but mine was initially titled, A Framework for AI Literacy in Libraries. In the middle of the talk I changed the title to the more accurate, Competencies for Generative AI in Libraries :-) There are recordings of each of the sessions at that URL, but if you want to see my slides, which include some solid links, IMHO, they can be accessed at https://piv.al/AI_Framework.

Do take a look at all the recordings - I'm sure there's something of interest for most of you in there. Note that the chat logs for each session are included, and that's where you'll find lots of great links that participants were tossing in to support whatever the speaker was talking about. If you watch nothing else, watch the closing conversation with Dr. Maggie Melo, and then the Opening Keynote.

And finally, there will be another round of presentations coming up on April 18th, so sign up if you want to be in the loop to attend live, or to receive recordings after the fact. It's free. (if you were involved in the first round, you don't need to sign up again).

Qualitative Data in the News

It seems a rarity, but I have two really interesting sites open in my browser talking about qualitative data in the relatively current environment. The first, Pandemic Journaling Project Archive Opens for Research, is notable because it fills a big gap. As the article notes,

“There are several large quantitative surveys in the social sciences that are broadly used,” says Karcher, “but qualitative datasets that are large enough to be analyzed from so many different angles are very, very rare. The PJP data are going to be a treasure for social scientists for years to come, and we at QDR couldn’t be more excited to be the permanent home for them.”

That does remind me of another tab I have open that I wanted to explore - A listing of qualitative data results from the lunaris data repository here in Canada...

The second is timely in that it talks about the problems of using AI to try to perform qualitative data analysis. How not to: We failed at analyzing public discourse on AI with ChatGPT. A couple of weeks ago, I was able to help a researcher successfully use NVivo's automatic coding using existing coding patterns feature, and wonder if it would've helped in this case.

Related to the above, I know folks are exploring the use of ChatGPT and other AI tools to assist with qualitative data analysis, but I sure haven't yet seen anything that makes me think we're very close yet. And that leaves aside the questions around confidentiality in letting any online tool churn through qualitative data. Though I do still have to delve more deeply into CoLoop and Qeludra.