Teaching an AI source criticism
Avoid hallucinations by grounding an AI in your own library, and getting it to flag what it can't verify. Also: are you in the model?
I just read a post promising to show how to avoid hallucinations. It was quite long before it got to the point, so my attention was already wondering, and it ended with links to software designed by the author to check for accuracy of references, and another link to a different software to check whether the references are suitable and appropriate.
And that got me thinking.
If someone promises me a tool to get rid of a problem, I don’t want to plow through a long preamble (and I am an academic, so clearly more used to doing that kind of thing than the average reader. I still find it annoying.)
People still have problems with hallucinated references? (I don’t mean editors or reviewers — I know full well that publishers have spectacularly failed to invest in the straightforward system to make academics’ lives easier. Shameful, but who’s surprised?)
Why would anyone go to two different pieces of software to do these checks, when your own AI is perfectly capable doing it for it you? (Yes, even if said AI has created these hallucinations — which means you did not set up your subscription service properly.)
So I decided to write a post about it.
(Also in this post, does you AI know you from their training? Scroll down.)
Catch-up service:
AI and Peer Review (almost entirely free)
How useful are AI tools for Academics? (by now a little outdated —consider this one as the update)
Is AI taking over social science — Part 1 and Part 2 (both almost entirely free to read)
Grounding your AI
Your AI draws on training data first, and will sometimes say they “inferred” or took it “from memory”. Be very wary when it does. If not properly set up, it may not even tell you. In that case, you have a problem.
Also, if you want to work seriously with an AI, make sure you have a subscription version. The free stuff won’t help you here.
Three things you need to do to get more out of your AI:
Create an instruction file (usually a free-text field within settings) that tells your AI how you want it to work with you.
Switch on memory. Check once in a while what is in there, and if that reflects what you want from your tool.
Familiarise yourself with skills/connectors/tools/plug-ins and make sure you have what you need here. Especially in Claude Co-Work, there are very useful ones available.
Read what releases are there and see what other users post about.
If you are very advanced, you may build your own. I am not there. Not even with vibecoding. But I sometimes like what people do when they make it available on Github.
From here on in, I can only tell you about my own set-up, which I find very good and suspect might be very good, period. On LinkedIn, a computing scientist from Germany asked why I did not just use Elicit or Consensus, which are built to do literature work. That was a good reminder to check these tools again; I gave up on them more than a year ago because I did not find them good enough. But they have really improved, so I endorse this message.
But I also said to the guy that I use MCP access to my own library and Scholar Gateway, which is pretty much the Swiss army knife when you work with and search for literature. I got no response on that. I suspect that, from my experience with German men (being German and growing up in Germany, living there until my mid-20s), this means he did not know you could do that with an MCP, and he was not going to say that publicly to a woman without a degree in computer science. But, you know, I might be wrong…
My Claude set-up
So here’s what I do with my Claude.
1. Global instructions
Tells it how I want to work and what my priorities are. For example, my file includes sentences such as these:
I value precision and thoughtfulness over obsequiousness or overpoliteness.
Always be clear about what you do not know, or where your conclusions are tentative.
Always check everything for hallucinated references and flag them.
Verify your own outputs as much as possible.
Avoid classic AI phrasing and uses
I have a colleague who uses the other major AI frontier model. He says he set it to “spikey” because he wants it to give him a bit of lip.
Each to their own. I use Claude, which I doubt could josh you even if you tried. It’s very serious. It’s like your very-well-read aunt whose sense of humour is at best occasional.
2. Switch on memory and search across chats
By switching on memory and allowing the tool to search across your chats, it builds up context across interactions and learning preferences you may be less aware of. When it works well, it draws on prior analysis and chats when I ask it a question without context, and it can establish a context and give a better answer.
If you use it for lots of personal stuff, I would be more concerned, and use incognito mode and regularly wipe memory. But I don’t. It’s my Swiss army knife for work.
You can see what is in the memory on Claude via settings. Check it once in a while.
I’ve noticed that Claude regularly points out that I might not have time to do something. Me, addicted to overcommitment? It’s got my number, alright.
3. There are tons of useful integrations
Too many to go into, in fact. But if you have favourite integrations, skills, etc., do note them in the comments! I always like to learn more.
I will tell you about two here that I think are essential for academics.
3.1 Scholar Gateway
Wiley launched Scholar Gateway. In Claude, it is known as a Connector. It’s really good because it allows Claude to pull on the publishing data in there to both search and confirm references. If you do nothing else, make sure you get this. It is available to subscribers via institutional access and as a limited free trial to researchers. There are other options for your AI, so it is not the end of the world.
3.2 Zotero MCP
Two things to consider:
This is by a private developer posted on GitHub. Not all of them are reliable, so it depends on your risk appetite. I’ve been using this for months, and it has been fine.
You need to be using Zotero as your reference manager for this to make any sense. Should be obvious.
For info, this is a discussion of various tools out there. I think I used Zotero MCP. It works well and has a lot of fine-grained controls on what you want to allow Claude to do.
So, here you go, you have the nuts-and-bolts now, as promised. After the paywall, my evaluation of how well it works, examples of what to do with it and when dedicated tools perform better than my Swiss army knife solution. Also, do the major AI models “know” you? That is, are you in their training data? Whether your anxiety is based on privacy or on status, you can find out now.
