A few weeks ago I attended a workshop on “AI in the Service of Research on Talmudic and Rabbinic Literature”. Very different from anything I have gone to before! Some of my Aberdeen colleagues presented some work we had done on authorship attribution, but I was also asked to give a general talk on NLP for humanities researchers who did not have much AI experience. This seemed well received, and below I summarise some of the “things to watch out for” that I mentioned in the talk.
As a side note, from my perspective, one of the most interesting challenges in this space is evaluation in a context where there is real uncertainty about the “gold standard” truth. For example, one task is reference disambiguation. Eg if a Talmud commentary says “Rabbi Hanan said X”, which individual is it actually referring to, since there are several sages named Rabbi Hanan? We cannot ask the authors, since the Talmud was written 1500 years ago. In 2026, sometimes expert scholars agree on correct disambiguation, but sometimes they do not; what kind of evaluation makes sense in this context? If we just evaluate on cases where experts agree, then our evaluation is only on the “easy” cases.
Anyways, below I summarise my advice. It is very simplistic from an NLP perspective, but hopefully can be useful to non-CS people.
Advice: Be skeptical and check AI
There were a few “gee whiz” talks at the workshop where people essentially showed output from Claude (etc) and said that it looked amazing. I tried to encourage such people to be skeptical and at least spot-check AI output for correctness and validity; I also warned them that the fact that Claude could do amazing things on complex tasks did not mean that it could automatically be trusted on simpler tasks.
NLP researchers are familiar with hallucinations and other LLM problems, but people with limited CS background can be “blown away” by responses that look very sophisticated and intelligent, and not check them carefully. This is very dangerous, and indeed recent news stories suggest that insufficient checking of AI almost led the US to attack a Chinese ship (CNN).
Advice: Remember that AIs are “sycophantic” and try to tell people what they want to hear
When researchers use AI for subjective tasks, they need to keep in mind that AIs try to tell people what they want to hear. Which can be dangerous in a research context, where unbiased analysis is absolutely essential.
A recent shocking example in the religion space (but not about research) was when a user thought GPT was Jesus Christ, and GPT agreed with him and indeed reinforced this delusion (Ars Technica).
Advice: Use open models if you care about replicability
Several speakers discussed how they choose which models to use. None of them mentioned replicability, and I pointed out to them that if they wanted other researchers to be able to repeat their work, they should use open models such as Llama or Qwen, because these will still be around in the future. In contrast, closed models such as GPT and Claude variants (which is what most people used) get deprecated and deleted, which makes replication hard.
Ie, if I used GPT4o in a 2024 research project, then replication in 2026/27 is hard because GPT4o is retired. Whereas if I used Llama3 in a 2024 research project, then I can still run the model today.
Advice: Remember that AI only knows what is on public internet
To a first approximation, AI systems only “know” what is on the Internet. In humanities fields which study ancient documents, some information may only be available in hard-copy books and manuscripts. Do not expect AI systems to know this information!
Advice: Data quality
If you ask an AI to analyse a text, try to make sure the text is high quality, clean, and standardised. For example, the project we presented involved analysing documents in ancient Syriac. We spent considerable effort on cleaning up and standardising material which we pulled from different sources; this made a big difference to the performance of our model. Addressing data quality issues is not exciting, but it is very important.
Advice: Beware of data contamination
If you ask an AI system to analyse a text, it may simply look up the analysis from a published source. For example, in an authorship attribution context, if we give the AI model a text and asked who authored it, it may simple recognise that it is chapter X of document Y, find a paper which claims that chapter X was written by Rabbi Jones, and then regurgitate the information, without actually analysing the text in the intended way. This is called data contamination.
I suspect that data contamination impacted several of the papers presented at the workshop. I advised people to check beforehand if the AI knew the answer (eg, ask the AI who authored chapter X, without giving it the text).
Advice: AI can steal your data if you are not careful
If you ask an AI system to analyse a text and are not careful, the AI may absorb the text into its training data. If you do not want this to happen, check how the AI uses data, or use a local open model.
In rabbinic studies, there have been cases in the past where scholars hoarded manuscripts and analyses and did not release them until relevant papers were published. Since writing and publishing papers takes time, some people are concerned that data shown to AI should not be available to other users.
Advice: AI performance is worse in low-resource languages
AI models are trained on the Internet, so they work much better in languages such as English which are common on the Internet (billions of web pages) than in languages which are less common such as Syriac (thousands of web pages). So if a model works very well in English or other “high-resource” languages, do not assume that it will also work well in “low-resource” languages, including the ancient Aramaic dialects used in parts of the Talmud.
Advice: Check anything written by an AI
If you ask an AI to help write a paper, including finding related work, check that what the AI has written makes sense, and that the referenced papers actually exist! AIs have a habit of making up references, and fabricated references can lead to papers being rejected and indeed a loss of academic reputation.
Final comments
As above, my comments are very simplistic from the perspective of NLP researchers! But researchers in many other fields are trying to use AI, and the above points may not be obvious to people with limited experience using AI.
