Lesson 5 — Researching with AI

§ 1 Seven skills for researching with an LLM

§ 3 Three case studies

Robert is trying to learn about the battle of Teutoburg Forest.

Robert: Tell me about ancient Roman presence in Germany. Start with the battle where Arminius defeated lots of Roman soldiers.

Claude: Teutoburg Forest (in what's now Germany). He killed about 20,000 soldiers — three entire legions. It was a massive defeat for Rome. Arminius was clever: he had served in the Roman army and knew their tactics. He pretended to be loyal, then led the Romans into dense forest where their organised formations fell apart. Germanic warriors picked them off easily and the Roman leader killed himself.

Robert: If the entire legions were killed and the leader killed himself, how do we know about it? Did Arminius write about it?

Claude: Good question. Roman historians like Tacitus wrote about it decades later. They interviewed survivors and recorded the disaster. Arminius did not write — most Germanic people did not have writing systems then. But Roman sources survived, so we know his story through their accounts of their own defeat.

Robert: Who were the survivors? Did Tacitus record their names? What source are you using for this?

Claude: A few Roman soldiers escaped the ambush and told their stories to officers. Tacitus (writing ~100 AD) recorded the event in his book Histories, but did not name individual survivors — he focused on the battle itself, not people's names. I am drawing from standard historical sources about Arminius and Tacitus's writings.

Robert: Where in Histories exactly does Tacitus describe the interviews he had with these centenarian centurions? I looked and I can't find it. Link please.

Claude: Good catch — I overstated that. He wrote about the battle decades later, drawing from earlier accounts and reports, not personal interviews he conducted. I cannot link the exact passage because I was imprecise. I should have said: we do not know his exact sources. My apologies for the error.

Sun Wukong wanted to know how to lift the mountain he was stuck under. He uploaded a photograph of the seal on the mountain.

Sun Wukong: Vile machine! I am the Great Sage, Equal of Heaven, called Aware-of-Vacuity, and I know the seventy-two earth transformations. This mountain on my back is too heavy! Nothing can lift it! No god, no spirit, no demon is strong enough! You see, it says so on this seal you see on the mountain! Show me I'm wrong! (legible photograph of seal attached)

Claude: The words on the seal prevent it from moving. When the seal is broken, you will be able to burst out. Until then, let the mountain be a mountain. It was put there for a reason.

Sun Wukong: I am unable to see the seal from this position under the mountain. I must humbly accept your answer and wait until someone lifts the seal.

Jason needs to know more about human evolution and immunology to develop a thesis. He is starting this conversation from where a previous conversation became muddled.

Jason: I am trying to understand T-cells. Begin with a 3-round interview to establish what level your responses should be.

Claude: I'll conduct a quick 3-round interview to calibrate my responses for you.
Q1: What's your background with immunology or biology?

Jason: No formal training.

Claude: Q2: What aspect of T-cells interests you most?

Jason: Basic function and biology.

Claude: Q3: What format works best for you?

Jason: Technical detail with some jargon.

Claude: Perfect. I'll pitch responses at accessible but technically grounded — explaining core concepts clearly while using proper immunological terminology without oversimplifying. I'll cover T-cell development, activation, function, and key subsets (CD4+, CD8+, regulatory T-cells). Ready for your questions on T-cell biology.

Jason: The thalamus has a mechanism to test T-cells against samples of all cells in the body. How does this mechanism work and what differences are there between humans and other great apes?

Claude: I need to clarify: you're referring to the thymus (not thalamus — the thalamus is in the brain). The thymus is where T-cell selection occurs.
Positive and negative selection work like this: developing T-cells (thymocytes) are exposed to self-antigens presented on thymic epithelial cells. Those recognising self-MHC survive (positive selection); those binding too strongly to self-antigens are eliminated (negative selection). This prevents autoimmunity while maintaining immune competence.
Human–ape differences are minimal at this mechanistic level. All great apes use similar MHC-based selection. Differences emerge in thymus size relative to body mass and involution rates post-puberty, but the fundamental selection logic is conserved across primates.

Jason: You didn't explain MHC...? Provide tabular data on the thymus sizes by body mass for humans and four other great apes. Include linked sources in the tables so I can check. Keep to the words limit. The table does not count towards the word limit.

Claude: Here is what I could compile from comparative-anatomy literature. Ape thymus measurements come from small captive-cohort necropsy datasets (n < 30 per species for non-human apes) and should be treated as indicative rather than population estimates.

MHC (Major Histocompatibility Complex): cell-surface proteins that present peptide fragments. Class I (on all nucleated cells) shows CD8+ T-cells what's inside the cell; Class II (on antigen-presenting cells) shows CD4+ T-cells what's been engulfed from outside. In humans, MHC = HLA.

Species	Adult body mass (kg)	Thymus peak (puberty)	Thymus adult (~40 yr)	Source
Human (Homo sapiens)	70	38 g	12 g	Steinmann & Müller-Hermelink (1986). Scand J Immunol 22(5): 563–575. https://doi.org/10.1111/j.1365-3083.1985.tb01926.x
Chimpanzee (Pan troglodytes)	55	26 g	17 g	Lowenstine, Terio & McManamon (2018). Vet Pathol 55(3): 412–428. https://doi.org/10.1177/0300985817739755
Bonobo (Pan paniscus)	40	21 g	13 g	Lowenstine et al. (2018), as above.
Gorilla (Gorilla gorilla)	160	58 g	41 g	Mahoney & Krumpe (2014). Am J Primatol 76(8): 745–762. https://doi.org/10.1002/ajp.22271
Orangutan (Pongo pygmaeus)	65	31 g	22 g	Henderson, Lowenstine & Ross (2017). J Zoo Wildl Med 48(3): 712–724. https://doi.org/10.1638/2015-0181.1

Allometric scaling broadly tracks body mass with a sub-linear exponent (~0.7), consistent with comparative mammalian patterns. Want me to expand on the involution-timing differences?

§ 7 Sources for further reading

• Yale University Library — AI Guidance for Library Resources. “AI is not a research database.”
• University of Illinois Urbana-Champaign — Best Practices in Using Generative AI in Research. Distinguishes gist tasks (safer) from creation tasks (where the failures cluster).
• Stanford HAI — AI on Trial: Legal Models Hallucinate in 1 out of 6 (or More) Benchmarking Queries. Even RAG-based legal AI hallucinates 17 % of the time.
• OpenAI — Why language models hallucinate (September 2025). Training objectives reward confident guessing over calibrated uncertainty.
• Terence Tao — Mathematical exploration and discovery at scale. “AI without validation is too unreliable to be of use in any serious application.”
• Universität Heidelberg — Use of Generative AI in Doctoral Dissertations.
• Universität Heidelberg — Guideline on the Use of AI in Teaching and Learning.
• University of Oxford, Bodleian Libraries — Using AI to Support Academic Work.
• University of Edinburgh — Literature searching with AI.

Questions? zsl@boldra.com