Is AI a Good Research Assistant? [Wading Through AI - Episode 9]

Episode Focus

This episode investigates whether current large language models can serve as effective research assistants for historical research (not coding). The test is controlled, spoiler-free, and based on two recent talks by the host—one of which involves the origin of the phrase “premature optimization is the root of all evil.”

How the Test Was Designed (What They Actually Did)

• The researchers created a prompt/query using a one-paragraph description of the task provided by the host (the host is the subject of the hidden talk content).
• A second speaker then added minimal but operational instructions:
  • Consider source types including:
    • Academic journals
    • Conference proceedings
    • Online discussions
  • However, prioritize first-hand knowledge, not merely “people on Reddit/Hacker News think X.”
  • Run the work as a research process that produces two outputs:
    • A linear research log in the order sources are discovered
    • A second synthesized report reviewing the findings
• Each model was run once (roughly 20 minutes per model) using “work as much as it wants” style instructions—explicitly not an interactive/iterative loop or agent workflow.
• Models evaluated included:
  • OpenAI (ChatGPT / “OpenAI” in discussion)
  • Anthropic (Claude)
  • Google Gemini
  • Grok is mentioned, but it was not included in the core comparison.

Key Findings by Model

Gemini (Google)

• The host’s first impression was that Gemini’s output was “primitive/naive” compared to the others.
• It allegedly produced mostly obvious, generic statements (including repeating the maxim itself), leading them to judge it not useful for real historical research in this single-pass setup.
• Nuance: They argue Gemini might improve in a feedback/agent loop, but this test intentionally limited evaluation to a single pass with no external intervention.

OpenAI (ChatGPT)

• OpenAI was viewed as producing the most research-assistant-like framing:
  • It distinguishes between what a source actually supports versus what it merely claims.
  • It warns about evidence boundaries (e.g., treating blog claims as claims, not definitive proof).
• The host particularly liked that OpenAI’s report was consistent and less prone to overclaiming.

Anthropic (Claude)

• The host valued Claude’s presentation/compositional skills and reported no egregious hallucinations of quotes.
• For historical research reliability, Claude was criticized for:
  • Treating weaker sources (like blogs) as more conclusive evidence than they really are.
  • Rushing toward conclusions—using phrasing that suggests something is settled when it’s actually conditional or interpretive.
• Example: A section framed as “numerical overliteralism” for figures 97% and 3% was argued by the speakers to be based on a misunderstanding of what the referenced primary sources actually say.
• Overall assessment: Claude seemed “reasonable,” but less trustworthy in epistemic calibration—i.e., knowing when evidence is insufficient.

Common Technical / Epistemic Issues Discussed

• Single-pass vs agent loops
  • Performance can look worse in one-shot prompting.
  • Open models can become more competitive when used in iterative agent/feedback settings (tests plus corrections).
• LLM confidence and uncertainty
  • Models may not reliably know when to say “I don’t know” or “can’t verify.”
  • This connects to active applied research on certainty/epistemic reliability, including concepts for legal AI (they mention “Harvey” as an example).
• Mechanism critique
  • LLM generation is described as driven by “conversation-ending/completion” dynamics, which can seep into reasoning and cause oversimplified or overconfident outputs.
• Hallucination vs paraphrase vs inference errors
  • They observed relatively low risk of gross hallucination (fabricated quotes were not strongly seen).
  • Still, there were notable detail-level misunderstandings and interpretive leaps.

Practical Conclusions / What the Researchers Think AI Is Good For

• AI research reports can outperform “AI summaries” that appear at the top of a search result because they can provide:
  • Structured complexity, not simplistic answers
  • Outputs that, given enough time, may be readable selectively by a person rather than accepted blindly
• Major limitation: AI won’t automatically surface the most interesting narrative elements.
  • The host describes a “97.3 problem”: even if AI captures most of the macro structure, the most interesting details may be missing—the “needle in a haystack” ~3% still requires human reading.
• Proposed workflow
  • Use AI to generate a bibliography / link set for manual verification.
  • Use AI as a secondary “did it find things I missed?” tool, not as the final authority for narrative or interpretation.
• Personal example (math/topology task)
  • AI helped locate an Internet Archive scan of Hausdorff’s book that keyword search wouldn’t surface.
  • Caveat: AI can still produce incorrect translations/interpretations in some steps.

Notable Side Detail: “Process Transparency” They Liked

One model output included:

• A statement about future archival project priorities
• An explicit disclaimer framed like: “most reasonable conclusion given available records”

This was singled out as a positive example of how AI can communicate reasoning boundaries.

Main Speakers / Sources (As Mentioned)

• Dmitry Spanos (co-host / organizer; involved in AI testing and explanation)
• Casey (the “Casey gave me…” persona used in the prompt workflow discussion)
• OpenAI (ChatGPT)
• Anthropic (Claude)
• Google (Gemini)
• Grok is mentioned, but it was not evaluated in the main test