AI Tools for Research

emLab Land and People lab meeting

Gavin McDonald

2026-02-26

Welcome

  • AI is transforming the way research is done
  • AI is rapidly evolving - tools and capability changing daily (this talk will probably be out of date by next week!)
  • Today I’ll focus on tools I’ve personally used, but this is by no means exhaustive
  • Since tools are changing so rapidly, we’ll also talk about best practices
  • I’m no AI expert - but rather just an enthuastic (and optimistically skeptical) user
  • I also really want to hear from you

AI can help with every stage of the research process

  1. Idea generation
  2. Literature review
  3. Data analysis
  4. Writing and communication
  5. Collaboration and project management

A bit on different LLM options

  • Google’s Gemini
    • General purpose, good for lit review, idea generation, etc
    • Excellent data privacy protections through UCSB license
  • Anthropic’s Claude
    • Excellent for coding
    • Direct agentic integration with your IDE
  • GitHub Copilot
    • Designed for coding, integration with your IDE and Github
    • Can use many LLM backends (Gemini, Claude, GPT, etc.)

What LLM is best for coding?

Short answer: Claude (for R, at least for now…)

Check out Claude’s Constitution

Keep an eye on Posit’s AI Newsletter and R Benchmark tests

Data privacy

  • Always check the data privacy policies of any AI tool you use
  • UCSB’s Gemini license has excellent protections
  • Claude Code has settings for data retention, whether your code is used to train models, etc

Research process (1/5): Idea Generation

  • Gemini, Claude, GPT
  • Gemini Deep Research
  • Gemini Gems

Research process (2/5): Literature Review

  • Gemini Deep Research
  • Gemini Gems
  • Gemini Notebook LM
  • Google Scholar Labs
  • Specialized tools: Research Rabbit, Nature Research Assistant, Elicit, Consensus

Research process (3/5): Data analysis

Coding agents: GitHub Copilot, Positron Assistant, Positron Databot, Claude Code, etc

  • Integrated directly into your IDE
  • Have access to your codebase, file structure
  • Can operate in different modes: ask, edit, or agent, depending on your needs and comfort level
  • Great for data science workflows (but anything really)
  • Often “BYO-key”
    • Can use any LLM backend - Claude, Gemini, etc.
    • You’re also subject to that backend’s data privacy policies

Positron Databot

  • Developed by Posit team (formerly RStudio) for Positron (modern polygot successor to RStudio IDE; VS Code fork)
  • Allows you to interact with your data using natural language
  • Designed for exploratory data analysis (can do ML too)
  • Designed with responsible, human-in-the-loop use in mind (Databot is not a flotation device)
  • Uses a WEAR loop: Write code, Execute, Analyze, Regroup

“In my 30-year career writing software professionally, Databot is both the most exciting software I’ve worked on, and also the most dangerous.” –Joe Cheng, Posit CTO

Positron Assistant

  • Developed by Posit team for Positron
  • General coding assistant for wide range of tasks
  • Similar to Claude Caude, but specifically tailored for data science workflows, with specific R and Python tooling
  • Can be used for code generation, debugging, documentation, and more
  • Has access to your codebase and file structure, so it can provide more context-aware assistance
  • BYOK: Can use various LLM backends (Claude, GPT, etc.)
  • Can be used in ask, edit, or agent mode

Which LLM backend for Positron Assistant and Databot: Claude or GitHub Copilot with Claude?

  • Claude requires a paid account; Copilot has options
  • Copilot gives access to Claude models, but also others
  • Using Claude directly gives you access to full context window - it is capped when going through Copilot
  • Using Claude directly is faster
  • Using Claude through Copilot can quickly exhaust your Copliot credits (at least wotj the education account)

Now what about Claude Code?

  • Claude Code has a VS Code extenion, which is similar to Positron Assistant: both provide agentic capabilities
  • Assistant is tailored for data science with R and Python; Claude Code more towards general software engineering
  • Claude Code can be used in VS Code or Positron; Assistant can only be used in Positron (and is tightly integrated)
  • Assistant is BYOK; Claude Code only uses Claude models
  • Important: Both Positon Assistant and Databot require an API key; so this works with pay-as-you-go Claude, but currently not with Claude Pro monthly subscription (that might change though)

Getting fancy: Customizing your AI coding assistants

  • instructions.md: Specify custom “always-on” instructions: coding standards, style guides, etc to use across all scripts (specify the how) (also called claude.md or positron.md)
  • prompts.md: Define reusables prompts for tasks you commonly ask your assistant to do (e.g., “write a function that does X”, etc) (specify the what)
  • agents.md: Create custom agent personas that can perform specific tasks, such as data cleaning, EDA, or model training (specify the who)

Research process (4/5): Writing and Communication

  • Gemini, Claude, GPT
  • Specialized tools: Research Rabbit, Nature Research Assistant, Elicit, Consensus
  • Gemini Nano Banana for image generation (e.g., flowcharts, technical diagrams, etc)

“Nano Banana Pro is the first image model that can sometimes generate coherent technical diagrams”

- Sara Altman and Simon Couch, Posit (source)

Research process (5/5): Collaboration and Project Management

  • GitHub Copilot for project management and collaboration
    • Can generate issues, pull requests, documentation, etc
    • Can be used to review code and suggest improvements (either reviewing PRs, or even reviewing code before it’s committed!)
    • Can help with project organization and workflow
    • Can be done either on GitHub website, or directly through Positron or VS Code IDE
  • Various other AI-powered tools in Slack, Asana, Zoom, etc

Resources

Examples

(Thanks, Robert!)

Getting Started

Recommended First Steps:

  1. Try Gemini for literature summaries
  2. Become familiar with GitHub Copilot
  3. Try a coding assistant for your next data science task
  4. Try making your own custom instructions file or Gem

Best Practices

  1. Remain accountable

  2. Verify AI-generated content - keep a human in the loop

  3. Check sources

  4. Cite usage appropriately

  5. Maintain data privacy (as needed)

  6. Stay critical, skeptical, open-minded, and curious

Thanks!

This presentation created with Quarto, Positron, and GitHub Copilot

Reach out: Gavin McDonald

gmcdonald@bren.ucsb.edu

Discussion

  • What AI tools are you currently using in your research?
  • What challenges have you faced with AI tools?
  • How can we use AI ethically and responsibly for environmental research?
  • What are your best practices for using AI?