Suggested AI Uses for Research Tasks

Ideation

Planning code or projects
Recommending project names or acronyms
Generating hypotheses
(See this example)
Mining the literature for research ideas
(See this example)

Text

Summarizing documents or content
Combining documents
Transforming format (e.g. text to table)
Searching & navigating text
Adjusting communication

Visual Design

Creating flyers for outreach
Creating presentation graphics
Creating logos (Keep in mind that graphic artists are still likely to do a better job of creating something closer to your exact vision!)

Code

Planning steps
Refactoring code
Annotating code
Writing documentation
Checking for security or privacy concerns
Understanding someone else’s code
Understanding errors
Proposing code validation/testing ideas
Translating code to a new language

Project Management

Developing templates
Breaking down tasks
Assigning roles
Creating meeting agendas and summaries
Creating lab document drafts
- Code of conduct
  (See this resource)
- Mentee Agreement
  (See this resource)
- Lab Handbook
  (See this example)

Tips for AI Use

Don’t use prompts that would violate data privacy restrictions or licensing (most commercial tools are not private!)
Be specific and iterate on prompts
Where possible work in an environment that allows for projects to retain the context
Typically asking AI to improve existing work is better than starting from scratch
Check everything! Ask the AI to question itself and have a human (maybe you) review text and code

example of more precise ai prompt - can you improve the paragraph is vague, can you make the paragraph more concise is better, can you make the paragraph more concise and appropriate for a highschool student is even better

Challenges and Mitigation Methods

Challenge	Mitigation Methods	Resources/Tools
Giving Credit Other people’s work was used to train models sometimes without permission.	● Choose tools that are more transparent about what data was used for training ● Be transparent about how you use AI tools, include tool versions ● Ask the tool for sources to credit and check what it gives you	● Check out IBM’s AI attribution toolkit ● Consider tools like OLMoTrace which show what training was used for the AI response
Data Privacy Some tools are using data that should not have been available	● Choose tools that are more transparent about what data was used for training ● Look into those data sources and see if they followed data privacy regulations. ● Never use private data or info in a prompt for a commercial tool	● See if your institute offers private tools ● Models run locally may be more private
Environmental Impact AI tools use data centers that require lots of electricity and water for cooling.	● Consider if AI is the best tool for the task ● Choose AI tools that are transparent about energy use and attempt to improve efficiency ● Choose models that use a smaller number of parameters, designed with methods like parameter-efficient fine-tuning (PEFT), or simply designed for more specific tasks and therefore often requiring less memory usage ● Consider using a model locally on your computer ● Consider tools with data centers in cooler climate locations or those that use the heat that is generated for other uses	● GreenPT is a very eco-conscious tool ● Ollama can help you run models like Gemma 2, Mistral AI, Phi-4 locally ● Offset AI is a tool to help track and offset the environmental impact of your AI usage
Trust and Deskilling Using AI too much can degrade trust in yourself and potentially result in the loss of skills you once had.	● Use AI for more specific help, like polishing as opposed to writing things from scratch ● Use AI only when it is likely to save you time or tedium ● Check in with trainees about AI use	● Check out this paper where humans were shown to over-trust AI ● Check out this paper where developers thought AI made them faster
Distorted responses and Hallucinations Even simple requests can generate false or skewed responses	● Check for errors ● Ask tools to consider the potential for distorted responses in your prompts ● Challenge tools when they fail to recognize errors ● Recognize that human oversight is necessary	● Check out this resource from MIT

Resources

Read more about why small models may be the future for more sustainable AI
Check out more about PEFT for efficiency
Read more about the MIT AI data provenance initiative
These ITN courses discuss more about AI use for research and coding:
- AI for Decision Makers which includes a mini course about AI Ethics
- AI for Efficient Programming

Authors: Content for this cheatsheet came from Carrie Wright. It was summarized and formatted by Carrie Wright and reviewed by Kate Isaac. Icons from https://www.iconpacks.net. The cheatsheet was also inspired by a discussion that involved the ITCR OPEN community (https://www.itcrtraining.org/open).
This content is free for noncommercial reuse with attribution. CC-BY-NC