AI in research: fundamentals
How AI systems work
Generative AI tools include large language models (LLMs), which produce text by predicting likely patterns in their training data, and image-generation models, which create or edit images from prompts. These are powerful tools that can help with summarizing, coding, creating images, and structured reasoning - but they have predictable limitations, so understanding how to navigate them safely is key.
With appropriate oversight, AI can improve efficiency and analytical thinking. Without verification, it can introduce avoidable errors. Across these guide pages, assume that outputs can vary even with the same inputs; the model may agree with your framing, or it may invent or omit details when uncertain. Treat AI outputs as provisional and match your level of review to the potential impact if the output is wrong.
AI for researchers: getting started
Start exploring by asking the AI tool itself what it’s capable of. It’s often the quickest way to generate ideas, structure thinking, and test possibilities. Ready-to-copy prompts are included throughout the AI playbook for researchers to help you get started.
📑 Copy and paste prompt: |
|---|
Before we start, explain: - What kind of AI model you are (e.g., text, image, multimodal). - In simple terms, how you generate outputs. - What you are good at. - What you are not reliable at. - Common mistakes users make when using you. - How I should verify or review your outputs for work tasks. Keep the explanation concise and practical. |
Exploration should always be informed. AI systems have known and inherent error rates due to how they are designed and trained. Before relying on AI outputs, you must understand the potential risks, the level of impact your task carries, and the principles that ensure responsible use.
Because AI outputs can contain inaccuracies, fabrications, bias, or hidden assumptions, the context in which you plan to use the output is critical for assessing risk and impact. Not all uses of AI carry the same level of consequence.
For example, any AI-generated output that will be used, quantitatively, statistically, to inform formal conclusions or to support decision-making with real-world consequences - must be treated as high impact. In these cases, further independent verification and safety checks are critical.
As a starting point, you should only use an AI tool if you can confidently answer yes to all four of the following checkpoints:
Checkpoint | Confirmation |
1. Impact and oversight | I understand the likely impact if this output is wrong, and I am applying oversight that matches that impact. |
2. Policies and governance | The relevant policies allow AI use for my role and context (institution, funder, journal/publisher, ethics/data agreements, applicable law). |
3. Permitted inputs | The information I am inputting is permitted to share with this tool, given its data handling practices. |
4. Verification | I can verify the output before I (or others) rely on it. |
If you answer no to any of these, limit AI use to low impact/low risk tasks, or do not use it.
If you can proceed with AI, you must apply the BE WISE framework, maintaining the six principles of human oversight throughout your workflow.