AI is reshaping life science research, from literature search to protein structure prediction to experimental design. 

But sometimes, it feels like you need to speak a different language to use it successfully. From our research team to yours, here’s a breakdown of all you need to know to get started. 

What is AI?

Artificial Intelligence (AI): Software that performs tasks normally requiring human judgement, such as recognising patterns, summarising text, making predictions. It's a tool trained on data to do specific jobs, but it’s by no means sentient. 

🔬 In practice: Your email spam filter is AI. So is predictive text on your phone. At the more intelligent end, an AI might be able to scan microscopy images and flag abnormal cells.

Machine Learning (ML): A type of AI that learns from examples rather than being explicitly programmed with rules. You feed it data, it finds patterns, and it improves with more data.

🔬 In practice: A model trained on thousands of labelled cell images learns to distinguish cell types, without a programmer manually writing rules for every edge case.

Model: The trained system that takes an input and produces an output. When you use an AI tool, you're interacting with a model. Models vary enormously in capability depending on what data they were trained on and for what purpose.

🔬 In practice: When you ask ChatGPT a question, it uses its latest model to search the internet, then construct a response. Depending on the question and type of response you want, you might need to select a more robust model or allow the model more processing time.

Large Language Model (LLM): The technology behind tools like ChatGPT and Claude. Trained on enormous amounts of text, it can read, write, summarise, and reason in natural language (aka the language you use to ask it to do something).

🔬 In practice: LLMs can be used to search and summarise literature, draft reports, interpret experimental results, and answer scientific questions in plain language.

Generative AI: AI that creates new content (text, images, molecular structures, protein sequences) rather than just classifying or predicting from existing data.

🔬 In practice: The AI that writes a lab report summary is generative. So is an AI that designs a novel protein scaffold from scratch.

AI agent: An AI that doesn't just answer questions but takes actions, such as searching databases, running analyses or calling other tools, to complete a multi-step task autonomously.

🔬 In practice: An agent could be given the goal 'find all papers on CRISPR base editing from 2024 and summarise the key mechanisms' and execute that end-to-end without step-by-step instructions.

How do I use AI?

Prompt: The instruction or question you type into an AI tool. The quality of your prompt directly affects the quality of the output.

🔬 In practice: "Summarise this paper" is a prompt. "Summarise this paper for a non-specialist, focusing on the key finding and its clinical relevance" gets a far better result.

Prompt Engineering: The practice of crafting better prompts to get better outputs from an AI. Despite what the name might suggest, no coding or engineering background is required. Prompt engineering is all about knowing what context to include to make sure you get the most appropriate answer to your question.

🔬 In practice: “Act as an expert science communicator. Summarise this paper for a non-specialist audience. Focus on the key finding and its clinical relevance.” is a well engineered prompt that will change the depth and vocabulary used in the model’s response. 

Output / Response: What the AI produces in reply to your prompt. Could be text, a table, a summary, a prediction, or an image. You should always critically evaluate a model’s output, especially where factual or statistical claims are being made. 

🔬 In practice: In your prompt, you can ask the agent to structure its response in an easy to review format. For example, “Summarise your response in bullet point format, with a source link for each bullet. Provide no more than 10 bullet points.” 

Context Window: A model’s short-term working memory. This is the amount of text and data a model can process at once. 

🔬 In practice: If you’re continuously working off a single chat in ChatGPT or Claude, you might notice responses get worse or less relevant over time. This is because the model can no longer process the information from the start of the chat - it’s fallen out of the ‘context window’. 

Fine-tuning: Taking a general-purpose model and further training it on specialist data to make it better at a specific task.

🔬 In practice: Giving a general LLM a sample of your written content (blog posts, essays etc.) can fine-tune it and make it more reliable at producing text responses that match your tone of voice. 

How do I use AI responsibly?

AI tools are powerful, but they have real limitations. Understanding these isn't optional, especially in a research context. Uncritical AI use can compromise data integrity, reproducibility, and scientific credibility.

Hallucination: When an AI confidently produces something that is factually wrong or entirely made up. An LLM might cite a paper that doesn't exist, or calculate its own statistics without a factual basis. 

🔬 In practice: Always verify AI-generated factual claims against primary sources before using them in your work. Include “Cite all of your sources in your response” in your prompt to make fact checking easier. 

Bias (in AI models): When a model systematically skews its outputs because of imbalances in its training data. For example, a model trained predominantly on data from one cell line or one species may perform poorly on others.

🔬 In practice: Understanding bias is critical for knowing when to trust AI outputs. Always check what data was used to train a model, and don’t rely on general-purpose LLMs for highly specialised tasks. 

Explainability / Interpretability: How well you can understand why an AI reached a particular output - not just what it concluded. In a scientific context, knowing how a model reached a certain conclusion will help you validate and explain results in papers or presentations. It’s also crucial if you plan to use an output to guide future research questions. 

🔬 In practice: Keep track of which model you are using for a particular question or instruction and take some time to understand the limitations of that model. Including a phrase like “Explain your processing steps and any assumptions made in generating your response” in your initial prompt will also help you keep track of what a model is doing. 

Over-reliance and skill erosion: Using AI to interpret all your results or write all your text can erode the critical thinking and domain expertise that makes a good scientist. AI should augment your judgement, not replace it.

🔬 In practice: You can use AI to help you understand a topic (“Teach me about DNA synthesis. Tell me what videos to watch and articles to read to get up to speed as a first year undergraduate biology student. Give me a summary of the latest breakthrough experiments in this area.”), but never start by asking ChatGPT to do an assignment for you. Once you understand a topic or a process, you’ll be able to write much more efficient prompts and you’ll be able to critically evaluate and edit a response. 

Data privacy: Be careful what you paste into public AI tools. Unpublished experimental data, proprietary sequences, or personal data (such as demographic or medical data) should not be entered into consumer AI interfaces without checking your institution's data governance policy.

🔬 In practice: When linking your Google Drive to LLMs like ChatGPT or Claude, make sure it always asks for permission before reading a document. Be wary of complete open access controls. Some LLMs now have strict biosafety rules that will trigger if you are asking it to process sensitive biological data, so always check what context or data you are providing before running a prompt. 

Confident ≠ correct: AI outputs are often fluent and authoritative in tone regardless of accuracy. A well-written response is not evidence of a correct one. Calibrate your scepticism to the stakes: the higher the consequence of an error, the more rigorously you should verify. 

🔬 In practice: As in explainability/interpretability, ask the model or agent you are using to explain its assumptions and processing steps so you can evaluate the output. Students who use AI without input from traditional sources (like journal articles, teachers or peers) tend to overestimate how much they have learned from AI. If you’re unsure, double check what an AI has told you by conducting more traditional research, or phone a friend!

Join our AI literacy series

If you’re keen to take your AI skills to the next level, and learn how to apply AI in your research, keep an eye on our events page for upcoming events and workshops. 

If you have a special use-case for AI you’d like to discuss with us, get in touch!