Executive Summary: AI Integration in the Laboratory
The Laboratory of Quantitative Biology is actively integrating AI, particularly Large Language Models (LLMs), to significantly boost research efficiency and equip personnel with crucial skills for the evolving scientific landscape. Our approach is guided by two core principles: maintaining researcher agency and taking full ownership of AI outputs.
We primarily utilize ChatGPT, Gemini, and NotebookLM (for source-grounded work), with an eye on Microsoft Copilot’s growing integration. LLMs are invaluable for enhancing scientific writing, performing deep literature searches, and accelerating coding tasks. While we strictly discourage LLMs for raw data analysis, they are powerful for understanding methodologies and debugging code.

The most critical skill for effective AI use is prompt engineering, which involves providing clear context, specific requests, and defined personas to optimize LLM responses and minimize errors. This strategic AI adoption aims to amplify our research capabilities and productivity.

The adoption of Artificial Intelligence (AI) in our laboratory is fully embraced and encouraged. We view AI as a transformative set of tools poised to significantly enhance our research efficiency, making its integration a clear imperative. In fact, developing proficiency in AI is becoming increasingly crucial in the current scientific environment, and neglecting this skill could prove to be a significant limitation. This guide is designed to introduce you to the specific AI tools we employ and outline the best practices for their use in our daily work.

Legitimate and responsible AI use

In our laboratory, we fundamentally distinguish between two main categories of AI, each with specific considerations for their legitimate use.

• You fully understand the underlying principles and what the tool is doing.
• The tool has been thoroughly tested on control data to validate its performance.
• The tool, and its application, are made available to others in the lab, and eventually the wider scientific community, to ensure full reproducibility of your findings.

2. Generative AI: Tools like Large Language Models (LLMs) fall into this category. The use of generative AI is equally subject to the same stringent rules of research integrity. However, due to its conversational nature, it might be easier to misunderstand what constitutes legitimate use.
Our overarching guidance for generative AI is simple and centres on two core principles:

Do Not Lose Agency. This is paramount. You are the scientist, and you retain ultimate control and responsibility. You decide what is scientifically sound, what information is included in your work, how it’s presented, and which hypotheses to pursue. Generative AI is a powerful assistant, designed to boost your efficiency either by:

• Accelerating tasks: Helping you complete work more quickly.
• Enabling deeper exploration: Facilitating searches into ancillary, yet highly useful, topics that might otherwise be too time-consuming to explore. The AI is a tool to amplify your capabilities, not to dictate your research

Own It. Any error made by generative AI becomes your error. While AI is impressive and can be incredibly exciting, it remains a tool, not a sentient collaborator. You must take full ownership of all its outputs. Failing to supervise AI-generated content rigorously can severely damage your reputation, and by extension, the lab’s reputation, due to inaccuracies in content, inconsistencies in style, or inappropriate language.

For instance, if an AI generates a biased or inappropriate statement and you incorporate it into your work, that statement becomes your responsibility, as you are obligated to review and validate all AI outputs.

Our AI toolset

The field of AI is advancing at an extraordinary pace, making any definitive list of tools quickly outdated. However, to provide practical guidance, it’s important to highlight the core generative AI tools we actively utilize in the lab today.

Currently, our primary tools are ChatGPT, Gemini, and NotebookLM. While the University may not yet provide institutional licenses for the advanced tiers of all these platforms, their free versions offer substantial functionality that we find highly beneficial. Furthermore, the more affordable paid tiers typically provide a generous expansion of capabilities and extended usage limits, which can be a worthwhile investment for increased productivity.

We are aware that the University does provide access to Microsoft Copilot, which is powered by OpenAI’s ChatGPT technology. We anticipate that Copilot’s integration into the familiar Microsoft 365 ecosystem will lead to a significant increase in its adoption and impact on our work in the future. As of the current date, while promising, Copilot’s capabilities are still evolving, and we continue to evaluate its effectiveness in significantly impacting our specific research workflows.

Leveraging Large Language Models (LLMs)

By now, the general utility of Large Language Models (LLMs) should be quite evident to most of us. Tools such as ChatGPT, Gemini, and Claude, alongside emerging products like Microsoft Copilot, are incredibly powerful for a range of applications. They are exceptionally useful for drafting or correcting text, refining code, and conducting advanced searches that synthesise information more effectively than traditional methods.

LLMs and Text Generation

Given their foundational design as language models, the application of LLMs for writing or correcting text is perhaps their most intuitive and straightforward use. While their capabilities in this area are impressive, I generally discourage using LLMs to generate extensive text from scratch. The primary reason for this stance is to ensure you maintain full intellectual agency and prevent the accidental inclusion of inaccuracies or unverified information.

However, there are two significant exceptions where generating text from scratch with an LLM can be highly beneficial:

1. For initial planning or brainstorming difficult content: If you’re grappling with initiating a challenging piece of writing—be it a complex grant section, a difficult discussion for a manuscript, or a detailed experimental plan—an LLM can serve as an invaluable sounding board. Think of it as seeking an initial opinion or a structured starting point, much like you would from a trusted colleague, friend, or supervisor. The LLM can help you overcome writer’s block by generating a preliminary outline, suggesting themes, or providing a first draft of an argument. Crucially, after this initial input, you then take full control: you write the actual content, shaping the narrative, style, and scientific accuracy to your precise requirements. Remember, this approach fully aligns with the principles of maintaining your agency and ownership over your final work.
2. For deep searches: LLMs can be incredibly powerful tools for complementing your existing knowledge within a field, or for rapidly gaining a foundational understanding of an entirely new area. Gemini’s “deep search” capabilities, for instance, are particularly strong (though ChatGPT is continually improving in this regard). These tools can synthesise extensive information from various sources into concise summaries, allowing you to grasp complex subjects in a remarkably short time. However, it is absolutely paramount that you verify all references and do not blindly trust the information presented. When discussing insights gained from LLM-assisted searches with other lab members or during presentations, ensure you clearly distinguish between knowledge derived from verified primary literature and information initially synthesised by an LLM. This transparency is crucial for conveying the level of confidence in your source, just as you would differentiate between information from an excellent journal article or a week yet interesting one.

LLMs for enhancing your writing

Perhaps the most common, though certainly not the least impactful, application of LLMs in our daily work is their utility in refining and enhancing your written output. Rather than simply asking ChatGPT or Gemini to rewrite your text, I strongly advise using them as sophisticated feedback mechanisms.

Think of an LLM as a highly attentive editor or a writing coach. You can provide your drafted text and then prompt the AI to:

• Provide feedback on clarity and conciseness: “Review this paragraph. Is it clear? Are there any redundant phrases or words that could be removed to make it more concise?”
• Assess scientific coherence and logical flow: “Does the argument in this section flow logically? Is there any part that feels disconnected or requires more detailed explanation to maintain scientific coherence?”
• Identify repetitions and unnecessary content: “Scan this document for repeated ideas or phrases. Are there any paragraphs that could be condensed or are unnecessary to the core message?”
• Suggest alternative phrasing or word choice: “I’m struggling to express this complex idea simply. Can you suggest alternative ways to phrase this sentence for better readability without losing scientific accuracy?”
• Check for consistency: “Ensure that all acronyms are defined upon first use and used consistently thereafter. Also, check that gene names are formatted correctly throughout the manuscript.”
• Targeted grammatical and stylistic improvements: Beyond basic spell check, you can ask for suggestions on improving sentence structure, verb tense consistency, or adherence to a specific academic style.

The versatility of LLMs extends to very specific and high-stakes writing tasks. For instance, you can upload a draft of your manuscript alongside the referees’ comments and your rebuttal letter. Then, you can prompt the LLM to assess whether your responses clearly and fully address the referees’ critiques. Similarly, if you’re drafting a grant proposal, you can provide the proposal text and the funding call guidelines, asking the LLM to analyse how well your proposal aligns with the call’s specific requirements and priorities. With the right prompt engineering, the insights you can gain from LLMs on these complex documents are substantial.

Consider LLMs as an always-available, highly knowledgeable senior colleague who can offer rapid advice and a fresh perspective on your work. However, it is absolutely critical to understand that LLMs are not a substitute for genuine human interaction and collaboration. Direct engagement with your peers, mentors, and more senior colleagues within the lab and the broader scientific community remains indispensable. These human interactions foster critical thinking, provide nuanced feedback, introduce diverse perspectives, and are fundamental to your intellectual growth, the robust revision of your work, and the collaborative spirit of scientific discovery.

LLMs and literature reviews

For those of us who remember, the traditional approach to literature review involved hours in the library, poring over index cards or abstract books to initiate searches. This was a painstaking and slow process, yielding a curated selection of papers that you’d then meticulously study. The references within those papers would, in turn, serve as the springboard for subsequent rounds of exploration. While this method was incredibly valuable for fostering a deep, focused understanding of the literature, the sheer time commitment was substantial.

The advent of the internet and the maturation of search engines fundamentally transformed this process. Today, our literature searches are primarily conducted through powerful databases like PubMed and Google Scholar. We still navigate from paper to paper by following citations, but now at the instantaneous speed of clicking a DOI link. The unprecedented speed and vast breadth of information we can now access online are invaluable. However, this accessibility is also countered by the limited time we have available to truly delve into every relevant research paper. Consequently, most of us inevitably miss some papers, quickly skim through many others, and are only able to properly study a select few.

This is precisely where generative AI becomes a game-changer. After some initial, admittedly “glitchy” starts, tools like ChatGPT and Gemini now possess remarkable capabilities to identify, search, synthesize, and analyze academic papers. While it remains crucial to complement our searches with established databases like PubMed and Google Scholar to mitigate potential biases inherent in LLM training data, these AI models are truly incredible new tools for engaging with the literature.

You can leverage standard chat interfaces to query LLMs for specific topics and relevant literature. Additionally, you can request “deep searches”—a feature where Gemini currently holds an edge over ChatGPT—to delve more comprehensively into a subject. You can even instruct LLMs to restrict their searches to specific criteria, such as “reputable academic literature,” to refine the quality of results. Critically, LLMs can efficiently and accurately summarize individual papers or even entire collections of papers that you might otherwise not have the time to read in full. This allows you to quickly grasp the essence of many studies, reserving your in-depth reading and critical study for only the most pivotal papers, as you normally would.

NotebookLM

NotebookLM is a remarkable tool developed by Google and powered by its Gemini models, stands out as an exceptionally powerful and unique AI resource for our lab. Unlike general-purpose LLMs such as ChatGPT or Gemini (when used without specific source grounding), NotebookLM’s core design principle is to restrict its knowledge space exclusively to the documents and audio files you upload. This “source-grounded” approach ensures that all its responses, summaries, and generated content are directly derived from your specific materials, minimising “hallucinations” and maximising relevance.

This capability makes NotebookLM invaluable for managing and extracting insights from your curated research library. You can upload vast collections of papers relevant to a project, your own drafts and writings, audio recordings of lectures or interviews, or even detailed guidelines for grant applications. You then interact with this personalised knowledge base using natural language queries.

Beyond simple querying, NotebookLM provides highly effective tools for:

• Summarising Content: It can distill complex documents or entire collections into concise summaries, helping you quickly grasp key arguments and findings without having to read every word.
• Producing Briefing Documents: You can generate compact briefing documents, FAQs-like summaries, or structured study guides based solely on your uploaded materials.
• Creating Audio Overviews: One particularly innovative feature is its ability to transform your sources into engaging, podcast-like audio overviews. This allows you to “ingest” complex content while commuting, exercising, or performing other routine tasks, providing a flexible and efficient way to stay on top of your literature.

NotebookLM essentially transforms your personal collection of research materials into a dynamic, queryable, and summarizable database, allowing for a depth of engagement that was previously impossible.

ChatGPT projects

ChatGPT Projects provide a powerful framework for enhancing your interaction with the LLM by creating dedicated workspaces. This feature lets you group related chats, maintain persistent context, and infuse the LLM’s general knowledge with project-specific instructions and documentation.

Within a project, you can:

• Maintain consistent instructions: Set overarching guidelines for the LLM across all related conversations, ensuring consistent understanding for long-term tasks.
• Embed project-specific documentation: Link or upload relevant files for the LLM to reference, providing it with your specific research data or protocols.
• Keep conversations organised: Group all project-related chats for easy tracking and continuity.

While projects allow specific documentation, their knowledge space is not as strictly restricted as NotebookLM’s. ChatGPT will still draw on its broader knowledge, but your project’s context and files will significantly guide its outputs for focused research.

Custom GPTs (ChatGPT) and Gems (Gemini): tailoring AI behavior

Both ChatGPT and Gemini offer powerful features to customise their behaviour to specific needs: Custom GPTs in ChatGPT and Gems in Gemini. These tools allow you to provide detailed instructions and upload specific documentation, enabling the LLM to act as a specialized assistant for particular tasks or domains.

A significant advantage of Custom GPTs is their shareability. Unlike Gemini’s Gems, which are currently private to individual users (even within an organisational workspace), Custom GPTs can be shared with others, making them excellent for collaborative lab efforts.

We leverage Custom GPTs widely for both teaching purposes and administrative tasks. For example, we’ve developed a Health & Safety GPT that aggregates all relevant University policies and lab-specific templates. This provides an instant, reliable resource for our team, ensuring consistent adherence to safety protocols and supporting our day-to-day laboratory operations efficiently

Microsoft Copilot

While Microsoft Copilot might currently appear to lag behind some other LLMs in certain aspects, we anticipate it will soon undergo a significant qualitative transformation. Just as Gemini is progressively integrating into Google’s vast ecosystem of business productivity tools, we expect the same deep integration for ChatGPT and Microsoft products through Copilot.

Given that our University primarily utilizes Microsoft tools, and with Copilot becoming increasingly embedded across applications like Word, Excel, and PowerPoint, we predict a substantial uptake of Copilot for many tasks we currently perform with ChatGPT. This synergy means that capabilities developed within ChatGPT will likely be inherited by Copilot, making it an incredibly powerful and pervasive assistant.

Our lab is actively planning to migrate our Custom GPTs into Copilot once the integration at the university level matures. We believe this move will further streamline our workflows and enhance collaboration within the familiar Microsoft environment.

Images, video, and audio: expanding our creative toolkit

Our adoption of AI extends beyond text to visual and auditory mediums, opening new avenues for communication and outreach.

Video generation is rapidly advancing, and although it remains relatively expensive for complex projects, its capabilities are truly incredible. We anticipate being able to generate high-quality videos for lab outreach or presentations with unprecedented ease soon, transforming how we visually share our research. I have done some experimentation with Gemini (Veo3), their Google Flow dedicated platform and also with Google AI studio.

Audio generation is a much more mature area. Tools within Gemini, with its “audio overviews,” and NotebookLM are particularly robust. NotebookLM, in combination with platforms like Spotify (by uploading generated audio), allows us to produce compelling audio content for outreach, education, and initiatives like CenGEM. This offers an efficient way to consume and disseminate information, even while multitasking.

Image generation is now remarkably easy, with ChatGPT being our primary tool, but with Gemini’s most recent model (nano banana) cathing up and even better for specific tasks. We regularly use ChatGPT to create visually appealing cover art for projects and public-facing materials, as well as artistic representations for presentations. AI-generated images are increasingly appearing in our “products,” excluding, of course, our scientific manuscripts, where primary experimental data is paramount.

Data analysis: where LLMs fit

While our lab actively uses and develops deep learning for image analysis, it’s crucial to clarify our stance on LLMs for data analysis. We do not use, and strongly discourage the use of, LLMs for raw data processing, statistical computations, or drawing direct scientific conclusions from datasets. LLMs aren’t built for numerical rigour and can “hallucinate” with complex data, compromising research integrity.

However, LLMs are incredibly useful for investigating and understanding data analysis methods. You can query them to learn about statistical tests, explain complex algorithms, or even generate/debug code snippets for data processing. Essentially, you can use LLMs to understand how to analyse data, but perform the actual analysis with dedicated statistical software and established pipelines.

Coding with LLMs

We are still exploring the use of LLMs for coding in our lab; initial tests are very promising. LLMs are increasingly integrated with platforms like GitHub, and we anticipate this will seriously impact our operational efficiency.

Both ChatGPT and Gemini can directly execute Python code. Gemini also offers direct export to Google Colab, which is highly useful for our data-driven projects, like the data scraping from Google Scholar we’ve used for identifying new papers for social media.

We’ve also used ChatGPT for Matlab code generation. While it successfully accelerated certain coding aspects, we found that as code grew in length or complexity, interacting with ChatGPT became considerably more challenging.

Lately, we have started to use Visual Studio for Code with CODEX (OpenAI/ChatGPT) and Gemini. Both are quite slow because they are not directly integrated into Visual Studio but rather interact with the files in the folder through PowerShell. However, they have been particularly useful to code the new CenGEM’s website. While the initial coding might not be faster, changes are then implemented as if using a content management system, but in natural language.

Like other applications, LLMs are excellent for getting hints, debugging, and as replacements for extensive manuals. While we still handle our main software development more traditionally, we are confident LLMs will soon adeptly manage very complex coding projects.

Mathematics and Physics

Our adoption of generative AI has significantly accelerated due to its impact on mathematical work. In photon statistics, for instance, we’ve trained ChatGPT on our past research, enabling us to replicate weeks of work in mere hours. This vastly increases our productivity in this specialised field.

The same efficiency gains apply to certain areas of physics. While I wouldn’t rely on LLMs to design a microscope from scratch, they are excellent for retrieving equations and relevant literature to verify assumptions. Their impact on physics research isn’t yet as profound as in math, but they serve as an incredibly smart search engine for all quantitative information.

Incidentally, ChatGPT has a strong knowledge of online optical catalogues like Thorlabs and Edmund Optics. This means it can effectively assist in finding information or recommending the best optical elements for a project

AI Agents: Our Next Frontier

We’re currently exploring the capabilities of AI agents within the lab, which represent a significant leap beyond traditional LLMs by autonomously planning and executing complex tasks. However, our experience with these advanced systems is still nascent, and we’re in the early stages of understanding their full potential and practical applications in our specific research workflows. For now we have a single-use agent for specific tasks (e.g., harvesting information from institutional pages of colleagues to create a new webpage) or – with varying success – for periodically screening funding calls to match ideas with opportunities.

Prompt engineering: mastering AI interaction

Prompt engineering is arguably the most crucial skill we need to develop in this new AI era. Think of it as how we “program” LLMs: the way we phrase our questions directly dictates the quality and relevance of their responses. While we aren’t responsible for their underlying code or training, our expertise in a topic combined with a deep understanding of how to effectively interact with LLMs is essential to unlock their maximum potential. Without this skill, we simply won’t get the most out of these powerful tools.

Before even crafting the prompt itself, there are crucial foundational steps you can take to significantly enhance your LLM interactions:

Instruct on specific behaviours (Contextual customization): For more targeted use, Custom GPTs and ChatGPT Projects (in ChatGPT) or Gems (in Gemini) accept specific instructions. These instructions guide that particular instance of the LLM to act in a defined role—for example, as an editor, a researcher, or to always fact-check its outcomes, ensuring its responses are tailored to your immediate needs.

Engineer their knowledge space: As discussed, strategically limiting an LLM’s knowledge base (e.g., via NotebookLM) or providing specific documentation (e.g., in ChatGPT Projects or Custom GPTs/Gems) ensures more relevant and accurate outputs.

Instruct on specific behaviours (Global customisation): For overall consistency, you can customise the general behaviour of your ChatGPT. For example, you can set a default instruction like: “Use a formal, professional tone. Be fact-based and try to double-check your observations. Tell it like it is; don’t sugar-coat responses.” (Note: This specific global trait-option isn’t currently available in Gemini in the same way).

Crafting Effective Prompts: The Three Pillars

When you’re ready to construct the actual prompt, its effectiveness hinges on three critical elements. These ensure the LLM understands exactly what you need, minimising “hallucinations” and maximising relevant, high-quality output:

1. Clear Context: Explicitly provide the background that best frames your query. This helps the LLM understand the situation, purpose, and audience for its response.
   • Example: “I’m a scientist writing an introduction for a manuscript (which I’m uploading, along with the title and abstract I’m confident in). The introduction needs to provide a good overview of the topic accessible to any biologist, even non-specialists. I plan to submit this manuscript to Journal X.”
2. Clear Request: Be precise about the task you want the LLM to perform. While LLMs can do many things, specifying your exact requirements, including any dos and don’ts, is vital.
   • Example: “Check the introduction for clarity. Provide specific examples of what I should change while preserving both my style and narrative. Do not rewrite the entire section, just suggest improvements.”
3. Specific Persona: Tell the LLM who it should “be” when answering. This can significantly influence the tone, depth, and perspective of its response, allowing you to tap into expertise that might not be immediately available to you.
   • Example: “Answer as if you’re an expert colleague with extensive experience publishing in Journal X. Be helpful without sugar-coating any criticism.”

By meticulously defining these three aspects, you effectively “program” the LLM to deliver the most valuable assistance for your specific needs. To illustrate the power of combining these elements, consider the stark difference between a vague request and a fully engineered prompt. A highly effective prompt, incorporating Context, Request, and Persona, might look like this:

“I’m a scientist writing an introduction for a manuscript (which I’m uploading, along with the title and abstract I’m confident in). The introduction needs to provide a good overview of the topic accessible to any biologist, even non-specialists. I plan to submit this manuscript to Journal X. Answer as if you’re an expert colleague with extensive experience publishing in Journal X. Be helpful without sugar-coating any criticism. Check the introduction for clarity, provide specific examples of what I should change while preserving my style and narrative, and do not rewrite the entire section, just suggest improvements.”

This is vastly different, and infinitely more effective, than simply typing: “Revise my text.” The detailed prompt guides the LLM to deliver precise, actionable feedback tailored to your specific needs and goals.

Beyond better feedback: the broader impacts of excellent prompts

Crafting clear, specific, and instructive prompts does more than just elicit professional and precise feedback from LLMs. These thoughtful prompts and your targeted follow-up questions also have two other critical benefits:

1. Minimizing Hallucinations: LLMs are known to “make things up” when they don’t have the correct answer. While newer models show improved reasoning and a reduced propensity for these “hallucinations,” it’s still common for them to attempt an answer even when uncertain. You can actively mitigate this by including instructions like: “Do not provide suggestions if the text is already of high quality,” or “Include in the search only authoritative sources.”
2. Countering Sycophancy: Some LLM versions have a tendency to “please” the user, offering flattering responses or agreeing with your (potentially incorrect) assumptions, especially in follow-up questions where you identify initial shortcomings. You can “box the sycophant” by explicitly granting the LLM permission to disagree or critically assess your input. For example: “I think your previous answer is incorrect because of Y. Double-check my observation, regenerate your answer, and explain why we disagreed. Consider I might be wrong, but use only scientific evidence to provide your answer.” This encourages a more objective and critical response.

Conclusion: Embracing the AI era

The landscape of AI is evolving at an incredible pace, and we’re discovering new applications for these powerful tools week after week. We are, without a doubt, in a period of rapid transition towards a fully integrated AI era. For now, this guide provides a foundational overview of how we can effectively leverage AI in our lab to enhance efficiency and accelerate our scientific discoveries. Let’s continue to explore and innovate responsibly together.

Wait… how did I use AI for this page?

First, I have written the text and the examples. I have asked both ChatGPT and Gemini to revise the page providing the link. However, despite using a prompt I thought was good, both tools were changing the text substantially. I am sure I could have improved the prompt further but I opted to revise paragraph by paragraph asking to simply revise the text for clarity. Each section was then checked by me and I have corrected a few sentences that changed meaning in the process.
The resulting text is clearer albeit a bit longer. Then I asked for an executive summary that I copied and pasted (after checking it) as provided by Gemini.
Lastly, I asked to generate a shorter and catchy blog post out of it… which is now public after my revision.