Taking up AI one step at a time
A writer or an organisation takes up a language-model tool in a cycle of steps, from understanding the technology to bringing a chosen tool into the work, and the cycle begins again as the technology changes.
The adoption cycle
An AI integration framework breaks the adoption of the technology into manageable steps that ensure an informed and thorough approach, with which individuals and corporations can gain the benefits of the evolving technology while minimising its risks. The cycle has five steps — AI literacy, considering use cases, assessing risk, evaluating tools, and implementing with risk management — and is relevant for leaders, employees and freelances alike; as the technology evolves and new applications become possible, the cycle begins again.
AI literacy
Successful application of language models to medical writing hinges on a clear understanding of their benefits and risks. AI literacy, the first step, gives a foundation for informed decisions and appropriate expectations of what AI can do. Understanding how a model works helps a user predict which tasks are appropriate for it.
Writers may meet language models in a standalone chatbot such as ChatGPT, as a function in software such as Microsoft's Copilot, or in an internet browser such as Chrome. A chatbot, standalone or inside an application such as Word, typically presents a blank box, which leaves the decision about which tasks are suitable and reliable entirely to the user. AI literacy training on which tasks suit a model lets writers use the tools effectively, raising productivity while keeping the high standards of accuracy and context sensitivity the field needs.
Use cases
AI literacy inspires creative exploration of which uses would suit AI tools. With a general sense of how models work, a user can think of routine challenges in their work that a model could help with, but should avoid AI solutionism, the feeling that a model can solve everything, where an expert system or regular software would do a better job. The ideal use cases are those that can benefit from a model's unique capabilities. Because a model can give inaccurate information, the output of a use case should be externally verifiable.
Large language models and large multimodal models can augment a user by taking on simple tasks, assist a user step by step, and amplify a user by expanding their skill set.
- To augment is to delegate a specific task to the model and review the outcome. Examples are formatting a list of abbreviations by fixing capitalisation and spelling with awareness of proper nouns, aligning references to an example style, and converting photographs of handwriting, slides or scanned documents to editable text.
- To assist is to collaborate as the model helps write, edit and create content. Examples are drafting a presentation script for a slide from its title and key points, turning text prompts into pictures for presentations or social media, and walking a user step by step through common computer problems, where the user should wait for a human to help if the model recommends entering an administrator password or editing registry files.
- To amplify is to gain new capabilities from the model, such as coding, creating a data visualisation or teaching a new concept. Examples are converting a request into a PubMed search string with Boolean operators, writing a Visual Basic for Applications macro for Microsoft Office and taking the user through the steps to run it, and basic data analysis, parsing large data sets and providing charts to visualise them, where the user should be mindful that models do not clean data automatically.
Risk assessment
Before tools built on language models are used, understanding their potential risks and planning how to mitigate them is paramount. Establishing a risk profile helps identify the tools that fit it. The risks of language models can be due to training data limitations, human factors, functional limitations or implementation challenges.
- Training data limitations: a model's knowledge is rooted in its training data, and gaps or weaknesses in that data can produce hallucinations — false information invented by the model — outdated output or biased responses. Examples are a data set with nothing after the training cut-off, unless the model is connected to internet browsing; outdated practices and language, with the model unaware which practices are now preferred or required; and missing context from material that was never digitised, such as conference presentations, sits behind a paywall, such as journal articles, or is in an inaccessible format, such as regulatory guidances in PDF.
- Human factors: models are very different from any technology previously available, which can introduce risks from their human users. Examples are automation bias, the assumption that machine-generated content is accurate; distrust, leading to loss of interest from readers, attrition of employees or client dissatisfaction; supplying the model with an incorrect or outdated source; and damage to the model from poisoned training data or from manipulative prompts, known as prompt injection.
- Functional limitations: the machine learning process is the root of some risks in the way models function; models are probabilistic systems that cannot be predicted or entirely understood. Examples are an opaque decision-making process; inconsistent output even with identical prompts; no ability to cite sources unless the model is equipped with retrieval-augmented generation; and the possibility of leaking a user's inputs, such as proprietary data or protected health information, into other users' outputs.
- Implementation challenges: implementing AI systems presents challenges that should be considered as part of a risk assessment. Examples are cost, including employee time and opportunity cost; obsolescence as the technology quickly advances; and finding legal, nonproprietary training data.
Risk evaluations should be captured together with the planned mitigation steps in a risk management plan, for which the National Institute of Standards and Technology's AI Risk Management Framework and Playbook is a resource.
Tool evaluation
With the use cases and the risk profile defined, a tool is chosen that both satisfies the use cases and mitigates the critical threats. Evaluating a tool should establish which model it uses and what modifications, if any, have been made to that model, since a tool built on an earlier model may be a lot less capable than more recent models. For an expert system with language-model features, the evaluation should establish which features rest on the model and which on more deterministic programming, so as to judge whether the tool matches the risk profile.
Because machine learning involves various degrees of learning, it can help to think of a tool's capabilities in potential layers, each like a stage of a medical writer's development.
- The foundation model, such as GPT-4 in ChatGPT, is like a medical writer's college education, giving a broad base of knowledge.
- Fine-tuning the model on additional specialty data sets such as clinical study reports is like the specialised knowledge a writer gains in a graduate or certificate programme.
- Giving the model access to the user's own data is like on-the-job training.
- A collection of proven prompts in a prompt library mimics the efficiency gained with work experience.
A task delegated to a beginning medical writer comes with more detail, instruction and follow-up than one delegated to an experienced writer, and the same logic applies to a tool. A tool with limited layers of capability needs a prompt with a lot of context and specific instruction, and its output may need substantial revision; a tool with several layers can take a more straightforward prompt, and its output will need fewer edits.
Implementation with guardrails
The last step is to integrate the tool into the work or the organisation, transparently and with appropriate guardrails. Transparency is critical before, during and after implementation, and it is vital to address the reservations stakeholders may have and to set guardrails for AI use. Partners and users should be told clearly what the tool can do, who will be authorised to use it, when AI use is appropriate, why the change is being made and how to use it appropriately. As of 2024, AI policies that record and share these principles were becoming a common business practice, and a policy can give rigid guardrails against the risks identified in a writer's or an organisation's own risk management framework.
Implementation is likelier to succeed when it starts with one or two use cases a model can consistently make easier. The feeling of productivity and success this brings can spark more interest in trying other use cases.