AI systems facing clinicians today

aArtificial intelligence is rapidly moving from the laboratory to everyday clinical conversation. Over the past two years, many clinicians have been experimenting with systems that allow them to draft letters, summarize research papers, and explain complex topics in plain language. These tools are powered by what are called large-scale language models (LLMs).

Despite the growing interest, the state of AI systems can feel perplexing. New models emerge every few months, each promising improved inference power, a wider context window, or increased speed. The important question for most clinicians is not which model is best, but what these systems are and how they differ from each other.

At its simplest, a large-scale language model is a type of machine learning system that is trained on vast amounts of text. By analyzing statistical patterns in language, the model learns to predict the most likely sequence of words given a particular prompt. This feature allows the system to generate consistent responses, summarize documents, and answer questions across a wide range of topics. It is important to emphasize that these models do not “understand” medicine in the human sense. Identify patterns in your data and generate responses based on probabilities. When used carefully, it can be a very useful assistant. If used uncritically, it can produce plausible but inaccurate information.

Various AI features
At one end are larger flagship models that are powerful systems that can analyze complex documents, create detailed reports, and manipulate images and code. Although these are very high performance, they are also computationally expensive. On the other side are lightweight models designed to respond to simple queries quickly and cheaply. These are often used in applications that require a quick response. In between these two extremes are middle-tier models, which provide a good balance of functionality and speed and can handle the majority of everyday AI tasks.

Finally, there are specialized models that are trained to perform specific tasks such as legal research, biomedical analysis, and coding. Understanding this spectrum helps explain why dozens of different models exist. They are just optimized for different jobs.

Closed model and open model
Another important distinction for clinicians is between closed (proprietary) and open or open weight models. Closed models are developed by commercial companies and accessed through a website or programming interface. Internal architecture and training data are typically not publicly available. In contrast, an open model makes the weights of the underlying model available. These can often be downloaded and run locally, allowing organizations to host them privately if desired.

This distinction is important in the healthcare field because it impacts issues such as data governance, privacy, and transparency. This makes open LLMs more compliant with EU AI law than opaque closed LLMs. Closed systems tend to be easier to use and often provide the best performance. However, open models offer greater control and can be deployed within a secure environment.

Common closed large language model
Several proprietary systems currently dominate the world of AI.

These models represent some of the most advanced general-purpose AI systems currently available and are widely used for writing, research, coding, and data analysis.

Common open or open weight models
In parallel to these commercial systems, an ecosystem of open models is rapidly growing.

For example, the BLOOM model was created by a French-led international research collaboration and was designed as a publicly accessible alternative to proprietary models. Trained on hundreds of billions of words across dozens of languages.

In recent years, open models, especially those developed in China and Europe, have become more competitive. For example, DeepSeek and Qwen have gained significant attention for their ability to provide strong performance at low cost. For healthcare organizations, open models offer the potential to run AI locally without sending data to external companies. This is an attractive option for institutions concerned about patient confidentiality.

Where clinicians may encounter these systems
Large-scale language models are already appearing in several clinical settings. They are used for:

• Summarize the medical literature.
• We will help you write a letter from the clinic.
• Explain complex symptoms to patients in easy-to-understand terms.
• Analyze datasets or guidelines.
• Support research and education.

Many clinicians are already experimenting with them informally, often asking AI systems to explain unfamiliar topics or summarize long documents. The key principles are the same as for other digital tools, and clinicians must remain responsible for interpreting and validating the output.

Introducing the AI ​​browser
Another notable development is the emergence of AI-native web browsers. Traditionally, clinicians accessed information through browsers such as Chrome, Edge, and Safari. However, AI systems are increasingly being integrated directly into the browsing experience.

New platforms such as ChatGPT Atlas, along with AI-enabled versions of existing browsers, aim to enable users to interact with the internet conversationally. Rather than searching for individual papers, clinicians can ask the browser to summarize evidence for a treatment or analyze research papers.

Microsoft is integrating Copilot AI into Edge, and Google is building Gemini into Chrome. These developments suggest that AI may soon become the standard interface for accessing medical information online.

A rapidly evolving landscape
One of the challenges of writing about artificial intelligence is that the field is evolving so rapidly. The models released today are likely to be replaced by more capable versions in the coming months.

But for clinicians, it’s not just about remembering the names of individual models. Instead, it’s important to understand broad categories such as powerful flagship models, lightweight models optimized for speed, open models that can be run locally, and specialized systems built for specific tasks. Additionally, those that are GDPR and EU AI law compliant.

Once you understand this framework, new models become much less scary.
Because, like any tool in medicine, artificial intelligence is most useful when clinicians understand what it can do and where not to trust it.

author
Professor Ray O’SullivanFRCOG FRCSEed MRCPI MSc DipCyber ​​DipBME. Associate Clinical Professor, Department of Obstetrics and Gynecology, St. Luke’s Hospital. Founder of VoxAI.