Blog header with presidents and a veterinarian

What the last 20 US Presidents can teach veterinarians about AI

There is a lot that the last 20 US presidents can teach veterinarians about Artificial Intelligence.    

No, I don’t mean the things that President Obama could teach you directly about AI, though I am sure he is a prompt master. I mean that the concept of “the last 20 US presidents” can help us understand the strengths and limitations of AI systems.  

Let me explain…  

I spend a lot of time talking to veterinarians about AI. I’ve written papers and position statements on the subject, consulted with large organizations on policy, presented on both diagnostic and general AI at conferences around the world and appeared on podcasts to discuss AI in vet med. And the more I talk to veterinarians about it, the more I realize that their view of AI is directly correlated to how well they understand how AI and LLMs work.

Veterinarians who know that there are different models available within each tool tend to have a more positive view of AI and use it more regularly. Whereas, those who think of ChatGPT as one thing tend to view AI more negatively and find themselves disillusioned with AI, seeing it as all hype. 

AI generated image of 20 US presidents with many inaccuracies, wrong pictures, wrong names, repetitions
A meme of US presidents with incorrect faces, names and dates generated by AI

This was crystalized to me by an interaction with a friend and colleague ahead of a recent lecture I was giving on AI. Without hesitation, they told me that ChatGPT sucked. As an avid user, I was surprised by this so I asked what they meant. Their reply surprised me: 

“I have a lot in common with most people. I’m a casual user who reads headlines that have consistently fed the message that AI is here to radically transform our lives. But this new technology that is here to revolutionize the way we work can’t even do a simple task such as produce a list of the last 20 US presidents with photos.” 

And it’s true. ChatGPT might have trouble with this task. But the problem here isn’t the AI. The problem is how someone chose to use it. Let’s explore the example of listing the last 20 US presidents with photos to teach veterinarians all about AI, with a focus on OpenAI to create a mini guide to using ChatGPT. 

How AI chat systems figure out who the last 20 Presidents are

When you access an AI tool through ChatGPT, Gemini, Claude or others, the most common interface is “chat.” When you type a message, this is known as a prompt. The prompt is then received by a large language model (LLM). You can think of the LLM as the artificial brain you are interacting with. You can ask a question like “who were the last 20 US presidents” and the model responds.    

There are different versions of LLMs. They differ not only in who created them (OpenAI made and operates ChatGPT, Google has Gemini, and Anthropic has Claude) but also when they were trained, how they were trained, what their strengths and weaknesses are, and perhaps now most importantly what kind of tools they have access to.  

Different models are kind of like different people.  If you asked me to list the last 20 US presidents, I would do a very bad job because my “training data” (Canadian upbringing) and recall ability is poor.  I would correctly identify the presidents I remember in my lifetime and start to get tripped up on who immediately preceded George HW Bush.  LLMs don’t have the same issue, since their training data is the entirety of the internet. 

However, a key limitation of an LLM is the cutoff date of its training data. While you and I have an active working memory of the current US president, the LLM is a brain stuck at a certain point in time. For the most recent OpenAI models (GPT-5) the training cut off is June 2024. So, if we ask it to list the last 20 US presidents it might get stuck… 

A list of presidents made by ChatGPT where the AI is unaware it is 2025 and Donald Trump is president.

Just like I have access to tools (the internet) and can look up who the last 20 US presidents are, modern LLMs have internet access and can look things up to bring their knowledge base up to the current date.  

The new GPT-5 models have been billed as “auto-switching”, which means the model contextualizes what tool to use and how much effort to apply (more on that to come) based on what you ask. Unfortunately, in my experience, the auto-switching often works poorly as you can see in the example above. It’s a simple query, so it did not trigger a check for the current date and the current US president, even though it probably should have.  

Users can ask the models to use their tools. This can be done through prompting or tool selection. To prompt the model, you often must explicitly tell it what to do, e.g. “use the internet and search for the current date and current president, then make a list of the last 20 US presidents.” There are some short cut techniques like telling it to “think harder” which tend to have mixed results. 

If you don’t know what the training cutoff date of the model you are using is, you can usually ask it. 

chatgpt message asking about the training cut off and getting it to use the internet
Asking ChatGPT its training cut off and getting it to use the internet to determine who the current president is.

Different tools can also be explicitly selected from the chat menuThis menu will look different depending on if you are logged in and if you have upgraded to a paid plan. Upgrading to the Plus or Pro version from OpenAI provides additional access to some usage limited features such as deep research (5 per month for free tier), opens up other options such as “Agent mode” and allows the user to explicitly select different models such as GPT-5 Instant, GPT-5 Thinking Mini, GPT-5 Thinking and legacy models such as GPT 4o and GPT o3. 

Language and Vision in modern LLMs

Most of the current LLMs have options to provide responses in both written text and with images. This makes them actually Large Multimodal Models (LMMs) but this distinction is pretty pointless today.  

Image generation models are a tool within the chat interface and are similar to their language counterparts but instead of giving a response as text, they create an image based on the prompt given. They can create art in different styles when requested.  

This is the type of model used to create the hilariously inept response my friend shared.  But just as I couldn’t list all the past 20 US presidents, most people also cannot reliably draw from memory the last 20 US presidents.  The issue isn’t with the system, it’s with what you are asking it to do. In the example below, I asked it to create an image of all 20 presidents but asked for a single headshot. It did its best to comply and did a remarkably good job despite the lack of clarity in my prompt. 

Asking ChatGPT to create an image of presidents in ultra-realistic black and white. Inconsistencies in the tense (asking for a single close up but also 20 of the last presidents) makes the model guess at what we want.

Its wild to realize that is a fake image of Donald Trump.  It looks 100% accurate.  Hence the challenge with “deep fakes” today. But the more you ask of the image creation model, the more it pushes at the edge of what is possible and starts to break down. 

Refining the prompt from above to indicate there should be 20 headshots. Notice there are only 16.

But this doesn’t mean AI sucks or that it can’t reliably give you a list of the last 20 US presidents.  For that we just need to use the right tools. 

Thinking mode, Deep Research and Agents

When ChatGPT was first released in November 2022, everyone was impressed by the speed at which it could give you realistic answers. But users quickly pushed the limits of it.  It appears the key to success with a LLM such as ChatGPT is patience -a thing my mother always reminded me “is a virtue.” These new tools (thinking mode, deep research and agents), all take longer to give a response but tend to give better, more accurate, and more fulsome answers when compared to the fast models.  

Thinking mode is when the AI focuses on step-by-step reasoning. Once prompted it indicates “thinking” and shows some of its process. You can expect longer, structured answers (lists, pros/cons, worked examples) that help you reason through a problem. These models are strong on logic and synthesis, but you should still verify factual claims. If you want to try this for your clinic try a prompt like “Help me think through the pros and cons of offering ultrasound services in a small clinic step by step.”  

Deep research is when the AI actively gathers, compares, and cites up-to-date sources (papers, websites, news) and synthesizes them into a concise summary. This mode is good for finding current facts, citations, product specs. It is suggested to use it for preparing evidence-based material like papers, blog posts, or grant backgrounds. However, in my experience the sources are often lacking. While it does provide direct citations or links, I am always careful to consider those. If you want to try this, select Deep Research and prompt “Provide me a review of the evidenced based practices associated with artificial intelligence in veterinary medicine.  Focus on the pros and cons of use and select peer reviewed sources when available.”  

Agent mode behaves like a helper that can take multi-step actions (search the web, run tools, fetch files, draft and format documents). Think of it as delegating a small project. I find it is useful for repetitive workflows, actions that need sequential steps, or when you want the system to run a process for you (for example: compile a literature list and format slides). This is still a new process, and I admit while I have played around with it, I have not pushed the boundaries of veterinary use cases. If you subscribe to Plus or Pro you can give it a shot with something like: “gather five recent peer reviewed papers on veterinary AI, save their citations, and produce a 1-page summary.  Bundle the PDFs of the papers and the summary into a single file.”  

With these tools and skills veterinarians can get more of AI. Not just making lists of the last 20 US presidents but using the right tool for client communications, using thinking mode to assist with differential diagnosis generation, deep research to support investigating new topics, and agent mode to accomplish tasks like downloading articles or other mundane internet related work.  

Thinking mode was able to pull together an accurate list of the last 20 US presidents and embed their actual images from wikicommons. 

This leads me to my last point.  ChatGPT can create a reliable list of US presidents and their photos in less than 2 minutes. But it’s still important to verify the information, because AI systems can still make mistakes. In the example below agent mode did not confirm the date and, it forgot to include Donald Trump as the current US president, which is easy to spot. But let’s imagine this was something more obscure, and it forgot a differential or got a dose wrong. Would you notice?   

A similar prompting exercise as above but using ChatGPT Agent who did not verify what year it is before completing the task.

Related Articles