Until fairly recently, the field of artificial intelligence (AI) has focused on specific domains, such as image analysis, general information look-up, or very specific models, such as financial risk prediction or basic primary care triage. The explosion of publicly available information alongside access to cost-effective processing time on cloud computing infrastructures has enabled researchers to train vast generalpurpose neural network models. Eventually, these models will allow us to ask a search engine a question and for it to provide us with a curated response, as opposed to simply generating a list of related websites.
However, these models go far beyond just answering questions – they can help us understand complex topics and can generate full articles or inspire us to dig deeper into a subject.
We will look at two publicly available AI engines – one that can generate quite detailed responses to questions (ChatGPT), and another that can synthesise original artwork based on a description (DALL-E).
ChatGPT
ChatGPT (Chat Generative Pre-Trained Transformer) was released to the public in November 2022 by the OpenAI Research Laboratory in San Francisco, US. It has been trained using deep learning techniques to be able to reply to questions by analysing vast amounts of information sources, including publicly available books, articles and websites. The ChatGPT model was fine-tuned by repeatedly altering parameters and testing the output until it produced responses indistinguishable from those one might receive from a human.
» Not only can ChatGPT give well thought-out answers to questions, but it is also capable of generating articles based on the questions it is asked, and refining that answer based on subsequent conversations with the user «
The results are ground-breaking; not only can ChatGPT give well thought-out answers to questions, but it is also capable of generating articles based on the questions it is asked, and refining that answer based on subsequent conversations with the user.
Using ChatGPT is relatively straightforward. You need to navigate to https://chat.openai.com/ and use your Google account to register in order to use it for free. You will be presented with a user interface similar to a messaging application. Here are a few questions you can try:
- ►Write an essay on clinical leadership versus management for NHS GP practices
- ►What will be the impact of the economic recession on the aesthetics industry?
- ►What is the socioeconomic profile of Manchester?
- ►What are the wellbeing benefits of aesthetic wrinkle reduction?
- ►Write an upbeat article about Eurovision coming to Liverpool
- ►Write a critical article about Eurovision coming to Liverpool
- ►Write an article about <your street>,<your town>
- ►Write a limerick about aesthetic clinics.
In all cases, ChatGPT will generate a well thought-out article of around 300 words. It is also capable of adapting the article based on further feedback. For example, on the Eurovision questions above, if you ask ChatGPT to refine or expand an article, such as including Ukraine as a topic, it will rewrite the article accordingly. However, always fact check your article, as there is a risk of exaggerating or getting facts wrong.
DALL-E
DALL-E is another AI project named after the surrealist artist Salvador Dali and Pixar's animated character WALL-E. Whereas ChatGPT creates text, DALL-E generates original artwork from your description.
Again, using DALL-E is relatively straightforward. Once you navigate to https://labs.openai.com, you can type in a description of the image you would like to see and in what style. Here are a few examples you can try:
- ►Blackpool Pleasure Beach in the style of Dali
- ►An oil painting of Dunham Massey at sunset
- ►A pastel painting of a kitten playing on the beach
- ►A relaxed patient undergoing a wrinkle reduction procedure as a photograph
- ►Medieval woman using an iPhone as an oil painting
- ►A bowl of soup that is also a portal to another dimension.
As with ChatGPT, the more expressive the description is, the more unique the image will be. Unlike ChatGPT, the number of images that can be generated for free is limited, but it is possible to buy more credits for just $1.00 USD each, making it a credible competitor to stock photograph sites if the images are not used for commercial purposes. However, as it is a research programme and has not been formally released, please check their website for updates to their policy.
Marketing opportunities
ChatGPT can get you started on publishing regular blog articles for your social media feeds;. however, you will still need to think of engaging topics for your audience. I recommend creating a list of topics alongside target publication dates for your social media sites. The output from ChatGPT will only be as good as the questions and refinements you feed into it.
DALL-E is capable of producing artwork based on the keywords you use for your blog topics, and by consistently specifying the style you want the image rendered as, you will end up with a consistent look and feel for your images. DALL-E works quite well for abstract images, but I suggest still using professional artwork for images portraying clinical procedures.
Combined, these tools can dramatically increase your marketing productivity and inspire you to try new approaches. However, I still strongly recommend monitoring the efficacy of the articles you publish, as articles that do not take into account your audience preferences will fail, regardless if their interests are ignored.
Ethical considerations
As a test, I asked ChatGPT to proofread this article, to which it made the following observation:
“However, I did notice that the introduction mentions the year 2022, which hasn't happened yet.”
The above error highlights one of the limitations of ChatGPT – notably, the model it is trained on is >14 months old, and it is only as good as the data it was trained on.
An example of well-trained models in a more clinical setting becoming out of date can be seen in the APACHE II/III predictive mortality models developed in the early 1990s for determining the efficacy of intensive care. The APACHE models use a severity scoring model based on the first 24 hours of admission to the intensive care unit. A coefficient based on the admission reason and calibrated on a large dataset can provide a relatively accurate risk of death (ROD) score. However, changes to how deteriorating patients are treated prior to admission causing lead time bias to the score, and advances in care and case-mix changes have made that original set of coefficients based on outcomes from the 1990s inaccurate, and revised models have been generated (Desai and Gross, 2019).
Furthermore, the models are trained on the internet, which is an essentially unmoderated platform containing a lot of misinformation and bias. This can produce inaccurate, discriminatory or offensive responses. If the AI learns from the questions a user poses it, there is a risk of a positive feedback loop where the responses become increasingly biased towards the attitudes and beliefs of the user. In addition to storage of the questions being asked, there is a risk of hacking which could be used to manipulate responses or spread malicious content.
Conclusions
The AI methods discussed above can significantly increase your marketing output, help you get past writer's block and develop your insight into a given topic, and I encourage you to experiment with it.
It is an exciting area with the potential to empower you to be more productive, but it still needs your skilful direction and monitoring. However, it should not be used as a replacement for clinical studies. You – not the AI are always responsible for ensuring the validity of your work, so always validate the output before publishing it.