At the end of 2022, something called ChatGPT burst into our lives. It wasn't new, we had been flirting with machine learning and some chatbot models for years but they had never really caught on beyond the IT or geeky areas (like me).

But Chat GPT was the most successful penetration of a new system to date. It reached the figure of 1 million users in 5 days, when it had taken years to reach this milestone.

Many people have wondered why, and it will certainly be the subject of studies and theses in the near future. The idea that it was essential to prove it spread like wildfire. And we all wanted to ask our questions to this new model of Large Learning (or LLM). And then other tools, integration plug-ins, extensions, updates, ... have been appearing. A whole new ecosystem of tools within our reach that have come to improve our day-to-day work. 

Although many people did 2-3 tests, they had a discreet success (or a complete failure) and have not touched it again. 

I am not going to dwell on the importance of creating good prompts (the questions we ask the AI), or of recovering high school notes to place the ideas in the correct syntactic order. We must never forget that we are dealing with language models, so it is very important that we master the language to convey the right idea to the model. More than ever it is important to ask the right questions.

When I am in my role as a toxicologist, my day-to-day work is to search for information and data, read technical and scientific articles, calculate and draw conclusions. And it's exciting, but it's very time consuming. So the idea of having an AI assistant is a very appealing idea. So I decided to tell you about my experience so far, in case someone could use it in their field.

The first thing I have to say is that since they came out, the models have evolved a lot. Their answers are getting better and better. So if it's been more than 6 months since your last query, I recommend that you go back in. 

The second thing is that they are models that are fed by the information contained on the internet, so they ‘drag’ the problems that the network has. And one of them, and for me the most important one, is disorder.

Let me give you an example. If I need to consult a data in official databases, it is very difficult for me to transmit it to the model. The model interprets ‘official’ as ‘.gov’ domains. But for example the ECHA database is not .gov but .eu, so it excludes it from the search. So either I list one by one the databases or pages where I want it to search, which generates a rather complex prompt that sometimes doesn't work, or I ask it to search the pages one by one, which doesn't save much more time than searching myself.

This is one of the points I am refining, so that I can be 100% confident in the searches I do.

Another of the problems that ‘drags’ from the internet is a certain degree of confusion on complex issues. In other words, the model does not know how to distinguish between facts and opinions. And it is difficult to train it, especially on controversial issues.

I will give you another example. When looking for information on the sensitising potential of substances, one of the pearls that came back to me was that people with a high tendency to skin sensitisation should prefer essential oils to other substances. Since essential oils are the products with the highest allergenic potential, I was very surprised. Obviously this is due to the large amount of content on the web that directly or indirectly relates natural to safe and the model has no tools to discriminate them. So when I search for information on specific substances, the answers obtained are quite reliable, as they are based on data contained in databases, but when the search is about chemical families or groups of substances, the answers are much more contaminated by currents of opinion as it also draws on the rest of the content on the internet and at the moment it is not able (or I do not know how) to make weight of evidence evaluations.

A few days ago I read that some people filter searches as ‘before 2015’ to avoid content heavily contaminated by SEO strategies. But this strategy is not valid for scientific content, so discarded.

I haven't solved this point, so I'm still trying.

But obviously AI is not only useful for asking you questions. There is already an app, for example, that converts scientific papers into audio, and also separates them into chapters. So you can listen and take notes, as if it were a lecture, instead of reading and writing. This is if you don't want to use one of the countless scientific text summarisers available. In addition, we can incorporate AI to the extensions to manage bibliographic searches and help us recommending sources, detecting plagiarism, and so on.

And as I said at the beginning, this all started just over a year ago. I cannot imagine what we will be able to do in a few more years. However, it is important not to be blinded by the lights and to be aware of the shadows. 

I summarise it as, on the one hand, the lack of consistency in the identification of sources, that domain extensions are not used properly so that the models cannot discriminate the origin of the information.

On the other hand, and derived from the previous one, when we talk about controversial issues such as the safety of substances, the model feels lost and values the presence of an idea by volume. 

So, for the moment in toxicology, AI is a great help for someone who already has knowledge and can detect these biases.

But what will happen when we have become so accustomed to using these tools that we no longer know how to do things without them? It's no laughing matter, nowadays we are so used to GPS that few people have maps or even know how to use them. What would happen if the GPS information was unreliable? What if it decided to indicate that Finland doesn't exist? Or that there is land between Galicia and Normandy? I'm not making this up, there are websites on the internet that say the same thing. If the GPS instead of being guided by data integrates AI, it could happen.

So, for everyone's sake, we should debug these little things together before we can't live without it and we have virtual doctors recommending that if we have allergies we use lots of essential oils.