Giles Thurston
In 2016 Microsoft launched their Tay chatbot on Twitter. This lasted less than a day before being taken down, after users managed to get it to respond with racist and homophobic comments. For the next few years, chatbots were relegated to being the first line of defence for online customer service portals, with a limited set of bland and typically off target responses. I am sure I am not alone in finding these a hinderance rather than help when looking for a solution to my problem.
Roll forward to Nov 15, 2022 and Meta launched their own Galactica platform focussing on scientific text but this too was taken down within days, as it was unable to discern fact from fiction. In eight years nothing seemed to have changed, proof if any was needed that artificial intelligence, natural language processing and conversational AI was hard.
Avatar originaly used by the Tay chatbot.
Homepage of Meta AI’s Galactica website while the demo was still active.
Just a few weeks later however, at the end of November 2022, San Francisco based OpenAI launched ChatGPT (Chat Generative Pre-trained Transformer) and within two months had gained over 100 million users worldwide, making it the fastest growing internet service in history. Had someone finally cracked it and, more surprisingly, was that someone other than the usual behemoths in the technology world?
Articles, podcasts and videos have been flying around the Internet proclaiming the success and failures of the tool and its underlying GPT-3 model. Unlike its predecessors, OpenAI looked to have taken a significant step forward with ChatGPT and we were all taking notice.
And not just the public, the behemoths previously mentioned have been plugging away at this for years keen to have a piece of the action. Within weeks, Microsoft made a ten billion dollar investment in OpenAI and is already looking to integrate ChatGPT into services such as Office and its Bing search engine.
To help keep our feet firmly on the ground, in early February 2023, Alphabet (Google’s parent company) announced Bard, powered by LaMDA (language model for dialogue applications), as their take on ChatGPT.
Unlike ChatGPT this had a very bumpy start to public life, with it giving a wrong answer in a promotional video and reports from the limited users able to try Bard, that its responses have been, well, less than perfect. Alphabet’s stock took a massive hit as a result, loosing more than $100 billion in its value in a single day, as traders feared that Microsoft (with OpenAI) had stolen a march on Google.
Truth is that ChatGPT didn’t just come out of no where. Like all good technology solutions, it was refined iteration by iteration, until the version we see today was launched. It is not perfect but even with its deficiencies, many are asking whether the dawn of conversational AI has finally arrived?
So why is Big Tech sinking so much money into these technologies and what does this really mean for Pharma and drug safety in particular?
Companies like Google, Microsoft and Facebook are always looking for the next big thing, the way to differentiate themselves and make their offerings more compelling. They already use AI in their technologies, such as trying to interpret the search queries we type, but the chance to provide that true “Star Trek experience” of us being able to converse with our computers and rely on their answers is tantalising.
Voice assistants, such as Alexa and Siri are increasingly becoming part of our daily lives but are mainly limited to simple tasks or instructions. What is the weather? Play the latest song from X or what is coming up next in my calendar?
Where our commands move beyond these simple instructions or requests to cloud services, all vendors are keen to be able to provide more. You will have heard the response “here is something I found on the web …” as our friendly assistant tries to find the answers from another source, with some or little success.
Big tech is betting that large language models, such as those underpinning ChatGPT, Bard and Galactica are the answer to this problem. These enormous models are fed vast quantities of text from across the Internet, to teach them how to take a question and formulate a simple, coherent response and follow-up answers if required.
They also believe this will be the next big thing in search. As we increasingly use smaller and smaller devices to access the Internet, such as our smartphones and watches, the limited screen real estate makes scrolling through long lists of information less appealing. Concise and accurate responses, by text or voice has to be the answer, right?
The problem is that the language models that underpin these technologies don’t actually understand what it is that we are asking and the meaning of their response. They are using their models to identify the most suitable word or phrase to follow the current one and through this build up a coherent sentence or conversation.
It’s a very simplistic way of looking at it but what can look really impressive, in many cases is actually just a probability problem, with the entire Internet at its disposal to try and solve it. They are mindless mimics, chewing through all the relevant content they can find and then repeating back the bits or combinations of bits that seem to make the most sense.
And there in lies the issue. There is no fact checking, no validation of the responses and no reference to where they are getting the information from.
When we use search engines, we understand that we are interacting with a computer and we have to then use our own thought processes to sift through the results and find the right answer. Conversational AI changes this dynamic, with the illusion that it understands what we are asking. We subconsciously take its responses in that context and somehow find them more believable, even when what they are saying maybe untrue.
And the truth is that in many cases there is not a simple answer. As Potthast, Stein and Hagen highlighted in their 2020 paper, the dilemma of the direct answer, regardless of how good the technology is, the response to most questions is “it depends”.
But how can we tell if what we are being told is the truth? While it is easy to identify errors when there are claims to the existence of Space Bears, where it relates to subjects the majority of us are less familiar, such as our healthcare or the pharmaceutical products we use, this can be more challenging.
Take the example below, where I asked ChatGPT some questions related to my generic hay fever medication.
As you can see, the responses were legible and, to my mind at least, sounded perfectly reasonable. A quick check on the product label confirmed what it had said, with pregnancy listed as a potential risk and alcohol not.
However it omitted to highlight any other risks or areas of consideration. Yes I asked specifically about alcohol and pregnancy, reasonable questions to ask but wouldn’t a medical professional also have highlighted other risks on the label, such as the need to be aware of any other medications I may be taking alongside it or dizzy spells before driving machinery?
Granted ChatGPT repeatedly highlights the need to speak to a medical professional but taken how believable these type of technologies can be, would we bother if we receive the answer we want to hear or believe?
This is also quite a simple drug with minimal risks associated with it, imagine a more complex therapy and associated risk management plan.
Technologies such as ChatGPT are interesting and it is really exciting to see the developments in this space. However we do need to approach these with a sense of caution and the plans of Google and Microsoft to integrate these into their search engines is a concern. In a world where disinformation and fake news is increasingly problematic, technologies such as conversation AI, as they stand today, look like they could make matters worse and not better.
For Pharma, where there is a regulatory commitment to engage with patients and healthcare professionals about their products and the safety guidance around them, you could argue they are better to stay away from the Internet. I would argue the exact opposite.
Now more than ever, we need trusted sources of truth online and as these technologies develop, narrowing their focus on these sources, with clear attribution when relevant, has to be the way to make them more useful and trustworthy Without that information online, in both a human and machine readable form, services like ChatGPT will be left to gather their information from who knows what sources.
The middle way has to be the best approach in the coming months and years. Ignoring them and steering clear of the Internet is not the answer, neither is wholeheartedly embracing them. As more and more investment is made, we will hopefully see these services improve and find use, particularly where there is a specific focus and Pharma will have a role to play here and content to contribute to this process.
Paul Simms says in his 2023 Predictions, “… ChatGPT (will) become recognised for misinformation and be kept away from healthcare due to its flaws.”. Paul may well be right but the secret Trekkie in me does hope he is proved wrong in the longer term or, at the very least, they eventually find a useful place somewhere in our lives.