There is a lot of hype in the news about the step change in AI technology with major companies talking about how AI will be a game changer for their business. The challenges facing many organizations today are: how will we incorporate AI into our products to drive growth and productivity, and how long will it take?
What is Generative AI is and what are its limitations?
The biggest player in the news today is OpenAI’s ChatGPT. ChatGPT is an advanced generative AI bot that has been trained on a lot of textual content (billions of words from websites, articles, and books), take textual input from a user, and generate a textual response. The demonstrated capability is very impressive and will certainly improve significantly in the near future, but what are the limitations?
We need to remember that it is not “intelligent” – it is a very good natural language processor. In the UK, during the 1900’s, they used to screen children as they moved from primary to secondary schools with an examination called the “11+” that was supposed to identify the top 20% of students based on IQ. When I took the test it was broken down into verbal, mathematical and non-verbal reasoning. I’m sure that ChatGPT would do very well on the verbal reasoning part answering questions like “apple is to fruit as bat is to ?”. We have already seen videos of ChatGPT answering questions on English Literature. Other AI products already do legal case preparation, generate code, etc. In all these cases, the AI is analyzing natural language text and generating a response. For a factual question the response is likely to be good.
Using generative AI to help formulate an opinion or moral perspective is more problematic and the AI industry has been working hard at attempting to address bias in generative responses. Consider a heavily debated subject, such as abortion. Many people have strong views on the subject that are either pro-x or pro-y. You can add the choice/life labels yourself. To take an extreme, let’s say the AI has only been trained on pro-x text. Anybody asking an abortion question will get a pro-x response. For a pro-x user this is an acceptable response. If you were pro-y, you wouldn’t be happy! Let’s say we attempt to eliminate the bias and train the AI on as many pro-x articles as pro-y. Again, a pro-x user may still see a pro-x bias response just because of how they phrased the question. The AI doesn’t really understand the question and it doesn’t know what life is in the same way a human does. It doesn’t “understand” the ethical or moral arguments. The conclusions are highly subjective and, in the case of ChatGPT, the AI has been programmed to deliberately refrain from generating a conclusion:
Another limitation is the potential for the user’s prompts to steer the AI to generate a false conclusion. ChatGPT has been programmed to attempt to produce a response to your questions that are appealing and it will avoid disputing you about the authenticity of its generated response. In the example below, ChatGPT was “convinced” that 15+20=45! We often assume that since we are talking to a computer, that it must always provide a mathematically accurate answer. But that’s not how generative AI works. Generative AI is formulating a molded answer based on multiple inputs and conditions.
Another limitation that is perhaps the most dangerous is the condition described as “AI hallucination”. AI hallucination is when the AI produces a response that is entirely false, but “appears” correct. ChatGPT summarizes the definition well:
This can sometimes produce bizarre answers that are factually wrong, but could be easily interpreted as true. For example, in Reid Hoffman’s “Fireside Chatbots” podcast, he asked ChatGPT who has been guest speakers on his podcast. ChatGPT listed some people that have never been on his podcast (but Reid admitted would be excellent people to invite!). In another example, Pete Cashmore (founder or Mashable) asked ChatGPT to summarize his biography. The response was a mashup between him and someone else with his name that also happened to have worked at The Guardian (ChatGPT summarized that Pete was a battle rapper!). Many others have had similar experiences, including scientists that are subject matter experts in their field finding responses that are factually wrong, but could have been easily mistaken to be correct by someone not as deeply familiar with the subject. As a side note, for anyone using ChatGPT for research it is very important to ask ChatGPT for citations and to validate the information is accurate.
Having recognized these limitations (as they are today), what are the practical use cases for AI tools, such as OpenAI ChatGPT? and why are we going to include it in the next release (due Q2/23) of our trellispark UX and CRUD virtualization platform? We see ChatGPT and other generative AI tech as a new tool that will improve the productivity of our users by automating some of the tasks they perform. And because trellispark virtualizes the UX and CRUD functionality, adding this capability to be available across any user experience is easy and with little effort!
OpenAI ChatGPT is already very good at generating textual content for use cases like drafting emails, letters, or other documentation. For this reason we will initially be adding it as an optional service for all multi-line string, and HTML string, fields in our dynamic page builder components for web, desktop and native devices. As the technology develops, and other use cases emerge, we will incorporate AI into more complex workflows.