Artificial Intelligence – The great sophist

Sophistry is ChatGPT’s greatest skill.

• ChatGPT’s great skill is in its ability to make nonsense and falsehoods sound convincing which is really the only threat that it represents and does not represent a breakthrough moment for AI.
• ChatGPT is an extension of the GPT-3 language model that has been sanitised and made available for public consumption which has again ignited the debate about the pros and cons of AI.
• It exists as a web instance where registered users (over 1m) can input any question or comment that they can conceive of and receive a response from the system.
• It can write articles, make up poems, and comment on views but it is not surprising that when it gets into trouble, it dodges the question.
• It does this by saying that it does not have enough information to answer a request, or it simply refuses to answer because it is a neutral agent.
• It is here where one immediately finds the first shortcoming of this system.
• With 175bn parameters, this system has access to more information than any human could ever assimilate over the period of his or her entire life and yet humans can answer the question and ChatGPT cannot.
• Furthermore, it is fairly easy to demonstrate bias in the ChatGPT which has crept in as a result of the views of the software engineers who worked on the algorithm.
• For example, although ChatGPT claims to be politically neutral, it is not very difficult to get the system to voice views from both ends of the political spectrum on certain topics.
• This, combined with ChatGPT’s tendency to produce nonsense, demonstrates that at its heart, ChatGPT is a pattern recognition system with no causal understanding of what it is saying.
• This is why OpenAI based ChatGPT on a snapshot of the world sometime in 2021 and as such it does not know that Russia invaded Ukraine or that Queen Elizabeth II has passed away.
• Freezing the data source in time makes it both finite and stable and gives OpenAI the ability to clean the data to give ChatGPT the highest probability of success.
• However, even with both of these advantages, the system is quite capable of spewing meaningless rubbish or commentary that is plain wrong.
• It also often provides completely different answers to the same fact-based question when the exact same question is asked for a second time.
• The danger of ChatGPT is that it is able to dress up nonsense so effectively that it sounds plausible to anyone who has no knowledge of the subject under discussion.
• This is where the problems lie as these systems have the potential to substantially increase the spread of false and misleading information that is picked up and mistaken for fact.
• I am less concerned about the use (or misuse) of this system by students to write their essays for them.
• This is because the examiner will be an expert in the field and should be able to spot ChatGPT-generated content a mile off.
• However, where it will be of some use will be in the generation of generic, unopinionated text such as instruction manuals.
• For example, the request “tell me how to operate a bicycle” results in a pretty good list of instructions of what to do.
• Hence, I don’t think that this newsletter has very much to fear from ChatGPT, but people who write instruction manuals or draft standard documents should probably start thinking about a career change.
• ChatGPT and its descendants will have some use, but they are very far from being intuitive or creative meaning that as it stands today, ChatGPT is an entertaining distraction.
• Hence, there is still no sign of the machines taking over or any real spark of real intelligence leaving AI still struggling to deal with the shortcomings that I have been highlighting for several years.