There’s no escaping artificial intelligence (AI), with many new solutions on the market for advisers.
But among the myriad messages, it’s rarely clear what type of AI is being promoted and whether it is suited to your needs.
There are two main types of AI today and it’s important to distinguish which one is being offered to you.
Before you commit, you must really consider what you want the tech to do for you.
Predictive AI
Often considered ‘traditional’ AI, this class of machine learning is trained to recognise patterns in data, text or speech.
Humans (data-annotation specialists) manually mark up data records with what they mean, and data scientists feed this training data into a ‘model’ – a statistical engine they have designed, usually based on some form of neural network technology.
Generative AI is only useful when a human is iteratively interacting and checking every response
This model is then used on new, previously unseen, data to predict what it should be labelled as.
Predictive AI can be used to distinguish whether a picture is a cat or a dog, or if a sentence concerns a client’s financial objectives, their emergency funds or their attitude to risk.
It can be made very accurate and, more importantly, it can be tested as to its accuracy. Its results are repeatable and any biases in its training can be removed.
In short, predictive AI is very good if the time is taken to train it well.
Generative AI
This is the ‘new’ AI – like ChatGPT, or at least the underlying GPT models from OpenAI, Facebook, Mistral, Baidu, Anthropic, Google and others.
You’ll often see these models described as large language models – or LLMs – although that’s a misused coinage, as there are many LLMs in the predictive AI family.
When buying any AI product, be explicit about what you want it for and ask each vendor to explain why their method is best suited to that
The clue to the purpose of generative AI is in the name. This class of AI is designed to generate new data, such as pictures, prose or speech.
It has no capability to understand or analyse your data, merely creating new content based on instructions or prompts. For example, if given an instruction like ‘draw me a parrot’ or ‘write a poem about the sea’ it will do so.
In simple terms, it does this by generating a likely start word from a limited random selection, picking a good next word that is statistically likely, then a third word and so on.
It doesn’t know what it is saying. It simply churns out a sequence of words statistically related to the prompt provided, based on what it has seen before – the data the GPT vendors have trained it on, mostly large portions of the internet.
So, generative AI is good at creating content, whereas predictive AI is good at identifying content.
The confusion
Many vendors are promoting generative AI that appears to understand or identify content.
In reality, they are first performing a simple search into your content to try and find relevant information, then using the first few search results within their GPT query prompt in order to formulate an answer.
It’s rarely clear what type of AI is being promoted and whether it is suited to your needs
Vendors use generative AI as a shortcut to painstakingly labelling data and training a predictive AI model suited to your needs.
The problem with this is that the search request itself is automatically generated, then only a handful of findings are used in generating an answer. The answer is then based on the standard GPT method of statistically generating one word at a time.
That’s plenty of opportunities for errors to creep in, inconsistencies to arise and even hallucinations to appear. In short, you can’t rely on the outcomes (not verified by a human) if you want to use this information for any kind of decision making.
Generative AI is only useful when a human is iteratively interacting and checking every response, such as in a chat or search application.
If you need reliable AI that will consistently identify relevant content or what it means, there is no shortcut to using predictive AI, especially if you want to limit the need for humans to check every answer.
When buying any AI product, be explicit about what you want to use it for and ensure you ask each vendor to explain why their chosen method is best suited to that and, most importantly, how they can guarantee its accuracy and data reliability.
Joe Norburn is chief executive at TCC Group and Recordsure
