Artificial Intelligence has become the major topics of discussion in most organisations and businesses since the release of OpenAi’s ChatGPT Large Language Model suite of tools and capabilities. All of a sudden it is possible to ask general questions and get a written detailed response, create art and pictures, and get your code written. There is a race now between Microsoft (betting on OpenAi and Bing) and Google with its Bard Tool. As with many new technology debates the hyperbole is substantial with talk of 30 – 50% of jobs being at risk. our world is to be shaken to its roots.
Reality is slightly different.
Background
- Ai has a long history – product pricing tools, predictive text on phones, route planning software has used forms of Ai for many years. Organizations should already have a strategy that includes developing technology solutions and AIs should already form a part of that.
- Large Language Models take information requests in the form of Prompts (creating a new role called Prompt Engineers), processes that information internally, and then produces a written, visual or spoken output. Large Language Models despite their apparent speed and capabilities are probabilistic models calculating the probability of the next word in a response. Hence sometimes models can produce responses that make little sense or are hallucinations; they use probability to make stuff up.
- There are three stages in any model that uses probabilities such as Large Language Mocdels
- entering data into the system through prompts. Ensuring the prompts are asking the right questions is critical. Small changes in prompts can produce significantly different output.
- The actual production of the response within the machine. using a vast library of data the tool calculates the probability of whatever the next word should be. If the library is not compete the moel will make up (or Hallucinate) an answer and
- a review of the output. Does it make sense and does it support the prompt entered.
- This is exactly the same process regulators go through when considering the use of pricing tools in say an insurance context.
Risks and Issues for Boards and Executives
1) Any development of Large Language Models or AI needs to be undertaken in the context of the strategy of the business. Models can help improve efficiency, customer service, quality, but as with most IT projects can rarely do all three. It is important that any development aligns with the existing strategy and direction of the business and can meet actual and real identified business challenges.
2) Particularly in regulated industries Boards and Executives need to be able to explain how the model(s) works – how data gets into the machine, quality control around the data, for what purpose is the data being used. Regulators then look for an understanding of how the machine actually works internally which is currently a significant challenge as their is little clear understanding outside of research papers on what actually happens. And finally regulators need to understand how the output is checked to ensure quality, relevance and a lack of bias. And then what is the output actually used for. \
While this is an issue for regulated industries the rise of ChatGPT and the ongoing interest in AI in the press is making their use more transparent to consumers. It would be wise to adopt a mechanism that allows businesses to be able to describe the items in 2 above to whoever asks.
3) Large Language Models and AI in general work by reviewing and ‘learning’ from training data, The data in some of the Large Models is said to be equivalent to the published data on the internet. this produces two distinct but important challenges
A) The data on the internet trends towards an average meaning that any responses from a Large Model is likely to be an average of the data that the system has learned from. As a business do you want to be providing average responses.
B) The data on the internet is biased – it is a reflection of the underlying society and as such if you want to produce unbiased responses you will need a mechanism to train or use data that isn’t biased.
4) There is one further significant risk that needs to be considered, Large Language models can be bought and used in house using a companies own data. This is potentially the most significant risk as the companies own data will be biased and there is potential for an echo chamber effect leading companies to believe entirely in their own data and as such reduce external sourcing of insight that will challenge in house produced results.
Further Reading
Board Notes New Zealand - Risk, AI, Sustainability 