Article • • 5 minute read

AI Driver's License

When dealing with Artificial Intelligence (AI) systems, it is important to have a basic understanding of how they work to use them securely, responsibly, and efficiently. This AI Driver's License is intended to give you this basic knowledge so that you can start using AI tools as part of your daily work as well as a launch point for you to continue learning. 

In this article

  1.  Security 
  2. Responsible AI 
  3.  Efficiency 

There are three main types of AI, however we have only achieved one type of AI: narrow AI. Narrow AI are very adept at performing extremely specific tasks. Of these, the most common implementation you will encounter to help in your workday is Large Language Models (LLMs) such as BingChat® or ChatGPT®, which are both built on OpenAI's GPT models, a generative AI. Generative AI is a class of AI that is used to generate text based on the input or prompt you provide. You interact with these models much like you would have a conversation with a coworker, using natural language, specifically English. 

  Security 

 When interacting with Generative AI models, it is important to know how your data is being transmitted, stored, and used. Just like any online tool you may use, it is important to make sure we are not exposing internal, sensitive, confidential, or customer intellectual property (IP) outside of where they belong. Be aware of the data sensitivity of the information you are working on and be sure that sensitive data is not exposed. 

 When you are interacting with LLMs, there is a computing cost associated with generating your answers, where smaller models take up less compute space than larger models. The model for GPT 4 is orders of magnitude larger than the model for GPT 3.5, so the compute cost is also increased. When you are not paying for this compute cost, be extra mindful of how your data is being used. Even if you buy a personal license where the company claims to not use your data, you must still be careful of what you put into the system, and never put anything for work in unapproved tools. 

 Like any other asset we may use for work, ensure you are compliant with policies around what are acceptable tools to use. These should be followed when interacting with AI tools and we should only be interacting with the tools that are on approved lists of technologies. 

 Responsible AI 

 Whether we are providing AI tools to our customers or are using existing tooling for ourselves, we must ensure that the AI solutions are built responsibly. What this means is that we need to consider how these tools will be used, how they will positively affect people, and if there are any negative outcomes to their use. In addition to going through established risk assessments, we should also include reviewing how these risks expand to involve AI when this technology is included. 

 Artificial Intelligence is not something that magically exists in the world, it is something that must be made by humans, and in being human-caused, it will have these human-caused biases. Even if it is not consciously programmed into the AI, these biases will be present in the data we use to train them. Even if the data is curated and more programming is added to avoid these biases, how AI learns to make its associations can very effectively notice the biases that we try not to acknowledge. This is something that has been seen time and time again when developing AI systems, from systems that immediately reject resumes from women when it has been trained on historical tech hiring trends despite their best efforts, to learning too quickly to include racist verbiage in responses when trained from live interactions in a culture where being kind on the internet is not the social norm, to even over-diversifying images and including inaccurate diversity in historical image generation where there was none in an effort to combat bias. 

  Efficiency 

 There is a learning curve involved in effectively and efficiently using AI tools, and it is important to use the right tool for the job. When using Generative AI, the terminology for interacting with these models is often called Prompt Engineering, which is just a fancy name for wording your inputs in a way that the AI can understand. Each model will have a communication style that will work best, and models are constantly becoming easier to work with. While there are general guidelines you can use for this, sometimes you will just need to use the tool to figure out how best to get your answers. 

  

Here are some general tips and tricks for interacting with Generative AI: 

  • Be specific when you are asking your questions and specify what information you would like.
  • Give context about who you are and how the response should be phrased.
  • Iterate on the responses by continuing the conversation.
  • Recognize that AI does not understand you, it can only approximate the world, and Generative AI, to put it super simply, is fancy auto-complete.
  • Own your output since AI is just a tool to help you, you are ultimately responsible for ensuring the accuracy of the information coming out of the AI.