Treading Carefully: Best Practices When Adopting AI

The prevalent feature of human innovation is to allow us to mess things up at a larger scale faster than ever before. We can manufacture things with flaws faster. We can drive cars faster. We can lay waste to entire natural ecosystems within minutes. We can crash the financial stability of the world with a tweet. We can eradicate every living thing on the face of the planet with the push of a few buttons.

This understanding is critical to coming to grips with the potential for disaster that the fervor behind the use of newer AI projects has. We are still (hopefully) a generation away from achieving true consciousness with AI. However, the foundational elements for that consciousness are being developed now. The main cause for concern is that these elements are currently *rife* with bias and error. For example, facial recognition can’t distinguish between people of color. Generative imaging AI creates inappropriate outputs when no inputs of such were part of the data set. Large language models (LLMs) will make up answers, called “hallucinations,” or produce results that contain known errors.

Aside from flaws in the technology itself, we are not always using the technology in a safe way. Users are entering sensitive data either in their queries or into the data set. Users are already relying on AI to produce answers that they don’t have the experience to vet. Some parties are creating deep fakes to assist them in disinformation campaigns. Others are using deep fakes to exploit people’s likenesses without their knowledge or consent. AND, because of the developmental momentum of these technologies, organizations are prioritizing the increasing adoption and monetization of AI rather than solving these issues first.

We don’t have a grasp on the long-term effects of use of AI on society or the world around us. We haven’t standardized any sort of governance to provide guardrails around the creation and use of these technologies. We don’t have transparency into how some of these AI models make decisions. We haven’t done significant research into the ecological effects of increased storage and compute needs of growing and more computationally complex data sets. We don’t have regulations around the inclusion of protected property in AI results. We haven’t resolved issues of liability for the misuse of AI or hallucinated results. We are barreling into a blind corner when it comes to the long-term effects and outcomes of our current efforts with AI. If we do not apply the brakes soon enough, we may be seeing the bottom of a chasm before building those guardrails!

With all these valid concerns in mind, AI is being adopted in more and more areas and products, and that’s unlikely to change. There are several large tech companies with competing projects and products, each vying to be the AI of choice. Try as we might, we cannot resist the current of dollars flowing into AI. However, we can ride that current through thoughtful selection and use of AI in our practices and products. To assist our navigation, here are some considerations for your practice and use of AI that should help you navigate these waters safely:

• Never input personal identification information (PII), credentials, secrets or other sensitive information into your models or chatbots: This should be self-evident, but there have already been incidents around this. Treat information going into your models and chatbots as if it were going to be available to the public.
• Be cautious when using AI photo editors: Be aware of the potential risks, such as the spread of misinformation, copyright and intellectual property concerns, and data privacy.
• Always verify the data you get from an AI tool before using it: Make sure to double-check the data you receive from an AI tool to ensure its accuracy. Accepting an answer as correct without verifying the results could have lasting impacts on your business.
• Understand the AI applications landscape: Understand the different applications of AI and choose the one that best suits your needs. There is no shortage of vendors in the AI landscape and decision paralysis is real. Choosing the right one for your use case is a decision that should not be taken lightly.
• Data preparation and management: Ensure that the data you use is accurate, relevant and up to date. Be cautious of ingesting your own mistakes into your models and data poisoning.
• Choose the right algorithms and models that are appropriate for your specific use case. This will keep you from trying to boil the ocean when it comes to modeling and presenting your data.
• AI development services: Use development services that are reliable and have a good track record.
• Training and evaluation: Train and evaluate your AI models regularly to ensure they are up to date and accurate. Are they producing the results you want consistently? How often are errors coming up and what is the nature of these errors?
• Deployment and scalability: Ensure that your AI models can be deployed and scaled effectively and sustainably. Gathering metrics on sustainability and the environmental impact of your modeling and AI usage will help you make better decisions around their use.
• Ethical and privacy considerations: Ensure that your AI models are ethical and respect user privacy. While this is nebulous, it helps if your organization establishes principles for AI ethics to guide your practices.

AI is here to stay, and it’s up to us to make sure it is incorporated into our world in a safe, ethical and sustainable way.