Things to Consider When Adopting AI

There is a popular online observation that Ultron, the villainous AI from the Marvel Universe, searched the internet for less than a minute and decided that humanity had to die. Skynet, the villainous AI from the Terminator series, came to the same conclusion, but we don’t know the timeline it took to formulate that decision. The Matrix explores a similar AI response, but we now know the truth is much less terror-inducing and, frankly, a bit sad. 

AI is far more likely to create images of a six-fingers Jesus-shark hybrid rescuing a weirdly shaped flight attendant from a plane wreckage than to start looking for nuclear launch codes. And yes, that image actually exists as a shining example of AI not understanding that things that generate buzz separately should sometimes not be combined. 

By now, you have been inundated with articles and stories littered with the most recent buzzword: AI or Artificial Intelligence. Many companies are working hard to incorporate some version of AI in their products to make them sound more appealing. Some are not actually AI, while some are, but they are being implemented in ways that make them virtually unusable. Here are a few things that we have found that should be considered when looking to adopt an AI product.

1. Enhance, not replace. AI is in its infancy, and even if you follow the suggestions in this article, your AI is still going to make mistakes. You are going to need someone who understands the output to make sure that the results are correct and work properly. The Salesforce team at Bosma has used AI to help generate code. Each time we needed our developer to look it over and correct some of the mistakes that the system made. One such mistake would have halted our production environment immediately upon release if it had been deployed as is. So, using AI to assist is great, but make sure you are using it to enhance a person's or group's output and not to replace it. That could be a very costly mistake and will garner a negative view of your company or brand.
2. Who taught the AI? Any admin worth their salt will always remind you that you can’t expect pristine output if you are feeding your system garbage. This is true for reporting and is especially true for what was used to train your AI. A good example of this is a recent set of articles where users were asking questions and getting some amusing and obviously false information, such as adding non-toxic glue to your food to make it more tactile. Now, I’ve never eaten glue, but I can only assume that tastes about as appetizing as it sounds. Another search yielded results telling the user that it is perfectly okay to eat rocks and that you should at least eat 1 rock a day. Another said it was okay to have bugs living in body parts because the bug's name had a crude name for the body part. A deeper dive into why the system returned this information is because when the system was taught to scour the internet for information, it wasn’t told to ignore satirical sites or sites where the suggestions are dubious at best. Reddit may be a go-to site for game hints, but even that is filled with misinformation and random nonsense. 
3. Whose data is the AI using? Much like the previous thing to consider, there is the issue of copyright infringement, especially when generating visual content. We already know that AI struggles with fingers, but it also struggles with knowing what information it can use for its output and what it can’t. There are cases out there right now where content creators are suing AI generators that are using their data to help create things without giving credit or compensation to the original creator. Knowing where the data is coming from and if the original creators have agreed to their content being included will need to be something worth considering. On the opposite side of this is, is the AI getting enough information to accurately assist? Is it using just your data to determine some outcome? How detailed is the data you are feeding it?
4. Make sure you understand the outcome. There was a medical group that decided to see if they could teach AI how to spot cancerous skin growths. So, they fed the AI a bunch of pictures with the ones that were cancerous and the ones that were not. Once done, they tested it and found out that the AI was actually really good at determining which ones were cancerous and which ones weren’t. Suspiciously so. So, they took a deeper dive and found out that the AI had reached a conclusion that had not been accounted for. You see, when people take pictures of cancerous growths, they typically have a small L-shaped ruler in the picture to let the person viewing the growth how big it was. Typically, people don’t add rulers to bug bites and pimples. So, the system concluded that the best way to determine if a picture was of a cancerous growth was to look for a ruler. If there was a ruler, it would be cancerous. If there wasn’t, it wasn’t. So, when they ran test data, their data has those rulers. Once they removed the rules, the AI was horrible at determining cancer versus non-cancer. So, it is always good to have a clear understanding of how the AI is coming to these conclusions. Additionally, as we mentioned with the Jesus-shark hybrid, AI can understand that something is popular but not that they shouldn’t be combined to make them even more popular. 

While our science fiction stories are filled with tales of Artificial Intelligence becoming fully sentient and power-mad, the truth is that right now, like any infant learning their way, they are prone to mistakes and are sometimes at the mercy of liars and confused individuals. This should not keep you from using the powerful tool that is AI, but for the foreseeable future, we have less to worry about Skynet or Ultron and more to worry about it telling us to eat glue and rocks while giving us nightmare-fuel pictures with eight fingers and a confused understanding of human anatomy.  

If you are interested in learning how Bosma Enterprises IT Consulting can assist you with learning about AI, contact us at bosmaitconsulting@bosma.org.