Artificial intelligence (AI) has become a popular IT service offering. In this video, you’ll learn about AI integrations, appropriate and inappropriate uses of AI, challenges with AI bias and hallucinations, and comparisons between public and private AI implementations.
AI is an abbreviation for artificial intelligence. And the goal of AI is to meet or exceed human intelligence in the form of technology. This means that our computers would be able to learn, infer, and reason based on input that we provide. This is, obviously, not a new idea. We’ve been working on concepts of AI for a number of decades. And of course, you see AI mentioned in numerous science fiction stories.
But now, of course, AI has become much more prevalent. We have new tools that allow us to create new content such as text, audio, and video, by simply asking a computer to create it. And very quickly, we have seen AI become integrated into most applications that we use every day. For example, if you perform a search on a popular search engine, you may find not only your search results, but also a summary of those results that was created using artificial intelligence. The goal is to take information that normally you would need to find across multiple sites and present all of that information to you one, single AI summary.
We’ve also seen AI in our email clients and on our mobile devices. Instead of having to read through a long email, you can read an AI summary that explains the information that was contained in that message. And as you’ve probably seen in graphics editors, you can use generative AI to fill in or remove information from an image by simply highlighting it on the screen and asking the AI to remove it. We’ve also seen artificial intelligence engines that allow you to input a description of what you’d like an image to be, and the AI will create that image on the screen.
We’ve seen a number of different uses of AI that we could put into the category of appropriate AI use. For example, there may be extremely large data repositories that we would like AI to be able to sift through and provide us with a summary of that information. For example, there’s AI that can look through terabytes and terabytes of log files in an organization and find potential security issues that may be hidden in all of these different logs. This would certainly be time consuming and difficult for a human to do, which makes it perfect for AI. We’ve also started integrating AI into our scripting and our automation, so that AI can make decisions on what to do when a certain situation occurs. Being able to do this without human intervention means that we can spend our time doing something else.
We’re also seeing AI integrated more into healthcare. For example, AI can take the output from an MRI scanner and evaluate that to see if there may be something that would be concerning. We might also have AI go through healthcare records and determine if anyone is receiving multiple drugs that might have negative interactions. This means the healthcare professionals can work with us at a human level and have AI take care of everything at a digital level. And if you’re watching this video on YouTube, you may be getting real-time language translation thanks to the AI that’s built into our new communications apps.
But, of course, there are downsides to any technology, and AI has been used for things that may not be as appropriate. For example, we’ve seen AI used for fraud. AI can impersonate a real person and even create video of that person, something we refer to as a deep fake. We’ve also seen application code that’s created with AI without any knowledge of what the application is actually doing. Or, perhaps, someone is using AI to create an image and taking credit for that image creation. And we’ve seen AI used in schools and colleges to plagiarize works. Students are using AI to summarize existing works, but they’re not citing AI in that summary.
One of the challenges we have when working with AI is that it can sometimes have a bias. The AI only knows the information that we’re feeding into it, and it has a pre-defined set of algorithms on how to manage that data. And sometimes, I can make the wrong conclusions based on this combination of data and algorithms. For example, we may be feeding AI a very large data set associated with healthcare in the United States. But inside of that data set, we may have underrepresented statistics for a particular ethnic group or particular gender group. In this case, AI could potentially create conclusions that are biased one way or the other. We’re, obviously, not trying to create bias as part of this AI analysis, so we have to make sure that we create an AI algorithm that doesn’t build bias when we have these large data sets.
A good example of this is what Amazon found when they took 10 years of submitted resumes and put them into an AI engine. This grouping of resumes had a number of terms, such as “executed” and “captured” as part of the resume text. The artificial intelligence engine tended to pick resumes that had those terms as part of the resume itself. What we found is that those terms were more often used when it was a male submitting a resume versus a female. And therefore, the AI created a bias towards male resumes. Obviously, Amazon chose not to use this AI algorithm for their resume research.
There’s also occasions where AI can completely misrepresent the data that it was provided. You can think of this as a confidently incorrect AI, where it really knows the answer, but the answer is completely wrong. For example, let’s say that we’ve told the AI that any picture containing a snout, a tail, and four legs is a dog. Then we provide the AI with a picture of a cloud, and the cloud appears to have a nose, a tail, and four legs.
Therefore, the cloud is obviously a real dog. This is, obviously, a simplified way to describe an AI hallucination, but this is what tends to occur when we have this type of data misinterpretation. These are also not very unusual to find, and you may have seen an AI hallucination when performing your own tests with an AI engine.
A practical example of an AI hallucination was when Google and Microsoft announced their new AI technologies. Google announced Google Bard– this is now Google Gemini. And Microsoft introduced Microsoft Bing Chat. Both of these engines provided a summary of some articles that had completely wrong information inside of them. Unfortunately, this hallucination was in the announcement that Google made regarding their new AI engine. This AI engine took an input that said, “What new discoveries from the James Webb Space Telescope can I tell my nine-year-old about?” And then it gave a number of bullets that you could use to help describe the James Webb Space Telescope and its findings.
Unfortunately, one of these findings was that the James Webb Space Telescope took the very first picture of a planet outside of our own solar system. Bruce McIntosh replied to this and said, “Speaking as someone who imaged an exoplanet 14 years before the JW Space Telescope was launched, it feels like you should find a better example.” This points out that even the best AI engines can suffer from an AI hallucination.
Behind the scenes, artificial intelligence is creating a conclusion using a model that is provided. And sometimes, it makes the wrong conclusion based on the model that it was given. And researchers will often perform tests of an AI engine to see how accurate it might be. They will get predictions from the AI, and then they will compare those predictions to known test data. One of these accuracy tests was done by Originality.ai They did an AI fact checking accuracy study in August of 2024. And you can read through the study yourself at the URL that you see here.
They evaluated their own AI and a number of other AIs that were currently available. The Originality.ai engine got a 72.3% accuracy, GPT-4 got a 64.9% accuracy rating, and GPT-3.5 got a 58.6% accuracy rating. As you can see, our current iterations of AI are far from perfect, and in some cases fall well under what most people think AI is able to do.
There’s been a lot of work made on both public AIs and private AIs. A public AI would be one that is publicly available on the internet. Things like ChatGPT or Google Gemini are good examples of a public AI. But some companies are creating their own AI engines, and these would be engines that are private to that organization. These would commonly have proprietary company data, and the organization has complete control over the modeling used for that artificial intelligence engine.
The decision on whether you use a public AI or a private AI may be based on data security. Whenever you put information into an AI engine, there is a possibility that data could be retrieved. And if that data is sensitive or contains private information, you could be putting that information available for anyone on the internet to see. We’ve also seen cases where an AI was able to provide passwords, encryption keys, certificate details, and other sensitive information.
Of course, if you’re running a private AI, only people within that organization have access to that proprietary data, and therefore, it limits the scope of any type of breach. There’s also quite a big difference between a public AI and a private AI when it comes to how much data they’re able to evaluate. Obviously, a private AI is only going to have data from a single entity, but a public AI is collecting data from everybody who’s on the internet.
In those cases, it may be that your public AI is more accurate than the private AI because the public AI has access to so much more data. But it is that large amount of data that does create concerns about privacy. Some of that information that’s being stored by a public AI is information about you. The AI knows where you live. It might even know your previous addresses, has information about your habits and things that you do on the internet, and it may have details about any memberships and groups that you belong to. Let’s say you’re applying for a job, what type of summary would the AI create of you, and what would be contained within your profile?