Generative AI, a field of artificial intelligence, has existed since the 1960s. However, since November 2023, when the first iteration of ChatGPT was launched, the generative AI surge has been unstoppable.
(Listen to our podcast on generative AI)
"Generative AI as a concept just refers to any kind of AI algorithm or model that can generate new content or data based on patterns and examples found in existing data," says Cognistx Data Scientist Justin Waltrip.
With chatbots, data scientists train generative AI algorithms to grab (or scrap) information from the internet. These algorithms pick up patterns in the text and images so that it can reproduce a response when an individual asks it a question it hasn't seen before.
Generative AI Use Cases
Over the past year, at Cognistx we’ve witnessed the emergence of numerous business-use cases as people gain a better understanding of the applications and benefits of generative AI. Among these, content generation stands out.
Content-generation apps, platforms and engines allow the creation of articles, blog posts and social media content tailored to specific topics or individual users. Generative AI may not be used to generate that entire content, but it perhaps assists in the process.
"So maybe it's helping to autocomplete things as you're typing, or maybe it's revising things for clarity,” Waltrip says. “You can think of it as a better version of a tool like Grammarly, where instead of just correcting your grammar, maybe now it's looking at sentence structure and paragraphs. [For] things like creating artwork, graphic designs, even architectural blueprints, using generative AI has proven to be helpful and a source of inspiration.”
With image generators such as DALL-E or Mid Journey, generative AI uses images that have been scraped from the internet. The AI model learns to pick up patterns and understand what various things look like. It knows what a dog looks like. It knows what a cat looks like. So, when you ask an AI image generator to generate a new image, it relies on that old data to create something that synthesizes those existing patterns.
"There's been a tremendous explosion in how powerful these tools are, but also, people are finding all these new and really creative ways to use them," says Waltrip.
People now use generative AI to develop video games. It could be used in drug discovery to create new compounds when there's a lack of existing data or when the data is very private, such as in the healthcare industry.
However, as generative AI’s spectrum of applications expands, its limitations are becoming increasingly apparent. Scalability and ethical concerns arise alongside the possibilities.
Generative AI Limitations
"AI has the potential to generate vast amounts of content across various sectors, yet the technology is not without its boundaries,” Waltrip says. “We have to remember that generative AI is a tool—a means to enhance human capability, not replace it—and this requires a mindful approach to its deployment."
A significant concern with generative AI is the lack of control over what's being produced by the model.
If you ask a chatbot for a specific drink recipe and it has yet to see that drink in its data set, it's going to draw its own conclusions from the information it does have. There's no way to control exactly which data it's pulling from or which data was in the data set.
"In the end, you aren't always super sure that what you're getting is actually what you asked for,” Waltrip says. “And it's the same with image generation. You can ask it to do things, but the more complicated your prompt gets, the less likely [it is] that the AI model is going to actually be able to reflect that."
Common Generative AI Problems
As with any technology, generative AI is not without its share of common issues, such as these:
● Lack of consistency: Occasionally, users may inadvertently input the same prompt twice. Many have encountered this situation with ChatGPT. You start with a familiar prompt, then you return to it a month later only to receive a vastly different response.
● Verification: It's often impossible for users to freeze the model and ensure consistent responses every time because AI developers may tweak the model or add constraints. This unpredictability could pose challenges, particularly in a business environment.
● Ethical concerns: There's been a lot of talk about deep fakes of celebrities on Twitter or X, especially with people like Taylor Swift. It's easy to use an open-source image generator such as Stable Diffusion to train or refine the generator specifically with images of her. This way, you can create high-quality images of her in various settings, even ones she may not have explicitly agreed to.
"We don't want to generate [information or images] that other people might think are real, and we have no way to really control the spread of that misinformation," Waltrip said.
How AI Developers Offset Issues
At Cognistx, the staff works to ground an AI model's responses by giving it access to specific sources of information. If you're writing an article for a business, you could provide it with access to articles that your company has already written. An AI model can analyze previous blog posts and articles to grasp your writing style, information presentation and critical points of focus. This guides the model’s creation of new articles with precision and relevance.
Ethics & Bias
Many universities are emphasizing the importance of computer science students taking ethics classes because, with the speed of technological development, laws have yet to catch up to the pace of technology.
"When you have the power to develop these really critical systems that millions or billions of people are using on a daily basis, it's important that you understand a lot of the ethical limitations with this technology and you just don't go full steam ahead and just say, ‘Oh, just because we can do it, we should,’” says Waltrip. "That's not always the right approach. When I look at a new technology, especially one that's as powerful as generative artificial intelligence, it's really important to consider the potential harms in that situation."
Moreover, it's crucial to acknowledge that language models, such as those trained on internet data, can inherit biases from the source material. For example, if the training data originates from user-response platforms like Reddit, various biases, including sexism or racism, may manifest in user responses.
When in training, a language model learns to recognize and amplify these patterns, including any biases present in the data. This means that if the training data includes sexist such as portraying women and minorities in lower-paying jobs or “underrepresented data of women or minority groups can skew predictive AI algorithms.”
Therefore, it's crucial to acknowledge these potential limitations and take proactive steps to address them when working with such models. Awareness and efforts to mitigate these issues are essential for responsible usage and to promote fairness.
Future of Generative AI
Understanding the presence of bias paves the way for a broader conversation about the future of generative AI. As we look ahead, there's vast potential for growth and improvement in this field.
Contrary to the negative scenarios of how AI is often portrayed, at Cognistx, we foresee a scenario where AI enhances efficiency. Just like how a calculator eliminates the need for manual arithmetic.
"We are basically trying to remove parts of your job that aren't interesting,” says Waltrip. “Our goal isn't to replace someone's job, but it's to automate a lot of the parts of that job that most people don't find all that interesting. That's kind of the first direction, and that's what we're focused on when we talk about our enterprise product SQUARE."
Listen to our AI-Driven podcast on generative AI.
