Vince Kellen understands the well-documented limitations of ChatGPT, DALL-E and other generative AI technologies — that answers may not be truthful, generated images may lack compositional integrity, and outputs may be biased — but he’s moving ahead anyway. Kellen, CIO at the University of California San Diego (UCSD), says employees are already using ChatGPT to write code as well as job descriptions.
OpenAI’s text-generating ChatGPT, along with its image generation cousin DALL-E, are the most prominent among a series of large language models, also known as generative language models or generative AI, that have captured the public’s imagination over the last year. The models respond to written requests to generate a variety of responses ranging from text documents and images to programming code.
Kellen sees ChatGPT-generated code as a productivity-enhancing tool in much the same way that compilers were an improvement over assembly language. “Something that produces libraries and software is no different than searching GitHub,” he says. “We also use it to write job descriptions that are sensitive to our content and formatting. You can then move on to editing very quickly, looking for errors and confabulations.” While the technology is still in its early stages, for some enterprise applications, such as those that are content and workflow-intensive, its undeniable influence is here now — but proceed with caution.
Ready for the right applications
Generative AI is ready for use in coding, administrative workflows, data refinement, and simple use cases such as pre-filling forms, says Oliver Wittmaier, CIO and product owner at DB SYSTEL GmbH, the wholly owned subsidiary of DB AG and digital partner for all group companies. And in the transportation industry, he says, “AI can directly or indirectly impact the avoidance of transport, the steering of transport, and the management of transport.”
Oliver Wittmaier, CIO and product owner at DB SYSTEL GmbH
DB SYSTEL GmbH
Content generation is also an area of particular interest to Michal Cenkl, director of innovation and experimentation at Mitre Corp. “I want contextual summarization and refinement via dialog, and that’s what these large-language models deliver, he says. Currently his team is looking into two use cases in the knowledge and expertise domains. “The first is if I want to write an email to one of our sponsors that summarizes the work we’ve done that’s relevant to them—and write it in the context of communications we’ve already had with them. That’s incredibly powerful.”
The second is for project staffing. Normally Cenkl reviews résumés and searches by skills tags to find the right people for a project. Generative AI can facilitate that. “For example, I might want to ask, ‘What can Michael do on this project,’ based on what he’s doing now, and get a summary of what he could do without me having to construct that from a résumé.”
And over at used car retailer CarMax, they’ve been using generative AI for over a year, leveraging OpenAI’s APIs to consolidate customer review text to summaries that are more manageable and readable. But CIO Shamim Mohammad says his team has expanded its use of the technology into other areas as well.
One application, in vehicle imaging, was conceived as a way to improve customer experience. AI optimizes images for every vehicle the company adds to its inventory, which at any given time includes between 50,000 and 60,000 vehicles, he says. “We make every image as realistic as possible without losing the validity of it,” he says. For example, its data scientists created a “digital sweeper” model that replaces a photo of a car sitting on a dirty floor with an image that shows the car sitting on a clean one. “It’s still the same car, but looks better and it’s a better experience for the customer,” he says.
Similarly, Nike has been using generative AI to generate product prototype images, says Forrester analyst Rowan Curran. “You can use a text-to-3D modeler, test in 3D space, and get a much more visceral feel for how it will look in the real world — all with very little effort,” he says.
Applications with the greatest potential payback
Creating code and improving customer experience are the main areas companies can take advantage of today using generative AI, and they have the greatest potential payback in terms of efficiency gains, Mohammad says.
Shamim Mohammad, CIO, CarMax
Gary Jeter, EVP and CIO at TruStone Financial Credit Union, says these are areas his developers have also pursued with GitHub’s implementation of OpenAI’s Codex. And, he says, using generative AI for coding has worked well. Cenkl adds that generative AI models work better on coding than on human language because programming languages are more structured. “It can tease out that structure, and that’s why it works,” says Cenkl.
CarMax is experimenting with GitHub’s Copilot, where he says engineers in some cases could potentially generate up to 40% of their code. “This is evolving quickly,” Mohammad says. “But you have to make sure there’s no copyright infringement, fake content or malware embedded if you’re using it to create software.” You can’t just plug that code in without oversight.
Other areas ripe for enterprise applications, says Curran, include generating marketing copy, images, designs, and creating better summaries of existing data so people can consume it more effectively. “Some people are even using these large language models as a way to clean unstructured data,” he says. And in the coming year, generative AI capabilities may begin to appear in some enterprise software ranging from help desk software to Microsoft Office applications.
Don’t trust, verify
Aside from the benefits, CIOs deploying the technology need to be aware of potential intellectual property issues regarding generated outputs, CarMax’s Mohammad cautions. Generative models, such as DALL-E that trains on data from the Internet, have generated content that may infringe on copyrighted content, which is why Getty Images recently sued Stability AI over its AI-driven art generation tool Stable Diffusion.
Michal Cenkl, director of innovation and experimentation, Mitre Corp.
The technology also needs human oversight. “Systems like ChatGPT have no idea what they’re authoring, and they’re very good at convincing you that what they’re saying is accurate, even when it’s not,” says Cenkl. There’s no AI assurance — no attribution or reference information letting you know how it came up with its response, and no AI explainability, indicating why something was written the way it was. “You don’t know what the basis is or what parts of the training set are influencing the model,” he says. “What you get is purely an analysis based on an existing data set, so you have opportunities for not just bias but factual errors.”
Wittmaier is bullish on the technology, but still not sold on customer-facing deployment of what he sees as an early-stage technology. At this point, he says, there’s short-term potential in the office suite environment, customer contact chatbots, help desk features, and documentation in general, but in terms of safety-related areas in the transportation company’s business, he adds, the answer is a clear no. “We still have a lot to learn and improve to be able to include generative AI in such sensitive areas,” he says.
Jeter has similar concerns. While his team used ChatGPT to identify a code fix and deploy it to a website within 30 minutes — “It would have taken much longer without ChatGPT” — and he thinks it’s useful for drafting terms and conditions in contracts, it’s not entirely proven. “We will not expose any generative AI to external members,” he says. “TruStone will not be bleeding edge in this space.”
Gary Jeter, EVP and CIO, TruStone Financial Credit Union
TruStone Financial Credit Union
When TruStone eventually starts using the technology for the benefit of its members, he adds, it will monitor conversations through human and automated review to protect its members and the brand.
Today, the key to successful deployment is still having a human in the loop to review generated content for accuracy and compliance, says UCSD’s Kellen. “Making sure the machine makes the right decision becomes an important litigation point,” he says. “It’ll be quite a while before organizations [use it] for anything that’s high risk, such as medical diagnoses.” But generative AI works fine for generating something like review summaries, provided there’s a human overseeing them. “That slows us down a bit, but it’s the right thing to do,” he says. Eventually, he adds, “We’ll find automated ways to ensure that quality is good. But right now, you must have a review process to make sure the content generated is accurate.”
Vince Kellen, CIO, UCSD
Another well-documented risk, in addition to accuracy, is the potential for bias in the models introduced from the data used to train them. This is especially problematic when generative AI is using content from the Internet, as ChatGPT does, but that may be less of an issue when training the model against your own private corporate data that you can review for potential bias, Kellen says. “The more you get to the enterprise, where the class of data is more constrained and more mundane, the more generative AI shines,” he says.
The thing to understand about large-language models, says Cenkl, is these machines are to some degree savants. “They don’t understand, but they’re very good at computing,” he says.
Changes in job responsibilities, roles
“Technology has made things better, but it’s also created a lot of extra work for us,” says Mohammad. However, he believes generative AI is different. “It’s exciting because it’s going to take away some of the stuff we don’t like to do and make us more intelligent,” he says. “It will augment humans.”
But Curran points out that there’s no expectation that generative AI will completely replace any role in the short term. “It may reduce the number of people needed to execute a role, such as in content development, product information management or software development,” he says. “But there will always be the need for a human in the loop.” And Mohammad adds that even if the technology can write and summarize, human intelligence will always be needed to ensure quality, and to control what’s been generated to make it better.
Steps to get started
Now is the time to get up to speed on generative AI technology and start experimenting, says Kellen. “CIOs have to get their heads inside this puzzle before they’re bamboozled by vendors who are embedding the technology into their enterprise software offerings,” he says. “If you spend the next year procrastinating, you’ll be behind the curve.”
It’s important to get educated and go deeper than the public discussion on ChatGPT in order to understand that this technology is much more complex than one application, says Curran. Then start considering use cases where generative AI might improve the efficiency or quality of existing processes. Finally, ask what types of capabilities you’ll need and whether you should acquire that from a vendor or build it yourself.
From there it’s a matter of testing the technology and consider potential use cases. “A lot of your systems, whether they use structured or unstructured data, will have at least some component of natural language and conversational interface,” says Cenkl. “Think about the data you have and what parts of that can be augmented by these technologies,” and then demonstrate the potential. For example, Jeter says he generated a template of terms and conditions and sent it to his compliance department to show how they could use it.
Generative AI models are large, and training them from scratch is expensive, so the best way to get started is to use one of the cloud services, says Curran. CarMax, for example, uses Microsoft’s Azure OpenAI Service with GPT 3.5. “The data we load is our own — it’s not shared with others,” Mohammad says. “We can have massive amounts of data and process it very quickly to run our models. If you have a small team or business problem that might take advantage of generative AI technology, give it a shot.”