The onslaught of ChatGPT and generative AI news is not necessarily helpful to today’s enterprise decision-makers.

Early indications are that conversations with OpenAI’s trained language model can take some strange paths. Still, innovators need to take care, explore and probe this AI technology’s use cases.

That’s the word from Rowan Curran, analyst with Forrester Research. While the great scale of generative AI datasets brings new complexity, the same basic rules that already guide sound AI governance are likely still to apply, according to Curran, who recently co-authored a report with fellow Forrester analysts on generative AI and the enterprise. Learning and experimenting is time well spent, he suggested.

Experimentation and excitement — and caution

While it is not yet easy to critically explore generative AI’s vast possible use cases, downplaying the technology would be a mistake, Curran told VentureBeat. Forrester is encouraging folks to embrace the experimentation and excitement of this space, he said, but to do so with the knowledge that what you are building out will very likely look quite different from ChatGPT and its brethren as seen today.

>>Follow VentureBeat’s ongoing generative AI coverage<<

“The important thing for leaders — whether it’s at the C-level or a couple clicks down — is to take a very investigative, skeptical and questioning approach in adopting generative AI,” he said.

Generative AI shows promise for content creation and summarization on both the textual and the image-generation side. In Forrester’s estimation, this can advance collaboration within organizations. It can begin to write code and support research on different programming schemes.

Still, as noted in Forrester’s Generative AI Prompts Productivity, Imagination, And Innovation In The Enterprise report: “Generative AI can go horribly wrong and there is much we don’t know yet about how generative AI models will perform at scale.”

What’s inside the box?

Some early experiences also suggest that these models’ large-scale datasets can take unwanted bias to new levels.

AI as a black box — one that creates unexplainable results — has long been an issue of concern to CEOs, technologists and society in general. VentureBeat asked Curran if generative AI leapfrogged these black box issues in any way.

“Absolutely not,” he responded. “We still have the same issues with data quality, bias and making sure that these models perform in a way that is acceptable. One of the current challenges with them is that, when we’re talking about large language models (LLMs), they are very much like black boxes in a lot of ways.”

Yet, a lot of work is being done to determine what is happening within these LLMs. But, the fact that there is no clear picture of their inner workings should not be a deterrent to cautious experiments, Curran said. And for many, the experience will be familiar.

“We have been using neural networks in a variety of different use cases for years, and understanding what a neural network is doing is still very hard,” he said. “Large language models are a black box, but that should direct how we apply them, not make us withdraw from them completely.”

But there are some differences too. Not only is the LLM itself something of a black box to the viewer, so are the datasets on which it works. And, large means large.

“The models’ very size makes it very hard for a complete review within a reasonable amount of time at a reasonable cost,” said Curran.

Weighing generative AI

Looking ahead, it’s up to enterprise decision-makers to discern what is sensible and doable with ChatGTP-style generative AI models. Failing to understand their pros and cons could be a costly mistake, Curran advised.

The pros include enhanced developer productivity, more extensive test sets for security hardening and expanding the breadth of human creative expression. The cons include a tendency toward bias, vulnerability to security attacks, a disarming human-like behavior and significant costs. 

Curran said it is important to look at these innovations as enterprise tools. He said that there’s no enterprise tool that solves every problem for everybody. “Taking an approach like that to generative AI is just going to end in disappointment,” he chided.

Each organization will need to study cases where they can take advantage of the strengths of the new tools in their organization. In the early going, said Curran, that may include content ideation and summarization.

VentureBeat’s mission is to be a digital town square for technical decision-makers to gain knowledge about transformative enterprise technology and transact. Discover our Briefings.