- GenAI faces growing skepticism as it struggles to deliver on high expectations
- Early excitement for ChatGPT and LLMs has shifted to concerns about costs, power consumption, and ethics
- The AI industry is now focused on finding practical use cases and addressing challenges in reliability, contextualization, and governance
The conversation around generative artificial intelligence (GenAI) is growing weary.
GenAI is a “hammer looking for a problem,” Gartner VP Analyst Bern Elliot told Fierce Network this month. “People believe it really can solve a lot of things that it can't. I mean, the unfortunate thing is that OpenAI was amazing. But it was useless. It can't really do anything.”
Initially, everyone was excited about ChatGPT, Elliot noted. By spring 2023, the focus shifted to Large Language Models (LLMs). By summer, it became clear that substantial software support was necessary. By fall, the attention was on prompting, retrieval-augmented generation (RAG) and vectors. Come winter, we started talking about governance as improper uses of these technologies emerged. Most recently, the buzzword has shifted to "agentic" AI.
All this, in the midst of growing concerns over the cost, power consumption and ethics associated with GenAI.
The threat of disillusionment
Many people thought GenAI would "remake business" in early 2023, noted Brad Shimmin, Omdia Chief Analyst for AI Platforms, Analytics and Data Platforms. However, the early expectations for GenAI are a far cry from current realities.
Among the factors driving early GenAI adoption and hype: GenAI is more accessible compared to predictive AI, and allows for the same model to be used for multiple applications. GenAI also doesn’t require specialized expertise in natural language processing, Shimmin said.
But now, GenAI seems to be entering the "trough of disillusionment" — a phase in the Gartner Hype Cycle where interest wanes as a technology fails to deliver. During this period, the technology is often criticized and faces a decline in expectations until it matures and proves its practical value.
AI can be hard to place on the Hype Cycle because so many enterprises are in different parts of their implementation journey, Elliot said.
“AI is a really, really big topic,” Elliot told Fierce Network. Parts of it are always in the hype, and parts of it are past the hype and onto the “plateau of productivity,” a phase in Gartner's Hype Cycle where a technology's benefits become widely understood and accepted, resulting in stable and sustained use.
Other technologies have fallen into the trough, like 5G, which did so rather unexpectedly, causing a “5G winter” period that has affected the entire telecom industry.
Already, there have been examples of high expectations for AI leading to disappointment (cough, McDonald’s and IBM, cough). In another instance, the makers of Devin AI — an AI tool purportedly designed to automate the process of building software applications — ultimately failed to meet its ambitious claims for what it was capable of.
“A lot of this idea that [GenAI is] just a disappointment is that it's just not mature enough to have been properly used,” Elliot said. With AI specifically, the cycle of hype and speculation repeats every few months, requiring continual reassessment of use cases and control measures.
Technologies that fall into the trough of disillusionment sometimes don't come out, Elliot added. That isn’t likely with GenAI, but to get it moving toward the plateau of productivity, "expectations must become realistic with what can be delivered."
Looking for a killer use case
The advent of GenAI has catalyzed AI adoption in general, yet it has also highlighted the limitations of current AI technologies.
GenAI has showed adaptability in various areas, Shimmin noted, including entity extraction, sentiment evaluation and graph creation. But mostly, GenAI is only seeing traction with the “low hanging fruit,” he said.
Customer experience management and productivity gains have been primary areas of investment. Customer service chat agents, for example, have become popular, although not without some difficulties.
There have also been practical deployments in sales, like Salesforce's Einstein, a copilot for sales enablement.
Yet, there are challenges in applying GenAI to more complex tasks like advanced process automation, Shimmin said. Additional challenges remain with reliability, as well as contextualization and grounding of AI.
RAG and semantic search techniques are being explored to speed these models, but “people don't really get how complicated it can be,” Shimmin said.
Other concerns for current GenAI models include cost and latency issues. Questions are emerging over whether high AI budgets will be justified by return on investment (ROI).
Then there are the increasing governance and ethics concerns, and not to mention the question of where we will get all the power needed to keep AI infrastructure up and running. According to Goldman Sachs, AI is poised to drive 160% increase in data center power demand.
On the plus side
If there’s a silver lining, it’s that these challenges and disappointments are fostering innovation in the field, Shimmin said.
For one, the industry is starting to understand the importance of unbiased, well-prepared data for effective AI implementation. And techniques like vectorizing data, while nascent, are aiming at the challenges of handling unstructured data for GenAI.
RAG, a method of sourcing relevant information from a large dataset, has high hopes for improving data utility.
Other examples of emerging models and techniques, like smaller open-source models, quantization and agentic workflows, are also promising, Shimmin said.