Generative AI is hype. Its ability to accomplish basic tasks is limited.

Though the tech continues to improve, there are still glaring problems with generative AI. Figuring out how to ensure models that are supposed to be reliable don’t generate false information has vexed researchers. At Google’s big cloud computing conference earlier this month, the company offered a new solution to the problem: Customers using its tech to train AI models could let their bots fact-check themselves by simply looking things up on Google Search.

Some claims about AI’s near-magic ability to do human-level tasks also have been called into question. A new paper from researchers at the University of Massachusetts at Amherst, Adobe, the Allen Institute for AI and Princeton showed that AI models routinely made factual mistakes and errors of omission when asked to summarize long documents. Another recent paper suggested that a claim that AI was better than the vast majority of humans at writing bar exams was exaggerated.

The big improvement in AI tech showcased by ChatGPT that kicked off the boom came from OpenAI feeding trillions of sentences from the open internet into an AI algorithm. Subsequent AIs from Google, OpenAI and Anthropic have added even more data from the web, increasing capabilities further. Seeing those improvements, some famous AI researchers moved up their predictions for when they think AI would surpass human-level intelligence. But AI companies are running out of data on which to train their models, raising the question of whether the steady improvement in AI capability will plateau.

Training bigger and better AI models has another crucial ingredient — electricity to power the warehouses of computer chips crunching all that data. The AI boom has kicked off a wave of data center construction, but it’s unclear whether the United States will be able to generate enough electricity to run them. AI, coupled with a surge in new manufacturing facilities, is pushing up predictions for how much electricity will be needed over the next five years, said Mike Hall, CEO of renewable energy management software company Anza and a 20-year veteran of the solar power industry.

“People are starting to talk about a crisis — are we going to have enough power?” Hall said.

Ethan Mollick, an associate professor at the University of Pennsylvania who studies AI and business, said executives are seeing the benefits of AI for their businesses in early experiments and tests and are trying to figure out how to work it into their organizations more broadly.

Some of them are concerned that their AI tools might be made obsolete by future tech, making them hesitant to invest a lot of money right now, he said. OpenAI is set to unveil its latest AI model in the coming months, potentially offering a whole new set of capabilities, Mollick said.

“The underlying technology drumbeat hasn’t really stopped,” he said. “It’s going to take some time to ripple through.”