Take the example of Google’s AlphaGo computer system, which managed to defeat a human at the ancient Chinese board game Go. But applying a similar kind of artificial intelligence in other situations will be difficult, explained Oren Etzioni, the CEO of the Allen Institute for Artificial Intelligence, a technology research organization.

Etzioni made the comments Monday at an MIT Tech Review conference in San Francisco about cutting-edge technologies. He was joined by fellow artificial intelligence experts from Google (GOOG) and the Chinese search company Baidu.

Get Data Sheet, Fortune’s technology newsletter.

Go, the game, is incredibly complex with more possible positions on the board than the number of atoms in the universe, Etzioni explained. But Go is still a board game, whereas the nuances of human language is far more complex and is difficult for even the best computer systems to understand, Etzioni said.

Just because AlphaGo defeated a human doesn’t mean that humans were uninvolved in its victory, he said. Humans still built the so-called neural networks used to train AlphaGo, and they were the ones who decided to feed it the data it required to learn how to play, Etzioni said.

“Machine learning is still 99% human work,” Etzioni said, referring to machine-learning algorithms, a subset of artificial intelligence technologies that has grown in popularity in recent years.

Peter Norvig, Google’s research director, explained that the rise of machine-learning algorithms has radically altered the way developers build software. Software powered by machine learning, like Google’s own voice-recognition service that it uses to perform search queries based on the human voice, must constantly react to the continual flow of data that they ingest, Norvig said.

The need to constantly react to the changing data means that the conventional “step-by-step” process of building software is not as relevant, Norvig said. “Every second, that data changes the machine,” he said. “We have some ideas to manage this uncertainty and change, but we need better models to deal with that.”

For more on Google watch our video.

Software powered by machine learning is also more complex than traditional software, and coders often have more trouble finding software bugs in it, he explained. In more traditional software, coders can more easily pinpoint where there is a bug and isolate it so it doesn’t affect the rest of the system.

With machine learning software, however, “changing one thing changes everything else,” said Norvig, which makes it hard to isolate errors. So far, not even Google has the appropriate techniques that let coders fix one bug and leave everything else the same in this advanced software, Norvig explained.

For others to create more advanced and innovative machine-learning software that can learn from data, there will need to be a “better set of tools” for software development than what’s now currently available, said Norvig.

Andrew Ng, Baidu’s chief scientist, explained the importance of data in order to make artificial intelligence systems work.

Baidu, along with companies like Google and Facebook (FB), are considered by experts to be among the leading businesses involved in artificial intelligence research. It uses artificial intelligence to improve its search engine and understand images in photos.

Apple Might Be Missing Out on AI

However, Baidu’s work in artificial intelligence depends on the data it collects, which requires the company to constantly look for new ways to obtain information beyond just its search engine. Ng explained that Baidu occasionally introduces new products for the specific purpose of accumulating data for artificial intelligence research, although he did not mention any specific products or the kind of data the company seeks.

He said that companies that want to benefit from artificial intelligence must think of the kind of data they have as well as how to obtain new data. Companies born from the Internet like Baidu and Google have a huge advantage because they are constantly sucking in data from the web while others do very little.

“Some companies [will] wish they had figured out an Internet strategy earlier,” Ng said.