Today’s digital assistants can sometimes fool you into believing they are human, but vastly more capable digital helpers are on their way. Behind the scenes, Siri, Alexa and their ilk use sophisticated speech-recognition software to figure out what you are requesting and how to provide it, and they generate natural-sounding speech to deliver scripted answers matched to your questions. Such systems must first be “trained”—exposed to many, many examples of the kinds of requests humans are likely to make—and the appropriate responses must be written by humans and organized into highly structured data formats.

That work is time-consuming and results in digital assistants that are restricted in the tasks they can perform. The systems can “learn”—their machine-learning capabilities allow them to improve their matching of incoming questions to existing answers—but to a limited extent. Even so, they are extremely impressive.

At a higher level of sophistication, technologies are now being developed to allow the next generation of such systems to absorb and organize unstructured data (raw text, video, pictures, audio, e-mails, and so on) from myriad sources and then autonomously compose cogent advice—or debate an opponent—on a subject they have never been trained to handle. 

We have seen glimpses of this capability at Web sites offering chatbots that, out of the box, can answer natural-language questions covering a wide variety of data sets they have trained on. Such chatbots need relatively little or no training on specific questions or requests; they use a combination of preconfigured data and the emergent ability to “read” relevant background materials supplied to them. They do, however, require some training in recognizing words and intentions before they can give highly accurate responses.

In June my employer, IBM, demonstrated a more advanced version of the technology: a system carried on a debate with a human expert in real time without having prior training on the topic or the position to be argued. Using unstructured data (including content from Wikipedia, some of which was edited for clarity), the system had to ascertain the relevance and veracity of the information and organize it into a reusable asset that it could call on to form coherent arguments supporting the position it had been assigned. It also had to respond to the arguments of its human opponent. The system engaged in two debates during the demonstration and was judged more persuasive in one of the two by a large group of spectators.

The enabling technology—which included software that could not only understand natural language but also handle the harder challenge of detecting positive and negative sentiment—was developed over more than five years and is still very much a work in progress. Nevertheless, the win by an unscripted AI system against an acknowledged human expert opens the door for countless related applications that could appear in the next three to five years, if not sooner. Such systems could, for instance, help physicians to quickly find research relevant to a complex case and then debate the merits of a given treatment protocol.

These intelligent systems will be useful only for assembling existing knowledge, not for creating it the way a bench scientist or an expert would. Still, as machines become increasingly intelligent, they raise the specter of job losses. It behooves society to provide the next generation with the skills it needs to tackle problems that require human ingenuity to solve.