Viewpoints

Archived Viewpoints

2009

Before September 2009, the Artificial Intelligence technology area was Knowledge-Based Systems.

View the Knowledge-Based Systems Viewpoints archive

About This Technology

In some cases, when a task requires intelligence, artificial systems can do the job. Speech- and face-recognition software rely on techniques that emerged from AI research. Other everyday applications of AI include household floor-cleaning robots and artificial characters in video games that change their skill level to adjust to a user's abilities. AI is also important in the development of autonomous vehicles. Developers aim to solve increasingly difficult challenges to reproduce human abilities—for example, analyzing security-camera images, translating texts, delivering natural-sounding speech answers to natural-language queries, and guiding robots to walk on unpredictable terrain and manipulate objects. Other AI-development efforts aim toward goals that would be difficult or impossible for people to achieve: Credit-card-authorization systems make fast decisions about whether to allow a purchase at a retail store, automated financial-trading systems arbitrage multiple stocks that a human could not keep track of, and AI software advises doctors in diagnosis, relying on knowledge bases that are more comprehensive than a person can reliably memorize. AI may even be evolving into an uncanny capability that does not emulate human intelligence but instead seems to have a mind of its own—as in the case of chess-playing computers, which have styles of play that are distinctly different from those of human chess players.

Researchers and developers mix and match various approaches to produce an intelligent system; in some cases, advocates for one approach compete against advocates of another approach. Machine learning—perhaps the most common tool in today's AI tool kit—helps recognize speech, classify unknown snippets of DNA, and recommend products on websites. Developers build a system using good-quality training data, and systems are then able to match words to sounds, analyze gene sequences, and predict relevant products and ads. Machine learning is a tool that two researchers may agree on even if they otherwise use competing families of approaches. Notably, training data find use in systems that evaluate statistical probabilities and systems that loosely mimic the structure of human neural networks, yet both statistical and biomimetic approaches have the shortcoming of lacking an audit trail: Humans cannot validate every detail of a machine-learning system's "thinking." As a result, some researchers and developers continue to make use of strongly deterministic rule-based and case-based expert systems, logical reasoning, and semantic inferences. "Good old-fashioned AI" uses formal reasoning to emulate human reasoning, and "classical AI" enables basic robot control and goal-seeking behavior. R&D in a field that is closely related to AI—the Semantic Web (whose advocates insist that it is "not AI")—aims to encode web pages in machine-readable form such that machines can employ formal reasoning to provide natural-language (human-readable) answers to users' natural-language questions about entities and relationships that the pages describe. Nevertheless, a major direction for AI relies on combinations of all approaches or ad hoc approaches—anything that works—to perform tasks that people consider to demand intelligence.

Current progress in AI promises to enhance transportation and robotics, office productivity and collaboration, physical security and information security, transactions and logistics, entertainment and advertising delivery, and social interaction. Cars will rapidly respond to dangerous situations to avoid collision. Software will coordinate movements of service robots to perform useful tasks—perhaps gracefully. Personal-information agents will assist with scheduling and organizing meetings. Image-recognition systems will retrieve digital photographs, automate content production, alert people when security cameras capture events of interest, and provide context-sensitive information, entertainment, and advertising. Yet AI research also produces disappointments as developers encounter seemingly intractable problems such as language translation and commonsense reasoning. Visionaries and optimists may be disappointed with near-term projects to identify terrorists before they commit terrorist acts and to build robots that can provide safe and effective care of elderly and frail persons. But a good number of AI challenges are intermediate between intractable and feasible. Decision makers can benefit from monitoring advanced developments to reduce uncertainty about this large portfolio of intermediate-level problems. For example, different outcomes will result depending on the success of AI developers' efforts to manage risks in financial portfolios and credit decisions, to generate content and advertising that is truly relevant to individual users at specific times and places, to accept training from people who are not technology experts, and to train robots to navigate through doors and handle food and drinks. Different outcomes could result depending on whether users will find software agents to be trustworthy when executing transactions, taking control of cars, and monitoring security cameras; whether people will consider that agents provide believable responses to conversational queries; and whether interacting with software agents in virtual environments is worthwhile.