This article was reviewed by a member of Caltech's Faculty.
Whether or not artificial intelligence (AI) will be able to outperform human intelligence—and how soon that could happen—is a common question fueled by depictions of AI in popular culture. Key to this comparison is a clear understanding of "intelligence."
One common definition describes intelligence as the ability of someone or something to achieve goals in a wide variety of environments. We can compare how well computers and humans are able to meet this definition.
Computers start with many advantages. They have better memories, they can quickly gather information from numerous digital sources, they can work continuously without the need for sleep, they don't make mathematical errors, and they are better at multitasking and thinking several steps ahead than humans. This makes them superior to humans at achieving some goals, such as calculating complex mathematical problems or sorting through large amounts of data. However, most AI systems are specialized for very specific applications.
Humans, on the other hand, can use imagination and intuition when approaching new tasks in new situations. This makes humans more readily able to apply their intelligence to a variety of environments, such as walking along unfamiliar trails. This is something machines often struggle with.
Intelligence can also be defined in other ways, such as the possession of a group of traits, including the ability to reason, represent knowledge, plan, learn, and communicate. Many AI systems possess some of these traits, but no system has yet acquired them all.
Scholars have designed tests to determine if an AI system has human-level intelligence. One example is the Turing Test, in which an interviewer exchanges messages with two players in different rooms. One player is a human, while the other is a machine. To pass the test, the machine must make the interviewer believe that it is the human player. Some AI systems can do this successfully but only over short periods of time.
As AI systems grow more sophisticated, they may become better at translating capabilities to different situations the way humans can. This would mean the creation of "artificial general intelligence" or "true artificial intelligence," a primary goal among some researchers. Theoretically, this could result in artificial intelligence that transcends human intelligence. The term "singularity" is sometimes used to describe a situation in which an AI system develops agency and grows beyond human ability to control it. So far, experts continue to debate when—and whether—this is likely to occur.
Several milestones highlight the advancement of artificial intelligence relative to human intelligence:
|1948–1949||The first autonomous robots were created by William Grey Walter. They could navigate around obstacles using light and touch.|
|1950||Alan Turing published the seminal paper, “Computing Machinery and Intelligence,” in which he posed the question “Can machines think?” and developed the Turing Test to answer the question.|
|1951||Marvin Minksy and Dean Edmonds built the first artificial neural network.|
|1956||The Dartmouth Summer Research Project on Artificial Intelligence was convened. It is considered to be the birth of the field of artificial intelligence.|
|1959||Arthur Samuel coined the term “machine learning” when describing machines that can learn to play checkers.|
|1964–1966||ELIZA, developed by Joseph Weizenbaum, became the first natural language processing program able to simulate conversation.|
|1966||Shakey became the first intelligent robot. It was able to perceive its environment, plan routes, recover from errors, and communicate in simple English.|
|1969||An optimized method for a backpropagation algorithm was published by Caltech alumnus Arthur Bryson and Yu-Chi Ho. This algorithm was key to enabling AI systems to improve on their own using their past errors.|
|1978||Texas Instruments’ Dallas research laboratory introduces the Speak & Spell, an educational toy that used a single silicon chip to electronically duplicate a human vocal tract.|
|1982||John Hopfield, a former Caltech faculty member, developed a neural network model to help explain how humans recall memories. The model helped to advance deep-learning technologies.|
|1989||Chess master David Levy lost a game to a computer for the first time.|
|1991||The introduction of the internet enabled online connections and data to be shared quickly and easily. This boost in data sharing had a significant impact on the advancement of AI.|
|1996||IBM’s Deep Blue computer defeated world champion Garry Kasparov in the first of a six-game chess series.|
|2005||Five autonomous vehicles successfully completed DARPA’s 2005 Grand Challenge, a 212-kilometer off-road course through the Mojave Desert.|
|2007||Caltech Distinguished Alumna Fei-Fei Li conceived and led the ImageNet project, a database that includes millions of labeled images available for computer vision research, highlighting the critical importance of large datasets in advancing AI.|
|2010||Siri, a voice-controlled virtual assistant, was released.|
|2012||AlexNet, an image-recognition model, completed the ImageNet Large Scale Visual Recognition Challenge with far greater accuracy than its predecessors. The publication of the AlexNet architecture is considered one of the most influential papers in computer vision.|
|2016||AI system AlphaGo, created by Google subsidiary DeepMind, defeated Go champion Lee Se-dol four matches to one.|
|2018||Joy Buolamwini and Timnit Gebru published the influential report, “Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification,” demonstrating that machine-learning algorithms were prone to discrimination based on classifications such as gender and race.|
|2018||Waymo’s self-driving taxi service was offered in Phoenix, Arizona.|
|2020||Artificial intelligence research laboratory OpenAI announced the development of Generative Pre-Trained Transformer 3 (GPT-3), a language model capable of producing text with human-like fluency.|