There is currently no universally agreed-upon definition of artificial intelligence. The term "intelligence" is understood as a measure of a machine’s ability to successfully achieve an intended goal. Like humans, machines exhibit varying levels of intelligence subject to the machine’s design and training. However, there are different perspectives on how to define and categorize AI.
In 2009, a foundational textbook classified AI into four categories:
- Ones that think like humans;
- Ones that think rationally;
- Ones that act like humans; and
- Ones that act rationally.
Most of the progress seen in AI has been considered "narrow," having addressed specific problem domains like playing games, driving cars, or recognizing faces in images. In recent years, AI applications have surpassed human abilities in some narrow tasks, and rapid progress is expected to continue, opening new opportunities in critical areas such as health, education, energy, and the environment. This contrasts with “general” AI, which would replicate intelligent behavior equal to or surpassing human abilities across the full range of cognitive tasks. Experts involved with the National Science and Technology Council (NSTC) Committee on Technology believe that it will take decades before society advances to artificial "general" intelligence.
According to Stanford University’s 100-year study of AI, by 2010, advances in three key areas of technology intersected to increase the promise of AI in the US economy:
- Big data: Large quantities of structured and unstructured data amassed daily from e-commerce, business, science, government, and social media. As datasets increase in size and quantity, so too do concerns of data standardization, securitization, and privatization.
- Standardization: data provided by multiple parties from multiple sources need to be converted to a common format to allow for consistent collaboration and application by researchers and programs.
- Securitization: sensitive data must be protected from unauthorized access, manipulation, and application of data throughout the computing system where data is used and stored. A common form of ensuring data security is called “Authentication” where authorized users must verify their identity through multiple methods such as providing a password and a generated passcode sent to the user’s phone.
- Privatization: while a subset of data securitization, data privatization relates to efforts to prevent the disclosure of sensitive information contained in the data such as health, financial, and criminal records. Privatization efforts include anonymizing data as well as providing users transparent indication of who will have access to their data for what purposes.
- Quantum and high-performance computing: Greater storage and parallel processing of big data made possible by emerging computing methods.
- Quantum computing: whereas traditional computers rely on storing and reading information in binary bits, quantum computers make use of new understandings of quantum mechanics that allow information to be read and stored exponentially faster and simultaneously on non-binary quantum bits or “qubits”.
- High-performance computing: while quantum computing can exponentially increase the abilities of single computers, advancement in high-performance computing enables the simultaneous application of multiple sets of computers, called “clusters”, to solve problems. Both quantum and high-performance computing allow for faster and more efficient problem solving, however these new capabilities could also be applied to nefarious uses that will have to be guarded against.
- Machine learning: the basis for many of the recent advances in AI. Machine learning is a method of data analysis that attempts to find structure (or a pattern) within a data set without human intervention. Machine learning systems search through data to look for patterns and adjust program actions accordingly, a process defined as training the system. To perform this process, an algorithm (called a model) is given a training set (or teaching set) of data, which it uses to answer a question. For example, for a driverless car, a programmer could provide a teaching set of images tagged either “pedestrian” or “not pedestrian.” The programmer could then show the computer a series of new photos, which it could then categorize as pedestrians or non-pedestrians. Machine learning would then continue to independently add to the teaching set. Every identified image, right or wrong, expands the teaching set, and the program effectively gets “smarter” and better at completing its task over time.
- Machine learning algorithms are often categorized as supervised or unsupervised. In supervised learning, the system is presented with example inputs along with desired outputs, and the system tries to derive a general rule that maps input to outputs. In unsupervised learning, no desired outputs are given, and the system is left to find patterns independently.
- Deep learning is a subfield in machine learning. Unlike traditional machine learning algorithms that are linear, deep learning utilizes multiple units (or neurons) stacked in a hierarchy of increasing complexity and abstraction inspired by structure of human brain. Deep learning systems consist of multiple layers and each layer consists of multiple units. Each unit combines a set of input values to produce an output value, which in turn is passed to the other unit downstream. Deep learning enables the recognition of extremely complex, precise patterns in data.
Experimental research in artificial intelligence includes several key areas that mimic human behaviors, including reasoning, knowledge representation, planning, natural language processing, perception, and generalized intelligence:
- Reasoning includes performing sophisticated mental tasks that people can do (e.g., play chess, solve math problems).
- Knowledge representation is information about real-world objects the AI can use to solve various problems. Knowledge in this context is usable information about a domain, and the representation is the form of the knowledge used by the AI.
- Planning and navigation includes processes related to how a robot moves from one place to another. This includes identifying safe and efficient paths, dealing with relevant objects (e.g., doors), and manipulating physical objects.
- Natural language processing includes interpreting and delivering audible speech to and from users.
- Perception research includes improving the capability of computer systems to use sensors to detect and perceive data in a manner that replicates humans’ use of senses to acquire and synthesize information from the world around them.
Ultimately, success in the discrete AI research domains could be combined to achieve generalized intelligence, or a fully autonomous “thinking” robot with advanced abilities such as emotional intelligence, creativity, intuition, and morality. Such autonomous agents could open new ethical and legal complications that will need to be adequately assessed and planned for. For instance, autonomous agents or programs may, as a product of their autonomy, operate outside the expectations of their creators. In the event that the agent or program’s creators have not implemented comprehensive stop gaps, the agent or program may inadvertently cause unintended harm to allies or adversaries. Whether the creators of the agents or programs are liable for any harms, and whether the harms should be given the same status of acts of war, is yet to be determined.