Cognitive Computing

Cognitive computing is a term for artificial intelligence (AI) systems that take in large amounts of unstructured data, such as text, speech, and images, generate and weigh hypotheses about it, and support human decisions with ranked, evidence-backed recommendations. IBM popularized the term after its Watson system won the quiz show Jeopardy! in 2011, and through the 2010s it served mainly as a product and marketing label for systems that combined natural language processing (NLP), machine learning, knowledge graphs, and conversational interfaces such as chatbots. The systems it described are now simply called AI, and the vendors that used the term have mostly retired it.

Origins and Definition#

IBM Research used "cognitive computing" for SyNAPSE, a DARPA program launched in 2008 to build brain-inspired chips; it produced the TrueNorth chip (2014), with one million spiking neurons and 256 million synapses. After Watson's victory, the meaning shifted to software that reasons over documents, and John Kelly, director of IBM Research, and Steve Hamm titled their 2013 book about Watson "Smart Machines: IBM's Watson and the Era of Cognitive Computing." In 2014 Sue Feldman, co-founder of the Cognitive Computing Consortium, and analyst Hadley Reynolds published a working group's definition of cognitive systems as adaptive, learning as information and goals change; interactive, with users and other systems; iterative and stateful, asking clarifying questions and remembering earlier steps; and contextual, drawing on time, place, domain, user profile, and task.

How IBM Watson Worked#

In February 2011, Watson played two Jeopardy! games against the show's most successful champions and finished with $77,147, against $24,000 for Ken Jennings and $21,600 for Brad Rutter. It ran DeepQA software, led by David Ferrucci, on 90 servers with 2,880 processor cores, without an internet connection during play. DeepQA searched its stored sources for many candidate answers, scored the evidence for each with hundreds of algorithms, and combined the scores into a confidence learned from past games, all in about three seconds:

Watson buzzed in only when its confidence cleared a threshold, and the broadcast showed its top three candidate answers with their confidence levels. Watson was an open-domain question answering system, and IBM's later products kept the design of returning ranked answers with their supporting evidence, a form of explainability that let professionals check the sources behind a recommendation.

Applications and Setbacks#

IBM aimed Watson at oncology, banking, tax preparation, customer service, and other industries where professionals read large volumes of documents, with each product an artificial narrow intelligence trained for its own domain. Healthcare proved hardest: a 2017 University of Texas System audit found that MD Anderson Cancer Center had paid more than $62 million to outside firms, mainly IBM and PricewaterhouseCoopers, for a Watson oncology project that had not produced a system ready for clinical use. IBM sold its Watson Health data and analytics assets to Francisco Partners in 2022, where they became the company Merative. In 2023 IBM launched watsonx, a platform built around foundation models, and Microsoft renamed Azure Cognitive Services to Azure AI services. The Ultralytics IBM watsonx integration guide trains a YOLO26 detector in a watsonx.ai notebook, fine-tuning yolo26n.pt to find plastic litter, biological material, and remotely operated vehicles in underwater images.

Cognitive Computing vs. Related Terms#

• Artificial general intelligence (AGI): A hypothetical system able to learn any intellectual task a person can. Cognitive computing systems each served one domain, such as oncology or tax preparation.
• Symbolic AI: Reasons with hand-written rules and logic. DeepQA mixed symbolic sources such as taxonomies with statistical evidence scoring.
• Cognitive architectures: Research models of human cognition, such as Soar and ACT-R, that simulate memory, attention, and problem solving to explain how people think. Cognitive computing products used the word for business software that assists experts.
• Affective computing: Detects and responds to human emotions, which IBM offered as a cognitive service with Watson Tone Analyzer, an API that scored emotions and tones in text.