The importance of artificial intelligence in computer science today

What artificial intelligence is, as a field of computer science

Artificial intelligence is the branch of computer science that builds systems able to do tasks that seem to need intelligence: learning from experience, reasoning about facts, recognizing images and sound, and using language. John McCarthy, co-author of the 1955 proposal described below, defined it as "the science and engineering of making intelligent machines, especially intelligent computer programs". It is not one technique but a family of research areas. The arXiv subject taxonomy shows the shape of it: its artificial intelligence category (cs.AI) covers expert systems, theorem proving, knowledge representation, planning and reasoning under uncertainty, while machine learning (cs.LG), natural language processing (cs.CL), computer vision (cs.CV), robotics (cs.RO) and multiagent systems (cs.MA) each have a category of their own.

What changed computing is the part that learns. A conventional program is a set of rules: a developer writes down how to turn input into output. Machine learning turns that around. You supply examples, inputs paired with the outputs you want, and a training algorithm fits a model, a function with a very large number of adjustable parameters, that reproduces them and generalizes to inputs it has not seen. This is why AI matters to computer science in particular: it gives the field a way to build software for problems nobody can specify as rules, such as recognizing a spoken word, spotting a new strain of malware or choosing a good plan for a database query.

Three kinds of learning cover most systems in use. Supervised learning fits labelled examples (this email is spam, this one is not). Unsupervised learning finds structure in unlabelled data, such as groups of similar customers. Reinforcement learning learns by trial and reward: the system acts, observes the result and adjusts, which is how several of the compiler and networking systems below were trained. Deep learning is machine learning with artificial neural networks of many layers; the ACM credits it with the recent breakthroughs in computer vision, speech recognition, natural language processing and robotics.

Generative AI, the kind behind today's chat assistants, is one application of these methods. Canada's privacy commissioners describe it as a subset of machine learning in which systems are trained on massive information sets to generate text, code, images, video or audio in response to a prompt, and note that its output is probabilistic: the same prompt can produce different answers.

A short history, from the 1956 Dartmouth proposal to the Nobel Prizes

AI research began after the Second World War. McCarthy credits Alan Turing with a lecture on intelligent machines in 1947, and with perhaps being the first to decide that AI was best researched by programming computers rather than by building machines. The research program then took shape in a funding proposal. On August 31, 1955, John McCarthy (Dartmouth College), Marvin Minsky (Harvard University), Nathaniel Rochester (IBM) and Claude Shannon (Bell Telephone Laboratories) proposed a two-month study of artificial intelligence by ten researchers at Dartmouth College in Hanover, New Hampshire, in the summer of 1956. Their working conjecture was that every aspect of learning, and any other feature of intelligence, could in principle be described precisely enough for a machine to simulate it.

The proposal reads like a table of contents for computer science. Beside language and "neuron nets", it lists a theory of the size of a calculation (how to measure the complexity of computing a function), self-improving programs, abstraction, and the role of randomness in creative thinking. It also argues that the main obstacle was not the speed or memory of the computers of the day but the inability to write programs that used them well. From the start, AI was a computer science problem.

Much of early AI was built from hand-written symbols, rules and search. Neural networks were already on the Dartmouth list, and by the 1980s they were in use as a tool for recognizing patterns; by the early 2000s, according to the ACM, only a small group of researchers, among them Yoshua Bengio, Geoffrey Hinton and Yann LeCun, was still committed to them. What changed was hardware and data: the ACM points to powerful graphics processing unit (GPU) computers and access to massive data sets.

The 2012 result is the paper by Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton that paired a large network with an efficient GPU implementation. The Turing Award citation, for breakthroughs that made deep neural networks "a critical component of computing", is the importance of AI in computer science in one line.

Where AI works inside computer science's own subfields

AI is not only something computer science builds for other fields. Its methods now improve the discipline's own machinery, often in places users never see. The examples below come from production systems and published research, each with its date.

Compilers and systems

Compilers are full of heuristics: rules of thumb, tuned by hand, for decisions such as whether to inline a function. In 2021, Google engineers described MLGO, a framework that replaced LLVM's hand-written inlining-for-size heuristic with a model trained by reinforcement learning. It reduced code size by up to 7% compared with LLVM's -Oz setting, and it is available in the main LLVM repository.

Learning can also find algorithms outright. In a paper published in Nature on June 7, 2023, DeepMind's AlphaDev used deep reinforcement learning to discover routines for sorting three, four and five elements. After review they were included in the LLVM libc++ standard sorting library: the first change to those subroutines in over a decade, and the first time any component of that library was replaced by an algorithm discovered with reinforcement learning.

AI-guided search now reaches whole data centers as well. In May 2025, Google DeepMind described AlphaEvolve, a coding agent built on its Gemini models that pairs them with automated evaluators and an evolutionary search over candidate programs. A heuristic it found for Borg, the system that orchestrates Google's data centers, had by then been in production for over a year and recovers on average 0.7% of Google's worldwide compute resources.

Databases and query optimization

A database index is a model in disguise: a B-tree maps a key to the position of a record. A 2017 paper, The Case for Learned Index Structures, took that literally and replaced index structures with learned models. On several real-world data sets, its neural networks beat cache-optimized B-trees by up to 70% in speed while using an order of magnitude less memory.

Query optimizers are the other target. The optimizer decides how to run a SQL query: which indexes to use and in which order to join tables. Bao (2020) does not replace it. It gives the existing optimizer per-query hints, learned from the queries it has already run, by combining a tree convolutional neural network with Thompson sampling, a long-established reinforcement learning method, so it adapts when the workload, the data or the schema changes.

In production, the common form is self-tuning. Automatic tuning in Azure SQL Database, which Microsoft describes as based on AI and machine learning, watches the queries a database runs and can create indexes that would help, drop indexes that have gone unused for 90 days, and force the last good execution plan when a query regresses. It applies changes when CPU and I/O use is low, measures the result, and reverts any change that makes performance worse.

Security and malware detection

In security, machine learning answers a speed problem: new threats appear faster than traditional signature updates can ship. Microsoft Defender Antivirus pairs its on-device engine with cloud protection that uses AI systems driven by machine learning models to identify new threats, sometimes before a single endpoint is infected, and that Microsoft says delivers protection far faster than traditional security intelligence updates; cloud protection is on by default. Gmail's AI-enhanced spam filtering blocks nearly 10 million spam emails every minute, and Google says Gmail stops more than 99.9% of spam, phishing attempts and malware from reaching the inbox.

A classifier like these does not know the truth about a message or a file; it outputs a score or a probability, and the operator decides what that triggers: block, quarantine, warn or pass to a person. The pairing of a threshold with a review path is where much of the engineering judgment in an AI security product lives, and it is the pattern to copy wherever AI errors are costly.

AI now finds vulnerabilities as well as blocking attacks. On November 1, 2024, Google's Big Sleep team (Project Zero with Google DeepMind) reported that its language model agent had found an exploitable stack buffer underflow in SQLite; the developers fixed it the same day, before it appeared in an official release. In the final of DARPA's AI Cyber Challenge, whose results were announced on August 8, 2025, autonomous cyber reasoning systems worked across 54 million lines of code. They found 54 unique synthetic vulnerabilities in 63 challenges and patched 43 of them, and they also turned up 18 real vulnerabilities, which are being disclosed to the projects' maintainers.

Networking

Networks run on control loops, and learned controllers now compete with hand-tuned ones. A video player chooses a bitrate for every few-second chunk of video, trading picture quality against the risk of the playback buffer running dry. Pensieve, presented at ACM SIGCOMM 2017 by researchers at MIT's Computer Science and Artificial Intelligence Laboratory, trained a neural network with reinforcement learning to make that choice from what the player observes, and beat the best existing scheme by 12% to 25% in average quality of experience.

At the edge of the network, machine learning separates people from bots. Cloudflare gives each request a bot score from 1 (almost certainly automated) to 99 (almost certainly human). Deterministic heuristics handle the clear-cut cases, and its documentation says the machine learning engine, drawing on the billions of requests its network proxies daily, accounts for the majority of detections. Site owners then use the score in WAF custom rules to block, challenge or allow a request.

Human-computer interaction

For many people, AI is simply part of the interface: speech recognition, captions, translation and cameras that describe what they see. Apple's accessibility features announced in May 2022 used advances in hardware, software and machine learning to add Door Detection, which helps blind and low-vision users find a door in the last few feet to a destination, and Live Captions for any audio on iPhone, iPad and Mac, generated on the device so that, in Apple's words, user information stays private and secure.

Human-computer interaction research also had to work out how people should use systems that are sometimes wrong. Microsoft researchers' Guidelines for Human-AI Interaction, presented at CHI 2019, proposed 18 design guidelines and tested them with 49 design practitioners against 20 AI-infused products.

Software engineering

In software engineering, AI now sits inside the developer's editor. GitHub made Copilot generally available to individual developers on June 21, 2022, after a technical preview the year before, as an editor extension that writes and completes code. Testing changed too: in November 2024, Google's OSS-Fuzz team reported 26 new vulnerabilities found with AI-generated and AI-enhanced fuzz targets, including CVE-2024-9143 in OpenSSL. Generated code still needs the same review, tests and scans as any other code; our secure coding checklist covers what that review should look for. What the evidence says about productivity, research directions and careers is in our guide to the future of AI in computer science.

Why AI matters for people today

Outside the lab, the importance of AI shows in services people already depend on, usually without seeing the model.

• Safer inboxes and devices. The spam, phishing and malware filters above protect ordinary Gmail accounts and Windows computers by default.
• Accessibility. Live Captions and Door Detection turn an ordinary phone into an assistive device, and the captions are generated on the device itself.
• Health care. The US Food and Drug Administration publishes a list of AI-enabled medical devices authorized for marketing in the United States. In its version with content current as of September 4, 2026, the list has 1,614 entries, 1,230 of them (about three in four) in radiology; 335 had their final decision in 2025, against 6 in 2015. The FDA notes that the list is not comprehensive, because it finds devices mainly by AI-related terms in their summaries.
• Science. The 2024 Nobel Prize in Physics went to John Hopfield and Geoffrey Hinton for foundational discoveries and inventions that enable machine learning with artificial neural networks. Half of the Nobel Prize in Chemistry went to Demis Hassabis and John Jumper of Google DeepMind for protein structure prediction: according to the prize announcement, their AlphaFold2 model has predicted the structure of virtually all 200 million proteins researchers have identified, and more than two million people in 190 countries have used it.

Why AI matters for businesses today

For businesses, the clearest evidence is adoption, and official statistics show it rising fast. Statistics Canada reported in June 2026 that 19.2% of businesses surveyed in the second quarter of 2026 had used AI to produce goods or deliver services in the previous 12 months, triple the 6.1% of the second quarter of 2024. In the EU, Eurostat found that 19.95% of enterprises with 10 or more employees used at least one AI technology in 2025, up 6.47 percentage points from 2024.

Two things stand out. First, a lot of AI arrives inside products a business already licenses, such as the antivirus, email and database services above. Second, not every business needs more of it: 40.0% of Canadian businesses said AI is not relevant to them, and among EU enterprises that considered AI and did not use it, 48.83% cited data protection and privacy concerns. The useful question is not whether to "adopt AI" but where a model should make or support a judgment in a specific workflow, what data it may see, and who checks its output. If you are weighing that for your own work, our AI and automation service starts from one real workflow, uses a model only where a step needs judgment and keeps a person approving what matters; sometimes the answer is a feature you already pay for.

The responsibilities that come with AI

A system that learns from data inherits the data's gaps, can repeat what it memorized and can be steered by its input. These are now computer science problems as much as ethical ones, and they have technical answers.

Bias: errors that fall on some groups more than others

Algorithmic bias means a system makes more errors, or worse decisions, for some groups of people than for others. In 2018, the Gender Shades study tested three commercial gender classification systems on a new benchmark balanced by gender and skin type. Error rates reached 34.7% for darker-skinned women, against a maximum of 0.8% for lighter-skinned men. The two existing benchmarks it examined were 79.6% and 86.2% lighter-skinned subjects; when test data looks like that, a system can score well overall while failing the groups it rarely sees. NIST's framework names three categories of AI bias to manage (systemic, computational and statistical, and human-cognitive) and notes that each can occur without any intent to discriminate. The practical rule: measure error rates for each group a system affects, not only overall accuracy.

Privacy: models remember

Models can reveal what they were trained on. In a 2021 USENIX Security paper, researchers extracted hundreds of verbatim text sequences from GPT-2's training data just by querying the model, including names, phone numbers and email addresses, even when a sequence appeared in only one training document; larger models were more vulnerable than smaller ones. In Canada, the privacy commissioners' principles for generative AI, published on December 7, 2023, apply familiar privacy duties to it, including legal authority and consent, necessity and proportionality, openness, accountability and limiting the collection, use and disclosure of personal information to what the purpose needs.

Safety and security: inputs that steer the model

Language models add a new attack surface. In the OWASP Top 10 for LLM Applications 2025, prompt injection, where text in the input changes what the model does, is LLM01, followed by sensitive information disclosure (LLM02); improper output handling is LLM05. Treat a model's output as untrusted input to the rest of your system, and let anything that changes records pass a rule or a person first.

The NIST AI Risk Management Framework

NIST released the AI Risk Management Framework (AI RMF 1.0) on January 26, 2023, for voluntary use. The framework document describes trustworthy AI by seven characteristics: valid and reliable; safe; secure and resilient; accountable and transparent; explainable and interpretable; privacy-enhanced; and fair with harmful bias managed. Its core organizes the work into four functions:

NIST added a Generative AI Profile (NIST AI 600-1) on July 26, 2024, and on April 7, 2026 released a concept note for a profile on trustworthy AI in critical infrastructure. For organizations with users in the EU, the AI Act entered into force on August 1, 2024 and became applicable on August 2, 2026, with exceptions: its prohibitions and AI literacy duties have applied since February 2, 2025, obligations for general-purpose AI models since August 2, 2025, and rules for high-risk uses in areas such as employment and education apply from December 2, 2027. Those dates follow the AI Omnibus amendment, in force since July 27, 2026. The Commission says the vast majority of AI systems in use in the EU, spam filters among them, are minimal risk, a tier for which the Act adds no rules.

AI is important in computer science for the same reason it needs this care: it is no longer a separate research topic but a component inside compilers, databases, networks, security tools and the interfaces people use every day. More articles on the subject are collected under AI.