AI is changing computer science in three places at once. Programming is becoming the work of specifying, reviewing and testing code that assistants and agents write. Research is turning toward efficiency, reasoning, interpretability and AI that improves computing systems. And the profession is shifting from routine coding toward design, verification and judgment.
This guide takes each change in turn with the evidence available in September 2026: what vendors and independent researchers measured, where they disagree, what curriculum guidelines and US job projections say, and what nobody has solved. The short version: adoption is near-universal, the productivity evidence is mixed, and the part that still needs a person is deciding whether the code is right.
Where AI and computer science stand in 2026
For most of its history, artificial intelligence was one subfield of computer science among many, next to operating systems, databases and networks. It moved from search and logic to statistical machine learning, then to deep neural networks, and then to large language models (LLMs) trained on text and code. That last step changed its place in the field: a model that writes code reaches into the daily work of every other subfield.
The adoption numbers show how far that has gone:
- 84% of respondents to Stack Overflow's 2025 Developer Survey use or plan to use AI tools in their development process, up from 76% in 2024 and 70% in 2023. 51% of professional developers use them daily.
- 90% of the nearly 5,000 technology professionals in Google's 2025 DORA research report using AI at work.
- More than a quarter of all new code at Google was generated by AI, then reviewed and accepted by engineers, according to Alphabet's CEO in October 2024.
Trust has not kept pace. In the same Stack Overflow survey, more developers distrust the accuracy of AI tools (46%) than trust it (33%), and the most common frustration, cited by 66%, is "AI solutions that are almost right, but not quite." That gap between use and trust runs through everything below.
How programming is changing: from autocomplete to agents
AI help for programming has gone through three stages, and all three are in use today:
| Stage | What it does | Where the developer's effort goes |
|---|---|---|
| Code completion | Suggests the next lines as you type | Reading and accepting suggestions |
| Chat assistant | Explains code, answers questions and drafts functions on request | Asking precise questions, checking answers |
| Coding agent | Takes a task, edits files across a repository, runs commands and tests, proposes a change | Specifying the task, setting limits, reviewing it |
A coding agent works in a loop. It reads the task and the relevant files, edits code, runs the build and the tests in a terminal or sandbox, reads the errors and tries again, until the tests pass or it gives up. The result is a proposed change, usually a pull request, that a developer reviews like any other. Agents reach tools and data through interfaces such as the Model Context Protocol (MCP), an open standard Anthropic published in November 2024 for connecting AI assistants to content repositories, business tools and development environments.
Agents caught on among early adopters: METR, a nonprofit that evaluates AI systems, reports that open-source developers took up agentic tools such as Claude Code and Codex throughout 2025. Across all developers they are not yet the norm. In Stack Overflow's 2025 survey, about 14% of respondents used AI agents at work daily, while 52% either did not use agents or stuck to simpler AI tools.
How long a task can an agent finish?
METR tracks agent capability with a measure it calls the time horizon: the length of task, timed on skilled humans, that an agent completes with 50% reliability. In March 2025 it reported that frontier models such as Claude 3.7 Sonnet had a time horizon of about 50 minutes on its software and research tasks, and that the horizon had doubled roughly every seven months since 2019. With a larger task suite in January 2026, METR measured faster doubling since 2024, about every 89 days, and put Claude Opus 4.5 at around 320 minutes.
Read those numbers with METR's own caveats: the tasks are self-contained and well specified, closer to what a new contractor without context could do than to work in a large, familiar codebase; 50% success is far from production reliability; and METR says measurements above 16 hours are unreliable with its current task suite.
What controlled studies measured about developer productivity
Adoption is not productivity, and the careful studies disagree more than the marketing does. In randomized experiments, developers finished a small JavaScript task 55.8% faster with an assistant (Peng et al.), 96 Google engineers took about 21% less time on an enterprise task, and experienced open-source developers working in their own mature projects took 19% longer with AI allowed in METR's early-2025 trial, while believing it had sped them up. METR's 2026 follow-up produced estimates it calls only weak evidence, partly because developers increasingly refused to work without AI. At team level, DORA's 2024 report linked AI adoption to lower delivery throughput and stability, and its 2025 report to higher throughput but still lower stability: AI amplifies what a team already is.
Each study's design, why the results spread so widely and what they mean for the job are set out in our guide to the future role of software engineers, which owns that evidence. For computer science, the lesson is methodological: the tasks, the codebases and the developers chosen decide the answer.
Note
The only productivity number that applies to your team is one measured on your codebase, your tasks and your delivery pipeline, before and after, counting rework and incidents as well as speed.
Testing, code review and the security of generated code
If models write more of the code, more of engineering becomes checking it. Research on generated code agrees on what to check:
- Vulnerable suggestions. In a study presented at the IEEE Symposium on Security and Privacy in 2022, researchers prompted GitHub Copilot with 89 scenarios built around high-risk weakness types and found about 40% of the 1,689 generated programs vulnerable.
- Overconfident users. In a user study published at ACM CCS 2023, participants with an AI assistant wrote significantly less secure code than those without one, and were more likely to believe their code was secure.
- Correct but exploitable. The BaxBench benchmark (2025) asked models to build 392 backend applications. Even the best model tested, OpenAI o1, produced correct code only 62% of the time, and exploits succeeded against about half of the correct programs each model wrote.
- Invented dependencies. A USENIX Security 2025 study found code models recommending packages that do not exist (package hallucination), on average at least 5.2% of the time for commercial models and 21.7% for open-source ones. Attackers can register those names.
Agents add a new class of attack. An agent that reads web pages, issues and files can be steered by instructions planted in that content, which OWASP calls indirect prompt injection and ranks first (LLM01) in its 2025 Top 10 for LLM applications. OWASP says it is unclear whether any method prevents prompt injection completely, so the defences limit what a hijacked agent can do: least-privilege access, human approval for high-risk actions, untrusted content kept separate and marked, and regular adversarial testing.
In practice, that becomes a short list of engineering habits:
- Write or review the tests first, so the agent works against a specification you trust.
- Review generated changes like a new contributor's pull request, with extra attention to authorization, input handling and error paths.
- Run static analysis, dependency and secret scanning on every change, whoever or whatever wrote it.
- Check that every new dependency exists, is maintained and is the one you meant.
- Run agents in a sandbox with scoped credentials, and require approval before anything is merged, deployed or sent.
Our secure coding checklist covers the review practices in detail.
How researchers measure AI's coding ability, and why the tests wear out
Benchmarks are how the field tracks progress, and coding benchmarks have had short lives. SWE-bench, presented at ICLR 2024, asks a model to resolve 2,294 real GitHub issues from 12 Python repositories; when it was published, the best model, Claude 2, solved 1.96% of them. In August 2024 OpenAI and the SWE-bench authors released SWE-bench Verified, 500 tasks screened by professional developers, because some original tasks were hard or impossible to solve.
Scores on Verified rose fast, then stopped meaning much. In February 2026 OpenAI reported that the state of the art had moved only from 74.9% to 80.9% in six months, that at least 59.4% of an audited set of hard problems had tests that reject correct solutions, and that every frontier model it tested could reproduce original fixes or problem details for some tasks, a sign they had seen them in training. OpenAI stopped reporting the benchmark and recommended SWE-Bench Pro, which draws 1,865 problems from 41 repositories and keeps those from 30 of them out of public view, so they cannot leak into training data.
Evaluation has become a research problem in its own right: benchmarks built from public code end up in training data (contamination), and a fixed test saturates while models keep improving. Expect more private, refreshed and time-based evaluations, like METR's time horizon, and treat any single leaderboard number with suspicion.
Research directions that will shape the next decade
Four directions account for much of the work, and each is as much a computer science problem as an AI one.
Efficiency: more capability for less compute
Training and running large models is expensive, and the cost shows up on the power grid. A US Department of Energy report published in December 2024 estimated that data centers used about 4.4% of US electricity in 2023 and could use 6.7% to 12% by 2028.
Research attacks the cost from several sides. The Chinchilla study (Hoffmann and colleagues, 2022) showed that many models were undertrained: a 70-billion-parameter model trained on four times more data beat the 280-billion-parameter Gopher on the same compute budget. Distillation moves a large model's abilities into smaller ones; DeepSeek reported in 2025 that the reasoning patterns of its large R1 models can be used to improve smaller models. Quantization, sparse mixture-of-experts architectures and better compilers and kernels cut the cost of each answer.
Reasoning: from fluent answers to checked logic
Reasoning models spend more computation while answering. OpenAI's o1, announced in September 2024, was trained with reinforcement learning to work through a chain of thought, and OpenAI reported that its performance improves with more training and with more time spent thinking, called test-time compute. DeepSeek-R1, published in January 2025, showed that behaviour such as self-reflection and verification can emerge from reinforcement learning without human-written reasoning examples.
The results reached competition level in 2025. At the International Mathematical Olympiad in July, an advanced Gemini Deep Think model solved five of six problems for 35 of 42 points, a gold-medal standard, working in natural language within the 4.5-hour limit. A year earlier, DeepMind's AlphaProof and AlphaGeometry 2 had reached silver-medal standard, but experts first had to translate the problems into the formal language Lean. In September 2025, Gemini 2.5 Deep Think solved 10 of 12 problems at the ICPC World Finals, the university programming contest, a result DeepMind says would have ranked second among the 139 university teams. (Our explainer on Google Gemini covers the model family.)
Logical reasoning is where classic computer science comes back in. Language models produce answers without guarantees; logic tools such as proof assistants, SAT and SMT solvers and model checkers give guarantees but need precise input. The CS2023 curriculum authors note that generative models lack foundational guarantees of correctness and point to combining symbolic and subsymbolic methods as a promising fix. A model that writes a proof and a checker that verifies it is one such combination, which is why formal verification matters more, not less. The same pattern applies to code: a model proposes, and a type checker, a test suite or a verifier decides.
Interpretability: reading what a model computes
Mechanistic interpretability tries to explain a model's output from its internal computation, not from what the model says about itself. In May 2024 Anthropic used sparse autoencoders to extract interpretable features, internal patterns that correspond to concepts, from Claude 3 Sonnet. In March 2025 it traced circuits linking those features and found, for example, that the model plans a rhyme before writing the line that ends in it, and that its account of how it does arithmetic differs from what it actually computes.
The limits are real. Anthropic says its method captures only a fraction of the computation, even on short prompts, and that understanding the circuits for a prompt of a few dozen words takes a few hours of human effort. Making that scale is one of the field's open problems.
AI for systems and algorithms
A more direct change is AI improving the algorithms and systems computer science is built on. AlphaDev, described in Nature in June 2023, used reinforcement learning to find sorting routines that are now in the LLVM libc++ standard library, up to 70% faster for short sequences and about 1.7% faster above 250,000 elements.
AlphaEvolve, announced by Google DeepMind in May 2025, pairs Gemini models with automated evaluators in an evolutionary loop: the model proposes programs, the evaluators run and score them, and the best become the starting point for the next round.
It found a scheduling heuristic for Borg, the system that orchestrates Google's data centers, that has run in production for over a year and recovers on average 0.7% of Google's worldwide compute resources. It also found a way to multiply 4×4 complex-valued matrices with 48 scalar multiplications, improving on Strassen's 1969 algorithm in that setting. Both results depended on the same condition: a problem where a program can measure every attempt. Compilers, schedulers, databases and hardware design are full of such problems.
What AI means for computer science education
The official answer so far: keep the foundations, change how they are taught. CS2023, the curriculum guidelines from ACM, the IEEE Computer Society and AAAI, endorsed by the three bodies in January and February 2024, organizes the discipline into 17 knowledge areas, including artificial intelligence, security, and society, ethics and the profession. Its chapter on generative AI makes four points:
- The skills expected of students are changing; the foundational knowledge they need is not.
- Tasks that are intellectually challenging but mechanical will be automated, so students must learn to work at higher levels of abstraction.
- Learning to program will shift from writing code and memorizing syntax toward prompting, reading, verifying, modifying and testing it, with more weight on problem decomposition. Students still need to write programs, if only to verify what a model produced.
- Society, ethics and professional issues matter more, and must be taught actively and revisited as the technology changes.
For a student, that says where to spend effort: algorithms and complexity (CS2023 expects students to judge a generated solution's time and space trade-offs), systems and networking (to know what code does when it runs), security, testing, and the habit of reading code critically. None of that is new; what changes is its weight, because it is the knowledge that checks what a model produced. For choosing between a computer science and a cybersecurity program, and checking its accreditation, see computer science and cybersecurity programs.
Careers: what the job data says
The US Bureau of Labor Statistics' Occupational Outlook Handbook, updated August 27, 2026, projects employment for 2025 to 2035. For the occupations closest to computer science:
| Occupation | Projected change, 2025–35 | Median pay, 2025 |
|---|---|---|
| Computer and information research scientists | Up 22% | $140,300 |
| Software developers | Up 10% | $135,980 |
| Software quality assurance analysts, testers | Up 6% | $104,300 |
| Computer programmers | Down 7% | $100,390 |
| All occupations | Up 3% |
BLS expects programming work to keep being automated, while demand for developers comes from software for AI, the Internet of Things, robotics and automation, and research scientists will be needed to create new AI technologies. Payroll data from the Stanford Digital Economy Lab, through June 2026, puts employment of 22- to 25-year-olds in AI-exposed occupations 19% below trend, mostly through less hiring, and the authors present it as an early indicator rather than a causal estimate. What this means for a software career, and the skills that now matter, is covered in the future role of software engineers. These are US figures; other countries publish their own outlooks.
The open problems
| Problem | Why it is hard | What to watch |
|---|---|---|
| Reliability | Models produce plausible wrong answers; 46% of surveyed developers distrust their accuracy | Verification tools that check output, not just generate it |
| Evaluation | Public benchmarks leak into training data and saturate | Private, refreshed test sets such as SWE-Bench Pro's held-out problems |
| Agent security | OWASP says it is unclear whether prompt injection can be fully prevented | Least privilege, sandboxes and approval gates as defaults |
| Measuring productivity | More developers decline to work without AI, so control groups shrink | Observational data and fixed-task experiments |
| Energy and cost | US data centers could use 6.7% to 12% of US electricity by 2028 | Efficiency research and smaller, distilled models |
| Interpretability | Current tools explain a fraction of a model's computation | Methods that scale beyond short prompts |
| Accountability | Someone has to own the risk of an AI decision | NIST's AI Risk Management Framework (January 2023) and its Generative AI Profile (July 2024) |
What to do now
- Developers: use assistants and agents, and measure what they change on your own work. Invest in the skills a model cannot supply: reading code critically, designing systems, writing tests that express intent, and security review.
- Teams: strengthen the pipeline before adding more AI output to it. DORA's data says AI amplifies whatever testing, review and feedback you already have. Give agents the least access they need and a human approval step before anything ships.
- Students and educators: keep the foundations CS2023 lists, and practise the new loop: specify, generate, test, review.
If you are deciding where AI belongs in your own workflows or products, our AI and automation service builds workflows, assistants and agents that use AI where a step needs judgment, with a person approving what matters, and runs an evaluation set from your real cases before every change. For the software around it, our custom software development team puts every change through the same gates, including secret, code, dependency and container scanning.
