
AI-Assisted Coding and Creativity: What Changes for Engineers?
AI-assisted coding means building software with an AI model that suggests, writes or reviews code while the engineer stays accountable for the result. In Stack Overflow’s 2026 Developer Survey, 65.9% of respondents reported using AI coding assistants or coding agents at work. What changes most is not typing speed but judgment around the code.
What is AI-assisted coding, and how is it different from vibe coding?
AI-assisted coding is software development in which an AI model completes, generates, explains or reviews code, and the engineer decides what ships. Vibe coding, a term coined in February 2025, means accepting generated code without really reading it. The difference is not the tool but who understands and owns the code.
Three levels of assistance: completion, chat and agents
AI-assisted coding covers different levels of autonomy.
| Level | What AI does | What the engineer does | Main risk |
| Code completion | Suggests lines or code blocks | Accepts, edits or rejects suggestions | Small errors pass unnoticed |
| Chat and generation | Produces functions, tests and explanations | Specifies, reviews and integrates changes | Plausible but incorrect code |
| Agentic coding | Coordinates changes across files and may run tests | Sets boundaries and approves results | Large changes become difficult to review |
Editorial comparison
The distinction matters because an agent that changes several files introduces different review demands from an autocomplete suggestion. See also how AI agents operate in business workflows.
Vibe coding: where the term comes from and where it stops
Andrej Karpathy popularized vibe coding in February 2025 as an approach in which developers largely follow AI-generated output without closely examining the underlying code.
That can be useful for disposable prototypes and experimentation. It is a poor default for production systems that require security, maintainability and accountability.
AI-assisted coding does not require surrendering understanding. The engineer remains responsible for explaining what the code does and why it belongs in the system.
Does AI-assisted coding make engineers faster?
It depends on the task and the engineer. In a 2023 controlled experiment, developers with GitHub Copilot built a small HTTP server 55.8% faster. In a 2025 METR trial, experienced developers working on their own large codebases took 19% longer with AI, while believing it had sped them up. Speed gains are real but conditional.
Why a bounded task and a mature codebase give opposite answers
Three controlled studies illustrate why a single productivity figure is misleading.
| Study | Participants and task | Finding | Important limitation |
| GitHub Copilot, 2023 | Developers building a JavaScript HTTP server | 55.8% faster with AI | Narrow, relatively self-contained task |
| METR, 2025 | 16 experienced developers completing 246 tasks in mature repositories | 19% longer with AI | Early-2025 tools and specific open-source projects |
| Anthropic, 2026 | 52 mostly junior engineers learning a Python library | About two minutes faster, not statistically significant | Learning task, not general production delivery |
Editorial synthesis of three primary studies
These results are not contradictory. A small, clearly specified task differs substantially from modifying a mature system with established conventions, dependencies and hidden context.
There is also a crucial update: in February 2026, METR reported preliminary follow-up results suggesting newer AI tools may provide greater speedups. However, the researchers said measurement and selection problems made those findings weak evidence for the size of any improvement. The 2025 slowdown should therefore be treated as a historical result, not a claim about every AI tool in late 2026.
The perception gap: feeling faster is not the same as being faster
In METR’s 2025 experiment, developers expected AI to reduce task time by 24%. After completing the work, they still estimated a speedup, despite measured completion times being longer.
This illustrates why perceived productivity is not enough.
Teams should evaluate actual task completion, rework and quality rather than relying solely on developer satisfaction or the number of AI suggestions accepted.
Which parts of an engineer’s work change, and which stay human?
AI shifts effort from producing code to defining, checking and integrating it. Generating boilerplate, tests, refactors and explanations gets cheaper. Framing the problem, choosing trade-offs, reviewing for correctness and owning production behaviour do not. The engineer’s day moves upstream to specification and downstream to verification, with less time in between.
From writing code to specifying, checking and integrating it
| Engineering activity | Where AI can help | What remains the engineer’s responsibility |
| Problem framing | Suggest questions and approaches | Define the actual business problem |
| Requirements | Draft specifications and acceptance criteria | Resolve ambiguity with stakeholders |
| Architecture | Propose design patterns | Choose boundaries and trade-offs |
| Implementation | Generate functions and routine code | Integrate with existing systems |
| Testing | Suggest cases and test code | Decide what must be verified |
| Code review | Highlight potential defects | Approve correctness and maintainability |
| Debugging | Generate hypotheses | Identify and understand root causes |
| Operations | Summarize logs and incidents | Make production decisions |
Editorial task map
The role is shifting from producing every line to ensuring that the resulting software behaves as intended.
That does not eliminate coding expertise. Understanding code becomes particularly important when generated output is wrong.
Problem framing: the work that happens before the prompt
Before asking AI to implement a feature, engineers should answer four questions:
- What problem are we solving, and for whom?
- What is explicitly outside the scope?
- Which security, performance, cost and operational constraints matter?
- How will we know the solution works?
A detailed prompt cannot compensate for an incorrect problem definition.
Where does creativity live in engineering when AI writes the code?
Two engineering teams can use the same AI model and produce very different systems.
One might optimize for rapid delivery. Another might prioritize reliability, extensibility or lower operational complexity. Those differences originate in human decisions about the problem and its constraints.
Creativity also emerges during implementation, debugging and refinement. Developers may find an elegant abstraction, reuse an existing component in an unexpected way, simplify a complex algorithm or devise a new strategy for solving a difficult technical problem.
Research supports this broader view. A 2023 IEEE Software publication on creativity in software engineering highlighted clever reuse, rather than pure novelty, as an important part of how developers understand creativity. An empirical study presented at ICSE 2024, involving 77 participants, also found associations between creativity scores and different problem-solving strategies during programming tasks.
Creativity also depends on how teams challenge ideas, combine perspectives and experiment. These practices are explored further in Webellian’s article on creativity in agile development teams.
Creativity in software is framing and trade-offs, not keystrokes
Two engineering teams can use the same AI model and produce very different systems.
One might optimize for rapid delivery. Another might prioritize reliability, extensibility or lower operational complexity. Those differences originate in human decisions about the problem and its constraints.
Creativity also depends on how teams challenge ideas, combine perspectives and experiment. These practices are explored further in Webellian’s article on creativity in agile development teams.
The sameness risk: why AI suggestions pull solutions toward the average
A 2024 Science Advances experiment by Doshi and Hauser studied AI-assisted short-story writing. AI improved evaluated creativity, particularly for less creative participants, but AI-assisted stories were more similar to one another.
Science Advances
This was a writing study, not a software engineering experiment. Applying its findings to engineering is an editorial inference.
Still, it suggests a useful precaution: teams should avoid treating the first AI-generated design as the obvious solution.
Three practical habits help:
- Sketch an initial approach before consulting AI.
- Ask for alternative architectures with different trade-offs.
- Use design reviews to question why competing approaches were rejected.
Why does design and architecture judgment matter more with AI?
When code is cheap to generate, the expensive mistakes move to design: wrong boundaries, hidden coupling, unclear ownership of data and behaviour that nobody can change safely later. An assistant optimises the snippet it sees. Architecture judgment weighs the whole system, its constraints and the people who will maintain it.
Five engineering decisions that require human ownership and accountability
| Decision | What the team must evaluate |
| Module and service boundaries | Which teams will own and change each component |
| Data models | Business meaning, future requirements and compliance |
| External dependencies | Long-term vendor, security and licensing risk |
| Non-functional requirements | Performance, reliability and cost priorities |
| Architectural consistency | Fit with the existing system and technical strategy |
Editorial decision framework.
AI can generate plausible solutions to each problem, but it cannot independently establish the organization’s priorities.
DORA’s 2025 research described AI as an amplifier of existing organizational strengths and weaknesses. A team with weak architecture practices may simply generate problematic code faster.
State of AI-assisted Software Development 2025
The engineering challenge is to maintain coherent design while the cost of generating implementations decreases.
How should code review change when AI writes part of the code?
Review AI-written code as code from a fast, confident contributor who does not know your system. Stack Overflow’s 2025 survey found 66% of developers struggle with AI output that is almost right, and 45% say debugging it takes longer. Reviews should check intent, edge cases, security and fit, with tests proving behaviour.
Five questions to ask of AI-written code before approving it
AI-assisted coding increases the importance of understanding a change before merging it.
Reviewers should ask:
- Correctness: Does the change solve the requested problem without introducing unrelated behavior?
- Edge cases: Does it handle invalid input, concurrency, failures and boundary conditions?
- Security: Does it introduce secrets, unsafe dependencies or access-control weaknesses?
- Architecture: Does it follow the system’s design and coding conventions?
- Understanding: Can the author explain the implementation and its trade-offs?
The author of a pull request remains accountable for the entire change, including generated code.
This matters because AI output can appear polished even when its assumptions are incorrect. In the 2025 Stack Overflow survey, 46% of respondents distrusted the accuracy of AI output, compared with 33% who trusted it.
2025 Stack Overflow Developer Survey
Tests as the contract between the engineer and the assistant
Tests should define expected behavior rather than merely confirm that generated code executes.
A useful approach is to establish acceptance criteria before implementation, generate candidate tests, review them and then validate the implementation.
AI-generated tests require the same scrutiny as AI-generated application code. Otherwise, a model may produce tests that reinforce its own incorrect assumptions.
The objective is not to reject AI-generated code. It is to make correctness demonstrable.
How does AI-assisted coding affect how engineers build skills?
Used as a shortcut, AI can slow learning. In Anthropic’s 2026 randomized trial, developers learning a new Python library with AI scored 17 percentage points lower on a follow-up quiz, with the biggest gap in debugging. A separate exploratory analysis found that explanation-focused and conceptual-question interactions were associated with higher scores among AI users, but did not establish that these interaction styles caused better learning outcomes.
What a 2026 randomized trial found about learning with AI
Anthropic’s January 2026 study involved 52 mostly junior engineers learning the Python library Trio.
Participants using AI scored an average of 50% on a subsequent knowledge test, compared with 67% among those working without AI. The largest difference concerned debugging. The AI group’s small completion-time advantage was not statistically significant.
The researchers also identified different interaction patterns.
| Interaction pattern | Typical behavior | Learning implication |
| Full delegation | Ask AI to produce the implementation | Less opportunity to develop understanding |
| Progressive reliance | Begin independently, then delegate more work | Risk of losing engagement with the problem |
| Explanation-focused assistance | Ask why code works and how concepts connect | Better preservation of learning |
| Independent work with consultation | Write code while asking targeted questions | Supports active problem-solving |
Editorial synthesis of the study’s qualitative findings, not a separate experiment.
The distinction is important: AI assistance itself did not guarantee poor learning. The manner of interaction mattered.
Juniors, seniors and the apprenticeship problem
Junior engineers traditionally develop judgment through debugging, writing routine code and understanding existing implementations.
If AI handles these tasks before juniors learn the underlying concepts, they may have fewer opportunities to develop the skills needed to review generated output.
Engineering teams can address this through deliberate practice: some AI-free learning exercises, code reviews that require explanations and pairing between junior and senior developers.
This also makes onboarding and knowledge sharing more important when scaling an engineering team.
Will AI replace software engineers?
The evidence reviewed here shows substantial changes in software engineering tasks, but it does not support a reliable prediction about how many software engineering roles AI will replace. Trials show large gains on narrow tasks and slowdowns on complex real ones, while many developers do not fully trust AI output. AI is already changing how code is written, reviewed and integrated, but its long-term impact on engineering employment remains uncertain.
What the evidence supports, and what it does not
The research supports several conclusions:
- AI can accelerate certain software development tasks.
- Productivity gains vary with task complexity, context, experience and model capabilities.
- Verification, architectural judgment and debugging remain important engineering responsibilities.
- Heavy reliance on AI can interfere with learning, particularly for less experienced developers.
It does not establish that software engineering roles will disappear by a specific year, or that every organization will achieve the same productivity improvement.
AI also makes software creation more accessible to non-engineers. As business teams develop applications with AI and low-code tools, professional engineers still have a role in integration, security, architecture and production reliability.
The more useful question is not whether engineers will stop writing code, but which capabilities will make them effective when code generation becomes easier.
What should engineering leaders change in their teams?
Leaders should treat AI-assisted coding as a change in how the team works, not a tool rollout. DORA’s 2025 research describes AI as an amplifier of existing strengths and weaknesses. That means agreeing how AI is used, keeping engineers accountable for what ships, protecting learning, and measuring outcomes at team level with delivery metrics.
Six working agreements for AI-assisted teams
Engineering leaders can establish six practical rules:
- Use approved tools. Define which services are permitted and what data may be shared.
- Keep ownership clear. The pull request author is responsible for all code, regardless of how it was produced.
- Require verification. Generated changes must meet the same review and testing standards as manually written code.
- Make substantial AI use visible. Agree when contributors should disclose AI-generated sections in pull requests.
- Protect learning time. Include independent coding and debugging practice, especially during onboarding.
- Review the rules regularly. Adjust team practices using observed outcomes, not assumptions about AI productivity.
These agreements are a starting point, not a universal policy. Teams should adapt them to their systems, security requirements and engineering maturity.
Measure the change at team level, not keystroke level
Counting generated lines or accepted suggestions tells leaders little about whether software delivery has improved.
More useful indicators combine DORA’s five software delivery performance metrics – change lead time, deployment frequency, change fail rate, failed deployment recovery time and deployment rework rate – with complementary measures such as code quality, review effort, developer experience and relevant AI-specific indicators. AI usage or code acceptance rates can provide additional context, but they do not establish productivity gains on their own.
These should be interpreted together. Faster delivery accompanied by more incidents is not necessarily a productivity gain.
The same principle applies to distributed engineering teams: measurement should support learning and accountability, not individual blame.
DORA’s 2025 research reinforces this systems perspective. AI adoption creates the most value when supported by effective workflows, technical practices and organizational capabilities, rather than tools alone.
For engineering leaders, the priority is clear: make AI-assisted coding a governed, measurable part of software delivery while preserving the skills required to understand and maintain the result.
FAQ
What is agentic coding?
Agentic coding is a form of AI-assisted development in which an AI agent can plan and execute multistep tasks, such as editing several files, running tests and revising code. Engineers define the task, set permissions and review the outcome. Greater autonomy can save manual effort, but it also increases the amount of work that requires verification.
Is AI pair programming the same as AI-assisted coding?
AI pair programming is one form of AI-assisted coding. It usually describes an engineer working interactively with an AI assistant to generate, explain or refine code. AI-assisted coding is broader and includes code completion, test generation, review and autonomous agents. In every case, the engineer remains responsible for the software.
Is AI-assisted coding good for beginners?
It can be, provided it supports understanding rather than replacing practice. Anthropic’s 2026 experiment found weaker learning outcomes when participants relied heavily on AI to complete an unfamiliar coding task. Participants who asked conceptual questions and requested explanations generally preserved more learning. Beginners should combine AI assistance with independent coding, debugging and review.
Sources
https://www.anthropic.com/research/AI-assistance-coding-skills
https://survey.stackoverflow.co/2026/ai/data/ai-select
https://survey.stackoverflow.co/2026/ai/data/ai-trust
https://metr.org/blog/2026-02-24-uplift-update
https://arxiv.org/abs/2302.06590
https://dora.dev/research/2025/dora-report
https://research.google/pubs/dora-2025-state-of-ai-assisted-software-development-report
https://doi.org/10.1126/sciadv.adn5290
https://research.google/pubs/creativity-in-software-engineering