AI coding tools have made it faster to produce working code for many tasks. The evidence does not show that deciding which problem is worth solving has become easier at the same pace. The title’s comparison is a useful argument to test, not a measured finding. What the reporting does support is narrower: code generation has changed part of the job, and developers and researchers are now asking about the rest of it, including quality, provenance, authority, and security.
What “solved” can and cannot mean here
Calling the how-to-code problem solved overstates things. Writing reliable software still requires understanding requirements, designing interfaces, handling edge cases, testing, and debugging code that a tool produced but nobody fully read. What AI tools have changed is the cost of getting a first draft of an implementation, and for some routine tasks, the cost of getting a usable one.
As an Amazon Associate I earn from qualifying purchases.
That is a meaningful change, but it is a change in one stage of software development. The phrase “what to build” points at a different stage: deciding whether a product, feature, or tool should exist at all, for whom, and at what cost to maintain it. Those decisions were never primarily about typing speed, which is why faster implementation does not automatically improve them.
Implementation speed and problem selection are different decisions
Implementation asks whether a chosen approach can be built. Problem selection asks whether the approach solves a need that people actually have, and whether it is worth the ongoing work of keeping it running. A team can produce a polished prototype in an afternoon and still have no evidence that anyone will use it.
Faster implementation changes the economics of trying ideas. Cheap prototypes mean more ideas get built, and that can be useful. It can also mean more weakly justified software reaches users, and more maintenance obligations pile up before anyone has checked demand. The bottleneck moves from “can we build it?” to “should we have built it?”
The title’s claim that selection is harder than coding is an argument, not something the available evidence measures directly. No cited study compares the difficulty of choosing projects with the difficulty of implementing them. Readers should treat the claim as a hypothesis that their own experience can test.
What the productivity figures actually measure
Productivity claims about AI coding tools are often quoted without their source, year, or population. The figures below are the ones relevant to this argument, each with the qualifier that comes with it.
Free tools Windows power users keep installed
One-click scans. No signup required.
| Figure | Publisher and year | Population or scope | What it does and does not show |
|---|---|---|---|
| Up to 55% productivity increase among developers who use GitHub Copilot | GitHub, reported in 2024; article updated April 15, 2025 | GitHub’s summary of its own earlier research on Copilot users | A vendor-reported result for one tool. It is not a general rate for all developers, teams, or AI tools. |
| 81% expected AI coding tools to increase collaboration within teams | GitHub, 2023 survey | Survey respondents | Expectations, not measured collaboration outcomes. |
| 87% said Copilot helped preserve mental effort during repetitive tasks | GitHub, 2023 survey | Survey respondents | Reported experience, not an objective measure of effort or output. |
| Qualitative interviews with 25 developers | GitHub, 2024 article | 25 interviewees | Illustrative accounts, not a population statistic. |
| More than 100 hours of qualitative data and survey responses from nearly 5,000 technology professionals | DORA, 2025 State of AI-assisted Software Development report | Technology professionals in the report’s sample | Describes how AI interacts with existing team practices; it does not rank project choices. |
The pattern across these figures is consistent: they describe what tools do for implementation work and how people feel about that work. None of them answers whether the thing being implemented was worth choosing.
What developers ask for beyond code generation
Research on developer needs has moved past asking whether AI can write code. The more useful questions concern what developers need to trust, supervise, and govern AI support inside their work.
DORA’s “amplifier” framing
The DORA 2025 report states: “The research reveals a critical truth: AI’s primary role in software development is that of an amplifier.” The point is that AI magnifies whatever a team already does. A team with clear priorities and fast feedback may get more out of it; a team with unclear goals may simply produce more unclear work, faster.
Rank #3
That framing matters for project selection. If AI amplifies existing practice, then the quality of the decision about what to build is part of what gets amplified. An unexamined idea that is implemented quickly is still an unexamined idea.
Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Scan for outdated or missing drivers - takes under a minuteDriver Scan →Microsoft’s survey of developer desires and concerns
Microsoft Research’s 2024 study, Towards Effective AI Support for Developers: A Survey of Desires and Concerns, surveyed 791 Microsoft developers. Its value for this argument is the scope of what it asks about: developers’ desires and concerns, not only their preference for faster generation. The study describes a population of Microsoft developers, so it should not be read as a measure of prevalence across all developers.
Requirements for AI systems developers want
A 2026 Microsoft Research publication, as summarized on its publication page, covers 22 AI systems developers want across five task categories. The examples highlighted there are revealing because none of them is about writing code:
- Early quality signals, so problems surface before they reach users.
- Explicit authority scoping, so a tool’s permitted actions are clear.
- Provenance, so developers can trace where information or output came from.
- Uncertainty signaling, so a tool indicates when it does not know.
- Least-privilege access, so tools receive only the permissions they need.
Each of these is a question about judgment and control. Answering them requires knowing what the system is for, which again returns to the problem being solved.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.A framework for evaluating a candidate project
The framework below is an editorial decision tool, not a research-backed ranking. Its four criteria come from the practical concerns that the evidence above raises, and they can be applied to any idea before a line of code is written.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Scan for outdated or missing drivers - takes under a minute3Repair Windows errors before they cause bigger problems1. Evidence of user need
- Can you name the people who have the problem, and describe how they handle it today?
- Have any of them paid, switched, or spent time on a workaround?
- Would a prototype change what they do, or only what they say they would do?
2. Feasibility
- Can the core workflow be built and tested with the data, access, and skills you have?
- Which part is hardest, and have you built a small version of it?
3. Maintenance burden
- Who will fix it when dependencies change, data formats shift, or users report errors?
- What recurring cost does it add in hosting, support, and attention?
4. Risk
- What happens if the output is wrong, and how would anyone notice?
- What data, credentials, or permissions does it require, and can those be limited?
- Who is affected if it fails, including people who never chose to use it?
A candidate that scores well on need but poorly on maintenance may still be worth a small experiment. One that scores poorly on need should not be built regardless of how quickly it could be written.
Best Value
What evidence should justify building a solution
A proposed solution deserves to exist when its builder can show, before scaling it, that a specific group of people has a specific problem, that the proposed approach changes their outcome, and that someone will carry its ongoing costs. Generated code can help produce the prototype that tests those claims. It cannot supply the claims.
The fastest way to test them is usually the least glamorous: talk to the people affected, measure what they do without the tool, and compare the result with what happens after it is in use. That comparison, not the speed of the first draft, is what shows whether the thing was worth choosing.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
