If this year’s data showed developers using AI tools consistently while trusting them conditionally, this section digs into the “conditionally” of it all. Where does that trust actually come from, and what gives an answer credibility in the first place?
Source attribution
As AI-generated content becomes harder to distinguish from human-produced work, and as governance conversations gain more traction across the industry, source attribution has become a bigger priority for technologists. This year, 93% of survey respondents called source attribution at least somewhat important when deciding whether to trust AI-generated answers, with 79% judging it important or very important.
AI models, of course, are non-deterministic by design. That variability is part of what makes them useful, but it also introduces the potential for errors and hallucinations. Developers have clearly internalized this, because the overwhelming majority say that knowing where an answer came from matters a lot, not a little, when deciding whether to trust it.
For AI responses, people want proof: 93% of respondents feel the source of the information is at least somewhat important.
View and share | Respondents | Percent | |
|---|---|---|
| Very important | 7,269 | 52.1% |
| Important | 3,701 | 26.5% |
| Somewhat important | 1,956 | 14.0% |
| Not very important | 753 | 5.4% |
| Least important | 266 | 1.9% |
Context that matters most
That trust in sourcing extends naturally into a question of inputs: what context actually helps AI get things right in the first place? Respondents ranked project goals or requirements (85%) and the underlying code and docs (73%) as by far the most important context they need their AI to consider.
But if someone has a specific question, then the sources of context shift slightly. The most trusted source of information is still the people you work with (72%), followed by the code (63%), and documentation (60%). In other words, you first go to the people, then the code, then what people wrote about the code.
Project goals or requirements (85%) and code, repositories, or technical documentation (73%) are the top most important sources for context at work
The top two sources of AI context are overwhelmingly project goals or requirements (85%) and the code and docs themselves (73%). The work itself is the most important context.
View and share | Respondents | Percent | |
|---|---|---|
| Project goals or requirements | 10,806 | 85.2% |
| Code, repositories, or technical documentation | 9,297 | 73.3% |
| Prior decisions and rationale | 5,580 | 44.0% |
| Related tickets, tasks, or issues | 4,679 | 36.9% |
| Business goals or product strategy | 4,189 | 33.0% |
| Pull requests, commit history, or release notes | 3,332 | 26.3% |
| Customer or user feedback | 3,296 | 26.0% |
| Team norms, preferences, or ways of working | 3,183 | 25.1% |
| Compliance, security, or policy requirements | 2,498 | 19.7% |
| Current status or ownership | 2,490 | 19.6% |
| Past discussions in chat, email, or meetings | 2,000 | 15.8% |
| Data, dashboards, or reports | 932 | 7.4% |
| I do not usually need additional context | 266 | 2.1% |
| Other | 79 | 0.6% |
Tokenomics vs. quality
Are developers changing their behaviors as they start seeing the costs of high-performing models? When asked about whether they changed their prompts to improve quality or to reduce costs, overwhelmingly respondents picked quality (75%) over cost (38%), probably because they’re the folks actually building software, not managers who have to report ROI to the board. A small portion (16%) said their prompts are perfect as they are, thanks, and aren’t changing them for either reason.
That doesn’t mean organizations have wired their AI models directly to their bank accounts. A majority of respondents have some sort of organization cost controls or monitoring: 38% set limits on usage, 28% reviewed spending data, 27% use different models for different work, and 25% set limits on who gets to use AI. 21% of respondents said their organizations saw the results from AI and started spending more.
| Respondents | Percent | |
|---|---|---|
| Yes, I personally monitor my own use | 337 | 32.6% |
| Yes, I choose cheaper models when possible | 261 | 25.3% |
| Yes, usage is metered to manage costs | 246 | 23.8% |
| No, cost/limits have not affected my usage | 231 | 22.4% |
| I have not encountered cost or usage limits | 213 | 20.6% |
| Yes, I am more selective about tasks | 171 | 16.6% |
| My company/team has increased budget for AI | 132 | 12.8% |
| Not sure | 117 | 11.3% |
| Yes, my company/team has switched models/vendors that are more affordable | 71 | 6.9% |
| Yes, I choose faster models even when they cost more | 51 | 4.9% |
Many have changed how they prompt in order to reduce cost or improve quality
Quality optimization outpaces cost-cutting 2:1
View and share | Respondents | Percent | |
|---|---|---|
| Yes, to improve output quality | 422 | 75.4% |
| Yes, to reduce cost | 215 | 38.4% |
| No | 88 | 15.7% |
When it comes to their own AI usage, individual respondents took a generally cost-conscious approach: 33% personally monitor their usage, 25% choose cheaper models when possible, and 24% say their usage is metered. Not everyone is so constrained by budgets: 22% are unaffected by costs or limits to their AI use while 21% have yet to see limits and probably have an intimidating yellow aura.
Explore Knowledge data
See the full breakdown