Executive summary: The most significant finding is not the gap between junior and senior developers — it is where they converge. Both groups prioritize critical thinking and problem-solving over AI tool proficiency. The competencies the market values most in the AI era are the ones technical education has always taught. The question is how quickly learning environments can incorporate the real-world practice where those competencies are consolidated.
Much has been written about the impact of artificial intelligence on the technology labor market. Far less has been examined about how developers who trained with AI available from day one perceive their own learning and preparation. That is precisely what the Q2 2026 Dev Barometer inaugurates: the first edition in the series to simultaneously survey junior and senior developers about the same phenomenon. And the results produce a tension that is not easy to explain at first glance.
85% of junior developers say AI improved their understanding of software development — Q16, n=958
16% of senior developers confirm that juniors fully understand the AI-generated code they submit — Q8s, n=478
A superficial interpretation of these data might lead to the conclusion that junior developers are deceiving themselves. But the data does not support that hypothesis. When asked directly whether AI is making the junior developer role less relevant, 50.5% of juniors say yes, just three points below the 53.3% of seniors who hold the same view. There is no denial of risk. Juniors have a clear-eyed reading of their position in the market.
The explanation for the paradox runs deeper, and lies in a phenomenon that cognitive psychology has documented with precision. Flavell (1979) established that the ability to calibrate one’s own learning, what he called metacognitive monitoring, depends critically on the existence of a comparative reference point. Without having experienced the alternative, learners lack the internal instruments to detect their own gaps in understanding. A developer who learned to code with AI available from the start never traveled that path without assistance. For them, using AI is constitutive of understanding, not a shortcut around it. When the tool systematically suppresses the experiences of genuine error and difficulty, which Flavell identifies as the most informative metacognitive signals, the learner loses the raw data on which their own monitoring operates.
This becomes visible in another data point from the study. Twenty-four percent of juniors say that writing code from scratch is the task they feel least confident doing without AI. Yet only 5% consider it a critical skill for getting hired today. There is an internal coherence to that response. The market indeed no longer requires writing code without assistance, but it coexists with something the juniors themselves cannot fully gauge: the construction of deep conceptual schemas that later enable evaluative judgment. The literature on cognitive offloading (Risko & Gilbert, 2016) distinguishes between delegating cognitive tasks to free up capacity, which is adaptive, and delegating the construction of understanding itself, which erodes the foundation on which that judgment develops. The central pedagogical challenge of this era is precisely to design learning environments where AI amplifies deep understanding rather than substitutes for it.
The Point of Convergence and What Seniors Say in Their Own Words
With the nature of the paradox established, it is worth pausing on the finding that stands out most in the Q2 2026 survey: not the gap, but the point where the two groups meet. Despite assessing the situation from radically different vantage points, juniors and seniors prioritize the same competencies when asked what matters most in the AI era. And they do so across four questions with different formats, which gives the finding unusual robustness for a survey of this scale.
When junior developers are asked which skill matters most for getting hired today, 48% choose analytical thinking and problem-solving, nearly three times the 18% who choose proficiency in AI tools. Senior developers, for their part, identify the most frequently missing skills in juniors as understanding how systems work end-to-end (23.6%) and breaking down problems independently (23.4%). On the irreplaceability scale over three years, critical thinking and analytical reasoning scores the highest average of all 11 skills evaluated: 4.11/5, ahead of security (4.03), adaptability (3.84), and prompt engineering (3.69). And when asked to write freely, without any predefined options, which skill is most important for a junior developer in the AI era, the dominant cluster is critical thinking and evaluation, with 25% of responses. AI literacy and prompting appears in 18% of responses. This is a meaningful figure, but one that falls below critical thinking alone, and well below the 56% that critical thinking (25%), fundamental knowledge (16%), and problem-solving (15%) account for together. When no options frame the answer, what the market values most is confirmed with clarity.
Four Stances in the Seniors’ Own Language
The analysis of 343 valid responses to the open-ended question Q14s reveals something that multiple-choice numbers cannot show: the convergence on “critical thinking” conceals four distinct stances, each with different logics and underlying assumptions. It is worth distinguishing them.
- A. AI is the context; judgment is the skill (25%) — “Bullshit detector — knowing when to trust AI, when to question it, and when to ignore it entirely”
- B. Fundamentals first, AI second (16%) — “New IT workers don’t understand logic, algorithms, SQL. I learned how a processor works. So many don’t understand what is an LLM”
- C. Problem decomposition is the meta-skill (15%) — “The most important skill is the ability to break down problems independently. It’s the foundation for using AI effectively”
- D. AI literacy is the core skill (18%) — “Today this is not a specific language. It’s knowing how to use AI efficiently (AI literacy + critical thinking)”
Stances A, B, and C share a premise that stance D does not: deep understanding is a prerequisite for evaluative judgment. Together they account for 56% of responses, compared to 18% for stance D. What separates them from each other is whether that understanding can be built with AI or whether it requires having learned before it. Stance B, generationally marked, more conservative, doubts that a junior who has always used AI can reliably evaluate its output. Stance A trusts that they can, provided they develop the appropriate critical judgment. That difference is not resolved in the data and it is the central pedagogical question that the Q2 2026 inaugurates without closing.
The senior who captures it best does so in two words: “bullshit detector.” The colloquial expression describes the skill with precision: critically evaluating an output without assuming it is correct simply because it was generated. It is the operational definition of critical thinking applied to working with AI, and it is consistent with what Flavell would describe as the highest function of metacognitive monitoring: knowing when one’s own process of understanding is failing. Another senior articulates it at greater length:
“The developers who will grow are the ones who use AI to accelerate what they already understand, not to replace the understanding itself. Learn the concepts first, then let AI help you move faster.” — Senior developer, Q14s
This distinction, using AI to accelerate what one already understands, not to replace the understanding, is the core of the argument. Pellegrino and Hilton (2012) and the revised Bloom’s taxonomy (Anderson & Krathwohl, 2001) converge on the same point. Deep transferable competencies like analyzing, evaluating, and creating are those that persist across technological transformations. What the Q2 2026 shows is that the tech market, when speaking without predefined options, responds with the same categories that pedagogical literature has been identifying as durable for decades.
The Gaps Are About Format, Not Content
At this point, it is worth examining another finding in the study with direct implications for those involved in developer training. Thirty-six percent of senior developers say that formal education does not adequately prepare for the AI era; 70.6% of juniors, however, say they feel at least “somewhat prepared” by their training. This divergence does not speak to institutional failure. It speaks to the pace at which the technological environment is changing.
Generative AI transformed the competency profile demanded by the tech market in less than eighteen months. No training system, public or private, in-person or digital, long- or short-form, has mechanisms to respond at that speed. The concept of curriculum lag (Trow, 1973; Dede, 2010) captures this structural dynamic. The gap between what programs can update and the pace at which the professional practice environment changes. In Latin America, that gap has historically given rise to complementary training models such as bootcamps, certifications, or accelerated programs, and the Q2 2026 suggests the challenge remains open even with those complements in place. The issue is not only one of speed, but of the type of learning experience.
The gaps that juniors and seniors identify are nearly identical: working in large or legacy codebases (20.7% of juniors cite it as the biggest surprise upon entering the market; 20.5% of seniors flag it as the biggest gap in the training they received), system design at scale (19.2% / 18.0%), and critical evaluation of AI-generated code (18.5% / 10.3%). None of these gaps require changing the discipline being taught. All of them are applications of knowledge that training programs already provide, but in conditions of scale, ambiguity, and complexity that the classroom can rarely replicate. As Lave and Wenger (1991) argue, genuine professional knowledge is built by participating in real communities of practice, not simulations of them. The gap is one of format and exposure, not of content.
It comes as no surprise, then, that when junior developers are asked where their training should focus more, 49.2% request more real-world project experience — the most chosen response by a wide margin — and 20% ask for more internships and hands-on work experience. Nearly 70% prioritize where and how learning happens over what is being taught. Senior developers, when assessing a junior’s readiness to work, select real-world project experience (69.7%), internships (56.5%), and strong performance in a practical coding task (52.9%) as the most reliable indicators. What the data shows is not that formal credentials are losing value, but that the market is incorporating new complementary signals to evaluate a developer’s readiness. Real-world experience, internships, and practical coding performance are indicators that coexist with academic training; they do not replace it. Programs that integrate project work, internships, and theoretical training as a continuous whole, rather than sequencing them, strengthen their position in that equation.
Stepping back to what the Q2 2026 data shows as a whole, the most important risk this edition identifies is not the displacement of the junior role by AI, which both juniors and seniors already anticipate and process, but the possibility that a generation of developers reaches professional output levels without building the foundational understanding that later enables judgment in architecture, technical leadership, and strategic decision-making. As BairesDev’s CEO observes in this edition: the seniors of 2030 and 2035 are the juniors of today.
The competencies the market values most, like critical thinking, problem decomposition, systems-level understanding, are exactly the ones that technical training has always placed at its core. The tension is between the pace of training and the pace of technological change. And the question the Q2 2026 opens without closing, whether those competencies can be built learning with AI, or whether they require having gone through the process without it, is the territory where pedagogical research has most to contribute.
References
- Anderson, L. W., & Krathwohl, D. R. (Eds.). (2001). A taxonomy for learning, teaching, and assessing: A revision of Bloom’s taxonomy. Longman.
- Dede, C. (2010). Comparing frameworks for 21st century skills. In J. Bellanca & R. Brandt (Eds.), 21st century skills. Solution Tree Press.
- Flavell, J. H. (1979). Metacognition and cognitive monitoring. American Psychologist, 34(10), 906–911.
- Lave, J., & Wenger, E. (1991). Situated learning: Legitimate peripheral participation. Cambridge University Press.
- Pellegrino, J. W., & Hilton, M. L. (Eds.). (2012). Education for life and work. National Academies Press.
- Risko, E. F., & Gilbert, S. J. (2016). Cognitive offloading. Trends in Cognitive Sciences, 20(9), 676–688.
- Trow, M. (1973). Problems in the transition from elite to mass higher education. Carnegie Commission on Higher Education.
- BairesDev. (2025–2026). Dev Barometer Q3 2025, Q4 2025, Q1 2026, Q2 2026. BairesDev Research
