The new literacy of code

Programming in 2025 begins in an unfamiliar way: the cursor waits while the machine writes first.

The reversal has unsettled the quiet assumptions that once anchored computer science education. A functioning program used to signal comprehension. A neatly written loop once demonstrated that a student grasped the idea behind iteration. Those markers have slipped out of place. The work students now submit often arrives polished and syntactically correct, composed not by the student but by the model at their side. Instructors are left facing a question the field has largely avoided for decades: what does it mean to learn computer science when the routine parts of programming can be generated instantly?

A path through that question emerged during a conversation with Mary Shaw, a Professor of Computer Science at Carnegie Mellon University and a foundational figure in software architecture. She described the shift with a clarity that felt like the beginning of a new curriculum.

“We need to teach people to read code written by somebody or something else,” she said in an interview with IBM Think. The remark captured the center of gravity that she and her Carnegie Mellon colleague, Michael Hilton, a Teaching Professor in the Software and Societal Systems Department, have been tracing in their upcoming paper. Both said they believe the discipline is moving away from the act of writing code and toward the deeper, steadier work of supervising it.

The larger landscape reinforces that shift. According to the 2025 Stanford AI Index, 81% of K-12 computer science teachers in the United States believe AI should be part of foundational computing education. Yet, fewer than half feel prepared to teach it. Shaw said this mismatch is one reason universities are reconsidering what it means to teach programming. Hilton said the gap between what students can produce with AI and what instructors can reliably assess is widening quickly.

The job market adds its own pressure. Shaw said the demand for AI-literate engineers is rising faster than universities can adapt, a point underscored by the latest Stack Overflow Developer Survey. According to the 2025 results, 84% of respondents are using or planning to use AI tools in their development process this year, and 51% of professional developers say they use them daily.

Hilton said these labor market signals underscore why universities must adjust the fundamentals of how they teach. He added in an interview with IBM Think, “I think the gen AI revolution has entirely broken the hiring process, and it has lowered the bar for application … recruiters are getting inundated with tens and tens of thousands of applications for every position.”

A look inside the classroom reveals how quickly the old signals have lost their value. Shaw said instructors long relied on the assumption that a working program indicated conceptual understanding. She said this link has simply dissolved. Hilton said he sees this first-hand when students submit assignments composed entirely by generative systems, which weakens the connection between output and comprehension.

Efforts to recalibrate are visible across campuses. Shaw said instructors are shifting toward assignments that require explanation and reasoning, not just production. She said students must demonstrate an ability to evaluate behavior, interpret logic and identify when an AI-generated solution does not align with the problem. Hilton said paper-based or supervised assessments have returned as a way to understand what students actually know when the model is taken away. Shaw said these are early attempts to redefine what counts as evidence of learning.

The tension is not confined to academia. Anna Gutowska, an AI Engineer and Developer Advocate at IBM who is also completing a master’s degree in computer science at Stanford, said in an interview with IBM Think that she sees many learners “skip core CS foundations and jump straight into vibe coding AI applications.”

Gutowska said the abundance of tools allows newcomers to assemble something that appears to work without engaging the underlying principles that make software stable. She noted that algorithmic foundations, data structures and systems knowledge remain essential because they determine whether someone can use AI to build reliable systems rather than brittle ones that fail under pressure.

The shift in classroom habits, Shaw said, reveals a deeper gap in how the field teaches itself. She noted that most disciplines build judgment through reading before making anything of their own: writers work through literature, musicians study repertoire, mathematicians read proofs. Programming often reverses that order. Students, she said, are pushed to write code before they have meaningfully read any, which Hilton said leaves them without a sense of structure or design. Shaw added that AI now exposes the gap sharply, because students must read, critique and verify large amounts of code they never wrote.

The difference shows up in how students use AI. Hilton said he sees one group that keeps prompting until the model produces something that runs, and another that asks for smaller pieces, inspects them and adjusts as they go. He said the second group builds stronger conceptual models. Shaw noted that the first often ends up with polished answers that conceal fragile understanding. Hilton said the slower, more deliberate style forces students to examine behavior rather than trust the appearance of correctness.

The same pattern appears in professional development. Gutowska said that debugging, testing, monitoring and bias detection are becoming central, as AI-generated components introduce subtle errors that require careful evaluation. She said these tasks rely on conceptual grounding rather than on the convenience of automated output. She added that attempts to rely on one model to correct the mistakes of another often make the problem worse.

Shaw noted that real software work rarely resembles the tidy examples in textbooks. She explained that developers often write small fragments of code to clarify their thinking, expose ambiguities and revise assumptions. AI, she observed, can accelerate that exploratory loop by generating variations that uncover misunderstandings and push students to articulate their intentions more precisely. Hilton added that these moments reveal exactly where a student’s mental model needs strengthening.

Testing now anchors the supervisory model. Hilton said AI-generated code cannot be trusted without evaluation. He said this is why he emphasizes property-based testing, which expresses expected behavior in general terms rather than relying on a few specific examples; this teaches students to think about software conceptually. Shaw said AI can generate tests and code, but the test suites still require human oversight. Someone must decide whether they capture the real requirements or simply formalize an incomplete understanding.

Gutowska said the professional world reinforces the same message. Colleges will need to prepare students to choose between AI-driven and traditional approaches depending on context. Large language models, she explained, are powerful but inconsistent across domains.

The demands of early-career roles reflect the same shift. Hilton said entry-level programming tasks that once served as apprenticeships might now be drafted by AI systems. He said junior engineers spend more time assessing, integrating and correcting model-generated components than writing code from scratch. Shaw believes this transition emphasizes the value of judgment. She said engineers must understand what software is doing, not simply produce it.

Recruiters face similar challenges. Hilton said candidates can now produce polished résumés and project samples using AI, which masks the difference between genuine understanding and automated fluency. Shaw said employers have long been split between those who want tool-specific training and those who want adaptable reasoning. AI has widened that divide.

Gutowska asserted that the engineers who succeed will be those who combine conceptual grounding with the ability to interrogate and supervise AI-generated work. She said the industry increasingly values those who can design systems that incorporate AI without becoming dependent on it.

A slower read of the moment, Shaw said, shows something more measured than disruption. She noted that the field’s core work has always involved understanding how systems behave, not assuming that a working output reflects that understanding. AI simply makes that gap visible by generating code that must be read, questioned and verified. She pointed out that real software development has long relied on exploratory coding, where developers sketch ideas in small fragments before refining them, and she said students need to learn that habit directly. Faculty, she added, cannot pretend to have a perfect method for teaching in this environment. Instead, she described sitting with students, reviewing AI-generated work and walking them through the questions that reveal comprehension: whether the code will run, whether it will integrate, whether it introduces vulnerabilities and whether its tests capture the correct behavior.

That orientation points toward a future defined more by supervision than by production. Shaw said AI already produces both code and tests, which means students must learn to examine the output rather than trust it. Instruction will center on helping students develop the judgment to spot when a model has misunderstood a requirement and to identify the mistakes that only emerge during evaluation. She emphasized that this is not an endpoint for the field but a shift back toward the reasoning practices that have supported it from the start. Shaw closed with a reminder grounded in that perspective: “AI is not going to kill software engineering.”