Generative AI in Programming Education

In this talk, I present recent research at Google and other tech companies that investigate how developers perceive generative AI tools at work. One study that we discuss is research on developer attitudes towards generative AI and their career. The research shows that developers have very complicated feelings, responding with both positive and negative attitudes towards the technology. In the rest of the presentation we discuss related research from around Google and other technology companies that provide insight into these thoughts.

Generative AI (GenAI) is playing an increasingly influential role in computing education across all levels, offering new opportunities to support both teaching and learning. However, its effective integration raises critical concerns related to trust, academic integrity, and broader social and ethical implications. While substantial attention has been given to GenAI use in introductory programming courses (e.g., CS0/CS1), there remains a notable gap in research addressing its application in “upper-level“ computing courses such as software engineering, human-computer interaction, algorithms, operating systems, and theoretical computer science.

This presentation gives a brief overview of the early work 2025 ITICSE conference Working Group 2 has drafted for a report that will present two complementary studies: a systematic literature review of GenAI interventions in upper-level computing education, and a survey of computing educators on their practices and perspectives regarding GenAI integration in these contexts.

Motivation: Students learning to program often reach states where they are stuck and can make no forward progress – but this may be outside the classroom where no instructor is available to help. In this situation, an automatically generated next-step hint can help them make forward progress and support their learning. It is important to know what makes a good hint or a bad hint, and how to generate good hints automatically in novice programming tools, for example using Large Language Models (LLMs).

Method and participants: We recruited 44 Java educators from around the world to participate in an online study. We used a set of real student code states as hint-generation scenarios. Participants used a technique known as comparative judgement to rank a set of candidate next-step Java hints, which were generated by Large Language Models (LLMs) and by five human experienced educators. Participants ranked the hints without being told how they were generated. The hints were generated with no explicit detail given to the LLMs/humans on what the target task was. Participants then filled in a survey with follow-up questions. The ranks of the hints were analysed against a set of extracted hint characteristics using a random forest approach.

Findings: We found that LLMs had considerable variation in generating high quality next-step hints for programming novices, with GPT-4 outperforming other models tested. When used with a well-designed prompt, GPT-4 outperformed human experts in generating pedagogically valuable hints. A multi-stage prompt was the most effective LLM prompt. According to a fitted random forest model, the two most important factors of a good hint were length (80–160 words being best), and reading level (US grade nine or below being best). Offering alternative approaches to solving the problem was considered bad, and we found no effect of sentiment.

Conclusions: Automatic generation of these hints is immediately viable, given that LLMs outperformed humans – even when the students’ task is unknown. Hint length and reading level were more important than several pedagogical features of hints. The fact that it took a group of experts several rounds of experimentation and refinement to design a prompt that achieves this outcome suggests that students on their own are unlikely to be able to produce the same benefit. The prompting task, therefore, should be embedded in an expert-designed tool.

The seminar opened with reflections on the rapid and transformative developments in Generative AI and their implications for programming education. This is an area marked by both excitement and uncertainty, with leaders from academia and industry expressing diverging views. These range from predictions that traditional university education, particularly in technical disciplines, may become obsolete, to optimism that AI will enhance learning and professional productivity. These perspectives were noted as mirroring recent debates in Communications of the ACM, where opposing viewpoints have argued either for the end of programming as it is currently understood or for a new era of AI-augmented programming practice.

The organisers of the seminar, Natalie Kiesler, Michelle Craig, James Prather, and Paul Denny, were excited at the prospect of fostering collaboration among researchers and practitioners from computing education, software engineering, HCI, and related fields. Participants were invited to explore how Generative AI is reshaping the ways that programming is taught, and how pedagogical approaches, assessment practices, and curricula need to evolve to ensure that AI supports rather than replaces meaningful learning in programming.

The organisers paid a special tribute to Dr Brett Becker, whose early work and vision were instrumental in the establishment of this seminar. Dr Becker, who passed away in October 2024, was remembered with deep appreciation for his contributions to the field and his role in bringing this community together.

Agentic GenAI tools, such as Copilot Agent, have been recently integrated into IDEs such as Visual Studio Code. Using thse tools, software developers can orchestrate software development by describing software requirements to a GenAI agent in plain English. The agent then implements the requirements by directly modifying a codebase. This process resembles a computational steering model with a human in the loop to monitor development progress and steer the agent toward the desired solution. I reflect on one month of intensive use of a GenAI agent, identifying high-level abstractions that a developer uses in lieu of computer code. I propose a set of general student learning outcomes that computing educators might use to design and learning activities to facilitate this form of software development and assess student learning. A video demo of my interactions and reflections can be found at http://tiny.cc/GenAIAgents.

Edugator is a browser-based, AI-enabled tool designed to help instructors of introductory computing courses create and deliver interactive educational content. It streamlines the content authoring process by incorporating generative AI models into both the creation and delivery stages. Using this tool, instructors can create bespoke interactive computing lessons and programming problems by providing a prompt and a few clicks. They can also author templates and test cases in programming languages such as C++, Java, C, and Python. Additionally, instructors can validate programming problems by running them against an auto-generated solution, allowing them to refine the problems before releasing it to students, preventing misinformation or ambiguity. Students can complete lessons and solve programming problems in a browser-based text editor receiving immediate feedback. They can also interact with a large language model-powered AI chatbot that scaffolds a student on how to approach the problem without giving out solutions. Edugator is built using modern web frameworks and the goal of the tool is to accelerate the adoption of automated assessment tools by minimizing the challenges instructors face with such tools. It also supports Learning Tools Interoperability (LTI), allowing seamless integration with learning management systems (LMS). Tools like Edugator streamline assessment authoring and delivery for Instructors while supporting student learning by promoting learning-by-doing and providing meaningful, personalized feedback. More information about the tool can be found at https://edugator.app/.

One of the objectives of this Dagstuhl Seminar, and of many other initiatives on Generative AI, is to develop recommendations for the pedagogical integration of GenAI in programming courses. I aim to challenge this seemingly “inevitable” integration.⁵⁵5https://leonfurze.com/2025/04/28/the-myth-of-inevitable-ai/

Large Language Models are built on human labor exploitation, stolen data, and are detrimental to the environment [2]. Unfortunately, this is usually an afterthought in many papers and talks. It is impossible to ethically justify their use at such a large scale. A huge amount of money is involved, and companies’ interests are clearly not in human learning [11]. The hype is unprecedented. The main goal of the major AI companies is automation, ultimately replacing much of human work. Turning to the consequences of its use, there is evidence of reduced critical thinking when knowledge workers use GenAI [12]. The benefits for programmers are debatable⁶⁶6https://blog.glyph.im/2025/06/i-think-im-done-thinking-about-genai-for-now.html; a recent study has shown that AI makes developers slower, although they thought it was the opposite [9].

So is the promise of revolutionizing education false? Current meta-reviews are not convincing [3], and many claims are unsupported by evidence [4]. Teachers spend a lot of time “reviewing, repairing and sometimes completely reworking AI-produced outputs” [8]. In computing education, researchers have observed a “widening gap” between weaker and stronger students [1] and signs of “social erosion” because students no longer ask each other for help [5]. While many new tools have emerged, they are often not built on known pedagogy or learning theory [7].

In the end, do we decide, or the technology [13]? “AI is unavoidable, but not inevitable”.⁷⁷7https://marcwatkins.substack.com/p/ai-is-unavoidable-not-inevitable Pretending it is not there does not make any sense, but we can take the lead in deciding what to do with it. We can take a step back and resist the hype [6]. Put pedagogy first, provide guardrails and scaffolding for authentic learning experiences. When building tools, only use LLMs for specific use cases, combining them with proven techniques (e.g. [10]). Finally, let us go back to the real problems at hand, and focus on community and learning.

Generative AI (GenAI) keeps advancing rapidly, and the literature in computing education is expanding almost as quickly. Initial responses to GenAI tools were mixed between panic and utopian optimism. Many were quick to point out the opportunities and challenges of GenAI. Researchers reported that these new tools are capable of solving most introductory programming tasks and are causing disruptions throughout the curriculum. These tools can write and explain code, enhance error messages, create resources for instructors, and even provide feedback. In 2024, new research started to emerge, focusing on the effects of GenAI usage in the computing classroom. At the same time, a new class of tools is being developed that can provide personalized feedback to students on their programming assignments or teach both programming and prompting skills. With the literature expanding so rapidly, an ITiCSE working group aimed to summarize and explain what is happening on the ground in computing classrooms. In our lightning talk presentation, we provided the results of a systematic literature review; a survey of educators and industry professionals; and interviews with educators using GenAI in their courses, educators studying GenAI, and researchers who create GenAI tools to support computing education. The triangulation of these methods and data sources helps expand the understanding of GenAI usage and perceptions at this critical moment for our community.

In my lightning talk I discussed two projects that are relevant to the teaching and learning of computer science (CS) in the age of Generative AI. The first project looks at learning loss from student plagiarism within an introductory programming course [1]. We collected data in an intro Python course for non-CS majors before and after the wide availability of ChatGPT. We developed four indicators of plagiarism. We found that these indicators appearing on homework submissions were correlated with lower performance on a final exam taken within a testing facility, while controlling for a student’s performance on an exam early in the semester. This speaks to the ways in which plagiarism appears to decrease learning and the age of Generative AI provides easy opportunities for plagiarism. The second project looks at using physical objects to help students reason about foundational ideas within programming [2, 3]. This includes variable types, variable assignment, and function calls. This work builds on taken for granted practices within mathematics education of using physical objects to help students learn about quantity. Resources for this project are available at CSTeachingTIps.org/3D.

One promise of generative AI for programming is that it can help us build software by quickly translating specifications into implementations. If we follow the premise, then skills like analyzing the qualities of specifications and evaluating the correctness of implementations could be more important to emphasize. But this presumes we know what software we want to build in the first place. Our choices shape not only what skills students learn but also their disciplinary values interpretation: “a process by which students reflect on the values of a disciplinary domain, as well as who they are and might become in relation to the domain” [1].

My scholarship has explored redesigning technologies as a way to shape disciplinary values interpretation by drawing on design methods such as iterative design [2, 3]. Iterative design is a software development practice that involves prototyping, testing, analyzing, and refining technology. Instead of assigning students a complete specification of a program to implement and focusing on evaluating the qualities of the final product, students could showcase their software development process over time. We can then ask questions about each step of the process.

Generative AI is often framed as a productivity tool, but what does productivity free us to do? Teaching students design methods could empower them to ask bigger questions about their work and challenge them to reflect on what exactly they hope to achieve in their future computing careers [4].

In my talk, I presented the tool APFEL, an adaptive learning environment for programming education that integrated GenAI in a controlled and transparent wa.

Unlike systems that leave it to LLMs to “decide” what kind of feedback to generate or provide and when, APFEL ensures that the feedback process is traceable and understandable for learners – if they want to know “Why did I get this feedback here?”

At its core, APFEL uses a symbolic/rule-based layer to determine the type of feedback that should be provided, based on empirical findings and pedagogical concepts and principles. This context is then used to systematically assemble a prompt, which is further enriched with additional context from a domain model and a learner model using RAG. This ensures that the resulting feedback is personalized and also didactically grounded and traceable.

Humans are considered to be the agents responsible for actions when they use tools to complete those actions. When we say “I cut down a tree”, we attribute the act to the human rather than the axe. Similarly, when we create code using a tool such as GenAI, we should attribute the act of creation to the human rather than the machine.

Although this view challenges some academics, publishers such as ACM have already decided that only humans can be authors, and they are responsible for the product of GenAI use.

We conclude that users of GenAI tools are the authors of the output produced. Since these tools are ubiquitous, authentic assessment tasks should allow students to use GenAI, and the work product should be graded in the same manner as any other student work. To operationalise this approach, we adopt a two-lane approach in which we make a distinction between secure and insecure assessments.

For secure assessment, we may prohibit the use of GenAI to determine what the student can achieve independent of tools. For insecure assessment, students are permitted to use GenAI and must be graded with the view that they are the author, and are responsible for the product that they produce using tools.

Generative AI provides students with valuable support even late at night when their friends are asleep. However, as students rely more heavily on these tools for help, instead of their friends, what will be the impact? Based on semi-structured interviews with 17 undergraduate computing students, we found that help-seeking requests are now often mediated by generative AI, with students frequently redirecting questions from their peers to generative AI instead of providing assistance themselves. This has the potential to undermine relationships and create a transactional classroom environment where friends never have a chance to form and learning communities erode. Students also reported feeling increasingly isolated from their peers and demotivated as a result. Our work is call to action to refocus our efforts on fostering vibrant learning communities and centering social interactions in an age of automation.

Computing students increasingly rely on generative AI tools for programming assistance, often without formal instruction or guidance. This highlights a need to teach students how to effectively interact with AI models, particularly through natural language prompts, to generate and critically evaluate code for solving computational tasks.

To address this, we developed a novel platform for prompt programming that enables authentic dialogue-based interactions, supports problems involving multiple interdependent functions, allows manual code edits of generated code, and offers on-request execution of generated code.

Data analysis from students using the tool in introductory programming courses shows high engagement, with most prompts occurring in multi-turn dialogues. Furthermore, students were highly selective about the code they chose to execute. Reflection data also indicates that the majority of students preferred a balanced mix of natural language prompting and manual code editing when tackling harder problems.

The advent of Generative AI tools is rapidly transforming the landscape of software development, presenting both challenges and opportunities for computer science education. In this talk, we address the critical need to adapt introductory programming curricula to this new reality. We share the journey of creating the textbook “Learn AI-Assisted Python Programming” and developing a corresponding CS1 course that embraces AI tools. The core of our approach is to shift the focus from traditional coding syntax to a new set of essential skills for the modern developer, including prompt engineering, critical code evaluation, testing, debugging, and problem decomposition in an AI-assisted environment.

We provide a detailed account of our pilot “CS1-LLM” course at UC San Diego, which served over 500 students. We discuss the course’s design, which featured open-ended, creative projects across various domains to keep students engaged. We conclude with practical lessons learned from our teaching experience and an overview of our ongoing work to support educators and build a broader community around integrating AI into computer science education.

Google’s Tech Exchange program has been shaping and building engineers for many years. The advent of GenAI caused some panic among the creators, and it became a critical question about what to do. It was decided that the curriculum must be entirely reimagined. Google partnered with over a dozen universities to rework its curriculum to be GenAI forward from the very beginning. The redesigned courses were CS1, CS2, Applied Algorithms, Software Engineering, and Project Management. The results show that students engaged, worked hard, and learned traditional coding and computer science concepts. Students also increased their general self-efficacy over the test semester as well as their GenAI-specific self-efficacy. We believe this curriculum represents an effective first step in redesigning our courses to account for this fast-growing technology.

The University of Toronto Task Force on Al explored the impact of generative Al on all aspects of the University. I served on the main task force and co-chaired the Teaching and Learning Working group. During our consultations we heard many faculty despairing that unsupervised assessments no longer promote student development and learning since students can use GenAl tools to complete much of the work. The working group report highlights the need for GenAl literacy programming for both faculty and students, the importance of reconsidering learning outcomes across the curriculum, the challenge of providing formative feedback at scale, and noted that classroom dynamics are shifting as students and faculty alike bring their Al avatars to class with them. The task force reports recommend that the university invest in Al technology so that faculty and staff may experiment with and develop tools to support student learning. As we refined recommendations, we kept coming back to the idea that “good teaching is still good teaching.” In other words, best teaching practices remain resilient in the age of GenAl. Ultimately, the task force focused on the university as an “institution dedicated to the development of human potential.” Key takeaways include the need for greater transparency with students about how learning happens and a renewed determination to foster social interaction among students to enhance their learning.

In order to make the learning process visible, E-Portfolios have been utilized for many years. However, for both, teachers and students, this is considered an time-consuming and tedious task. As part of the project AISOP at the PH-Weingarten, these tasks have been simplified using AI methodologies.

Teachers define a concept map, which the tool uses to evaluate the created portfolios. Students get feedback on their current state before going into the examination, which only bases on the portfolio.

As code, it looks for the student like this:

  while(learning) {
    write_portfolio() //here reflection with learning happens
    gather_experience()
    update_and_evaluate_using_ai() // concept map and AI tools are triggered
    get_feedback()
  }
  prepare_for_exam()
  do_exam()

For the teachers, it could be represented as follows:

  define_concept_map()
  while(learning) {
    assign_tasks()
    gather_insights_from_ai() //using dashboards, etc.
    provide_additional_feedback() // human in the loop
  }
  check_portfolios_with_ai()
  //prior to examination, check for missing entries, unclear documentation, etc.
  do_exam_using_portfolio()

The project was defined and started in the pre-GPT era, with ChatGPT disrupting a lot of development. Many possible improvements exist, as higher quality evaluation is now possible. Currently, the project is being finished, and the idea pushed and feedback on it’s implementation as a general tool collected. We hope that the tool gets the opportunity to be finished and provided to other teachers. This will facilitate a different type of assessment that cannot be simply solved by memorization and and requires some time and effort for students. The time spent helps students to better understand larger concepts and emphasizes continuous learning

A summary of recent trends, stats, and beliefs regarding the use of generative AI within the tech industry.

In this talk, I discuss two separate GenAI accessibility topics.

First, that there’s an unequivocal pro of GenAI tools (in all of the needed pro-con discussions!): GenAI-powered tools are making visual-related accessibility improvements that I never thought possible. They are describing videos, making inaccessible apps accessible, helping with usability concerns on websites, describing graphs and other diagrams from papers, etc. Many of these challenges have been understood but not solved for decades (“add alt tags to your images!”); GenAI is allowing us to go beyond a mere description of visual material to full-on interaction.

Second, that we as a community need to ensure that the GenAI tools we ask our students are indeed accessible! In particular, I described the inherent accessibility of commandline-based GenAI tools that otherwise maintain a user’s existing workflow (rather than requiring them to use IDEs that may take weeks to learn and that themselves may not be accessible).

Traditional software development (SD) involved minimal AI and a lot of code writing and reading. While AI supported SD has been characterized by the use of manually triggered AI to do some of the code writing the recently emerged agentic SD enables AI to handle complex tasks such as planning, testing, and refactoring. Human developers still play a crucial role in defining goals, reviewing, fixing outcomes, and guiding the process. When the agent-generated codebase falls short on, e.g., functional or stylistic requirements. One can ask the agents to modify something. One can even suggest how to go about it (e.g., where in the codebase; what library to use). And one can iterate … At some point, one may give up and modify the code themselves. Hence, (1) there will likely be less natural opportunities for one to directly write and read code; (2) a developer will likely be responsible for handling increasingly larger and complex codebases; (3) the same or even greater expertise in software engineering, including reading and writing code, might be needed. Then, how can we ensure that future software engineers go about their everyday work in a way that systematically hones their expertise? In conclusion, there appears to be a fundamental human-AI interaction challenge: How do we design interactions between agentic SD frameworks and humans to ensure continued growth of a human as an expert?

Advances in Generative AI (GenAI) have changed the way we view programming and programming education. GenAI’s ability to write usable, working code has now sparked conversations and research regarding how computer science courses should be taught. Students are utilizing tools such as ChatGPT to produce coded solutions to tasks, but still require the ability to read and comprehend the code. Current research examines the implications of GenAI tools across the breadth and depth of CS courses and what it means for future iterations of CS curricula. What is lacking, however, is research addressing GenAI’s impact on future software developers on two fronts. On one front, we need to consider how differential access to GenAI tools and education in K-12 among specific populations can widen the gap in terms of academic progression in higher education, especially for historically marginalized groups. There is already a gap in access to curricula and teachers, as seen with the AP Computer Science Principles. GenAI may further widen this gap due to several factors: a lack of a curriculum adapted for GenAI, a shortage of teachers skilled in using, advising, and understanding GenAI, and limited access to computers and AI tools in schools. These factors can create disadvantageous situations for students entering college to pursue a degree in CS. On the other front, industry hiring practices are evolving unevenly, with AI being used both by candidates and evaluators, often without clear standards or awareness of the implications. To what extent may companies utilize GenAI tools to aid in the evaluation of candidates, and how will they try to mitigate bias in the hiring process? What are candidates allowed to delegate to an AI if applying for a position and what training do they require for it?

This position paper argues that without systemic changes to the entire education through hiring pathway, AI risks reinforcing existing biases, introducing new biases, and rendering traditional assessments obsolete. We call for a rethinking of programming education across all levels, so that students acquire the competencies to utilize AI, and avoid the creation of new biases and unethical systems by gatekeeping newcomers who do not have access to special tools.