Thursday, April 16th, 14-15:30
Tech Debt is spiraling out of control. A recent CAST report estimated that Tech Debt worldwide has amounted to 61 billion workdays. Gartner, in its “Predicts 2026” report, emphasizes the increasing risk AI agents which assist software engineering tasks pose to systems; that report goes as bold as predicting that “a new remediation market will emerge for specialized tools and consulting services that can audit, identify, and refactor the “AI-generated technical debt” that companies accumulate.” We are approaching a situation similar to where technical debt research began in 2010—marked by uncertainty, confusion, and a lack of clear direction for managing the growing Tech Debt facing developers and organizations, alongside an increasing number of tools promising efficiency and quality that overwhelm software engineers. Only this time driven by AI.
This discussion panel will examine how AI acts as a double-edged sword for Tech Debt: on the one hand AI can potentially provide the tools to automatically identify, measure, prioritize and repay Tech Debt; on the other hand, there is existing evidence that AI is causing the incurrence of massive amounts of Tech Debt through code generation and agentic development. Even if AI tools can provide locally optimized solutions, they often compromise overall design principles and system-wide quality attributes. What will it take to reap the benefits of Generative AI in Tech Debt management while mitigating the risk of making systems even more complex and unsustainable? What are the research directions to prioritize? What does industry really need to solve this problem?
Chairs: Paris Avgeriou and Ipek Ozkaya
Click here to see it in the schedule
Panelists
Dr. Dongmei Zhang is a Distinguished Scientist at Microsoft with extensive experiences in research and development. She pioneered and established “Software Analytics” as a new research direction, founding the Software Analytics Group at Microsoft Research Asia in 2009 to drive the integration and advancement of data, AI, and software engineering. She subsequently expanded her work into the field of data intelligence by creating the “Data, Knowledge & Intelligence (DKI)” research group. Over a career spanning more than twenty years at Microsoft, Dr. Zhang has led her teams to closely collaborate with product teams across the company on technology transfer. Her contributions include digital ink technologies in Windows 7 and Microsoft Math, as well as foundational research in Software Analytics, AIOps, and data intelligence that have enhanced and impacted products such as Power BI, Excel, Azure, and Microsoft 365. Currently, she is the Chief Scientist of the Microsoft Software Technology Center in China (STCA) while continuing to lead the DKI research group. Dr. Zhang received her Ph.D. in Robotics from the School of Computer Science at Carnegie Mellon University. She received her M.E. and B.E. from Tsinghua University.
Position statement:
Knowledge Debt – Challenges and Opportunities In the AI era, software engineering needs to treat knowledge debt as a first-class, cross-cutting concern within the broader technical debt landscape. Prior technical debt research has already shown that debt extends beyond code to artifacts such as architecture, tests, and documentation. The issue is not that the field failed to notice knowledge-related problems, but that it has not made knowledge itself explicit enough as the common liability that often cuts across these categories. The challenge is that the knowledge needed to understand, change, operate, and evolve systems often remains tacit, fragmented, stale, weakly linked, or weakly governed. Teams then pay for this later through slower onboarding, repeated misunderstandings, riskier releases, longer incident response, and greater coordination overhead, etc.
The AI era makes this both more urgent and more promising. It becomes more urgent because new knowledge-bearing artifacts, including prompts, retrieval pipelines, evaluation logic, agent workflows, and opaque automation paths, can themselves accumulate drift, inconsistency, and hidden coupling. At the same time, AI creates a major opportunity. It can help capture, refactor, operationalize, and continuously improve knowledge at a scale that was previously unrealistic. AIOps 2.0 offers one concrete exploration of this direction. Its core idea is to treat operational knowledge as a governed engineering asset, i.e. anchored to real system touchpoints, structured for both humans and agents, and continuously validated and evolved as systems change. In that sense, AIOps 2.0 is not only a vision for smarter operations, but also a concrete example of how knowledge debt can be surfaced, governed, and gradually reduced.
Dr. Heiko Koziolek is Corporate Research Fellow for Software Architecture at ABB Research in Mannheim, Germany. His work focuses on bringing cutting-edge IT technologies into ABB’s software landscape for industrial process control and industrial drives - across embedded devices, on-premises servers, edge gateways, and cloud back-ends. He advises ABB development teams on cloud-native practices and generative-AI solutions. He co-authored Modeling and Simulating Software Architecture (MIT Press) and has received “Most Influential Paper” awards from Elsevier Journal of Systems and Software, ICSA, and ICPE. Dr. Koziolek serves as Senior Associate Editor for Elsevier JSS, sits on the editorial board of IEEE Software, and is a long-standing program-committee member for ICSA, ECSA, and ICPE.
Position Statement:
Reports of rising technical debt from AI-assisted development abound, but it remains an open question whether these reflect fundamental limitations or merely the growing pains of an immature practice. Developers and architects are still learning to direct these tools effectively - understanding their capabilities, recognizing their boundaries, and configuring appropriate guardrails. Without explicit architectural constraints and governance, AI will optimize locally and generate debt accordingly. This is not a novel problem: it closely parallels onboarding junior developers, who require the same kind of guidance to respect overarching design principles. Nothing suggests AI cannot reach - and eventually exceed - that level of architectural compliance, given adequate context.
Conversely, AI’s potential to prevent and remediate technical debt remains largely untapped. Much self-admitted debt originates from time pressure and cognitive overload — precisely the constraints that agent teams, running unattended, can absorb: producing proper abstractions, generating useful documentation, eliminating the conditions under which such debt accrues. AI-driven automated refactoring, though nascent, holds significant promise for restructuring heavily indebted codebases at scale. Research should therefore pursue both objectives in parallel: hardening AI against introducing new technical debt, and developing methods and tooling that leverage AI to retire existing debt with minimal human intervention.
Dr. David Lo is the OUB Chair Professor of Computer Science and the founding Director of the Center for Research in Intelligent Software Engineering (RISE) at Singapore Management University. A champion of the AI for Software Engineering (AI4SE) field since the mid-2000s, he has demonstrated how AI—encompassing data mining, machine learning, information retrieval, natural language processing, and LLM agents—can transform software engineering data into actionable insights and automation. Through empirical studies, he has identified practitioner pain points, characterized the limitations of AI4SE solutions, and explored acceptance thresholds for various AI-powered tools. His contributions have earned over 25 awards, including four Test-of-Time awards and 17 ACM SIGSOFT/IEEE TCSE Distinguished Paper awards, with his work garnering over 47,000 citations. An ACM, IEEE, and ASE Fellow, as well as a National Research Foundation Investigator (Senior Fellow), Lo has served as General Chair for MSR’22 and ASE’16, and as Program Co-Chair for ICSE’25, FSE’24, and ASE’20. For more information, please visit: http://www.mysmu.edu/faculty/davidlo/.
Position Statement:
AI and Technical Debt: Friend or Foe? Technical debt is defined as “development shortcuts taken for expediency that cause the degradation of internal software quality” [1]. In managing and mitigating the impact of such debt, AI can be an ally to developers. For over a decade, research in AI for Technical Debt (AI4TD) has evolved from traditional text mining and deep learning to large language models and, more recently, collaborative agents. These technologies have been deployed to identify, classify, and repay debt ranging from code smells (maintenance debt) to critical security vulnerabilities (security debt). Recent studies, including ours [2, 3], have demonstrated the practical impact of these tools, successfully identifying hundreds of vulnerabilities significant enough for developers to patch—efforts that have resulted in more than 100 CVEs.
However, AI can also be a foe. Just months after the release of ChatGPT, we highlighted the diverse forms of technical debt that AI can introduce [4]. Our recent study on hundreds of thousands of self-admitted AI contributions to thousands of GitHub repositories found that AI agents introduced roughly as many debts as they repaid [5]. Moreover, our new benchmark uncovers that state-of-the-art AI agents generate correct and secure code less than 25% of the time [6]. Such agents also remain susceptible to a wide range of security attacks [7], rendering their usage a form of technical debt—offering short-term speed at the cost of significant future liabilities. These findings highlight a critical weakness that the software engineering, AI, and cybersecurity research communities must address collaboratively.
Moving forward, research should prioritize improving the reasoning capabilities of AI agents by developing structured reasoning curricula that leverage domain knowledge. We must also place greater emphasis on the non-functional properties of AI-generated outputs and work to understand the specific types of technical debt introduced by human-AI teams (including social debt). Furthermore, it is essential to investigate debts within the AI training and serving stack, as well as prompt and data debts that lead to correctness and security issues. Finally, we must address not only team-level debt but also planetary-scale debt, including energy consumption and the broader socio-economic impact on human roles within the software industry and beyond.
Dr. Markus Borg is Principal Researcher at CodeScene, a platform used by organizations worldwide to manage technical debt and improve code quality in practice. He is Adjunct Associate Professor at Lund University, Sweden. His current work focuses on how organizations adopt coding agents at scale, and the direct and indirect consequences for software systems and the people who build them.
Position statement:
Software engineering is clearly being disrupted by coding agents. Interestingly, the push to change how we work at the code level is now coming from boardrooms rather than bottom-up from developers. At a tools company focused on managing technical debt, this forces us to revisit many of our assumptions. AI is fundamentally changing the economics of software development: the cost of code creation is approaching zero. I argue that typing speed has never been the main cost driver. But will program comprehension remain as costly as before?
Without solid quality engineering practices in place, things can quickly spin out of control. Generative AI amplifies everything related to code, often manifesting as rapidly increasing code volume, and exposes new bottlenecks wherever humans remain in the loop. In many contexts, relying on human code review in pull requests will not scale. We need to shift quality assurance further left into the agentic workflows and the LLMs themselves. While technical solutions are emerging, I am more concerned about the socio-technical consequences of large-scale development on autopilot. But the genie is out of the bottle. We must now learn as we run.
Dr. Margaret-Anne Storey is a Professor of Computer Science at the University of Victoria and a Canada Research Chair in Human and Social Aspects of Software Engineering. She is co-author of the SPACE framework and a leading researcher in Developer Experience (DevEx). Her research focuses on how developers and teams understand complex software systems, and how tools, AI, and collaborative practices shape that understanding. Her recent work examines how generative AI is transforming software engineering by changing how understanding is created, shared, and maintained. She collaborates with industry partners including Microsoft and DX. She holds an honorary doctorate from Lund University.
Position statement:
Generative AI is rapidly changing how software is produced, and with it, the nature of the risks we face. Traditionally, technical debt has been a central concern, capturing how implementation shortcuts make systems harder to change over time. But as AI systems increasingly generate, modify, and repair code, many aspects of technical debt may become easier to manage or defer. Some even argue that developers may no longer need to care about the code itself. With these possible futures in mind, the more pressing question is whether we are focusing on the right form of “debt” in this new context.
I would like to propose a shift in emphasis toward two underexamined but increasingly critical forms of debt: cognitive debt and intent debt. Cognitive debt reflects the erosion of shared understanding across a team, when developers can no longer confidently reason about system behavior, even if the code appears correct. Intent debt refers to the absence or degradation of explicit rationale, goals, and constraints that both humans and AI systems rely on to safely evolve software. As AI accelerates code production, it risks outpacing our ability to maintain shared understanding and explicit intent. This raises a provocative possibility for the panel: in the age of AI, technical debt may no longer be the primary bottleneck. Instead, should we be more concerned with cognitive and intent debt?
Dr. Ciera Jaspan is the tech lead of the Engineering Productivity Research team within Developer Infrastructure. The team brings a data-driven approach to business decisions around engineering productivity. They use a combination of qualitative and quantitative methods to triangulate on measuring productivity. She received her B.S. in Software Engineering from Cal Poly and her Ph.D. from Carnegie Mellon, where she worked with Jonathan Aldrich on cost-effective static analysis and software framework design.
Position statement: A preamble: things are changing so fast right now, it’s hard to guess at what the future will hold. None of us have a crystal ball; we don’t know what exactly things will look like 2 years from now. But this will be my best current guess at what we need to get the benefits of AI…without the risk of having an unsustainable mess.
Ultimately, tech debt is the difference between what the code is and what the developer thinks it ought to be. Code that is never looked at has no debt. If AI can truly modify the code correctly, without the developer intervening at all, there is no debt. Notice that we don’t talk about “debt” in assembly; debt is in the language we use to describe the system. If the future is one where we implement software purely through an AI, then arguably that debt goes away.
But as soon as the developer needs to understand it, now there might be piles of debt. Will we be entering a future where instead of paying down debt iteratively over time, we fix technical debt in a single swoop, as first step of trying to understand the code so we can fix a particularly tricky bug? Yikes. It would be quite awful to tell management “I’m sorry, I need to refactor this entire system before I can fix the outage”.
Clearly, we need to have agents consider and maintain tech debt, at the least so that developers can comprehend it in an emergency. This means that agents need to think like a human. We need metrics that models can hill-climb on: cognitive complexity, software design, modifiability, and testability. This is a prerequisite to having agents that can consider technical debt.