What does Software Quality mean in the age of AI? | Thomas Steirer | TEDxTUWien

Summary

In this talk, the speaker delves into the intersection of artificial intelligence (AI) and software quality. Drawing from personal experience, the speaker discusses AI's rapid integration into everyday life and its implications for software testing. With AI systems becoming more prevalent, traditional quality assurance methods are no longer sufficient. The speaker raises important questions around setting realistic expectations for AI, ensuring accountability in case of failures, and establishing safeguards against biases and harmful impacts. The call to action is for society to consciously guide AI development, ensuring it aligns with human values and needs.

Takeaways

• 😀 The speaker has a deep passion for both Artificial Intelligence (AI) and software quality, with a long history of engaging with technology from a young age.
• 😀 The speaker's background includes studying computational intelligence, blending logic, heuristics, and machine learning, and a desire to see AI impact daily life, not just specialized industries.
• 😀 In addition to a hobby of video game testing, the speaker has worked in professional software testing, contributing to the functionality of systems we rely on, such as public transport and online banking.
• 😀 The concept of 'destructive creativity' is introduced, meaning a mindset where one constantly thinks about how software could break and how it can be improved.
• 😀 AI has rapidly advanced and is now capable of creating pictures, writing texts, generating music, and more, but it raises challenges regarding software quality and how we verify AI outputs.
• 😀 Unlike traditional software, AI outputs are not always clear-cut or easily measured. For example, asking AI for open-ended answers creates challenges in determining correctness.
• 😀 Traditional software has 'test oracles,' like mathematics, to verify results. However, AI outputs, such as historical explanations, require new methods for validation because they are not fixed answers.
• 😀 The speaker gives an example of autonomous driving, highlighting how expectations, like zero accidents, may be unrealistic and raising questions about accountability in case of failures.
• 😀 As AI becomes embedded in daily life (e.g., smart lights, recommendations, photo enhancements), it becomes more important to establish clear expectations and safeguards for its use.
• 😀 Safeguards for AI systems are necessary, and while methodologies from software testing can be applied, AI introduces complexities like bias and social impact, which require additional considerations for quality control.
• 😀 Ultimately, we need to ask ourselves what we expect AI to do, what we don't want it to do, and ensure that we are responsible in guiding AI's development and integration into society.

Q & A

What is the speaker's background with artificial intelligence (AI) and software quality?

The speaker has a long history with computers, starting from a young age when they modified programs their father created. They later studied computational intelligence at the Vienna University of Technology, which covered topics like logic, heuristics, and machine learning. They have a professional background in software testing and have contributed to various systems used in daily life, such as public transport, online banking, and flight systems.

What does the speaker mean by 'destructive creativity' in the context of software testing?

'Destructive creativity' refers to the mindset where, instead of merely appreciating a new product or program, the speaker looks for ways to break it. This approach is critical for identifying flaws and ensuring the software is robust and can withstand real-world use.

How does AI differ from traditional software systems in terms of interaction?

AI systems differ from traditional software because they respond to open-ended, conversational queries. Unlike traditional software, where responses are predictable and well-defined (like mathematical calculations), AI-generated responses can be text, images, videos, or other formats, which makes validation more challenging.

How does the concept of a 'test oracle' apply to traditional software, and how does it change with AI?

In traditional software, a test oracle is a predefined method to determine if a response is correct, such as using a mathematical formula. With AI, there is no such clear-cut oracle because the responses are more subjective, such as an exploration of ideas or concepts, which makes verifying accuracy more difficult.

What is the fundamental problem with validating AI-generated content?

The fundamental problem with validating AI-generated content is the lack of a predefined correct answer. For example, if an AI is asked about a historical topic, the response is an exploration of ideas, and it requires research and judgment to determine whether it is accurate, making validation challenging.

What are the challenges in setting realistic expectations for autonomous driving systems?

The challenge in setting expectations for autonomous driving is determining what level of performance is acceptable. For example, should the benchmark be zero accidents, or is it acceptable for the system to make occasional mistakes? The comparison to human drivers is complex because different drivers have varying skill levels and driving habits, which complicates the standard for self-driving cars.

What is the ethical dilemma regarding accountability when an accident occurs with a self-driving car?

The ethical dilemma involves determining who is responsible when a self-driving car causes an accident. Is it the person who gave the driving instruction, the taxi service that owns the car, the manufacturer, or even the car itself? The question of accountability becomes complex in the context of autonomous systems.

How does AI influence our daily lives beyond traditional software systems?

AI influences our daily lives in various subtle and profound ways. It controls systems like lighting, provides song recommendations, and enhances the quality of photos taken on smartphones by ensuring people look their best. These interactions with AI are often invisible but significant in shaping our daily experiences.

What are some of the safeguards and methods available to ensure AI systems function properly?

Safeguards for AI include techniques like human oversight, quizzes, and even using another AI to monitor the first one. There are also regulatory frameworks and standards that aim to ensure AI systems perform as intended, taking into account not just accuracy but also factors like bias and societal impact.

What is the central question that arises when considering the future of AI in our lives?

The central question is what we want AI systems to do and what we don’t want them to do. This encompasses not just the functionality and correctness of AI systems, but also their broader societal impact, ethical considerations, and the potential consequences of their decisions.