1.1: We have discussed the use of AI in algorithmic trading, supply chain and manufacturing lines.
a. What common features in workflows, or scopes of work exist where AI has demonstrated success?
đź’ˇ The intention behind this question is to provide a simple starting point, based on your comfort level, on what AI may have done for you.
b. How about the converse – where AI has not been able to demonstrate much success?
đź’ˇ This question is not so much one where you are supposed to derive a model answer, but it is meant for you to challenge what the limits of AI are. Frame these limits from the perspective of different types of AI we mentioned in Chapter 1.
1.2: This book has omitted the discussion of AI uses in popular culture such as meme creation. However, the Studio Ghibli controversy in 2025 1 has resulted in discussions over the extent to which such AI uses may result in copyright infringement of copyrighted works.
a. What actions can be taken to minimize the risk of such copyright infringement?
💡 This topic is still an active discussion. At the heart of the topic is the concept of “human input” in said AI work.
b. Given the different laws and guidelines applicable in different jurisdictions, what sort of AI governance policies should be adopted for such AI uses?
đź’ˇ A starting point for more context is to read up on how the usage of IP laws to derive and protect financial benefits that arise from creations. However, this is still a live discussion (as of time of writing of the question) in terms of precedence being set by various courts of law. Treat this as an exploratory question that requires an examiniation on the fundamentals of IP, the extent of human ingenuity required on the creation of material, as well as the incentive/ease of creation.
1 World Intellectual Property Organization. (2023). IP and Frontier Technologies. In frontier-technologies. Link2 Studio Ghibli vs AI: tribute or copyright infringement? (2025, April 15). IP Helpdesk. Link
1.3: We have discussed earlier that AI systems are statistical by nature.
a. Think of applications where (i) both false positives and negatives are tolerable, (ii) false positives are tolerable, (iii) false negatives are tolerable and (iv) neither false positives nor negatives are tolerable.
đź’ˇ Contrast the stakes of a bad movie recommendation with those of a missed medical diagnosis. How does the real-world cost of an error dictate our tolerance for algorithmic inaccuracy?
b. What impact does ultimate accountability over AI systems have, if these systems are statistical?
đź’ˇ We use quantum mechanics in our daily lives although, at the fundamental level, quantum mechanics is probabilistic. One does not necessarily need the system to be entirely deterministic to be able to take accountability for them. What knowledge, then, do we require, for a certain confidence that we understand these AI systems at different levels?
c. How would your answers in (a) explain what types of systems AI can never replace?
đź’ˇ Connect your risk tolerance matrix to the core essence of human responsibility. Are there high-stakes domains where the mere act of outsourcing the decision is unacceptable by the organization? One clear example, for instance, is the principle of democracy, where humans must ultimately exercise their vote for the political leaders of their choice.
1.4: Besides cybersecurity, there also exists an increase in AI slop. AI slop refers to a flood of low-quality content that is generated rapidly and cheaply by AI.
a. What are some societal effects of AI slop?
đź’ˇ Imagine a digital ecosystem where finding human-made truth or experiences feels like finding a needle in a haystack. How does this constant barrage of synthetic content erode public trust and reshape our online behavior? (This is an ongoing issue.)
b. What economic incentives encourage the creation of AI slop?
💡 Consider the mechanics of the attention economy, click-based ad revenue, and the extent to which it is possible to create “AI slop” that optimizes click-through rate. One perspective that is helpful to understand how cheaply AI slop can be generated is the rise of AI-generated micro-dramas3. In China, 95 percent of micro-dramas created in the first quarter of 2026 are AI generated!
c. To what extent do you think the adversarial ideas discussed in GANs can be used to train models to detect AI slop?
đź’ˇ In a generative adversarial network, a counterfeiter and a detective are locked in a continuous arms race. Could we weaponize this exact architecture to serve as an automated "immune system" for the internet? In Chapter 10, we will discuss the idea of bot-driven immunization, but we can give you the spoiler first4.
3 Xin Ying, B. (2026, July 5). AI is rewriting China’s filmmaking rulebook, but the script isn’t finished. CNA. Link4
Ng, L. H. X., & Carley, K. M. (2026). The dual personas of social media bots. Online Trust and Safety, 35–54. Link
1.5: We discussed how the lack of universality in contentious ethical and moral issues makes it challenging to codify ethical and moral norms in AI systems.
a. Go through the MIT Moral Machine exercise (ethical dilemmas involving self-driving vehicles) to get an idea of your own biases and tendencies. Would you say other members of society can agree with your biases and tendencies?5
💡 We naturally assume our own moral compass represents the default, and is a default mental model we use when assessing situations at hand. While experimenting with the MIT Moral Machine, consider how cultural, geographical, or generational differences possibly skew what society considers the “right” choice in a crisis?
b. Another aspect of codifying ethical and moral norms is the application of optimization strategies to highly sensitive topics. (For example: the valuation of human lives in self-driving cars.) Would you be comfortable with an AI delivering such judgement? (Record this answer now, and then return to it after you finish the book)
💡 Reducing human life to a mathematical weight in a cost function can feel deeply dystopian. What happens when the cold optimization logic of a machine collides with the perceived immeasurable value of a human life? Fun fact: there is a concept called “value of a statistical life” (VSL). VSL is a measure of the average quantitative value of a human life, which is used in public policy applications.6
c. Suppose your answer to (b) is no, what drives the discomfort? How would you reconcile the desire for autonomy, resulting in the need for decision-making, and the discomfort associated with said autonomy?
💡 Unpack the friction between our desire to let machines handle dangerous tasks and our deep fear of the choices they might make? Do we demand a level of moral perfection from AI that we don't expect from human experts? For that matter, will the existence of a “human experts” solve moral dilemmas?
d. Recall the fact that AI models are statistical. Is there an acceptable margin of error for the integration of AI into systems that can cause direct impact to human lives?
đź’ˇ Because AI models are fundamentally probabilistic, the probability of eventually rolling an undesired outcome will tend towards 1, with many rolls. What percentage of failure will you find yourself comfortable with?
5 Moral Machine. (2016). Moral Machine. In Moral Machine. MIT Link6 Keller, E., Newman, J. E., Ortmann, A., Jorm, L. R., & Chambers, G. M. (2021). How Much Is a Human Life Worth? A Systematic Review. Value in Health, 24(10), 1531–1541. Link