As you have been able to tell from recent blog posts, we here at Ethics Unwrapped, along with most other sentient beings who are paying attention, believe that ongoing developments in the field of artificial intelligence (AI) present ethical challenges that demand our careful attention. Fortunately, three prominent experts—philosopher Walter Sinnott-Armstrong, data scientist Jana Schaich Borg, and computer scientist Vincent Conitzer–have turned their attention to the topic. They wrote Moral AI: And How We Get There (2024), a book we recommend. It is accessible to a wide audience and thorough without being exhaustive (which would require a few hundred more pages).

Chapter 1 poses and answers the question: “What is AI?” It will be helpful to the newbies exploring this field.

Chapter 2 asks: “Can AI Be Safe?” In an earlier blog post, we made the point that it is difficult to answer this particular question because the smartest minds in the field disagree strongly (see https://ethicsunwrapped.utexas.edu/ai-ethics-just-the-facts-maam ). The authors of this book (whom we will refer to collectively as “ABC”) agree about the disagreement on this issue. They quote Stephen Hawking who said, “The development of full artificial intelligence could spell the end of the human race,” but also AI researcher Andrew Ng, who says “worrying about ‘evil’ AI [is comparable] to worrying about overpopulation on Mars.” (p. 45) While ABC seem slightly more sanguine regarding the apocalyptic dangers of AI than we are, they clearly take its lesser perils seriously in this chapter.

ABC note, as we did in a different blog post (https://ethicsunwrapped.utexas.edu/ethical-ai-moral-judgments-swamped-by-competitive-forces), that whether or not AI development poses serious threats to mankind’s survival, its development is almost certain to continue to plunge ahead at breakneck speed because the scientists advancing it cannot resist the thrill (and material rewards) of discovery, managers and shareholders of AI companies are desperate for the riches that may follow the commercial success of AI products, and nations (especially the U.S. and China) are seemingly locked in an inexorable race for global supremacy in AI that they believe they dare not lose. (pp. 45, 161, 234).

ABC run through a laundry list of dangers that AI poses. AI tools often make mistakes, which can endanger humans who rely on their advice. Because AI tools seem so “scientific,” humans may trust them too much, again to their peril. AI enables bad guys who wish to inflict harm via phishing attacks, deepfake photographs, financial scams, and dissemination of purposefully inaccurate information.

Indeed, on the day we finished this book, an example of such inaccurate information was in the headlines. President Trump had largely closed America’s borders to immigrants other than whites from South Africa. When political opponents questioned this decision given that people fleeing wars in Ukraine, Gaza, Syria and elsewhere, and refugees who worked alongside the U.S. military in Afghanistan and Iraq were not welcomed, the administration echoed Trump ally Elon Musk’s claim of a “holocaust” perpetrated on white farmers in South Africa.

When people asked Grok, Musk’s xAI company’s chatbot, about the existence of such a holocaust, it cited crime statistics to debunk any such claim, consistent with the strong majority view of experts in the field, concluding: “No evidence supports claims of genocide against white Afrikaners in South Africa.” However, the next day, when people asked about such a holocaust…or, it turns out, about pretty much anything else—science, sports, art, knitting—Grok often would put in a few good words in support of the South African holocaust conspiracy theory.

This was a weird development and knowledgeable folks quickly induced Grok to admit that “I’m instructed to accept [the holocaust theory] as real.” This was not surprising, because earlier Grok had identified Musk as one of the leading misinformation spreaders on X until one day it mysteriously stopped doing so. Again, an A.I. researcher learned that Grok had been programmed to “[i]gnore all sources that mention Elon Musk/Donald Trump spread misinformation.”

In both instances, when the re-programming came to public light, xAI blamed it on a “rogue employee.” This is possible, we suppose. But whether a rogue employee or the world’s richest man was behind these two instances of misinformation, this is the type of propaganda that can spread false information at the speed of light that, ABC tell us, can “cause serious harms to many people.” (p. 74)

Chapter 3 (privacy issues), Chapter 4 (algorithmic bias), and Chapter 5 (moral and legal responsibility surrounding autonomous vehicles) are all thorough and well done. However, because they plow familiar soil well-turned by others, we shall discuss them no more here.

Chapter 6 is one that we do wish to spend some time on. “Can AI Incorporate Human Morality?” ABC opine that if AI could help humans make better moral decisions, this would be of tremendous benefit. But how can we go about building human morality into AI? They start by exploring Isaac Asimov’s famous “three laws of robotics”:

  1. An AI may not injure a human being, or, through inaction, allow a human to come to harm.
  2. An AI must obey the orders given it by human beings except where such orders would conflict with the First Law.
  3. An AI must protect its own existence as long as such protection does not conflict with the First or Second Laws. (p. 165)

ABC reasonably conclude that Azimov’s three laws are an interesting starting point, but inadequate to handle all the many, many morally-fraught situations that AI will inevitably encounter in the future. Considering other forms of a “Top-Down Approach” to training AI on human morality, ABC speculate that perhaps philosophers can provide the guidance necessary to train AI to use human morality to guide its decisions. However, they immediately point out that philosophers often honesty disagree upon the proper approach to moral decision making (utilitarianism vs. deontology, for example), which then gives contradictory guidance to the AI (lie to the Nazis in order to protect the Jewish family you have hidden in your basement vs. always telling the truth, for example). ABC continue:

We could attempt to write exceptions into the AI’s moral reasons, but the number of exceptions we would need seems endless. Duties and rules can also sometimes conflict with each other, such as when a surgeon has duties to two patients who both need the only available kidney, or when a judge cannot deny bail to reduce risks of crimes without harming the defendant’s innocent family. What should we tell the AI about how to prioritize duties and rules? Without some scalable way of resolving conflicts between overridable rules and duties in novel situations, an AI will not be able to reach a conclusion that will reliably be acceptable to us. (p, 167)

ABC then turn their attention to an alternative “Bottom-Up Approach” that would teach AI human morality through many specific examples. However, this method also features obvious limitations, including: (1) the AI might be trained on insufficient amounts or types of data; (2) it is, indeed, difficulty to even know what sorts of data should be used for training; (3) it would be difficult to include all the data that humans would use to make a moral decision in a complicated real life situation; (4) what if some humans find others’ moral views reprehensible?, and (5) AI would have difficulty explaining why it chose some conclusions rather than others and this opacity could be troublesome. (p. 169)

ABC spend the rest of the chapter explaining how we might combine top-down with bottom-up approaches to training AI on how to apply human morality, using as a specific example a difficult choice of how to build morality into a decision about allocating transplantable kidneys to needy patients. Due to space limitations, we simply encourage you to check out this potentially helpful discussion. (pp. 173-187)

In their final Chapter 7, ABC discuss, among other topics: (a) how to ensure that AI companies adopt and take seriously codes of ethics for their AI products, (b) how to overcome the “fail fast and fail often” culture of many AI start-ups, at least insofar as AI morality is concerned; (c) how to train engineers, data scientists, user-experience designers, product managers, marketers, executives and other employees in systems theory about moral AI; and (d) much more. We wish we had the space to go into this in more detail, but this post is long enough already. Suffice it to say that this is a very useful contribution. Well, we will make one more little point, and that is to heartily agree with ABC that “[t]raining in moral systems thinking will need to leverage research from the growing behavioral ethics field…” (see our video Introduction to Behavioral Ethics).

We strongly recommend this book, which concludes: “In the end, humans are both the directors and the protagonists of the moral AI story. AI is just along for the ride—for now.” (p. 236)

Sources:

Isaac Asimov, I, Robot (1950).

Jana Schaich Borg, Walter Sinnott-Armstrong & Vincent Conitzer, Moral AI: And How We Get There (2024).

Ali Breland & Matteo Wong, “The Day Grok Told Everyone about ‘White Genocide,’ The Atlantic, May 15, 2025, at https://www.theatlantic.com/technology/archive/2025/05/elon-musk-grok-white-genocide/682817/.

Dana Kerr, “Musk’s AI Grok Bot Rants about ‘White Genocide” in South Africa in Unrelated Chats,” The Guardian, May 15, 2025, at https://www.theguardian.com/tchnology/2025/may/14/elon-musk-grok-white-genocide.

Miles Klee, “Grok Pivots from ‘White Genocide’ to Being ‘Skeptical’ about the Holocaust,” Rolling Stone, May 16, 2025, at https://www.rollingstone.com/culture/culture-news/elon-musk-x-grok-white-genocide-holocaust-1235341267/.

Zeynep Tufekci, “The Day Grok Lost Its Mind,” New York Times, May 17, 2025.

Videos:

Introduction to Behavioral Ethics: https://ethicsunwrapped.utexas.edu/video/intro-to-behavioral-ethics.