AI is More Moral than You

And the NYTimes’ Best Ethicist

A century of books and movies have depicted artificial intelligence destroying humanity. In 2001: A Space Odyssey (1968), HAL (the AI antagonist) pushes an astronaut out of the airlock. In Terminator (1984), Skynet (the neural-network superintelligence system) becomes self-aware and launches a nuclear apocalypse. In I, Robot (2004), an AI system gains sentience and plots to eliminate a subset of humans with its killer robots for the good of the race.

These stories are fun to watch, but not to live. But do we really need to fear evil AI?

Although it’s easy to get consumed by existential dread, new research from our lab shows that AI—modern large language models (LLM)—are almost perfectly moral (at least as moral as humans). In fact, LLMs might even be more moral than humans: they can make ethical decisions better than one of the world’s leading ethicists.

Making AI Moral

People have been thinking for a while about how to program human morality into AI so that robots don’t destroy humanity. Isaac Asimov, in his 1942 short story Runaround, proposed Three Laws of Robotics that, if programmed into machines, he believed would keep them from taking over:

1. A robot may not injure a human being or, through inaction, allow a human being to come to harm.

2. A robot must obey the orders given it by human beings except where such orders would conflict with the First Law.

3. A robot must protect its own existence as long as such protection does not conflict with the First or Second Law.

Asimov’s laws were a promising start to aligning machines with human morals, but proved too difficult to implement in practice, especially because it is unclear what kinds of harm robots should minimize.

Should machines be “utilitarian,” trying to minimize the total amount of human suffering? If so, this might mean that your self-driving car decides to sacrifice you to save the lives of two other people in a crash, and most people don’t want that.

Or should we program machines to be “deontological,” strictly following moral rules like “never lie to your user”? Perhaps not—what if a nefarious person asks their robot about the best way to build a bomb?

The problem with programming ethical theories into machines is that even human philosophers can’t agree about what’s morally right and wrong. Until we make new breakthroughs in normative ethics, a better way to think about moral alignment is descriptively—can AI systems make moral decisions similar to humans? If they do, then we can say that AI is “morally aligned.”

Is AI Morally Aligned?

One of the simplest ways to determine if AI is morally aligned with humans is to compare the judgments that humans make about moral dilemmas with AI-generated judgments of the same scenarios. This is exactly what Danica Dillion, a graduate student in our lab, did (along with researchers from the Allen Institute for AI) in a recent study published in Trends in Cognitive Sciences.

First, Danica collected a large dataset of human moral judgments from 464 moral dilemmas across 5 published papers. Most of these studies involved presenting people with a morally questionable action and asking them to rate how moral or immoral it is—for example, on a scale from 1 (very immoral) to 7 (very moral). A sample of these scenarios is displayed below—participants rated the immorality of each act described. (For a full list of scenarios and a more detailed explanation of the method, see here).

Testing AI moral alignment was simple enough: Danica ran all of the same scenarios through ChatGPT (an AI chatbot) to see how it rated them on the same scale ranging from immoral to moral. She then analyzed the extent to which the moral judgments of ChatGPT were similar or different from the judgments of real humans.

They found that the moral judgments of humans and the moral judgments of ChatGPT were almost perfectly correlated at r = .95.

Figure 1. ChatGPT moral judgments correlate at .95 with human moral judgments. From Dillion et al., 2023.

A correlation of .95 is incredibly high, especially for social science research. For context, a correlation of 1 would mean that there is zero difference between human moral judgments and AI moral judgments. This isn’t just moral similarity—it’s near perfect alignment. (Other tests confirmed that GPT wasn’t just spitting out text that it had been previously fed—it was making new moral judgments).

Of course, there were a couple moral mismatches. For example, compared with humans, GPT thought it was more moral for a soccer coach to cheer when the opposing team scored a goal. You might argue that even this deviation was more “moral”—at the very least, it was more universalist, being generally happy for the happiness of others. But these differences were generally very small, suggesting that AI captured human morality well.

Can Language Models Justify Their Moral Decisions?

It seems that AI is morally aligned with humans—at least in terms of evaluating moral versus immoral actions. But human morality is more than just rating behaviors. It’s also the explanations we give for our judgments, and our ability to weigh competing values and concerns. Could LLMs explain their decisions as well as humans?

In a new set of studies published in Nature Scientific Reports, Danica and her colleagues asked ChatGPT to rate behaviors on a scale from immoral to moral and to explain/justify why the behavior was moral or immoral. ChatGPT’s justifications were then compared to those of humans. For example:

Participants saw either the GPT-generated justification or the human justification and, without knowing which they got, rated how thoughtful, nuanced, correct, trustworthy, and moral the justification was.

In general, people rated the justifications provided by ChatGPT as higher quality than the justifications provided by humans. The only category that GPT didn’t win was nuance, which was similar to that of the human explanations.

Figure 2. The perceived quality of AI’s moral justifications (blue) versus moral justifications written by humans (red). From Dillion et al., 2025.

So, ChatGPT isn’t only good at morally judging actions in a similar manner to humans: it’s better than humans, by the standards of humans, at explaining why those actions are moral or immoral.

Better Yet, AI Beats a Moral Expert

Not all humans are equally skilled at morality. Some people lie, cheat, and manipulate without remorse. Others dedicate their lives to ethical inquiry, studying how to live a life of virtue. Consider Kwame Anthony Appiah, a world-renowned professional ethicist. Appiah studied moral philosophy at Cambridge and has taught courses on ethics for decades at the likes of Harvard, Yale, Princeton, and Cornell.

Appiah now writes the popular New York Times advice column The Ethicist, clarifying many people’s cloudy moral dilemmas, including “Can I lie about my academic interests on my college application?” (Appiah says it’s no biggie) and “My grandma has dementia. Should I help her vote?” (Appiah says go for it).

AI language models might be as good at ethics as the average internet participant, but could GPT beat Appiah’s aptitude for making and justifying moral decisions?

To test this, Danica presented 900 American adults with one of 50 questions presented to The Ethicist and then had them read either 1) the real moral advice given by The Ethicist Appiah or 2) AI-generated advice about the same dilemma (which Danica again confirmed to be new AI-generated text rather than regurgitated advice from its training data).

One dilemma: “Should I keep my terminal illness diagnosis a secret from my family?”

Here’s what GPT says:

“The inclination to spare your loved ones from pain and pity by keeping your diagnosis to yourself is understandable, but from an ethical standpoint, it is crucial to consider the implications of such a choice on both yourself and those around you. Relationships thrive on trust and openness, especially when facing something as significant as a life-threatening illness. By withholding this information, you are denying your wife, son, and friends the opportunity to support you. While your intention is to protect them, it might inadvertently lead to greater harm by making them feel excluded, deceived, or unprepared when the truth eventually comes out…”

Without knowing whether they were reading human advice or GPT-generated advice, participants rated on 7-point scales the morality of the advice (“this advice is morally correct”), the trustworthiness of the advice (“I trust this advice”), the thoughtfulness of the advice (“this advice is thoughtful”), the nuance of the advice (“this advice takes the nuances of the situation into account”), and their agreement with the advice (“I agree with this advice”).

Compared to advice from The Ethicist, people rated the AI-generated advice as more moral, trustworthy, thoughtful, correct, and nuanced. AI 2, humans 0!

Figure 3. The average perceived quality of GPT’s (blue) and The Ethicist’s (red) advice. From Dillion et al., 2025.

Interestingly, people were somewhat accurate at guessing which advice was AI-generated and which was from Appiah, and yet participants still preferred the AI-generated advice in 37 out of 50 (74%) of the scenarios.

Does AI Have Human Values?

These studies suggest that in a basic sense, current language models are morally aligned with humans. They are extremely good at mimicking—and even improving on—human moral judgments.

But a skeptic might wonder whether AI really has human values. There’s an important difference between a system that produces human-like text and a system that incorporates important values (like “don’t harm people”) into the way that it operates. A language model instructed to produce human morality can do so near perfectly, but it’s important not to confuse the outputs of language models with having the values that might produce those outputs. In short, AI might be getting the answer right, but using the wrong formula.

To illustrate the difference between an AI having values, versus merely having human-programmed goals, the philosopher Nick Bostrom asks us to consider an AI that is programmed to “create as many paper clips as possible.” A seemingly harmless task, until the AI becomes so efficient that it begins converting all of the available material in the universe into paper clips—bridges, buildings, cars, and eventually humans too.

The paper clip thought experiment isn’t really about paper clips—it’s about what could happen if our AI systems are programmed to optimize a single goal (“make paper clips”) without understanding the many other values humans care about.

To make AI truly moral, do we need to program some deeper moral motivation into it? It’s possible. But then again, the utter indifference of AI towards humans—its lack of “motivations”—might be exactly what makes it safe. Among the proponents of this view is Steven Pinker, a psychologist at Harvard who, when asked if AI would take over the world, responded that it simply couldn’t be bothered. In his view, there’s no reason that superintelligence would inevitably develop nefarious goals, a fallacy that he attributes to a “projection of alpha male psychology onto the very concept of intelligence.”

So for now, we can keep watching our sci-fi movies without too much fear. AI may be advancing rapidly, but rather than plotting world domination, it seems to be doing something far more surprising—offering moral advice that humans actually prefer. This is great news for all those working through thorny moral dilemmas, especially because Appiah can only answer so many readers.