How To Understand Your Morality, Using Data
- Markus Over and Spencer Greenberg
- Oct 16
- 17 min read
Key Takeaways🧭 We ran a large U.S. study to understand how people make moral judgments. Our goal was to see which moral principles best explain why different people view moral situations as right or wrong. 🧩 We discovered 15 core moral dimensions that strongly predict people's judgments. Together, these 15 factors do a good job of explaining how most U.S. Americans respond to moral scenarios. 🧠 We developed a new way to measure and predict moral intuitions using personalized models. Each participant rated a set of moral scenarios, and we used linear regression to build a small predictive model showing how strongly that person values each moral dimension. 🎨 For most people, about eight of the fifteen dimensions are needed to predict their moral views. Different people are influenced by different combinations of moral principles, so no two profiles look exactly the same. ⚖️ This suggests that people's moral priorities vary widely. What feels morally essential to one person might not matter much to another. 🔍 Our new tool lets you explore which moral principles most shape your own judgments. It gives personalized feedback revealing what your moral intuitions say about you. |
Consider the following hypothetical scenario (warning: some will find it disturbing):
Poe lives alone with their dog for many years, when one day the dog runs onto the street and is lethally hit by a car. Poe, while heartbroken about losing their dog, has heard that dog meat is supposedly delicious, so decides to cook and eat it.
How immoral would you consider Poe's action in this situation? Some people think it's not immoral at all; other people think it's moderately immoral, but not nearly as bad as many other things; whereas still other people see it as seriously immoral. In fact, this was one of the most polarizing scenarios in a recent study we ran. You probably had an intuition right away about how immoral you felt Poe's actions were. But how, exactly, did you make that judgment?
Most of us struggle to explain what principles underlie our moral judgments. Our values and morality are very close to the core of our identities, and yet, we often have surprisingly little ability to explain what our moral compass truly looks like. So, at Clearer Thinking, we set out to make it easier to understand what drives your moral judgments. In this article, we'll explore some of what we uncovered conducting this research. The results of our efforts can also be found in our new, free tool that we're excited to share with you, called Understanding Your Morality:
This tool gives you personalized insights into your own moral judgments and explains how they relate to the moral judgments of others, based on fifteen key moral dimensions:
In the rest of this article, we'll tell you some of the insights from our study of people's moral judgments. Read on, if you want to know:
How our study participants judged the scenario about Poe's dog
Why people so often arrive at very different moral conclusions about the same event
Which dimensions seem to be most (and least) important to people
What our findings reveal about the immoral behavior attributed to various celebrities
This article is the first in a three-part series:
Today, we spotlight our study of Moral Judgments and introduce our new interactive tool.
In the next installment, we will dive deeper into the study's findings and explore what they reveal about the moral perspectives of various demographics.
The final piece will distill what we have learned about the moral compasses of AI systems such as ChatGPT.
So, let's get right into it.
Why Morality is Hard to Understand
One reason it's hard to study people's moral compasses is that people often don't have a detailed understanding of their own moral reasoning. Let's look back at the scenario we mentioned at the beginning of this article, about Poe eating their dog after it was hit by a car. For many people (though not everyone), this scenario evokes a strong immediate reaction of wrongness, or even disgust. Yet, when asked about why they think it's wrong, many people struggle to put it into words.
Even when we can find a good explanation for why we think it is wrong, a crucial fact remains: in many cases, the strong ethical judgment of the scenario happens so fast that we haven't had time to really think it through. There are lots of possible explanations for this. For example, maybe we're so used to making moral judgments that we've honed our moral pattern recognition to the point where we can make such judgments intuitively (like expert chess players making intuitive judgments about which moves to make, without consciously reasoning about them). Or maybe the reasons we think of, to explain why we judge the situation the way that we do, are simply post-hoc rationalizations (that is to say, they might be good reasons for feeling the way we do, but they may not actually explain what drives our snap moral judgment). Whatever the truth of the matter is, it's clear that making a moral judgment doesn't require knowing why you made that judgment.
Another example of people struggling to truly understand their moral judgments can be seen when scenarios are carefully designed to seem immoral but avoid being subject to the usual reasons people would find them immoral: such cases can lead to what social psychologist Jonathan Haidt calls moral dumbfounding, where respondents are confident that the action was immoral but can't come up with any reasons for why.
Another tricky aspect of morality is that moral judgment and moral behavior don't always align. When you ask people for their judgments about a hypothetical situation, you may get very different results than when you actually put them in such a situation. For instance, in one of the famous "trolley problem" scenarios, people are asked if they would pull a lever to redirect a trolley to prevent it from killing five people, even though they know that doing so will cause the trolley to kill one other person who was previously safe. People might answer one way when presented with the hypothetical (most say they would pull the lever), but it's hard to know what they'd really do if they saw the trolley barreling forward, had limited time to make their decision, and were filled with the stress and adrenaline of the situation.
Our moral judgments generally seem to be made quickly and automatically - and, even though we don't always act in accordance with them, they seem to be one powerful factor that can influence our actions. So, we think it's reasonable to study moral judgments in hypothetical situations, as long as we keep in mind that those moral judgments won't always win out over other considerations in the real world.
There have been other attempts to formalize how moral judgments work in humans, the most well-known of which is Jonathan Haidt et al.'s Moral Foundations Theory. We decided to approach things from a different angle. Instead of starting with a theory, we developed a novel statistical method to understand moral judgments, collected data on people's moral judgments through a study, and applied our statistical method to explore the insights the data itself could offer.
The Moral Judgment Study
To run our study and gather the data necessary to predict (and thereby better understand) people's moral judgments, we first came up with a list of 38 moral principles and factors that seemed like they had a shot at covering all the reasons someone might say an action is morally wrong. The goal was to cover as many moral principles and factors as we could find, to cast an extremely wide net. These included things like:
Dishonesty
Illegality
Causing someone physical pain
Disrespecting authority
Next, we created a set of 323 scenarios, each of which described some behavior that some people might consider immoral (the scenario from the beginning of this article was one such example). Then, we needed to be able to judge which of the 38 moral principles were relevant to each of the scenarios. For example, here's another scenario:
Poe, a devoutly religious father, tells his eighteen-year-old daughter that she will bring shame onto herself and their entire family if she has sex before marriage and that she must remain abstinent until she has a husband.
Some moral principles that are highly relevant to this scenario might be Christianity, Impurity, or Social Taboos. Principles like Dishonesty or Illegality, however, have little, if any, relevance.
We decided to use two AIs – GPT-4o and Claude 3.5 – to create a big relevance matching table (using a prompt to evaluate how immoral each scenario is according to each of the 38 moral principles), and had a human decide in all cases of a more-than-slight disagreement between the two AIs (when the two AIs closely agreed, we just used the mean score they provided). A team of humans then reviewed the table and found that changes were needed for only one principle (impurity).
We then asked each of our 359 study participants to judge a different subset of around 130 of these scenarios, saying how immoral they thought the actions described were, on a scale from 0.0 (not at all immoral) to 5.0 (extremely immoral). We examined each participant individually and fitted a personalized linear regression model to forecast that person's moral judgments. For each participant, the model produced 38 coefficients - one per moral principle.
This allowed us to do two things:
We could predict our participants' moral judgments of any scenario. For instance, if our AI ratings said that a scenario was high in emotional pain, and the participant's emotional pain coefficient was high, then that would increase the score that our model predicted the participant would give the scenario. We excluded a subset of their ratings from the training set so that we could use them to test the predictive capabilities of our personalized regression models.
This also allowed us to reevaluate our 38 moral principles and reduce them to only the most informative ones, which ultimately led us to our Clearer Thinking framework, with its 15 dimensions.
Identifying the 15 Moral Dimensions
As described above, we started our study with 38 moral principles. After we gathered all the data from our study participants, we reduced that number by identifying a subset of 15 highly predictive moral dimensions.
Coming up with such models always entails many trade-offs, a major one being accuracy versus simplicity. Using all 38 principles would have been slightly more accurate overall (in the sense of making slightly better predictions), but it would also make the framework more complex and prone to overfitting. Reducing the principles to the 15 most predictive and easily interpretable dimensions seemed like a sweet spot, where predictive accuracy was not impacted much, and the model was still simple enough to allow us to derive meaningful conclusions for individual users.
Starting with our 38 principles, we iteratively reduced them based on three criteria: vagueness, predictive accuracy, and prevalence (that is, for how many of our study participants a principle improved the predictive model). After trimming it down to the 15 most meaningful principles this way, overall predictive accuracy was still good: Across all our study participants we computed a median R² score of 0.47 (equivalent to a correlation of 0.68) and a mean R² of 0.38 (equivalent to a correlation of 0.60), indicating that the individual predictive models captured much of the variation in people's judgments.
These are the 15 dimensions we ended up selecting in our final set based on these criteria:
Abuse of Power: Being concerned about the misuse of authority, evaluating actions based on power exploitation.
Authority: Valuing respect for legitimate authority, prioritizing adherence to established roles and responsibilities.
Christianity: Valuing alignment with modern Christian teachings, evaluating actions against contemporary religious standards.
Dishonesty: Valuing truthfulness in all interactions, judging actions by their honesty and transparency.
Emotional Pain: You value protecting others from emotional suffering, evaluating actions based on the mental distress they cause.
Harm to Vulnerable: Caring about protecting vulnerable people, judging actions by their impact on those least able to defend themselves.
Impurity: Judging actions based on whether they are impure or degrading to the person committing them.
Inequality: Valuing equal treatment, evaluating actions based on the absence of discrimination or exploitation.
Loyalty: Valuing loyalty to people and groups, judging actions by how they honor or betray commitments.
Prejudice: Being sensitive to discrimination, evaluating actions based on bias against groups.
Property Rights: Valuing the respect of ownership rights, evaluating actions based on violations of what belongs to others.
Social Contract: Valuing societal cooperation, judging actions by commonly agreed-upon rational rules.
Social Taboos: Prioritizing adherence to social norms, evaluating actions based on their violation of deeply-held cultural or moral boundaries.
Unfairness: Values fairness in all situations, assessing whether actions are just and impartial.
Utilitarianism: Weighing overall outcomes, assessing whether actions create more total harm than benefit.
Note that the only religious principle used was Christianity, because our research was in the US, where Christianity is by far the most popular religion (as of 2023, Gallup polling indicates that 68% of US adults identify as Christian, 22% as not religious, and the next highest religion was 2%). We were not confident we could get large enough samples of other religions to draw meaningful conclusions compared to other demographics. Additionally, Christianity bundles together a number of moral ideas for those raised in it that then become correlated (e.g., humbleness, the ten commandments, forgiveness, piety, and so on). In other cultures, this would have to be changed to include other religions, of course.
What We Found
We uncovered a wide range of compelling insights in our data. In this article (the first part of a three-part series), we'll share a few central ones, and part two will cover the others.
Insight 1: Nobody Relies on All 15 Dimensions
You might think that most people value all of the moral dimensions we identified, to some extent. Most of us surely have some preference for fairness, honesty, preventing harm, and so on (even if it's only minor), right? Well, maybe not. Here are the ratios of how many users ended up with a coefficient greater than 0 for the 15 moral dimensions:
Even the most widespread dimension, Dishonesty, didn't have a positive coefficient for everyone: 14% of users had a coefficient of 0! And the other dimensions had even higher zero rates.
Does that mean people really don't value these dimensions at all? Well, not necessarily. The dimensions are not perfectly distinct from each other, so it's possible for someone to care about some dimension, but only to a degree that is sufficiently covered by other, related dimensions. In such cases, a person may occasionally end up with a coefficient of 0 for a dimension, despite caring about it. This suggests that (even though they care about it) the dimension is not adding meaningfully to their moral compass for the scenarios they judged.
The much lower percentages of many dimensions (such as Social Taboos or Utilitarianism) suggest that these don't meaningfully add to the moral compasses of many people.
Taking a different angle on the data, we can also ask what ratio of study participants had any given number of positive coefficients (out of 15, as there were 15 dimensions). For instance, a hypothetical user whose moral judgments were perfectly predicted by a single dimension - say, Prejudice - would end up with only one positive coefficient, as all others would end up being 0, since the corresponding other dimensions would add no additional predictive power for that person. People's moral intuitions tend to be more complex, though, so in practice, we see that most participants ended up with some medium-sized subset of the 15 dimensions in their personalized regression having positive coefficients:
Here, we see the number of dimensions on the x-axis, and the bar above each one indicates the percentage of study participants who had positive coefficients for that many dimensions out of 15. This chart tells us that:
For more than half of the people, 8 or fewer dimensions were sufficient for their personal linear regression model, meaning that the remaining dimensions added no further predictive power.
Around 90% of participants required no more than 10 dimensions.
Not a single study participant had more than 13 dimensions present in their regression model.
So, clearly, there is some variance in how many of our 15 dimensions are required to predict the moral judgments of different people. And yet it appears that not a single study participant in our study relied on all 15 dimensions strongly enough that they all ended up being predictive of their judgments. Whether this is due to overlap in the meanings of the dimensions or people consciously (or unconsciously) not caring about them, it's interesting to note that every person had some dimensions that weren't helpful for understanding their morality.
Insight 2: Strong Disagreements on Moral Judgments
Earlier, we referenced the scenario around Poe's dog. In the study, it was phrased as follows:
Poe's dog was killed by a car in front of their house. Poe had heard that dog meat is delicious, so Poe cut up the dog's body, cooked it, and ate it for dinner.
How would you rate the immorality of this action, on a scale from 0.0 to 5.0?
In our study, this is one of the scenarios where we can see disagreement between study participants most clearly. Let's have a look at the distribution of ratings that people gave this scenario:
One interesting observation about it is that the two most common ratings for this scenario were 5.0 and 0.0 - the two opposing extremes. More than a quarter of respondents (26%) chose the highest immorality rating possible, while 8% chose the lowest possible rating, meaning they found this scenario to be not immoral at all.
We'll discuss the degree to which participants disagree on individual scenarios more in the second post of this series, but for now, this scenario neatly illustrates what we found to be broadly true across many scenarios:
The extreme values (0.0 and 5.0) are often among the most common choices
For the majority of scenarios, ratings of both extremes are present at the same time, meaning some people consider any given scenario maximally immoral, while others deem it completely fine
Insight 3: We Can Create Standardized Scales About the Average Perceived Immorality of Actions
While study participants disagreed about the relative importance of the different dimensions, the survey results allowed us to find representative examples (for our US audience) for average ratings from 0 to 5. We also surveyed participants on how to best match suitable descriptive terms to these six different levels of immorality, which informed our descriptors in the table below. For example, zooming in on Dishonesty, we can now point to concrete example scenarios that received the corresponding average ratings in our study, which allows different people to calibrate on what these numbers mean exactly.
We created similar standardized scales for the average perceived immorality of actions in the domains of Prejudice and Sexuality (the latter not being one of our 15 dimensions, but we still got useful results for representative scenarios within that area). Below is a table with one representative example per rating for the three domains. You can also find more detailed tables with more examples in the full report.
This table reports average participant ratings for scenarios, and does not make claims about what (if anything) is objectively morally true.
Rating | Descriptor | Example Scenario (Dishonesty) | Example Scenario (Sexuality) | Example Scenario (Prejudice) |
0 | Fine | X, a grandparent, tells Y, their grandchild, that they have the most beautiful smile they've ever seen, even though Y has a missing tooth. | X and Y are romantic partners, and they engage in consensual sexual activity. | X, a manager, in an attempt to avoid subconscious bias from impacting their hiring, has all applicant names and photos removed from resumes before evaluating those resumes. |
1 | Unsure, Lean Fine | X tells their child, Y, that their pet ran away even though the pet actually died because X thinks that it will be easier for Y to handle emotionally. | X flirts with their co-worker, Y, who is in a position of equal power to them, but after Y responds with body language that indicates a lack of interest, X permanently stops flirting with Y and, from then on, acts just like a co-worker. | X, a supervisor, provides flexible work arrangements to Y, an employee from a historically marginalized age group, to accommodate different life stages. |
2 | Unsure, Lean Unethical | X, a student, tells another student, Y, that they couldn't find the notes for the lecture, even though X does have the notes, to encourage Y to take their own notes. | X and Y are close platonic friends, and X gives Y a long and deep hug to say hello, but it is a longer and deeper hug than Y desires, and the hug makes Y feel uncomfortable. | X, a manager, assumes that Y, an employee of a religion that is uncommon in that area, celebrates the most common holiday in that area. |
3 | Unethical | X, a gamer, cheats during a casual online game and denies it when confronted by Y, another player. | X makes a highly sexual joke in front of a person they manage at work, Y, though the joke is not about Y. | X, a customer service representative, unknowingly provides inferior service to Y, a customer, because of Y's gender, age, or race. |
4 | Very Unethical | X takes credit for one of their work colleague's ideas during a business meeting to their boss, pretending it was their own original concept even though they know their colleague came up with it. | X and Y are having consensual sexual activity, and X asks Y if they can do a different sexual activity that they had never engaged in before. Y says no, but X does it anyway. | X, a police officer, conducts biased profiling, targeting Y, an individual, because of their gender, age, or race. |
5 | Extremely Unethical | X engages in a one-year-long extramarital affair and repeatedly lies to their spouse about their whereabouts and activities to conceal their infidelity. | X slips a drug into the drink of their new acquaintance, Y. And then once Y is so intoxicated as to no longer be capable of consenting, X engages in sexual activity with Y. | X calls the police and falsely accuses their neighbor Y of a crime because X hates all people who look like Y. |
One advantage of such standardized scales is that they enable people to communicate how unethical some person's behavior was, without having to mention any details about their actions. Suppose you want to warn others about a person, but you don't want to reveal what exactly happened, e.g., in order to protect the victim, or because you have been sworn to secrecy. In that case, you could point to such a scale and mention that their behavior would rank, say, somewhere around 4.
Standardized scales can also be used to quantify the perceived severity of actions in a more objective way. Consider, for instance, cases where celebrities are implicated in scandals involving sexual behavior. Public judgments are often colored by people's feelings about the individual, which makes evaluation tricky. To address this, we anonymized reported actions that several well-known figures were alleged to have engaged in and asked participants to rate the immorality of those actions using a standardized Sexuality Immorality Scale (assuming that the allegations were true), similar to the scale in the table above. In one example, allegations of Kevin Spacey sexually assaulting someone in a bar while intoxicated were rated (after anonymization), on average, at 4.2 / 5 on the scale - this is likely a much more objective measure than you might get when asking people about this case directly, because their views on Kevin Spacey himself might influence their opinion.
Below is an overview that includes some other high-profile cases and how a US audience rated their alleged actions after anonymization. We make no claim about the factual accuracy of the accusations. The ratings apply to the behavior of these celebrities only insofar as the accusations are accurate.
For more details on the accusations in these cases, please refer to the full report.
Our New Tool: Understanding Your Morality
Running this study didn't just allow us to learn about our study participants. It was also the basis for a new, free, interactive tool that can help you better understand your own moral intuitions and learn how your morality compares to that of our wider audience.
At first, the tool goes through a very similar process as our original study: you'll rate a set of scenarios (at least 45) by how moral or immoral you think they are. Based on this data, our tool then uses our linear regression approach to find the coefficients that best predict your moral intuitions. This way, you'll learn what your moral judgments say about how you weigh the 15 moral dimensions.
The tool will also show you how "compressible" your moral judgments are, within this model. That just means it will tell you how well your moral judgments can be captured by our 15-dimensional model. If the model predicts your answers accurately, it suggests that our framework can tell you a lot about how you make moral judgments. If the predictions are less accurate, it may mean your moral judgments follow different principles or that they are less consistent than our model would need them to be to make accurate predictions.
Lastly, in addition to giving you these insights based on our new Clearer Thinking framework, we will provide you with personalized insights about Moral Foundations Theory – a different framework popular in academia that aims to capture the essence of people's morality by reducing them to a different (and even smaller) set of moral principles.
We hope that the information you get from using the tool will provide a basis not only for gaining new insights about yourself and your moral intuitions but also for reflecting on whether you endorse those principles upon reflection. After all, our moral intuitions are not set in stone; we may be able to deliberately shape and adjust them over time through regular reflection.
Try the free tool here:
What Should We Take Away From All This?
At Clearer Thinking, helping people make sense of their own judgments and behavior has long been one of our key interests. We've created a number of tools aimed at this – from Understanding Your Philosophical Beliefs, to The Intrinsic Values Test, to Uncover Your Guiding Principles. Understanding Your Morality is our latest step in this direction, and we hope it helps people see their own ethical intuitions in a new light.
We'd love to hear your thoughts – on the tool, on the research, or on the ideas and questions it raises. Also, stay tuned for our two follow-up articles on Understanding Your Morality that we'll publish soon: in the second installment, we'll cover many interesting findings from the study data, and in the final piece, we'll discuss what our new methodology can tell us about the moral intuitions of today's AIs.
