1. The Gradualist Hypothesis

The gradualist hypothesis is of recent origin. John Stuart Mill writes that, ‘The creed which accepts as the foundation of morals, Utility, or the Greatest Happiness Principle, holds that acts are right in proportion as they tend to promote happiness, wrong as they tend to produce the reverse of happiness’ (Mill Reference Mill1863: 9, my emphasis). If we read this literally, it follows that acts that produce half as much happiness as the optimal alternative(s) are half-right. However, Mill's gradualism seems to have been a slip of the tongue. In the rest of his writings, he never discusses this implicit notion of degrees of rightness, and he never uses the idea that rightness comes in degrees for supporting other claims or for rebutting objections to his ethical theory.

Many modern consequentialists believe that an act is right just in case no alternative brings about better consequences, and that every act that is not right is wrong. This criterion of rightness is obviously incompatible with the gradualist hypothesis. No matter what the consequences are, every act will be either right or wrong in the binary sense. This holds true even if some consequences are incomparable or on a par, because an act is considered to be right as long as its consequences are not worse than those of any alternative act. However, some consequentialist theories do allow for gradable notions of right and wrong. According to the multidimensional account of consequentialism that I discuss in The Dimensions of Consequentialism (Peterson Reference Peterson2013), an act's rightness or wrongness depends on several irreducible aspects, such as the total wellbeing produced by the act and the degree of equality with which wellbeing is distributed in society. According to this theory, moral rightness comes in degrees whenever no act is optimal with respect to all moral aspects.

In the example in table 1 (see Peterson Reference Peterson2013: 13) the first act, A, is optimal with respect to total wellbeing but scores poorly with respect to equality. The opposite is true of C. Act B is not optimal with respect to any of these aspects, but it scores pretty well with respect to both of them. Multidimensional consequentialists believe that all three alternatives are somewhat right and somewhat wrong, but B is right to a higher degree than A and C all things considered. Multidimensional consequentialists may, for instance, argue that C is almost entirely right (right to degree 0.99 on a scale from 0 to 1) while A and C are half right and half wrong (right to degree 0.5).

Table 1 Multidimensional consequentialists believe that B is more right than A and C, but not entirely right.

Some deontologists also accept the idea that rightness and wrongness come in degrees. In Moral Uncertainty and Its Consequences, Ted Lockhart claims that ‘actions come in varying degrees of moral rightness between “right” and “wrong”’ (Lockhart Reference Lockhart2000: 75). He argues that this view is particularly attractive for those who accept W. D. Ross's theory of prima facie duties: ‘Even for some deontological theories, we have no great difficulty entertaining a many-valued concept of moral rightness. A prima facie duties theory, for example, may readily regard actions that, ceteris paribus, accord with more prima facie duties than others as having greater moral rightness. Moreover, for some prima facie duties, it seems that we can conform to or violate them to greater or lesser degrees. Examples of such duties in W. D. Ross's theory include beneficence, nonmaleficence, and justice’ (Lockhart Reference Lockhart2000: 80).

Lockhart's proposal is a revisionary extension of Ross's theory, which Ross himself would not accept. Ross is committed to what Rob Lawlor calls a ‘threshold account’ of rightness, just as mainstream consequentialists: acts that have the right kind of properties make it across the threshold, and those acts are right (or permissible) in the binary sense (Lawlor Reference Lawlor2009: 21). All other acts are wrong. For mainstream consequentialists, the right-making property is that of having optimal consequences, and for Ross the threshold a right act has to pass is to be fitting to perform in light of all applicable prima facie duties.

Somewhat surprisingly, Lockhart's own proposal is a hybrid view that collapses into a threshold account when only two alternatives are available: ‘x has greater degree of moral rightness than y in situation S for agent A just in case, if x and y were the only alternatives open to A in S, then x would be morally right for A in S and y would be morally wrong for A in S’ (Lockhart Reference Lockhart2000: 81).

Lockhart's thesis is compatible both with Hurka's commonsensical version of gradualism mentioned above, and with Mill's gradualist account of utilitarianism. However, as pointed out by Campbell Brown, Lockhart's reductive analysis is unable to compare degrees of rightness across different situations. Brown notes that, ‘a person who uses her mobile phone while watching a movie in a public cinema may be acting wrongly, but surely not as wrongly as one who does the same while driving a heavy lorry on a motorway thereby causing a major accident’ (Brown Reference Brown2016: 27). The problem for Lockhart is that the acts described by Brown are performed in different situations and can therefore not be compared according to his criterion. To address this problem, Brown develops a more general reductive analysis of degrees of rightness that tallies well with Mill's gradualist theory mentioned above (see Brown Reference Brown2016).

2. Three Experimental Studies

Should we accept the idea that right and wrong are gradable concepts? Almost everyone agrees that meaning tracks use in the sense endorsed by modern linguists. By studying how people use words and phrases we can, under normal circumstances, make reliable inferences about their meaning. As noted above, this is an epistemic thesis. Whether meaning is use in a strict semantic sense is another issue. The truth of the gradualist hypothesis can thus be assessed by testing the following empirical conjecture:

To test this conjecture, I distributed a series of online questionnaires to three sets of respondents: students at Texas A&M University taking classes in engineering ethics in spring 2019 (n = 715) and spring 2020 (n = 578), and a group of US citizens who voted in the 2016 election (n = 182). Students took the surveys for credit (about 0.5 percent of the total course grade) while respondents in the third study received between $2.65 and $3.00 in compensation. In all three studies, respondents were invited to answer up to twelve questions. Both the order of the questions and the answer options were randomized. At the request of the Institutional Review Board of Texas A&M University, which approved the study, no demographic information was collected.

Respondents were presented with up to five different types of tasks: abstract, semi-abstract, concrete, comparative, and open-ended tasks (detailed below). By using several different kinds of measures that test a single hypothesis the validity of the measurement instrument can be assessed. If the validity is high, we should expect the results to be roughly the same for all types of tasks.

3. Abstract Tasks

Subjects were invited to evaluate abstract statements about ethics, mathematics, colors, scientific evidence, and scientific facts on a seven-point Likert scale ranging from ‘strongly disagree’ (1) to ‘strongly agree’ (7). The following example pertains directly to the gradualist hypothesis: A1.

The following abstract statements were used as reference points for comparative purposes: A2.

All items were followed by a simple comprehension check. Responses from subjects who did not answer it correctly have been excluded from the analysis.

The data sets are not normally distributed. It is therefore appropriate to perform a Mann-Whitney U-test. This is a nonparametric test for the null hypothesis that the distributions of two data sets are identical and therefore have the same median value. Table 2 summarizes the results for Study 2, which is less likely than the others to yield significant results due to its smaller sample size. The Mann-Whitney U-test indicates that moral rightness and wrongness is judged to come in degrees to a significantly higher extent (p <0.01, one-tailed) than truth in mathematics is judged to come in degrees, but there is no statistically significant difference between A1 (moral rightness and wrongness come in degrees) and A3 (colors come in degrees), not even at p <0.05. This indicates that rightness and wrongness is judged to come in degrees to approximately the same extent as colors are judged to come in degrees. The results of Study 1 offer additional support to this conclusion: all comparisons in Study 1 yield significant results (p < 0.01, one-tailed) except that between colors and moral rightness. However, in Study 3 the comparison between colors and moral rightness also yields a significant difference (p <0.01, one-tailed), indicating that the extent to which colors and moral rightness is reported to come in degrees is not precisely the same.

Table 2 Mann-Whitney U-tests for data in Study 2.

The abstract tasks also included the following general statements about scientific evidence and scientific facts:

In all three studies, the Mann-Whitney U-test for A4 versus A5 indicates that scientific evidence is judged to come in degrees to a somewhat higher extent than scientific facts in all three studies (in Study 2, U = 2615.5, p-value < 0.00001; significant at p < 0.01; one-tailed). Moreover, scientific facts are also judged to come in degrees to a significantly higher extent than mathematical truths in all three studies (in Study 2, U = 1454.5, p-value < 0.00001; significant at p < 0.01; one-tailed), and moral rightness and wrongness are reported to come in degrees to a significantly higher extent than scientific evidence in all three studies (in Study 2, U = 3792.5, p-value = 0.00004; significant at p < 0.01; one-tailed). Finally, colors are reported to come in degrees to a significantly higher extent than scientific evidence in all three studies. (In Study 2, U = 2317, p-value < 0.00001; significant at p < 0.01; one-tailed).

These findings can be represented as a hierarchical order with four ordinal levels. Table 3 summarizes the results for Study 1 and 2. (As noted, in Study 3 colors come in degrees to a slightly higher extent than moral rightness and should therefore be represented on a separate sublevel.) The difference in agreement in table 3 between all pairs of levels is statistically significant at p < 0.01, except that between scientific evidence and scientific facts, which is significant at p < 0.05.

Table 3 Four ordinal levels of agreement in Study 1 and 2. Level 1 is the highest level, and all pair-wise differences between items at different levels are statistically significant at p < 0.01; one-tailed.

An alternative explanation of the results in table 3 could be that respondents tend merely to report their belief in how much disagreement there is in a certain domain, not that the phenomena themselves come in degrees. To control for this, the following items were included in Study (table 4):

Table 4 Moral disagreement does not explain the findings.

These numbers indicate that respondents do believe that there is more disagreement on moral issues than on mathematical issues (U = 768, p-value < 0.00001; significant at p < 0.01; one-tailed). This is not surprising. However, respondents also believe that moral rightness and wrongness come in degrees even when there is no disagreement, and they do so to a much higher extent than for mathematical issues. (U = 1587.5, p-value < 0.00001; significant at p < 0.01; one-tailed). This casts doubt on the alternative explanation, but fits well with the gradualist hypothesis. There is little reason to think that respondents merely reported their belief in how much disagreement there is in a certain domain.

4. Semi-abstract Tasks

All three studies included two semi-abstract tasks in which subjects were invited to complete moral statements by selecting one of a set of pre-defined alternatives. The first of these tasks (n = 242, 287, and 90) was formulated as shown in table 5:

Table 5 Responses to S1. Lying in a situation in which doing so would bring about the best consequences is . . .

The gradualist analysis was the most frequently selected answer option in all three studies. The differences in gradualist responses in Study 1 (88.2 percent), Study 2 (75.0 percent), and Study 3 (58.9 percent) can be explained by the fact that respondents were presented with four answer options in Study 1, five in Study 2, and six in Study 3. In Study 3 about 17.8 percent chose the relativist answer option, which was not available in the other studies. The more options respondents have to choose from, the less likely is it that everyone selects the same option. If right and wrong had been binary concepts many subjects could arguably have been expected to favor the Kantian answer option (‘always morally wrong’) or the utilitarian answer option (‘always morally right’).

Study 2 included an answer option designed to capture Hurka's notion of degrees, according to which some acts are more importantly right or more seriously wrong. This answer was selected by no more than 5.8 percent of respondents.

In Study 1, a separate group of respondents was presented with the answer option ‘sometimes a bit right and a bit wrong’ instead of ‘sometimes right to some degree, but also wrong to some degree’. About 83.6 percent (n = 116) selected ‘sometimes a bit right and a bit wrong’, compared to 89.3 percent (n = 244) for the group presented with the option ‘sometimes right to some degree, but also wrong to some degree’. The difference between 89.3 percent and 83.6 percent is not significant (Χ² = 8.396, the p-value is < 0.039; not significant at p < 0.01). This is an indication of robustness. The results reported here do not depend on minor alterations of the wording of the gradualist hypothesis.

Unsurprisingly, the vast majority also reported gradualist responses for the second semi-abstract item S2 (n = 119, 288, and 90), as shown in table 6:

Table 6 Responses to S2. Exceeding the speed limit in an emergency is . . .

A possible explanation of why so many respondents in Study 3 selected the binary response ‘sometimes right and sometimes wrong, but never a bit right and a bit wrong’ (23.3 percent) or ‘always morally right’ (15.6 percent) could be that speeding, but not lying, is viewed as less morally problematic by experienced drivers. US citizens who voted in the 2016 election are on average older than college students and are thus more likely to drive.

5. Concrete Tasks

Study 1 included six concrete tasks in which respondents were invited to assess brief descriptions of particular acts. As these tasks were not designed to study respondents’ views on moral relativism or Hurka's hypothesis, the number of answer options was limited to four. In Study 1 the following task was evaluated by all respondents (n = 715): C1.

Although relatively little information is provided in the vignette, over ninety percent reported that John's act was wrong. The next item (C2, n = 363) serves as a reference point for what seems to be a case of someone doing something right: C2.

The gradualist hypothesis states that some acts are somewhat right and somewhat wrong, not that all are. Therefore, the findings for C1 and C2 neither refute nor confirm the gradualist hypothesis. However, data for the following tasks, C3–C6, support it. For these items, the gradualist answer option ‘What [Denise/the captain/Anna/Miriam] did was right to some degree, but also wrong to some degree’ was the most frequently selected answer in all three studies (see table 7.

Table 7 Relative frequencies for items C3 to C6. The four answers options were the same as in C1 and C2.

Table 8 summarizes pair-wise chi-square tests for all combinations of C1–C6. The degrees of freedom for all comparisons is df = 3, so for p < 0.01 the critical chi-square value is 11.34, and for p < 0.001 it is 16.27. Note that all pair-wise comparisons are statistically significant. However, the chi-square values for C1 and C2 stand out: they are ten to one hundred times higher than the values for all other items. (See the dashed box in table 8.) From a statistical point of view, the explanation is that C3–C6 are items in which the gradualist answer option is the most popular one; therefore, C3—C6 have more in common with each other than with C1 and C2.

Table 8 Pairwise chi-square tests for all combinations of C1 – C6. For p < 0.01 the critical chi-square value is 11.34, and for p < 0.001 it is 16.27 (df = 3). Note that all comparisons are statistically significant at p < 0.001.

It is also worth pointing out that the chi-square values for C6 are four to ten times higher than the corresponding values for C3, C4, and C5. What could explain this? Note that in items C3, C4, and C5 some widely accepted norm is violated for a good reason: By violating the speed limit, or by lying, the agent brings about good consequences for others. A large majority reported that such norm violations are somewhat right and somewhat wrong. However, in C6 the agent violates a norm for what appears to be a selfish reason. Fewer subjects considered this to be somewhat right and somewhat wrong, and about twice as many considered it to be wrong.

Combining the findings for the abstract, semi-abstract, and concrete tasks I found that only 4 percent in the largest study (Study 1, n = 715) persistently used right and wrong as non-gradable terms.

6. Comparative Tasks

Study 1 included a fourth type of task designed to test the gradualist hypothesis without including gradualist terms such as ‘degree’ and ‘somewhat right’ in the vignettes. If gradualist terms appear in questions or answer options, respondents might be more willing to apply such terms than they otherwise would (see Messick and Jackson Reference Messick and Jackson1961; see also the discussion below.)

Subjects were asked to make pairwise comparisons between items C1–C6 and a seventh item, C7:

The comparative task was formulated as follows:

Pairwise comparisons of seven items, C1–C7, requires 21 comparisons. Each respondent was invited to make four comparisons, which yielded 2830 comparative data points (n = 98 to 141). To verify that subjects understood the comparative task correctly, three identical comparisons were included in the questionnaire, C1–C1, C6–C6, and C7–C7. The average dissimilarity reported for these items were, 0.1, 0.2, and 0.3, which indicates a good understanding of the task. Table 9 summarizes the results:

Table 9 Average degree of dissimilarity, ranging from 0 (very similar) to 6 (very dissimilar).

Dissimilarities can be interpreted as distances in an n-dimensional geometric space. The more dissimilar two items are, the farther apart is their location in space. The twenty-one comparisons listed in table 9 can range over twenty dimensions, but by applying multidimensional scaling techniques, the dimensionality of a multidimensional dataset can be reduced (see Kruskal and Wish Reference Kruskal and Wish1978).

The aim of a classical multidimensional scaling is to represent the original data set by a new set of points in a smaller number of dimensions such that the Euclidean distance between each pair of points in the new set approximates the distance in the original multidimensional data set. Ideally, each pairwise distance (similarity) in the original data set (table 9) should be exactly the same as the corresponding Euclidean distance in the new representation. However, as we reduce the number of dimensions some minor errors will typically be introduced into the new representation. This is acceptable as long as the errors are small.

Figure 1 shows a classical multidimensional scaling of table 9. The maximum error is 0.36 units, which is a relatively large error (see below). When interpreting figure 1, it is important to keep in mind that item C1 is almost unanimously (91.3 percent) considered to be an example of wrongdoing, whereas C2 is widely considered (74.9 percent) to be an example of an agent doing something right. It is thus not surprising that C1 and C2 are located far apart along the x-axis. C7 can also be taken to be an example of someone doing something right, so it is equally unsurprising that C7 is located close to C2. The locations of C1, C2, and C7 can thus be explained without invoking the hypothesis that rightness and wrongness come in degrees.

Figure 1. A classic multidimensional scaling of table 9. The maximum error is 0.36 units. Recall from above that a vast majority (91.3 percent) thought the agent acted wrongly in C1, whereas the majority (74.9 percent) thought the agent acted rightly in C2. For the options in between the most popular answer was that the agent acted somewhat right and somewhat wrong: 69.3 percent for C3; 71.5 percent for C4; 62.9 percent for C5; and 43.0 percent for C6. No data are available for C7.

However, the locations of C3–C6 cannot be easily explained without assuming that rightness and wrongness come in degrees. Although these items are located somewhat to the left in figure 1, they are not close to C1 along the x-axis. Moreover, recall that C3–C6 are items in which it is widely believed that the agent's act was somewhat right and somewhat wrong (69.3 percent for C3; 71.5 percent for C4; 62.9 percent for C5; and 43.0 percent for C6). This fits well with their locations in figure 1: items C3–C6 are, literally speaking, located ‘between’ the entirely wrong act in item C1 and the entirely right acts in C2 and C7.

If the binary hypothesis had been true, it would have been possible to represent all seven items along a single dimension, and all items would have been clustered in two distinct areas in the figure: right and wrong. However, the findings in table 9 cannot be represented in such a way and are thus incompatible with the binary hypothesis.

That said, it remains to explain why C6 and C3 in figure 1 are located above C4 and C5 on the y-axis. Note that C3 and C6 are cases in which the agent violates the speed limit for a good reason, whereas C4 and C5 are cases in which the agent lies for what seems to be a good reason. This suggests the following somewhat speculative interpretation: The more similar the agent's reason for some acts are, the closer are their locations on the y-axis. If so, the underlying similarities between C3 and C6, and C4 and C5, seem to be visible in the figure, which indicate that data in table 9 have a reasonable degree of validity.

A drawback of the two-dimensional scaling is that the maximum error in figure 1 is, as noted, relatively large. It is therefore appropriate to consider a three-dimensional scaling. See figure 2. In this representation, Kruskal's stress value is 1.07 x 10^-6, which indicates a good fit. Figure 2 confirms the conclusion of figure 1: C1 (the entirely wrong act) is located far apart from C2 and C7 (the entirely right acts), but C3–C6 are located between the entirely right and entirely wrong items. In figure 3, it is thus also reasonable to interpret the x-axis as a visual representation of an act's degree of rightness. The interpretation of the y-axis is the same as in figure 1, but we can leave it open how the z-axis is to be interpreted as that is of no importance of the present discussion.

Figure 2. A three-dimensional metric multidimensional scaling of table 9 Kruskal's stress value is 1.07 x 10^-6. Note that C3–C6 are located ‘between’ C1 (the entirely wrong act) and C2 and C7 (the entirely right act).

Figures 1 and 2 are based on the assumption that all similarities in table 9 can be represented in a metric space. Because it is hard to know if this assumption is true, nonmetric multidimensional scalings are also worth considering. In this type of representation distances are interpreted as ordinal orderings: the aim is to preserve the ordering among the original points in a lower number of dimensions. Figure 3 shows a two-dimensional nonmetric multidimensional scaling of data in table 9. Kruskal's stress value is 8.5 x 10^-5, which indicates a good fit. This figure confirms the previous conclusions: C3–C6 is located between C1 and C2 and C7, which tallies well with the conclusion that the act C1 is entirely wrong, while the acts C3–C6 somewhat right and somewhat wrong, and C2 and C7 entirely right.

Figure 3. A two-dimensional nonmetric multidimensional scaling of table 9. Kruskal's stress value is 8.5 x 10^-5. Note that C3–C6 are located ‘between’ C1 (the entirely wrong act) and C2 and C7 (the entirely right act).

In summary, all three multidimensional representations fit well with the gradualist hypothesis, but they are incompatible with the binary one.

7. A Socratic Midwifery Effect?

Study 3 included an open-ended task in which half of the respondents were invited to write a couple of sentences about a case without relying on any predefined answer options. The other half was invited to select one of the following answer options: ‘right’, ‘wrong’, ‘right to some degree, but also wrong to some degree’, and ‘I don't know, I would need more information’. The purpose was to study to what extent gradualist answers occur naturally. The study was conducted in May 2021, during the COVID-19 pandemic. Below is the case:

The group asked to answer the question by writing a couple of sentences without relying on any predefined answer options (n = 87) submitted 3,453 words. All responses were analyzed and categorized manually. Gradualist conclusions were expressed by 16.1 percent of respondents, compared to 46.5 percent (n = 95) in the group presented with predefined answer options. (In the first group 29.0 percent reported that it was right to pressure Anouska's father to take the vaccine, compared to 16.1 percent in the second group; 26.0 percent concluded that it was wrong to pressure Anouska's father to take the vaccine, compared to 33.3 percent in the second group; and 29.0 percent presented with open-ended tasks stated views that were ambiguous or could not be reasonably classified as all-things-considered verdicts, compared to 4 percent in the second group who selected ‘I don't know, I would need more information’.) Below are some examples of spontaneously submitted gradualist answers from respondents in the first group:

• • ‘What she did was morally grey, but the end result is positive’.

• • ‘From a moral standpoint, it would probably be both right and wrong, but erring toward wrong’.

• • ‘It's a little right and a little wrong’.

• • ‘It is a bit right and wrong to do what she did’.

It is not surprising that respondents used gradualist terms spontaneously, but it is surprising that gradualist responses occur almost three times as often (46.5 percent versus 16.1 percent) if a gradualist conclusion is included among a set of predefined answer options. This difference is significant at p > 0.01, one-tailed. What explains the difference?

One explanation could be that respondents’ willingness to express gradualist views might be stimulated by a bit of philosophical midwifery. Socrates taught us that philosophers can help people articulate ideas they are unable to express clearly themselves, so I call this the Socratic midwifery effect. According to this explanation, respondents have a tacit but somewhat underdeveloped understanding of the gradualist hypothesis, which is clarified by the presence of a clearly stated gradualist answer option and therefore selected to a higher extent.

However, the findings reported in this paper do not permit us to state any definitive conclusion about the possible significance of a Socratic midwifery effect. About half of the respondents in Study 3 (n = 95) were presented with tasks C1, C3, and C4 described above. The rest (n = 87) were presented with versions of these tasks in which the original answer options (‘a bit right and a bit wrong’, and so on) had been replaced with more complex ‘midwifery’ answer options: ‘There are reasons for, but also against, doing what [the agent] did. On balance it was neither entirely right nor entirely wrong; it was a bit right but also a bit wrong / it was right / it was wrong / I don't know’. For task C1, the percentage of gradualist answers increased from 10.5 percent to 27.6 percent (which supports the Socratic midwifery effect), but for C3, the percentage of gradualist answers decreased from 49.5 percent to 12.6 percent, and for C4 it decreased from 55.8 percent to 19.5 percent. This speaks against the Socratic midwifery effect and there is no obvious explanation of this anomaly. The significance of the Socratic midwifery effect could be a topic for future research.