EVIDENCE AND VALUE FREEDOM
in these postpositivist times, the slogan “science is value free”
is frequently rejected disdainfully as a vestige of a bygone age. But, as of-
ten happens with babies and their bathwater, there may be something
worthwhile in this slogan that we should try to identify and retain. To
this end, I’ll begin with two absurdities:
(1) Scientists aren’t inﬂuenced by their ethical and political values
when they do science.
(2) Scientiﬁc inference is independent of values.
I don’t know if any philosopher ever believed either of these proposi-
tions. Proposition (1) is false for the simple reason that scientists are peo-
ple, just like the rest of us. Perhaps they often strive to leave their ethical
and political values at the laboratory door, but who ever thought that all
of them have this aim, and that those who do succeed 100% of the time?
This, by the way, does not mean that we get to assume that scientiﬁc ac-
tivity can be explained solely in terms of the ethical and political values
that scientists have. Rather, recognizing the absurdity of (1) should lead
us to approach such psychological and sociological questions on a case-
by-case basis. Scientists may vary among themselves, and a single scientist
may be more inﬂuenced by these values in some contexts than in oth-
ers. Maybe some scientiﬁc work proceeds completely independently of
these values and other parts of science are entirely driven by them, and
perhaps a good deal of the real world falls somewhere between these
Proposition (2) is absurd because scientiﬁc inference is regulated
by normative rules. Scientists try to construct good tests of their hy-
potheses, they judge some explanations good and others bad, and they
say that some inferences are ﬂawed or weak and others are strong. The
words I have italicized indicate that scientists are immersed in tasks of
evaluation. They impose their norms on the ideational entities they
construct. However, the obvious falsehood of (2) leaves it open that a re-
stricted version of that proposition might be on the right track:
(3) The fact that believing a proposition would have good or bad ethi-
cal or political consequences is not evidence for or against that
proposition’s being true.
Is this proposition, or some reﬁnement of it, the kernel of truth in the
frequently misstated idea that “science is value free”?
We should not accept proposition (3) just because it “sounds right.”
After all, the evidence relation often connects facts that seem at ﬁrst
glance to be utterly unrelated. The proposition that the dinosaurs went
extinct 65 million years ago because of a meteor hit and the proposition
that there now is an iridium layer in certain rocks may appear to have
nothing to do with each other. How could the presence of iridium in
present-day rocks bear on the question of why the dinosaurs went ex-
tinct so long ago? Well, appearances to the contrary, there may well be
such a connection (Alvarez and Asaro 1990). Why, then, should we be
so sure that the ethical consequences of believing a proposition have no
bearing on whether the proposition is true? This is a good question, and
in the absence of a good answer, we should not complacently assume
that (3) is correct.
Sometimes people believe (3) because they think there are no ethi-
cal truths in the ﬁrst place. If there are no ethical truths, then ethical
truths don’t provide evidence for anything. Whether or not nonfactual-
ism is correct, I think it fruitful to consider proposition (3) on the
assumption that there are ethical facts. In our everyday lives, we treat
110 EVIDENCE AND VALUES
ethical statements as if some of them are true. Perhaps this is a mistake,
but for present purposes, let’s take that practice at face value. If some
normative ethical statements are true, what is to prevent them from
standing in evidential relations with nonethical, scientiﬁc propositions?
To investigate this question, let’s begin with a useful example of the
evidence relation at work—smoke is evidence of ﬁre. This relation
holds because the probability of ﬁre if smoke is present exceeds the
probability of ﬁre if smoke is absent. The two events ±smoke and ±ﬁre
are correlated. This is a perfectly objective relation that obtains between
smoke and ﬁre; it obtains whether or not anyone believes that it does.
Understood in this way, the evidence relation has an interesting prop-
erty: It is symmetric. If smoke is evidence of ﬁre, then ﬁre is evidence of
smoke. This means that proposition (3) entails a further claim:
(4) The fact that a proposition is true is not evidence that believing the
proposition would have good or bad ethical or political conse-
If (3) is true, so is (4). But surely there are counterexamples to proposi-
tion (4). Consider a physician who will give a drug to her patients if she
thinks the drug is safe but will withhold the drug if she thinks it is not.
Suppose that the drug will provide signiﬁcant health beneﬁts if it is safe.
And suppose further that the physician is a pretty good judge of whether
the drug is safe. We then have a causal chain in which earlier links raise
the probability of later ones:
the drug is safe →the doctor thinks the drug is safe →the patients receive
the drug →good consequences accrue to the patients
In this instance, correlation is transitive. The nonethical statement “the
drug is safe” is therefore evidence for the ethical statement “good con-
sequences accrue to the patients.” An ethical and a nonethical fact are
evidentially related, just like smoke and ﬁre. I conclude that (4) and
therefore (3) are false (Stephens 2000).
It might be replied that (3) and (4) can be saved from refutation by
focusing exclusively on the two propositions under discussion. Consider
just the two statements “the drug is safe” and “good consequences ac-
crue to the patients.” In the absence of any further information, there is
no saying whether these statements are positively evidentially relevant
EVIDENCE AND VALUE FREEDOM 111
to each other, negatively relevant, or entirely irrelevant. We assumed in
our story that the physician is well meaning and discerning. This as-
sumption was enough to bring the two statements into a positive rela-
tion. However, if the doctor were malevolent or a very bad judge of drug
safety, the two statements would stand in a relation of negative rele-
vance or be mutually irrelevant. With no further information, the rela-
tion of the two statements is indeterminate. This is a probabilistic ana-
log of Duhem’s thesis (Sober 1988).
The trouble with this reply is that what is true for the pair of state-
ments about the physician is true for practically any pair of statements.2
The evidence relation isn’t binary; it has at least three places. When one
statement is evidence for a second, this usually is due to the mediation
of a third, which provides background information.3Given this, it is
hardly surprising that ethical facts can provide evidence about scientiﬁc
propositions; they can do this if one’s background assumptions include
other, ethical claims.
How is this criticism of (3) and (4) related to Hume’s famous dic-
tum that an ought cannot be inferred from an is? Although Hume was
talking about deduction, it is natural to generalize his thesis to a claim
about evidential relationships that are nondeductive and probabilistic.
What one needs to say here is that an is-statement and an ought-statement
are not evidentially relevant to each other, unless one’s background as-
sumptions include other ought-statements. But because one’s back-
ground assumptions often do include such “bridge principles,” evidential
reasoning can run from is to ought and from ought to is.4
How might propositions (3) and (4) be reﬁned? What is needed is a
way to separate the pattern exempliﬁed by the physician from a second
class of examples, which William James (1897) addressed in his famous
discussion of the will to believe. James argued that believing in God can
provide substantial psychological beneﬁts. Faith in God can help peo-
ple feel that their lives are meaningful, which may save them from de-
pression and allow them to lead worthwhile lives. It is debatable how
general this point is; there are enough happy atheists and depressed the-
ists around to make one suspect that, at least for many people, mental
health is independent of theological conviction. But let’s restrict our at-
tention to people who ﬁll James’s bill. These people would beneﬁt from
believing that God exists; I’ll go further and say that it is a good thing for
112 EVIDENCE AND VALUES
EVIDENCE AND VALUE FREEDOM 113
P is true P is false
S believes P w x
S does not believe P y z
Table 5.1. Comparing utilities as a function of what is believed and what is
these people to embrace theism and thus save themselves from the
slough of despond. However, I still want to claim that the fact that they
would beneﬁt from believing in God provides no evidence that God in
fact exists. It is examples like this that make proposition (3) sound so
plausible. What distinguishes James’s theist from the well-meaning
We can separate these cases by considering the two-by-two table
above (table 5.1). The entry in a cell represents utility—how good or
bad the consequences are of being in that situation. In the physician
case (where P =“the drug is safe”), there are both “vertical” and “hori-
zontal” effects. The well-being of the patients is affected both by
whether the drug is safe (w >x)5and by what the physician believes
(w >y and z >x). In the case of James’s believer (where P=“God exists”),
however, there are only vertical effects. As far as the individual’s psycho-
logical well-being is concerned, the only thing that matters is that he
or she believes in God (w >y and x >z); whether God actually exists
doesn’t matter (w =x and y =z).
This, I think, provides the key to revising propositions (3) and (4).
Our question—when do the ethical consequences of believing a propo-
sition provide evidence as to whether the proposition is true?—can be
represented by using the tools of decision theory. We begin by identify-
ing the expected value of each of the two “acts.”6
EV[Believe P] =wp +x(1 −p)
EV[Don’t Believe P] =yp +z(1 −p).
Here pdenotes the probability that the proposition P is true (and I as-
sume that acts are independent of states of the world). When does the
fact that EV[Believe P] >EV[Don’t Believe P] provide information about
the value of p? A little algebra reveals that
EV[Believe P] >EV[Don’t believe P] if and only if p/(1 −p) >(z −x)/
Suppose that the left-hand inequality is true. When will ﬁlling in the
values for the utilities w, x, y, and z provide a nontrivial lower or upper
bound on the value of p? This fails to happen in the case of James’s the-
ist because (w −y) is positive while (z −x) is negative. With these values,
all that follows is that p/(1 −p) must be greater than some negative num-
ber; this is entirely uninformative, since no ratio of probabilities can be
negative. The case of the physician is different. Here (z−x) and (w−y)
are both positive; therefore, their ratio provides a nontrivial lower bound
for the value of p.7Thus, science and ethics are not always as separate as
propositions (3) and (4) suggest; sometimes information about the ethi-
cal consequences of believing a proposition does provide information
about the probability that the proposition is true.
I have represented the ethical consequences of believing a proposi-
tion in terms of the expected utility of doing so. This may sound as if it
requires a consequentialist ethics, but in fact it does not. A nonconse-
quentialist also can enter payoffs in the four cells. Furthermore, the ar-
gument is general. Utilities may be calculated in accordance with an
ethical theory or on some nonethical basis. What we have here is a gen-
eral format in which judgments about which acts are better, plus infor-
mation about the utilities of outcomes, can have implications about the
probabilities of propositions.8
Before phoning the National Science Foundation with the news
that ethics can be a source of evidence for scientiﬁc claims, we should
reﬂect on the fact that the expected utility of an action is a composite
quantity, built up from probabilities and utilities. If one’s ethical judg-
ments about which actions one should perform are based on comparing
their expected values, then ethical judgments require information about
probabilities for their formulation. If so, how can ethical judgments be a
source of information about probabilities? This point does not rescue
propositions (3) and (4) from counterexample, but it does suggest a new
way to think about the problem.
114 EVIDENCE AND VALUES
There is a special circumstance in which it is possible to decide which
of the two actions (believe P, or don’t) is better without any information
about the probability that P is true. This is the case in which one action
dominates the other. James’s argument is of this type: He contends that
belief in God is beneﬁcial, whether or not God in fact exists.9However,
we have already seen that the inequality of the expected utilities pro-
vides no information about the probabilities in James’s case. Without
dominance, no conclusion about which action is better can be reached
unless one already has information about the probabilities. What we
have here is an instance of the maxim “out of nothing, nothing comes.”
If comparing the ethical consequences of believing P and of not believ-
ing P has implications about the probability of P, this must be because
the description of the ethical consequences already has built into it
some information about those probabilities. Thus, ethical facts (about
expected utilities) and scientiﬁc facts (about probabilities) are con-
nected, contrary to what (3) and (4) assert. However, the problem is that
this connection is useless; we can’t use ethical information to gain infor-
mation we don’t already have about probabilities.
The situation would be different if we were able to discover which
actions are better than which others, when dominance fails, without al-
ready having to have information about probabilities. For example, if
there were an infallible guru who would simply tell us what to do, and
who would reveal the utilities that go with different states of the world,
we could use these inputs to obtain new information about probabili-
ties. But in the absence of such an authority, we are left with the con-
clusion that our access to information about what we should do must be
based on information about probabilities (except when there is domi-
nance). When an ethical conclusion requires information about proba-
bilities (as in the case of the physician), that conclusion can’t be a
source of new information about those probabilities. And when the eth-
ical conclusion can be reached without information about probabilities
(as in the case that James describes), the conclusion tells us nothing
about the probabilities. This suggests the following dilemma argument:
The question of whether believing P has better ethical consequences than
not believing P either depends for its answer on information about the
probability of P, or it does not.
EVIDENCE AND VALUE FREEDOM 115
If it does so depend, then we need information about probabilities to an-
swer the ethical question, and so the ethical judgment cannot supply
information about probabilities that we don’t already have.
If it does not so depend, then the ethical judgment has no implications
about the probability of P.
(5) Judgments about the ethical consequences of believing P cannot
supply new information about the probability of P.
The conclusion of this argument, proposition (5), is a reasonable suc-
cessor to the failed propositions (3) and (4).
The argument just presented is reminiscent of Rudner’s (1953) well-
known argument that the scientist qua scientist makes value judgments.10
Rudner describes a physician who must decide whether a drug is safe and
argues that this decision must be based on considering ethical features of
the four possible outcomes depicted in table 5.1, which we have already
discussed. Rudner’s argument elicited two criticisms. Levi (1967) con-
tended that accepting a proposition and acting on one’s belief are distinct
and that the former should not be based on ethical values; Jeffrey (1956)
maintained that science is not in the business of accepting and rejecting
but merely seeks to assign probabilities to hypotheses. My own argument
is neutral on Rudner’s position. Perhaps deciding what to believe depends
on ethical values; perhaps it does not. My point is that no matter how one
decides what to believe, one still can consider what the ethical conse-
quences are of that decision. My question was whether this ethical con-
sideration has implications concerning the probabilities of hypotheses. It
does in the case of the physician but not in the case of James’s theist.
In the example about the physician, and in many other examples of
moral deliberation about which action to perform, one’s ethical deci-
sion depends on matters of scientiﬁc fact. In terms of table 5.1, the ubiq-
uitous pattern is that an inequality gets reversed as one moves from the
ﬁrst column to the second. Although one’s ethical decision thus de-
pends on a judgment about a matter of scientiﬁc fact, it is possible to
form a judgment about the scientiﬁc facts without having a commit-
ment, one way or the other, on the ethical question. The physician can’t
decide whether to administer the drug without knowing something
about its probability of being safe, but it is perfectly possible to discover
whether a drug is safe without having a view, one way or the other, on
116 EVIDENCE AND VALUES
whether unsafe drugs should be withheld from patients. Moral ignora-
muses can assess the weight of evidence, but scientiﬁc ignoramuses can-
not make good moral decisions (when those decisions depend, as they
almost always do, on scientiﬁc matters of fact).
Let’s review our progress from propositions (1) and (2) through (3)
and (4) and then to (5). Proposition (1) concerns the behavior of scien-
tists, whereas (3) and (4) concern the logic of various scientiﬁc concepts.
Proposition (2) ambiguously straddles this distinction; “scientiﬁc infer-
ence” can be taken to refer to what scientists do or to the formal proper-
ties of various types of argument. Proposition (5) addresses the concept
of evidence; it does not assert that scientists are immune from political
and ethical inﬂuence when they decide whether one proposition is evi-
dence for another. More speciﬁcally, the claim is not that ethical values
(represented by the expected utilities of actions and the utilities of out-
comes) have no implications about the probabilities of hypotheses, but
that ethical inputs are not needed to estimate those probabilities. This is
why looking to ethics for evidence concerning the truth of scientiﬁc hy-
potheses is to place the cart before the horse.
1. I am grateful to Ellery Eells and Dan Hausman for useful comments.
2. The exception arises when one statement deductively entails the other; then they
must be positively relevant or of zero relevance, and negative relevance is ruled out.
3. I say “at least” three places because in most scientiﬁc contexts, what one can dis-
cuss is whether the evidence discriminates between a pair of hypotheses, given a set of
background assumptions. See Sober (1994a) for discussion.
4. Does the ought implies can principle undermine Hume’s thesis? The princi-
ple’s contrapositive asserts that if it is impossible for an agent to perform an action, then
it is false that the agent ought to do so. Here an is implies the negation of an ought.
Hume’s thesis can be preserved by insisting that the negation of an ought-statement is
not itself an ought-statement. Similar reasoning is required if one wishes to reconcile
Hume’s thesis with the fact that philosophers have presented philosophical arguments
(of varying quality) for the claim that normative ethical statements lack truth values.
These arguments do not contain premises that are normative ethical statements. For ex-
ample, Harman (1977) and Ruse and Wilson (1986) each present parsimony arguments
for the nonexistence of ethical facts; see Sober (2005) and (1994b), respectively, for dis-
cussion of each.
5. I take it that it doesn’t matter whether the drug is safe if the doctor doesn’t believe
that it is (y =z), because patients won’t receive the drug in that situation regardless of
whether the drug would be good for them.
EVIDENCE AND VALUE FREEDOM 117
6. It might be suggested that believing a proposition is not an action, in the sense
that it is not subject to the will. This point is sometimes used against Pascal’s wager, but it
is an objection that Pascal successfully addressed: He says that if absorbing his argument
does not instantly trigger belief, one should go live among religious people so that habits
of belief will gradually take hold. Believing a proposition is like other “nonbasic actions”:
Being president of the United States isn’t something one can directly bring about by an
act of will, but this does not place it outside the domain of decision theory. For further dis-
cussion, see Mougin and Sober (1994).
7. Symmetrically, if not believing P were the better action, this would impose a
nontrivial upper bound on the value of the probability.
8. Decision theory from early on has taken an interest in describing the interrela-
tionships of expected utility, utility, and probability. For a brief introduction, see Skyrms
(2000, pp. 138–43).
9. Pascal’s wager, when the payoffs are ﬁnite, does not have this property. One needs
some information about the probability of God’s existing to reach a decision about whether
believing is better than not believing. In fact, the theist contemplating Pascal’s wager
(with ﬁnite payoffs) is in the same qualitative situation as the physician deciding whether
to believe that the drug is safe. See Mougin and Sober (1994) discussion.
10. The problem that Rudner addresses is the one that James (1897) and W. K. Clif-
ford (1879) debated. It also was central to the debate between the “left” and “right” wings
of the Vienna Circle. Neurath argued that evidence does not determine theory choice
and that ethical and political values can and should be used to close the gap; Schlick,
Carnap, and Reichenbach countered that the intrusion of such values into theory choice
is both undesirable and unnecessary: It would compromise the objectivity of science, and
scientiﬁc inferences can be drawn without taking such values into account. See Howard
(2002) for discussion.
Alvarez, W., and F. Asaro. 1990. “What Caused the Mass Extinction? An Extraterrestrial
Impact.” Scientiﬁc American, 263, pp. 78–84.
Clifford, W. K. 1879. “The Ethics of Belief,” in Lectures and Essays, vol. 2, pp. 177–211.
Harman, G. 1977. The Nature of Morality. New York: Oxford University Press.
Howard, D. 2002. “Philosophy of Science and Social Responsibility—Some Historical
Reﬂections,” in A. Richardson and G. Hardcastle, eds., Logical Empiricism in North
America. Minneapolis: University of Minnesota Press.
James, W. 1897. “The Will to Believe,” in The Will to Believe and Other Essays in Popular
Philosophy. New York: Longmans Green.
Jeffrey, R. 1956. “Valuation and Acceptance of Scientiﬁc Hypotheses.” Philosophy of Sci-
ence, 33, pp. 237–46.
Levi, I. 1967. Gambling with Truth. Cambridge, MA: MIT Press.
Mougin, G., and E. Sober. 1994. “Betting against Pascal’s Wager.” Nous, 28, pp. 382–95.
Rudner, R. 1953. “The Scientist Qua Scientist Makes Value Judgments.” Philosophy of
Science, 20, 1–6.
118 EVIDENCE AND VALUES
Ruse, M., and E. Wilson. 1986. “Moral Philosophy as Applied Science.” Philosophy, 61,
pp. 173–92. Reprinted in E. Sober, ed. 1993. Conceptual Issues in Evolutionary Biol-
ogy, 2nd ed. Cambridge, MA: MIT Press.
Skyrms, B. 2000. Choice and Chance, 4th ed. Belmont, CA: Wadsworth.
Sober, E. 1988. Reconstructing the Past—Parsimony, Evolution, and Inference. Cam-
bridge, MA: MIT Press.
Sober, E. 1994a. “Contrastive Empiricism,” in From a Biological Point of View. New York:
Cambridge University Press.
Sober, E. 1994b. “Prospects for Evolutionary Ethics,” in From a Biological Point of View.
New York: Cambridge University Press.
Sober, E. 2005. Core Questions in Philosophy. Upper Saddle River, NJ: Prentice-Hall.
Stephens, C. 2000. “Why Be Rational? Prudence, Rational Belief and Evolution.” PhD
dissertation, University of Wisconsin, Madison.
EVIDENCE AND VALUE FREEDOM 119