In his essay “Two Faces of Common Sense” (1972) Popper wrote ‘We are seekers for truth but we are not its possessors’. In the same essay, Popper argues ‘The quest for certainty, for a secure basis of knowledge, has to be abandoned.’
Popper wanted to separate (un)certainty and truth and he believed that science should seek truth. Two quotes from Popper: ‘science has nothing to do with the quest for certainty or probability or reliability’ and ‘we too see science as the search for truth’.
For Popper, truth is fully independent of the believer. I have always struggled with the concept of ‘truth’ as ‘independent of the believer’. Maybe Popper did not want to replace God with Truth, but to me it could appear that way.
My interest in the philosophical ideas of Frank Ramsey (1903-1930) followed from listening to a podcast ‘A tale of truth’ by Simon Blackburn in 2023. Frank Ramsey’s redundancy theory of truth appealed to me. Popper himself once thought the correspondence theory of truth was dispensable, and noted in “Conjectures and Refutations” that Ramsey had suggested it might be empty altogether. If you adopt the redundancy theory of truth, science becomes managing uncertainty and assessing universal laws for reliability.
Reading the philosophical ideas of Frank Ramsey also addressed another issue I was struggling with. In his essay “Conjectural Knowledge” (1972), Popper’s rejection of induction is total at the logical level, while at the same time Popper argued that this creates no clash with rationality, empiricism, or scientific practice. My view is that people do reason non-deductively. Inductive reasoning has traditionally been interpreted through a deductive lens: an argument that has a conclusion and supporting premises. This interpretation works very well for deduction, but I think it is incorrect for induction. The re-interpretation of induction by Frank Ramsey provided an ‘a-ha’ moment to me.
Finally, the Cox-Jaynes theory is widely used in Bayesianism. The Cox-Jaynes theory is founded on three desiderata. I discovered that Frank Ramsey provides an alternative account of Bayesian updating based on the coherence of betting behaviour.
In this essay I will describe Ramsey’s epistemology and highlight the implications of adopting a ‘Ramseyian’ approach on how to deal with universal laws and induction.
Frank Ramsey on Truth, Degrees of Belief and Inductive Reasoning
Note on sources: This essay draws on the following works by Frank Ramsey: 'Facts and Propositions' (1927), 'Truth and Probability' (1926), 'Law and Causality' (1928), and the two 1929 notes 'Knowledge' and 'Probability and Partial Belief.' Where a section combines positions from more than one essay, or draws an inference that Ramsey does not state explicitly, this is noted in the text. Direct quotations are taken from Ramsey's own words. The section on the reinterpretation of induction combines Ramsey's betting framework with his treatment of variable hypotheticals in a form he does not himself state, but each step draws on positions he does hold.
Truth
Ramsey argues that truth is not a property that needs its own theory. Saying 'it is true that Caesar was murdered' adds nothing to saying 'Caesar was murdered.' This is now called the Redundancy Theory of Truth.
This contrasts with the correspondence theory, where a statement is true if and only if it corresponds to the facts, and truth holds independently of any mind that may or may not believe it.
Ramsey does not deny that facts exist. He is a realist about the world. What he denies is that 'true' names a relation between a proposition and a fact. That connection is carried by the proposition itself. 'True' is not what connects language to the world. It is a device for reasserting and generalising what is already asserted.
Beliefs and Degrees of Beliefs
Beliefs are instruments for guiding action. Frank Ramsey understands belief as a mental state that tends to produce certain actions, and degree of belief as how strong that tendency is: not a felt intensity, but a measure of what a person would do or be disposed to do on the basis of holding it.
He distinguishes beliefs that are consciously held and expressed in words or images from mere habits of response, such as an animal avoiding food it has learned is harmful. It is this kind of belief that can be assessed. Such a belief is, in Ramsey's own words, 'a map of neighbouring space by which we steer.'
Degrees of belief, for Ramsey, govern rational choice under uncertainty. A person uncertain about a proposition weights possible outcomes by their degree of belief and acts on whichever option produces the greatest expected value. His 'crossroads' example illustrates this: how far a traveller is willing to go to verify a route depends directly on how confident they are in it. Degrees of belief must conform to the probability calculus, since incoherent degrees of belief expose an agent to what Ramsey calls having 'a book made against him by a cunning better,' a guaranteed loss regardless of outcomes.
Ramsey distinguishes two ways of evaluating belief. Coherence governs how degrees of belief relate to one another and is captured by the probability calculus. Reliability concerns the causal habits by which beliefs are formed, assessing whether they tend to produce beliefs that are borne out. This distinction is drawn across several essays rather than stated in a single place by Ramsey.
Ramsey's account of degrees of belief rests on three connected steps: assigning a scale to value, treating mathematical expectation as a psychological law, and proving that consistent degrees of belief must obey the probability calculus.
Step 1 Assigning an interval scale to value
Ramsey defines degrees of belief as willingness to take a bet, where the stakes are measured in units of value. Value is a measure of how much an agent prefers one outcome over another. Ramsey constructs a scale of value in which differences between positions are meaningful but the origin and unit are arbitrary. Later decision theory would call this an interval scale. Ramsey begins with the special case in which an agent's degree of belief is one-half, making the agent indifferent between betting either way, and uses such indifference conditions in constructing a scale of value.
From that anchor point, the rest of the interval scale is constructed by presenting the agent with a series of choices between a certain outcome and a bet whose result depends on whether a proposition is true or false (a conditional bet). Each time the agent is indifferent between the certain outcome and the bet, the difference in value between the outcomes of that bet is mapped onto the interval scale.
Ramsey is not concerned with the absolute value of any outcome. What the interval scale captures is the difference in value between outcomes, mapped relative to the anchor point. The anchor point establishes that two outcomes with the same value for the agent occupy the same position on the scale. Differences in preference between outcomes are mapped as gaps on the interval scale, with larger gaps corresponding to larger differences in preference.
What Ramsey called value later became utility, the standard term in economics and decision theory.
Step 2 Psychological law
Ramsey treats mathematical expectation as a psychological law: a descriptive approximation of how agents could behave when reasoning under uncertainty, on the basis that all deliberate action involves, in some sense, acting on a bet about how the world will turn out.
An agent who reasons in accordance with mathematical expectation takes each possible outcome, multiplies the difference in value that outcome represents by the degree of belief attached to the proposition on which that outcome depends, sums those products across all outcomes for a given option, and selects the option with the highest total. Since value is measured as differences in position on the interval scale rather than as fixed numbers, the calculation does not require exact values for any outcome. The arbitrary origin and unit of the interval scale cancel out, and what matters is how the differences between outcomes compare to each other.
Ramsey acknowledges that this approximation cannot account for all the facts of human behaviour. An agent might choose to reason differently.
Step 3 Mathematical proof
If an agent does reason and act in accordance with mathematical expectation, then Ramsey proved mathematically that the agent's degrees of belief must obey the probability calculus or the agent will be exploitable. An agent is exploitable if a shrewd opponent can construct a set of bets, each acceptable to the agent at the agent's own stated odds, that guarantees the agent a loss regardless of how the propositions turn out.
The probability calculus includes the multiplication law, which states that the degree of belief in two propositions both being true equals the degree of belief in the first multiplied by the degree of belief in the second given the first. The multiplication law is a static relationship between degrees of belief at a given moment.
Ramsey formulates what later became known as conditionalization, but he does not offer an independent derivation of it. Instead, the update rule appears as a natural consequence of his treatment of conditional belief and the multiplication law. Conditionalization is the dynamic rule that tells an agent how to update degrees of belief when new evidence arrives. Within Ramsey's framework, updating by conditionalization is the consistent way of revising beliefs in light of new evidence, given the multiplication law. The multiplication law provides the foundation, while conditionalization follows as its consequence for belief revision.
If an agent chooses to reason like a betting man, their degrees of belief must obey the probability calculus and they must update those beliefs with new evidence using Bayes' theorem.
Reinterpretation of Induction
The following combines Ramsey's betting framework from Chapter 4 with his treatment of variable hypotheticals in Chapter 7. Ramsey does not state the argument in this form, but each step draws on positions he does hold.
Assigning probabilities to universal propositions is problematic
A Ramsey-inspired objection to assigning probabilities directly to universal laws is that, within a betting interpretation, such propositions do not admit the same straightforward settlement conditions as ordinary particular propositions.
The reasoning is as follows:
a. Within a betting interpretation, a probability assignment is only meaningful if it corresponds to a wager with a determinate settlement procedure.
b. Such a wager requires that both winning and losing outcomes be reachable in principle.
c. A universal proposition such as "All swans are white" ranges over every instance without exception, including those never encountered. No finite set of observations covers all instances. A bet on such a proposition has no determinate settlement procedure. The winning outcome, confirmation across all instances, is not reachable. The losing outcome, a single counterexample, is.
d. A universal proposition therefore does not satisfy the settlement conditions in (a) and (b).
e. On this reconstruction, assigning probabilities directly to universal propositions or universal laws is problematic. Particular propositions derived from experience do not raise the same difficulty.
Induction and Expectation-Generating Rules
A standard textbook definition of induction has been ‘the process of reasoning to a general or universal conclusion on the basis of a number of particular observations’. This would imply that from the observation ‘I saw 1,000 white swans’ the universal proposition ‘All swans are white’ could be inferred. And induction has traditionally been evaluated using propositional logic.
On that view, the problem of induction is the problem of justifying the inference from a finite number of observed cases to a universal proposition covering all cases, including those not yet observed. No finite number of observations can conclusively verify a universal proposition, and so the logical gap between evidence and conclusion has never been fully closed. The previous section argued, drawing on Ramsey's betting framework, that addressing this gap by assigning probabilities to universal propositions is problematic.
Frank Ramsey reframed induction:
Not: ‘I saw 1,000 white swans’ inducing the universal proposition ‘All swans are white’.
But: ‘I saw 1,000 white swans’ inducing the expectation-generating rule ‘Next time I encounter a swan, I expect it to be white’
Ramsey named the expectation-generating rule a ‘variable hypothetical’.
On a reconstruction combining Ramsey's account of partial belief with his treatment of variable hypotheticals, induction is not a matter of inferring from particular propositions to a universal conclusion. It is a matter of adopting or revising expectation-generating rules on the basis of particular experience. The degree of belief that is open to assessment is not the degree of belief in a universal conclusion, but the degree of belief in the next particular case derived from a variable hypothetical. The agent bets on the next swan being white, not on all swans being white. The variable hypothetical generates that bet, and is assessed by whether its expectations tend to be borne out.
On this reconstruction, induction is best understood as the formation, revision and assessment of variable hypotheticals that generate expectations from experience.
Assigning probabilities to a variable hypothetical is as problematic as assigning probabilities to universal propositions. Induction should be assessed on the reliability of the variable hypothetical: do the expectation-generating rules tend to produce expectations, expressed as propositions, that are borne out.
This connects to Ramsey's distinction between coherence and reliability: coherence governs how degrees of belief in particular propositions relate to one another, while reliability governs whether the expectation-generating rules from which those propositions are derived tend to produce expectations that are borne out.
Knowledge
In his 1929 essay, Ramsey states that a belief is knowledge if it is (i) true, (ii) certain, and (iii) obtained by a reliable process. He does not elaborate on the truth condition. He devotes most of the essay to worrying about what reliable process means, eventually preferring the phrase 'formed in a reliable way.'
That certainty is not immune to doubt. Ramsey acknowledges Russell's point that 'all our knowledge is infected with some degree of doubt' and does not reject it.
The redundancy theory, developed in 'Facts and Propositions,' suggests that truth adds no independent requirement beyond what the proposition itself asserts. Whether Ramsey intended these two positions to be reconciled, or considered the tension worth addressing at all, the essays do not say.
A belief is formed in a reliable way when it is caused by what it is about through a rule for judging (a variable hypothetical) that can generally be relied upon to produce beliefs that are borne out, and where any intermediate beliefs in the causal chain are themselves borne out.
Logic
'Logic must then fall very definitely into two parts: (excluding analytic logic, the theory of terms and propositions) we have the lesser logic, which is the logic of consistency, or formal logic; and the larger logic, which is the logic of discovery, or inductive logic.' From Ramsey, F. P. (1926) 'Truth and Probability'
‘The logic of consistency’.
The logic of consistency deals with propositions and degrees of belief in propositions. A proposition is the kind of thing that can be asserted or denied, borne out or not. It includes mathematics and the probability calculus. It assesses whether beliefs cohere with one another. It carries a necessity of assertion: if one asserts p, one is bound in consistency to assert whatever follows from p. The logic of consistency asks: are my degrees of belief coherent with one another? It governs rational organisation of uncertainty: given what one believes, what else is one bound to believe.
‘The logic of discovery’
The logic of discovery deals with variable hypotheticals as well as propositions. A variable hypothetical is not a proposition in the primary sense: it cannot be asserted or denied as a proposition can, and the standards of consistency that govern degrees of belief in propositions are not the appropriate standard by which to assess it. It can, however, be disagreed with, adopted or abandoned, and assessed by whether it reliably generates expectations that are borne out. The logic of discovery includes induction. It assesses whether habits of belief formation track the real world. Individual beliefs are then assessed derivatively, by reference to the habits that produce them. One is bound to revise a habit that proves unreliable, on pain of forming beliefs that are not borne out. The logic of discovery asks: do my habits of belief formation track the real world? It governs which habits of expectation are worth trusting, given how the world has behaved.
Summary
Ramsey's positions on truth, degrees of belief and induction are connected.
His redundancy theory holds that 'true' adds nothing to a proposition: it functions as a device for reasserting and generalising what is already asserted rather than naming a relation between a proposition and a fact.
Degrees of belief, measured by willingness to bet in terms of utility, must obey the probability calculus or the agent is exploitable, and must be updated by Bayes' theorem for the same reason.
Induction is reframed through variable hypotheticals (expectation-generating rules). The traditional approach to assessing induction through propositional logic is not adopted. Observations or reasoning from analogues do not generate universal propositions such as 'All swans are white', but an expectation-generating rule: 'If I encounter a swan next, I expect it to be white.' Induction is therefore not a matter of assigning probabilities to universal laws but of assessing whether those expectation-generating rules are reliable.
Taken together, coherence assesses how degrees of belief relate to one another and is captured by the probability calculus. Reliability assesses whether the expectation-generating rules an agent adopts, such as expecting the next swan to be white, tend to be borne out in practice.
Ramsey's epistemology is built around these two distinct but related tasks: assessing the coherence of degrees of belief, and assessing the reliability of the expectation-generating rules from which those beliefs derive.
If one rejects Humean or Popperian demands for a deductively justified theory of induction, Ramsey offers a replacement in which induction is understood as the formation and revision of variable hypotheticals that guide expectations rather than infer universal propositions. And if one is uncomfortable with Cox-style derivations of probability, Ramsey also provides an alternative justification of Bayesian updating based on coherence of betting behaviour rather than functional representation theorems.
Cox-Jaynes Bayesianism vs Ramseyian Bayesianism
The Cox-Jaynes framework and Ramsey's account of probability share a common starting point but diverge in three respects that matter for how each handles objectivity, universal laws, and induction. The first divergence concerns the degree to which the probability calculus constrains not just the relations between beliefs but their content. The second concerns how each framework handles universal propositions and universal laws. The third concerns the treatment of induction itself, and whether Bayesian updating is the right formal expression of inductive reasoning or whether the assessment of induction belongs outside Bayesian updating.
Both the Cox-Jaynes framework and Ramsey's account share a commitment to consistent reasoning. Both hold that degrees of belief must satisfy the probability calculus if they are to be mutually consistent. The probability calculus covers the sum rule, the product rule, and Bayes' formula. These rules fix how probability assignments must relate to one another and how they should be revised in light of new information. They are not matters of personal preference.
Cox-Jaynes goes further. Desiderata IIIb and IIIc make probability assignments, in Jaynes' words, "completely 'objective' in the sense that they are independent of the personality of the user." They are a means of encoding "the information given in the statement of a problem, independently of whatever personal feelings you or I might have about the propositions involved." If the information state specifies relevant values such as base rates, likelihoods, and false positive rates, those values are fixed for any agent reasoning according to the desiderata. Two agents with the same information ought to arrive at the same probability assignments. Anyone who reaches a different conclusion, Jaynes states, "is necessarily violating one of those desiderata."
Ramsey's account contains no corresponding principle. The probability calculus constrains the relations between an agent's beliefs, but it does not constrain the inputs themselves. An agent may use the values given by the information state or adjust them in light of personal belief, and remain fully coherent either way. Ramsey holds that asking what initial degrees of belief are justified is, in his words, "a meaningless question," and that formal probability theory cannot determine what prior probabilities an agent ought to hold. His framework provides consistent constraints on the coherence of beliefs, but not prescriptions concerning their content.
If one is uncomfortable with Cox-Jaynes derivations of probability, Ramsey provides an alternative justification of Bayesian updating based on coherence of betting behaviour rather than the three desiderata.
A further divergence concerns universal propositions, where the two frameworks take different approaches.
Both Cox-Jaynes and Ramsey face difficulty in assigning probabilities directly to universal propositions. Cox-Jaynes excludes them on formal grounds. The framework requires a defined and bounded hypothesis space. A universal proposition tested against an open-ended class of alternatives has, in Jaynes' words, a probability that "is simply undefined because the class of all conceivable theories is undefined." Within a bounded set of specified alternatives, Cox-Jaynes supports Bayesian updating directly. Given two or more defined universal propositions, Bayes' theorem updates their relative probabilities in light of evidence in the normal way.
Ramsey's measurement procedure takes a different path. His framework grounds degrees of belief in bets, and a bet requires a settleable outcome. A universal proposition cannot be conclusively verified in finite time, so the bet cannot be closed. Ramsey's framework therefore assigns probabilities to the observable consequences of competing universal propositions and updates on those. The bet is placed on the observable outcome rather than on the universal proposition directly. Universal propositions enter Ramsey's framework indirectly, through their testable implications, rather than as direct objects of degree of belief.
If one is uncomfortable with comparing universal laws and updating confidence in universal laws using Bayes' formula, Ramsey provides an alternative route, assigning probabilities to the observable consequences of competing universal laws rather than to the universal laws themselves.
A third divergence, related to the above, concerns the treatment of induction itself.
Jaynes treats Bayesian updating as the quantitative expression of inductive reasoning. In his words, "modern Bayesian analysis is just the unique quantitative expression of this reasoning format; the inductive reasoning that philosophers like Hume and Popper held to be impossible." Within a bounded, well-defined set of alternatives, repeated observations raise or lower the probability of competing hypotheses via Bayes' theorem. Successful predictions increase confidence in a hypothesis; failed predictions reduce it. Induction, on this account, is a quantitative process conducted within the probability calculus, with Bayesian updating as its formal mechanism.
Ramsey takes a different route and redefines induction. He treats universal propositions not as propositions at all, but as variable hypotheticals: expectation-generating rules applied to particular cases rather than statements with determinate truth conditions.
Because these rules are not propositions in the primary sense, they fall outside the betting framework and outside Bayesian updating. Their assessment rests on reliability, specifically whether the expectations they generate prove correct over time. That reliability assessment sits outside the probability calculus and cannot be reduced to Bayesian updating.
Observed frequencies can inform that assessment. An agent may form a degree of belief in the proposition that a given rule has succeeded in a certain proportion of cases. That degree of belief is directed at a proposition about observed frequencies rather than at the variable hypothetical itself. The rule, as a variable hypothetical, is not the kind of thing that can be the direct object of a degree of belief within the betting framework.
The asymmetry is precise. Jaynes treats Bayesian updating as the formal expression of inductive reasoning. Ramsey places the assessment of the variable hypotheticals that guide induction outside Bayes' theorem altogether.
If one is uncomfortable with treating induction as 'the process of reasoning to a general or universal conclusion on the basis of a number of particular observations', Ramsey provides an alternative route, treating inductions as expectation-generating rules applied to particular cases, assessed by their reliability rather than their logical justification.
Example: The Swan and the Variable Hypothetical
An agent observes 1,000 white swans. Induction, in Ramsey's reframing, does not generate the universal proposition ‘All swans are white’. It generates the variable hypothetical: ‘The next time I encounter a swan, I expect it to be white.’ This expectation-generating rule can be treated as a bet.
The bet is: ‘The next swan I see will be white.’
The prior reflects the agent's willingness to bet, given background knowledge. Having observed 1,000 white swans, the agent assigns a high but uncertain prior -- between 0.5 and 0.95, depending on location, season, and other background knowledge. A prior of 1 is not advisable. An agent can never be certain.
In this first version of the example, the likelihood and false positive are degenerate:
· Likelihood: P(observe white swan | next swan will be white) = 1
· False positive: P(observe white swan | next swan will not be white) = 0
If the evidence is a white swan, the posterior is 1: the bet is won. If the evidence is a black swan, the likelihood becomes 0 and the posterior becomes 0: the bet is lost. The agent could revise the reliability of the variable hypothetical accordingly.
This degenerate case illustrates the structure of the Ramseyian approach. The agent bets on the next particular case, not on the universal proposition. The universal proposition "All swans are white" is never the object of the bet and never needs to be assigned a probability.
The example becomes genuinely probabilistic when background knowledge enters. Suppose the agent learns two things before placing the next bet: the worldwide proportion of white swans is 85%, and the next observation will be made in the southern hemisphere, where black swans are native to Australia.
These two pieces of information pull in opposite directions. The worldwide base rate favours white. The location pulls against it. The agent can no longer justify a prior of 0.9. A prior closer to 0.15 is more defensible, depending on precisely where in the southern hemisphere the observation will take place.
The likelihood is no longer 1. Even if the next swan will be white, the agent might observe a white swan that is an introduced European species, an escaped captive, or an albino. P(observe white swan | next swan will be white) remains high but not certain. A value of 0.95 is reasonable.
The false positive is no longer 0. P(observe white swan | next swan will not be white) is small but positive. A white swan sighting in the southern hemisphere is not impossible under the not-white hypothesis. A value of 0.05 reflects this residual possibility.
The prior was 0.15. Observing a white swan raises it to roughly 0.77, because a white swan sighting is surprising under the not-white hypothesis. Uncertainty remains.
The posterior from this bet informs the prior for the next one, but does not determine it mechanically. New background knowledge may have entered: a report of introduced species, a change of location, a different season. The agent exercises judgement in setting the new prior. As Jaynes notes in Probability Theory: The Logic of Science, the prior for any new problem encodes all relevant prior information, not merely the output of the last calculation. This preserves the subjective character of the prior that Ramsey insisted upon: the prior reflects the agent's actual willingness to bet, not a formula applied blindly to previous results.
The two examples together show the Ramseyian approach at different levels of realism. The variable hypothetical generates the prediction in both cases. In the first, the probabilities are degenerate and the bet collapses into a deductive verdict. In the second, new background knowledge acts as evidence that reshapes the prior through Bayesian updating, and the agent enters the bet with a revised degree of belief. Induction, on this account, is the formation and revision of expectation-generating rules under uncertainty, not the inference of universal propositions from particular observations.
The swan example has no scientific content. It does not explain why swans are likely to be white. It illustrates the structure of the Ramseyian approach but nothing more.
A scientific universal theory such as Newton's law of gravity carries both explanatory and predictive power. Observations have established that Newton's law is reliable at low velocities and weak gravitational fields. Einstein's theory of relativity has proved reliable at relativistic velocities and under strong gravity. Both theories generate variable hypotheticals that can be assessed for reliability across different conditions. The shift from Newton to Einstein is not a matter of one universal proposition replacing another, but of one expectation-generating rule proving more reliable than its predecessor across a wider range of conditions.
Conclusion
Ramsey's epistemology offers a coherent alternative to the Popperian picture on three connected points.
On truth, the redundancy theory removes the need for a correspondence relation that holds independently of any believer. Truth adds nothing to a proposition beyond what the proposition itself asserts. Science, on this account, does not seek a target that stands apart from inquiry. It manages uncertainty and assesses whether expectation-generating rules are borne out.
On induction, Ramsey's reframing bypasses the logical problem entirely. The traditional interpretation treats induction as an inference from particular observations to a universal conclusion, and then struggles to close the logical gap. Ramsey dissolves the gap by changing what induction produces. The output is not a universal proposition but a variable hypothetical: an expectation-generating rule assessed by its reliability over time, not by its logical justification.
On Bayesian foundations, Ramsey provides an account of Bayesian updating grounded in the coherence of betting behaviour. The probability calculus constrains the relations between an agent's degrees of belief, and Bayes' theorem follows as the consistent update rule. This does not require the Cox-Jaynes desiderata, and it does not prescribe what prior probabilities an agent ought to hold. The prior reflects the agent's actual willingness to bet, informed by background knowledge and judgement. Assigning probabilities to universal propositions is more restrictive on the Ramseyian account than on Cox-Jaynes. Cox-Jaynes can compare and update between defined universal theories directly within a bounded hypothesis space. Ramsey cannot: a bet on a universal proposition has no settleable winning outcome, so probabilities attach to observable consequences rather than to the universal proposition itself.
A further divergence concerns induction. Cox-Jaynes treats Bayesian updating as the formal expression of inductive reasoning. The two are unified within the probability calculus. Ramsey separates them. Bayesian updating governs the coherence of degrees of belief in particular propositions. The assessment of variable hypotheticals (the inductive rules that generate those propositions) sits outside the probability calculus and cannot be reduced to Bayesian updating. For Ramsey, coherence and reliability are distinct tasks answered by distinct methods.
Taken together, Ramsey's positions shift the question. The Popperian asks: is this theory true, and are we getting nearer to truth? The Ramseyian asks: are our expectation-generating rules reliable, and are our degrees of belief coherent? These are not the same question. Ramsey's variable hypotheticals occupy a similar role to Popper's conjectures: both are provisional rules that guide expectation and are open to revision. The difference lies in how revision is triggered. Popper assesses conjectures by attempted falsification. Ramsey assesses variable hypotheticals by reliability. Whether the shift from Popper to Ramsey is a gain or a loss depends on whether one thinks science needs a target that stands independently of the inquirer, or whether that target is a regulative fiction and explaining and tracking the world reliably is what science does.
I was taught to evaluate an inductive argument by ‘Are the premises true?’ and ‘Is the argument strong?’. Now I evaluate an inductive argument by ‘How reliable is the expectation-generating rule’ and ‘What should be the priors for an expectation in a particular case’
In his essay “Two Faces of Common Sense” (1972) Popper wrote ‘We are seekers for truth but we are not its possessors’. In the same essay, Popper argues ‘The quest for certainty, for a secure basis of knowledge, has to be abandoned.’
Popper wanted to separate (un)certainty and truth and he believed that science should seek truth. Two quotes from Popper: ‘science has nothing to do with the quest for certainty or probability or reliability’ and ‘we too see science as the search for truth’.
For Popper, truth is fully independent of the believer. I have always struggled with the concept of ‘truth’ as ‘independent of the believer’. Maybe Popper did not want to replace God with Truth, but to me it could appear that way.
My interest in the philosophical ideas of Frank Ramsey (1903-1930) followed from listening to a podcast ‘A tale of truth’ by Simon Blackburn in 2023. Frank Ramsey’s redundancy theory of truth appealed to me. Popper himself once thought the correspondence theory of truth was dispensable, and noted in “Conjectures and Refutations” that Ramsey had suggested it might be empty altogether. If you adopt the redundancy theory of truth, science becomes managing uncertainty and assessing universal laws for reliability.
Reading the philosophical ideas of Frank Ramsey also addressed another issue I was struggling with. In his essay “Conjectural Knowledge” (1972), Popper’s rejection of induction is total at the logical level, while at the same time Popper argued that this creates no clash with rationality, empiricism, or scientific practice. My view is that people do reason non-deductively. Inductive reasoning has traditionally been interpreted through a deductive lens: an argument that has a conclusion and supporting premises. This interpretation works very well for deduction, but I think it is incorrect for induction. The re-interpretation of induction by Frank Ramsey provided an ‘a-ha’ moment to me.
Finally, the Cox-Jaynes theory is widely used in Bayesianism. The Cox-Jaynes theory is founded on three desiderata. I discovered that Frank Ramsey provides an alternative account of Bayesian updating based on the coherence of betting behaviour.
In this essay I will describe Ramsey’s epistemology and highlight the implications of adopting a ‘Ramseyian’ approach on how to deal with universal laws and induction.
Frank Ramsey on Truth, Degrees of Belief and Inductive Reasoning
Note on sources: This essay draws on the following works by Frank Ramsey: 'Facts and Propositions' (1927), 'Truth and Probability' (1926), 'Law and Causality' (1928), and the two 1929 notes 'Knowledge' and 'Probability and Partial Belief.' Where a section combines positions from more than one essay, or draws an inference that Ramsey does not state explicitly, this is noted in the text. Direct quotations are taken from Ramsey's own words. The section on the reinterpretation of induction combines Ramsey's betting framework with his treatment of variable hypotheticals in a form he does not himself state, but each step draws on positions he does hold.
Truth
Ramsey argues that truth is not a property that needs its own theory. Saying 'it is true that Caesar was murdered' adds nothing to saying 'Caesar was murdered.' This is now called the Redundancy Theory of Truth.
This contrasts with the correspondence theory, where a statement is true if and only if it corresponds to the facts, and truth holds independently of any mind that may or may not believe it.
Ramsey does not deny that facts exist. He is a realist about the world. What he denies is that 'true' names a relation between a proposition and a fact. That connection is carried by the proposition itself. 'True' is not what connects language to the world. It is a device for reasserting and generalising what is already asserted.
Beliefs and Degrees of Beliefs
Beliefs are instruments for guiding action. Frank Ramsey understands belief as a mental state that tends to produce certain actions, and degree of belief as how strong that tendency is: not a felt intensity, but a measure of what a person would do or be disposed to do on the basis of holding it.
He distinguishes beliefs that are consciously held and expressed in words or images from mere habits of response, such as an animal avoiding food it has learned is harmful. It is this kind of belief that can be assessed. Such a belief is, in Ramsey's own words, 'a map of neighbouring space by which we steer.'
Degrees of belief, for Ramsey, govern rational choice under uncertainty. A person uncertain about a proposition weights possible outcomes by their degree of belief and acts on whichever option produces the greatest expected value. His 'crossroads' example illustrates this: how far a traveller is willing to go to verify a route depends directly on how confident they are in it. Degrees of belief must conform to the probability calculus, since incoherent degrees of belief expose an agent to what Ramsey calls having 'a book made against him by a cunning better,' a guaranteed loss regardless of outcomes.
Ramsey distinguishes two ways of evaluating belief. Coherence governs how degrees of belief relate to one another and is captured by the probability calculus. Reliability concerns the causal habits by which beliefs are formed, assessing whether they tend to produce beliefs that are borne out. This distinction is drawn across several essays rather than stated in a single place by Ramsey.
Ramsey's account of degrees of belief rests on three connected steps: assigning a scale to value, treating mathematical expectation as a psychological law, and proving that consistent degrees of belief must obey the probability calculus.
Step 1 Assigning an interval scale to value
Ramsey defines degrees of belief as willingness to take a bet, where the stakes are measured in units of value. Value is a measure of how much an agent prefers one outcome over another. Ramsey constructs a scale of value in which differences between positions are meaningful but the origin and unit are arbitrary. Later decision theory would call this an interval scale. Ramsey begins with the special case in which an agent's degree of belief is one-half, making the agent indifferent between betting either way, and uses such indifference conditions in constructing a scale of value.
From that anchor point, the rest of the interval scale is constructed by presenting the agent with a series of choices between a certain outcome and a bet whose result depends on whether a proposition is true or false (a conditional bet). Each time the agent is indifferent between the certain outcome and the bet, the difference in value between the outcomes of that bet is mapped onto the interval scale.
Ramsey is not concerned with the absolute value of any outcome. What the interval scale captures is the difference in value between outcomes, mapped relative to the anchor point. The anchor point establishes that two outcomes with the same value for the agent occupy the same position on the scale. Differences in preference between outcomes are mapped as gaps on the interval scale, with larger gaps corresponding to larger differences in preference.
What Ramsey called value later became utility, the standard term in economics and decision theory.
Step 2 Psychological law
Ramsey treats mathematical expectation as a psychological law: a descriptive approximation of how agents could behave when reasoning under uncertainty, on the basis that all deliberate action involves, in some sense, acting on a bet about how the world will turn out.
An agent who reasons in accordance with mathematical expectation takes each possible outcome, multiplies the difference in value that outcome represents by the degree of belief attached to the proposition on which that outcome depends, sums those products across all outcomes for a given option, and selects the option with the highest total. Since value is measured as differences in position on the interval scale rather than as fixed numbers, the calculation does not require exact values for any outcome. The arbitrary origin and unit of the interval scale cancel out, and what matters is how the differences between outcomes compare to each other.
Ramsey acknowledges that this approximation cannot account for all the facts of human behaviour. An agent might choose to reason differently.
Step 3 Mathematical proof
If an agent does reason and act in accordance with mathematical expectation, then Ramsey proved mathematically that the agent's degrees of belief must obey the probability calculus or the agent will be exploitable. An agent is exploitable if a shrewd opponent can construct a set of bets, each acceptable to the agent at the agent's own stated odds, that guarantees the agent a loss regardless of how the propositions turn out.
The probability calculus includes the multiplication law, which states that the degree of belief in two propositions both being true equals the degree of belief in the first multiplied by the degree of belief in the second given the first. The multiplication law is a static relationship between degrees of belief at a given moment.
Ramsey formulates what later became known as conditionalization, but he does not offer an independent derivation of it. Instead, the update rule appears as a natural consequence of his treatment of conditional belief and the multiplication law. Conditionalization is the dynamic rule that tells an agent how to update degrees of belief when new evidence arrives. Within Ramsey's framework, updating by conditionalization is the consistent way of revising beliefs in light of new evidence, given the multiplication law. The multiplication law provides the foundation, while conditionalization follows as its consequence for belief revision.
If an agent chooses to reason like a betting man, their degrees of belief must obey the probability calculus and they must update those beliefs with new evidence using Bayes' theorem.
Reinterpretation of Induction
The following combines Ramsey's betting framework from Chapter 4 with his treatment of variable hypotheticals in Chapter 7. Ramsey does not state the argument in this form, but each step draws on positions he does hold.
Assigning probabilities to universal propositions is problematic
A Ramsey-inspired objection to assigning probabilities directly to universal laws is that, within a betting interpretation, such propositions do not admit the same straightforward settlement conditions as ordinary particular propositions.
The reasoning is as follows:
a. Within a betting interpretation, a probability assignment is only meaningful if it corresponds to a wager with a determinate settlement procedure.
b. Such a wager requires that both winning and losing outcomes be reachable in principle.
c. A universal proposition such as "All swans are white" ranges over every instance without exception, including those never encountered. No finite set of observations covers all instances. A bet on such a proposition has no determinate settlement procedure. The winning outcome, confirmation across all instances, is not reachable. The losing outcome, a single counterexample, is.
d. A universal proposition therefore does not satisfy the settlement conditions in (a) and (b).
e. On this reconstruction, assigning probabilities directly to universal propositions or universal laws is problematic. Particular propositions derived from experience do not raise the same difficulty.
Induction and Expectation-Generating Rules
A standard textbook definition of induction has been ‘the process of reasoning to a general or universal conclusion on the basis of a number of particular observations’. This would imply that from the observation ‘I saw 1,000 white swans’ the universal proposition ‘All swans are white’ could be inferred. And induction has traditionally been evaluated using propositional logic.
On that view, the problem of induction is the problem of justifying the inference from a finite number of observed cases to a universal proposition covering all cases, including those not yet observed. No finite number of observations can conclusively verify a universal proposition, and so the logical gap between evidence and conclusion has never been fully closed. The previous section argued, drawing on Ramsey's betting framework, that addressing this gap by assigning probabilities to universal propositions is problematic.
Frank Ramsey reframed induction:
Not: ‘I saw 1,000 white swans’ inducing the universal proposition ‘All swans are white’.
But: ‘I saw 1,000 white swans’ inducing the expectation-generating rule ‘Next time I encounter a swan, I expect it to be white’
Ramsey named the expectation-generating rule a ‘variable hypothetical’.
On a reconstruction combining Ramsey's account of partial belief with his treatment of variable hypotheticals, induction is not a matter of inferring from particular propositions to a universal conclusion. It is a matter of adopting or revising expectation-generating rules on the basis of particular experience. The degree of belief that is open to assessment is not the degree of belief in a universal conclusion, but the degree of belief in the next particular case derived from a variable hypothetical. The agent bets on the next swan being white, not on all swans being white. The variable hypothetical generates that bet, and is assessed by whether its expectations tend to be borne out.
On this reconstruction, induction is best understood as the formation, revision and assessment of variable hypotheticals that generate expectations from experience.
Assigning probabilities to a variable hypothetical is as problematic as assigning probabilities to universal propositions. Induction should be assessed on the reliability of the variable hypothetical: do the expectation-generating rules tend to produce expectations, expressed as propositions, that are borne out.
This connects to Ramsey's distinction between coherence and reliability: coherence governs how degrees of belief in particular propositions relate to one another, while reliability governs whether the expectation-generating rules from which those propositions are derived tend to produce expectations that are borne out.
Knowledge
In his 1929 essay, Ramsey states that a belief is knowledge if it is (i) true, (ii) certain, and (iii) obtained by a reliable process. He does not elaborate on the truth condition. He devotes most of the essay to worrying about what reliable process means, eventually preferring the phrase 'formed in a reliable way.'
That certainty is not immune to doubt. Ramsey acknowledges Russell's point that 'all our knowledge is infected with some degree of doubt' and does not reject it.
The redundancy theory, developed in 'Facts and Propositions,' suggests that truth adds no independent requirement beyond what the proposition itself asserts. Whether Ramsey intended these two positions to be reconciled, or considered the tension worth addressing at all, the essays do not say.
A belief is formed in a reliable way when it is caused by what it is about through a rule for judging (a variable hypothetical) that can generally be relied upon to produce beliefs that are borne out, and where any intermediate beliefs in the causal chain are themselves borne out.
Logic
'Logic must then fall very definitely into two parts: (excluding analytic logic, the theory of terms and propositions) we have the lesser logic, which is the logic of consistency, or formal logic; and the larger logic, which is the logic of discovery, or inductive logic.' From Ramsey, F. P. (1926) 'Truth and Probability'
‘The logic of consistency’.
The logic of consistency deals with propositions and degrees of belief in propositions. A proposition is the kind of thing that can be asserted or denied, borne out or not. It includes mathematics and the probability calculus. It assesses whether beliefs cohere with one another. It carries a necessity of assertion: if one asserts p, one is bound in consistency to assert whatever follows from p. The logic of consistency asks: are my degrees of belief coherent with one another? It governs rational organisation of uncertainty: given what one believes, what else is one bound to believe.
‘The logic of discovery’
The logic of discovery deals with variable hypotheticals as well as propositions. A variable hypothetical is not a proposition in the primary sense: it cannot be asserted or denied as a proposition can, and the standards of consistency that govern degrees of belief in propositions are not the appropriate standard by which to assess it. It can, however, be disagreed with, adopted or abandoned, and assessed by whether it reliably generates expectations that are borne out. The logic of discovery includes induction. It assesses whether habits of belief formation track the real world. Individual beliefs are then assessed derivatively, by reference to the habits that produce them. One is bound to revise a habit that proves unreliable, on pain of forming beliefs that are not borne out. The logic of discovery asks: do my habits of belief formation track the real world? It governs which habits of expectation are worth trusting, given how the world has behaved.
Summary
Ramsey's positions on truth, degrees of belief and induction are connected.
His redundancy theory holds that 'true' adds nothing to a proposition: it functions as a device for reasserting and generalising what is already asserted rather than naming a relation between a proposition and a fact.
Degrees of belief, measured by willingness to bet in terms of utility, must obey the probability calculus or the agent is exploitable, and must be updated by Bayes' theorem for the same reason.
Induction is reframed through variable hypotheticals (expectation-generating rules). The traditional approach to assessing induction through propositional logic is not adopted. Observations or reasoning from analogues do not generate universal propositions such as 'All swans are white', but an expectation-generating rule: 'If I encounter a swan next, I expect it to be white.' Induction is therefore not a matter of assigning probabilities to universal laws but of assessing whether those expectation-generating rules are reliable.
Taken together, coherence assesses how degrees of belief relate to one another and is captured by the probability calculus. Reliability assesses whether the expectation-generating rules an agent adopts, such as expecting the next swan to be white, tend to be borne out in practice.
Ramsey's epistemology is built around these two distinct but related tasks: assessing the coherence of degrees of belief, and assessing the reliability of the expectation-generating rules from which those beliefs derive.
If one rejects Humean or Popperian demands for a deductively justified theory of induction, Ramsey offers a replacement in which induction is understood as the formation and revision of variable hypotheticals that guide expectations rather than infer universal propositions. And if one is uncomfortable with Cox-style derivations of probability, Ramsey also provides an alternative justification of Bayesian updating based on coherence of betting behaviour rather than functional representation theorems.
Cox-Jaynes Bayesianism vs Ramseyian Bayesianism
The Cox-Jaynes framework and Ramsey's account of probability share a common starting point but diverge in three respects that matter for how each handles objectivity, universal laws, and induction. The first divergence concerns the degree to which the probability calculus constrains not just the relations between beliefs but their content. The second concerns how each framework handles universal propositions and universal laws. The third concerns the treatment of induction itself, and whether Bayesian updating is the right formal expression of inductive reasoning or whether the assessment of induction belongs outside Bayesian updating.
Both the Cox-Jaynes framework and Ramsey's account share a commitment to consistent reasoning. Both hold that degrees of belief must satisfy the probability calculus if they are to be mutually consistent. The probability calculus covers the sum rule, the product rule, and Bayes' formula. These rules fix how probability assignments must relate to one another and how they should be revised in light of new information. They are not matters of personal preference.
Cox-Jaynes goes further. Desiderata IIIb and IIIc make probability assignments, in Jaynes' words, "completely 'objective' in the sense that they are independent of the personality of the user." They are a means of encoding "the information given in the statement of a problem, independently of whatever personal feelings you or I might have about the propositions involved." If the information state specifies relevant values such as base rates, likelihoods, and false positive rates, those values are fixed for any agent reasoning according to the desiderata. Two agents with the same information ought to arrive at the same probability assignments. Anyone who reaches a different conclusion, Jaynes states, "is necessarily violating one of those desiderata."
Ramsey's account contains no corresponding principle. The probability calculus constrains the relations between an agent's beliefs, but it does not constrain the inputs themselves. An agent may use the values given by the information state or adjust them in light of personal belief, and remain fully coherent either way. Ramsey holds that asking what initial degrees of belief are justified is, in his words, "a meaningless question," and that formal probability theory cannot determine what prior probabilities an agent ought to hold. His framework provides consistent constraints on the coherence of beliefs, but not prescriptions concerning their content.
If one is uncomfortable with Cox-Jaynes derivations of probability, Ramsey provides an alternative justification of Bayesian updating based on coherence of betting behaviour rather than the three desiderata.
A further divergence concerns universal propositions, where the two frameworks take different approaches.
Both Cox-Jaynes and Ramsey face difficulty in assigning probabilities directly to universal propositions. Cox-Jaynes excludes them on formal grounds. The framework requires a defined and bounded hypothesis space. A universal proposition tested against an open-ended class of alternatives has, in Jaynes' words, a probability that "is simply undefined because the class of all conceivable theories is undefined." Within a bounded set of specified alternatives, Cox-Jaynes supports Bayesian updating directly. Given two or more defined universal propositions, Bayes' theorem updates their relative probabilities in light of evidence in the normal way.
Ramsey's measurement procedure takes a different path. His framework grounds degrees of belief in bets, and a bet requires a settleable outcome. A universal proposition cannot be conclusively verified in finite time, so the bet cannot be closed. Ramsey's framework therefore assigns probabilities to the observable consequences of competing universal propositions and updates on those. The bet is placed on the observable outcome rather than on the universal proposition directly. Universal propositions enter Ramsey's framework indirectly, through their testable implications, rather than as direct objects of degree of belief.
If one is uncomfortable with comparing universal laws and updating confidence in universal laws using Bayes' formula, Ramsey provides an alternative route, assigning probabilities to the observable consequences of competing universal laws rather than to the universal laws themselves.
A third divergence, related to the above, concerns the treatment of induction itself.
Jaynes treats Bayesian updating as the quantitative expression of inductive reasoning. In his words, "modern Bayesian analysis is just the unique quantitative expression of this reasoning format; the inductive reasoning that philosophers like Hume and Popper held to be impossible." Within a bounded, well-defined set of alternatives, repeated observations raise or lower the probability of competing hypotheses via Bayes' theorem. Successful predictions increase confidence in a hypothesis; failed predictions reduce it. Induction, on this account, is a quantitative process conducted within the probability calculus, with Bayesian updating as its formal mechanism.
Ramsey takes a different route and redefines induction. He treats universal propositions not as propositions at all, but as variable hypotheticals: expectation-generating rules applied to particular cases rather than statements with determinate truth conditions.
Because these rules are not propositions in the primary sense, they fall outside the betting framework and outside Bayesian updating. Their assessment rests on reliability, specifically whether the expectations they generate prove correct over time. That reliability assessment sits outside the probability calculus and cannot be reduced to Bayesian updating.
Observed frequencies can inform that assessment. An agent may form a degree of belief in the proposition that a given rule has succeeded in a certain proportion of cases. That degree of belief is directed at a proposition about observed frequencies rather than at the variable hypothetical itself. The rule, as a variable hypothetical, is not the kind of thing that can be the direct object of a degree of belief within the betting framework.
The asymmetry is precise. Jaynes treats Bayesian updating as the formal expression of inductive reasoning. Ramsey places the assessment of the variable hypotheticals that guide induction outside Bayes' theorem altogether.
If one is uncomfortable with treating induction as 'the process of reasoning to a general or universal conclusion on the basis of a number of particular observations', Ramsey provides an alternative route, treating inductions as expectation-generating rules applied to particular cases, assessed by their reliability rather than their logical justification.
Example: The Swan and the Variable Hypothetical
An agent observes 1,000 white swans. Induction, in Ramsey's reframing, does not generate the universal proposition ‘All swans are white’. It generates the variable hypothetical: ‘The next time I encounter a swan, I expect it to be white.’ This expectation-generating rule can be treated as a bet.
The bet is: ‘The next swan I see will be white.’
The prior reflects the agent's willingness to bet, given background knowledge. Having observed 1,000 white swans, the agent assigns a high but uncertain prior -- between 0.5 and 0.95, depending on location, season, and other background knowledge. A prior of 1 is not advisable. An agent can never be certain.
In this first version of the example, the likelihood and false positive are degenerate:
· Likelihood: P(observe white swan | next swan will be white) = 1
· False positive: P(observe white swan | next swan will not be white) = 0
If the evidence is a white swan, the posterior is 1: the bet is won. If the evidence is a black swan, the likelihood becomes 0 and the posterior becomes 0: the bet is lost. The agent could revise the reliability of the variable hypothetical accordingly.
This degenerate case illustrates the structure of the Ramseyian approach. The agent bets on the next particular case, not on the universal proposition. The universal proposition "All swans are white" is never the object of the bet and never needs to be assigned a probability.
The example becomes genuinely probabilistic when background knowledge enters. Suppose the agent learns two things before placing the next bet: the worldwide proportion of white swans is 85%, and the next observation will be made in the southern hemisphere, where black swans are native to Australia.
These two pieces of information pull in opposite directions. The worldwide base rate favours white. The location pulls against it. The agent can no longer justify a prior of 0.9. A prior closer to 0.15 is more defensible, depending on precisely where in the southern hemisphere the observation will take place.
The likelihood is no longer 1. Even if the next swan will be white, the agent might observe a white swan that is an introduced European species, an escaped captive, or an albino. P(observe white swan | next swan will be white) remains high but not certain. A value of 0.95 is reasonable.
The false positive is no longer 0. P(observe white swan | next swan will not be white) is small but positive. A white swan sighting in the southern hemisphere is not impossible under the not-white hypothesis. A value of 0.05 reflects this residual possibility.
The posterior, given a white swan sighting, is:
P(H | white swan) = (0.15 × 0.95) / [(0.15 × 0.95) + (0.85 × 0.05)] = 0.1425 / 0.185 ≈ 0.77
The prior was 0.15. Observing a white swan raises it to roughly 0.77, because a white swan sighting is surprising under the not-white hypothesis. Uncertainty remains.
The posterior from this bet informs the prior for the next one, but does not determine it mechanically. New background knowledge may have entered: a report of introduced species, a change of location, a different season. The agent exercises judgement in setting the new prior. As Jaynes notes in Probability Theory: The Logic of Science, the prior for any new problem encodes all relevant prior information, not merely the output of the last calculation. This preserves the subjective character of the prior that Ramsey insisted upon: the prior reflects the agent's actual willingness to bet, not a formula applied blindly to previous results.
The two examples together show the Ramseyian approach at different levels of realism. The variable hypothetical generates the prediction in both cases. In the first, the probabilities are degenerate and the bet collapses into a deductive verdict. In the second, new background knowledge acts as evidence that reshapes the prior through Bayesian updating, and the agent enters the bet with a revised degree of belief. Induction, on this account, is the formation and revision of expectation-generating rules under uncertainty, not the inference of universal propositions from particular observations.
The swan example has no scientific content. It does not explain why swans are likely to be white. It illustrates the structure of the Ramseyian approach but nothing more.
A scientific universal theory such as Newton's law of gravity carries both explanatory and predictive power. Observations have established that Newton's law is reliable at low velocities and weak gravitational fields. Einstein's theory of relativity has proved reliable at relativistic velocities and under strong gravity. Both theories generate variable hypotheticals that can be assessed for reliability across different conditions. The shift from Newton to Einstein is not a matter of one universal proposition replacing another, but of one expectation-generating rule proving more reliable than its predecessor across a wider range of conditions.
Conclusion
Ramsey's epistemology offers a coherent alternative to the Popperian picture on three connected points.
On truth, the redundancy theory removes the need for a correspondence relation that holds independently of any believer. Truth adds nothing to a proposition beyond what the proposition itself asserts. Science, on this account, does not seek a target that stands apart from inquiry. It manages uncertainty and assesses whether expectation-generating rules are borne out.
On induction, Ramsey's reframing bypasses the logical problem entirely. The traditional interpretation treats induction as an inference from particular observations to a universal conclusion, and then struggles to close the logical gap. Ramsey dissolves the gap by changing what induction produces. The output is not a universal proposition but a variable hypothetical: an expectation-generating rule assessed by its reliability over time, not by its logical justification.
On Bayesian foundations, Ramsey provides an account of Bayesian updating grounded in the coherence of betting behaviour. The probability calculus constrains the relations between an agent's degrees of belief, and Bayes' theorem follows as the consistent update rule. This does not require the Cox-Jaynes desiderata, and it does not prescribe what prior probabilities an agent ought to hold. The prior reflects the agent's actual willingness to bet, informed by background knowledge and judgement. Assigning probabilities to universal propositions is more restrictive on the Ramseyian account than on Cox-Jaynes. Cox-Jaynes can compare and update between defined universal theories directly within a bounded hypothesis space. Ramsey cannot: a bet on a universal proposition has no settleable winning outcome, so probabilities attach to observable consequences rather than to the universal proposition itself.
A further divergence concerns induction. Cox-Jaynes treats Bayesian updating as the formal expression of inductive reasoning. The two are unified within the probability calculus. Ramsey separates them. Bayesian updating governs the coherence of degrees of belief in particular propositions. The assessment of variable hypotheticals (the inductive rules that generate those propositions) sits outside the probability calculus and cannot be reduced to Bayesian updating. For Ramsey, coherence and reliability are distinct tasks answered by distinct methods.
Taken together, Ramsey's positions shift the question. The Popperian asks: is this theory true, and are we getting nearer to truth? The Ramseyian asks: are our expectation-generating rules reliable, and are our degrees of belief coherent? These are not the same question. Ramsey's variable hypotheticals occupy a similar role to Popper's conjectures: both are provisional rules that guide expectation and are open to revision. The difference lies in how revision is triggered. Popper assesses conjectures by attempted falsification. Ramsey assesses variable hypotheticals by reliability. Whether the shift from Popper to Ramsey is a gain or a loss depends on whether one thinks science needs a target that stands independently of the inquirer, or whether that target is a regulative fiction and explaining and tracking the world reliably is what science does.
I was taught to evaluate an inductive argument by ‘Are the premises true?’ and ‘Is the argument strong?’. Now I evaluate an inductive argument by ‘How reliable is the expectation-generating rule’ and ‘What should be the priors for an expectation in a particular case’