The core idea
Scientific inquiry develops models and explanations that face evidence. Falsifiability asks whether a claim rules out possible observations, but actual tests depend on background assumptions and measurement. Reliable inquiry combines critical tests, uncertainty assessment, causal reasoning and openness to better-supported alternatives.
1. Description, prediction and explanation differ
A description records what happens; a prediction states what to expect under conditions; an explanation makes some aspect of the event intelligible. These tasks overlap without being identical. A calendar can predict a scheduled bell without explaining how its electrical circuit produces sound. A causal explanation identifies relevant processes or dependencies, while some scientific explanations appeal importantly to mathematical structure or constraints. Philosophers therefore debate whether one universal pattern captures all explanation. The deductive-nomological tradition examines derivation from laws and conditions; later accounts emphasise mechanisms, causes or unification. The lesson is not that anything counts as explanation. Ask what question is answered, why the cited factors matter, and what difference their presence makes. A restatement of the observation in more technical vocabulary is not automatically an explanation.
Sources: 20th Century Theories of Scientific Explanation, Stanford Encyclopedia of Philosophy ↗ · Causal Approaches to Scientific Explanation, Stanford Encyclopedia of Philosophy ↗
2. A useful model selects and simplifies
A model represents a target system for a purpose. It may be an equation, a diagram, a physical construction or a simulation. Abstraction leaves some features aside; idealisation can deliberately distort others, as when friction is set to zero. Such simplification can reveal an important dependence while limiting accuracy or scope. The right question is not only whether the model contains a false assumption, but whether that assumption undermines the use being made of it. A model suitable for estimating a short journey may fail for long-term planning under changing conditions. State the target, variables, assumptions and intended range before judging success. Two models of one system can emphasise different aspects without being interchangeable or equally good for every question. Comparing predictions with evidence helps establish where each earns trust.
Sources: Models in Science, Stanford Encyclopedia of Philosophy ↗
3. Falsifiability makes a claim vulnerable to evidence
Karl Popper emphasised a contrast between claims that forbid possible observations and claims adjusted to accommodate every result. Falsifiable does not mean already false; it means that some possible evidence would conflict with the claim. In logical form, theory T together with auxiliary assumptions A predicts P. If not-P is established, the conjunction of T and A has failed. It does not follow by logic alone that T, rather than an instrument assumption or initial condition, is the faulty part. This matters because real experiments test organised packages of commitments. Checking an auxiliary assumption can be responsible inquiry. Repeatedly inventing uncheckable excuses to protect a favourite theory is different. A revision becomes informative when it has independent support or yields further predictions that can themselves be examined.
Test a theory together with its assumptions
(T ∧ A) → P
¬P → ¬(T ∧ A)
| Time | Model prediction | Invented observation |
|---|---|---|
| 0 min | 65°C | Starting value |
| 10 min | 45°C | 50°C |
| 20 min | 35°C | No observation given |
4. Statistical evidence does not behave like a simple contradiction
A probabilistic model often permits every individual outcome while assigning very different probabilities to patterns. Under a fair independent coin model, ten heads in ten tosses has probability 1/1024. It is unusual, but not logically impossible. The observation therefore does not deductively prove that the coin is biased. It can nevertheless be evidence worth investigating, especially when a test and comparison were specified in advance. Repeated testing, selective reporting and choosing a pattern after seeing the data change how surprising the result should be considered. Measurement uncertainty also matters: a tiny difference may be compatible with the model's stated precision. Falsifiability alone is consequently not a complete recipe for scientific judgment. We also need clear statistical assumptions, reliable records and an honest account of uncertainty and alternative explanations.
Sources: Karl Popper, Stanford Encyclopedia of Philosophy ↗ · How science works, Understanding Science, University of California, Berkeley ↗
5. Worked application: testing an invented cooling model
Suppose a model for a particular setup says that a liquid's temperature excess above a constant 25°C room halves every ten minutes. Starting at 65°C, the excess is 40°C; after ten minutes it predicts 45°C, and after twenty minutes 35°C. These are stipulated example values, not reported experimental data. Now an observation at ten minutes is 50°C. If the discrepancy exceeds measurement uncertainty, the model-plus-assumptions needs investigation. Check elapsed time, the initial reading, room stability and the measurement method. Do not simply announce that all thermal theory is false, or silently replace the prediction after seeing the result. If independent checks confirm the conditions, the proposed halving rate may be wrong for this setup. Estimating a new rate should be followed by a prediction tested on additional observations.
Sources: Models in Science, Stanford Encyclopedia of Philosophy ↗ · Karl Popper, Stanford Encyclopedia of Philosophy ↗
6. Worked application: from a growth difference to a causal claim
An invented classroom study reports mean growth of 8 cm in one plant group and 12 cm in another. The second group received both more light and more water. The difference is 4 cm, but attributing it to light alone ignores the changed watering condition. A useful next study could vary a specified light treatment while holding other relevant conditions comparable and allocating plants to groups randomly. Repetition and enough independent units would help assess variability; one unusually vigorous plant is weak support for a broad claim. Even a well-supported treatment effect is not automatically a complete mechanism explaining the biological pathway. This separates three achievements: noticing a pattern, identifying a causal effect under stated conditions, and explaining how that effect is produced. Each can require additional evidence.
Sources: Causal Approaches to Scientific Explanation, Stanford Encyclopedia of Philosophy ↗ · How science works, Understanding Science, University of California, Berkeley ↗
7. Critical inquiry is organised, revisable and plural
Actual science moves between questions, observations, models, tests and discussion rather than following one inflexible staircase. Independent checks and criticism help expose errors that an individual can miss. Values influence which problems receive attention and what risks are acceptable, while factual conclusions still require evidence. An ethical question about which harms to permit is not settled by a measurement alone. Likewise, a metaphysical teaching is not confirmed merely because its vocabulary resembles a scientific model. Classical Nyāya's concern with a reason's relevance offers a useful philosophical comparison, but should not be relabelled as identical to modern experimental science. Across these distinctions, a responsible explanation states what is known, how it was tested, which assumptions matter and where confidence should stop. Revisability is a discipline, not a licence to accept every alternative equally.
Sources: How science works, Understanding Science, University of California, Berkeley ↗ · Nyāyasūtra: Sanskrit text, GRETIL, University of Göttingen ↗
PUT IT INTO PRACTICE
Practice: reconstruct and improve a test
- Use the invented cooling model. From an initial 57°C and a constant 25°C room, calculate a 32°C excess and a ten-minute prediction of 41°C.
- Suppose the reading is 44°C. Record the 3°C discrepancy and ask whether it exceeds the instrument’s uncertainty; the size alone cannot answer that question.
- List independently checkable assumptions: starting temperature, elapsed time and stable surroundings. Explain that a failed prediction challenges their conjunction with the model.
- If the checks hold, revise the proposed rate and reserve new observations for testing it. Fitting the old observation alone does not establish predictive success.
Check your understanding
Can an idealised model still be useful?
Yes, when its simplifications preserve the relationships needed for a stated purpose. Its limits and conditions must remain explicit.
What follows from T and A predicting P, but not-P being established?
The conjunction of T and A fails. Logic alone does not identify which component is responsible.
Why are ten heads not a logical disproof of a fair coin?
The fair independent model assigns that sequence probability 1/1024, not zero. Unusual evidence can matter without being impossible under the model.
What makes a revised assumption more than an excuse?
Independent support or new testable predictions make the revision accountable. An adjustment that evades every possible test does not.
Why cannot the 4 cm plant difference be assigned to light alone?
Watering changed too. The comparison has not separated the effects of the two factors or assessed other relevant variation.
Does identifying a treatment effect fully explain its mechanism?
No. Evidence that changing a factor affects an outcome can precede evidence about the intermediate processes producing that outcome.