‘Before science, there is just thinking. The question is what kind.’
Intuitive (System 1) thinking is fast, automatic, and mostly unconscious; it uses pattern recognition, prior experience, and social signals to arrive at conclusions without deliberate effort. It feels certain. It is often wrong.
Intuition defaults to authority bias (believing what a credible source says), confirmation bias (seeking evidence that fits what we already believe), and availability heuristic (judging probability by how easily examples come to mind).
The problem is not that intuition is irrational, it is often adaptive. The problem is that it is non-transparent: you cannot audit its reasoning, reproduce it, or know when it has failed.
Science begins when you notice your intuition might be wrong and decide to check
The shift to hypothetical-deductive thinking
Science is not the absence of intuition it is intuition made falsifiable
Observation→Question→Hypothesis→Prediction→Test→Revise or retain
Deductive reasoning: from general rules to specific predictions. “If all swans are white and this is a swan, it must be white.” The conclusion is guaranteed by the premises but the premises must be independently justified.
Inductive reasoning: from specific observations to general rules. “I have seen 1,000 white swans, therefore all swans are white.” The conclusion is never guaranteed that one black swan falsifies it. This is the fundamental problem of induction (Hume).
Abductive reasoning: inference to the best explanation. Given what we observe, what hypothesis most parsimoniously explains it? Not guaranteed, but productive it generates testable candidates. This is how clinical diagnosis works.
The hypothetical-deductive method (Karl Popper) does not ask “what proves this?” but “what could disprove it?” This reframing is the core of scientific thinking.
Rationalism vs empiricism
Rationalism holds that knowledge can be derived through logic and reason alone, without direct observation. Judging a race winner from how confident they look is rationalism you are reasoning from prior beliefs, not observing the race.
Empiricism holds that knowledge must ultimately come from experience and observation. A claim is only knowledge if it has been tested against reality.
Modern science is empirical with rational scaffolding: experiments must be observable and measurable, but their design, analysis, and interpretation require rigorous reasoning.
What makes a hypothesis scientific: the rules
Not every claim is a hypothesis. A scientific hypothesis must satisfy specific criteria before it earns the right to be tested.
THE CORE CRITERIA
Falsifiability (Popper): a hypothesis is scientific only if there exists a possible observation that could prove it wrong. “All swans are white” is falsifiable; one black swan refutes it. “Everything happens for a reason” is not falsifiable, nothing could ever refute it. A theory that explains everything explains nothing.
Specificity / operationalisability: the hypothesis must be statable in terms of measurable quantities. “Exercise improves health” is too vague. “Thirty minutes of moderate aerobic exercise five days a week reduces resting heart rate in sedentary adults over twelve weeks” is testable.
Equipoise: the researcher must genuinely not know which outcome is more likely. Equipoise means you are in honest uncertainty you have not already decided what you want to find. A document written to win rather than to test is not in equipoise; every framing favors the hypothesis.
Parsimony (Occam’s Razor): between two hypotheses that equally explain the data, prefer the simpler one. Do not multiply causes beyond necessity. The simplest theory that still explains everything is the better working hypothesis.
The demarcation problem (Popper again): what separates science from non-science is not the content but the structure is it falsifiable? Astrology makes predictions, but its practitioners adjust the framework to absorb any failure, making it unfalsifiable and therefore non-scientific.
SELECTING AMONG COMPETING HYPOTHESES
Get Jatin.’s stories in your inbox
Join Medium for free to get updates from this writer.
Apply the Principle of Parsimony first to eliminate the unnecessarily complex.
Rank remaining hypotheses by prior plausibility (what does existing evidence suggest?), testability (which can be tested with available resources?), and consequence (which, if true, would matter most?).
A hypothesis about mechanism is stronger than a hypothesis about association if you can explain why, not just that, you have a better scientific claim.
Designing the test
Correlation is not causation. Ice cream sales correlate with drowning rates; the confounder is summer. Two variables moving together tells you nothing about which causes which, or whether a third variable drives both.
Confounding variables are factors that influence both the independent and dependent variable, creating a spurious association. Identifying and controlling for confounders is the central challenge of observational research.
Randomisation is the cleanest solution randomly assigning participants to conditions that distributes known and unknown confounders equally across groups. Without randomisation, you must argue exhaustively that no relevant confounders differ between groups.
Surrogate endpoints vs direct endpoints: we often measure what we can rather than what we want to know. Cholesterol level is a surrogate endpoint of what we actually want to know is cardiovascular mortality. Surrogate endpoints are valid only if the relationship between the surrogate and the outcome of interest is itself established.
Hill’s Criteria for causation (for observational data where randomisation is impossible): strength of association, consistency across studies, specificity, temporality (cause precedes effect), biological plausibility, dose-response relationship, coherence with existing knowledge, experimental evidence, analogy. Temporality is the only necessary criterion cause must precede effect.
Observer effects / Hawthorne Effect: the act of observation changes what is being observed. Knowing you are being watched alters behaviour. Study designs must account for this blinding and control conditions exist precisely to neutralise it.
Interpreting evidence what the numbers actually mean
A p-value does not tell you the probability that your hypothesis is true. It tells you the probability of observing data this extreme if the null hypothesis were true. p < 0.05 is a threshold, not a fact it means roughly one-in-twenty chance of a false positive even if nothing is real.
Statistical significance ≠ practical significance. A study of 100,000 people can produce p < 0.001 for an effect that is meaninglessly small in the real world. Always ask: what is the effect size?
Bayesian updating: new evidence should update prior beliefs proportionally to how well the evidence distinguishes between competing hypotheses. A single significant result in a low-prior-probability domain (e.g., homeopathy) should move your belief less than a significant result in a high-prior-probability domain.
Base rate neglect: we routinely ignore background frequencies when interpreting new information. A highly accurate test for a rare disease will still produce many false positives because the disease is rare and the base rate matters enormously.
What corrupts the process
The scientific method is only as good as the humans running it. Several systematic failures have been documented and named.
P-hacking: running analyses in multiple ways, adding or removing participants, or testing multiple outcomes until p < 0.05 is achieved. Each additional test inflates the false-positive rate. Pre-registration committing to a hypothesis and analysis plan before collecting data is the antidote.
Publication bias / the File Drawer Problem: journals publish significant results and reject null results. This means the published literature systematically over-represents positive findings. The game is structurally rigged toward false positives. Open science and registered reports are partial solutions.
Confirmation bias: selectively attending to evidence that supports the hypothesis and downweighting evidence that contradicts it. A document written to win rather than to test exhibits this structurally; every framing favors the hypothesis, no concessive language, unidirectional interpretation.
Certainty inflation: communicating results with more confidence than the evidence warrants. “This drug works” for a single small trial. “This study suggests” for a decade of replicated evidence. Scientific language must match the epistemic weight behind the claim.
Lead-time bias / survivorship bias: systematic errors that arise from how samples are constructed. Survivorship bias means you only see the winners, the companies, treatments, or people who made it without accounting for the identical cases that failed silently.
Algorithmic bias: AI systems are not neutral; they learn from historical data that encodes existing biases. Garbage In → Bias Out. A model trained on biased data produces biased outputs, and the bias is systematic (not random), making it harder to detect and more damaging at scale.
How science self-corrects
Reproducibility vs Replicability: reproducibility means re-running the same analysis on the same data and getting the same result (a minimum standard). Replicability means running a new study with the same method on a different sample and getting the same result (a stronger standard).
The Replication Crisis: large-scale replication projects (particularly in psychology and medicine) have found that a substantial fraction of published findings fail to replicate. This is not a failure of science, it is science working: the correction mechanism identified the error. The problem is that the original incentive structure (publish positive results, get funded) made the errors predictable.
Pre-registration: publicly committing to your hypothesis, sample size, and analysis plan before data collection. Converts confirmatory analysis (p-hacking risk) into genuinely confirmatory analysis (pre-specified). Registered reports go further; the study is accepted for publication before results are known, based on the quality of the question and design alone.
Open Science: sharing raw data, analysis code, and null results. Transparency is the structural solution to the file drawer problem. If other researchers can audit your pipeline, errors surface faster.
Peer review: not infallible it does not catch fraud, does not reliably detect p-hacking, and is subject to confirmation bias in reviewers. It remains valuable for basic sanity checking, but it is not a quality guarantee. Post-publication critique and replication are more important than pre-publication review.
Scientific thinking is not a checklist. It is a disposition of the willingness to be wrong, to hold conclusions proportionally to evidence, and to design tests that could genuinely threaten your hypothesis rather than merely support it. Every rule above falsifiability, equipoise, parsimony, pre-registration is a formalisation of that one underlying commitment.
