The Replication Crisis Explained: What Broke and What's Fixed

16 October 2026

The replication crisis is the realisation, which hit hard around 2011 to 2015, that a large share of published scientific findings could not be reproduced when other researchers repeated the same experiment. It hit psychology and parts of medicine hardest, though the underlying problems exist everywhere. Since then, several concrete fixes have changed how research gets done, even if the problem is not solved.

What actually happened

In 2011, a Cornell psychologist named Daryl Bem published a paper claiming evidence for precognition, the idea that people could sense future events before they happened. The statistics in the paper were conventional and the methods looked ordinary. That was the problem. If an unremarkable, normal-looking study could produce evidence for something impossible, then the normal methods themselves must be able to produce false positives far more often than anyone had assumed.

That paper became a trigger rather than a cause. It pushed a group of researchers to start systematically checking whether published results could be repeated. The biggest of these efforts, a project involving over a hundred researchers, tried to redo around 100 psychology studies using the original materials wherever possible. Less than half produced the same result the second time.

Around the same time, a similar pattern turned up in parts of medicine and biology. Pharmaceutical companies trying to build on published cancer biology research found that many foundational results would not reproduce in their own labs. The specific numbers varied between these projects, but the direction was consistent: a surprising number of textbook-ready findings fell apart under a second look.

Why so many studies failed

The core issue was not fraud, though a handful of high-profile fraud cases did surface during this period. The bigger issue was a set of ordinary, widespread habits that made it too easy to find something that looked like a real effect even when there was nothing there.

One habit was running many different analyses on the same data and only reporting the one that worked, a practice sometimes called p-hacking. Another was deciding what to test after seeing the data, then writing it up as though the hypothesis had been there from the start. Both practices are tempting because data is messy and almost always contains some pattern if you look hard enough.

A second problem was publication itself. Journals wanted interesting, positive results, not studies that found nothing. That created pressure to produce clean, exciting findings, and a reason for null results to stay in a filing cabinet rather than reach print. If a hundred labs test a weak effect and five find something by chance, those five get published and the other ninety-five vanish. Anyone reading only the published record would think the effect was real.

A third problem was small sample sizes combined with high research costs. Psychology experiments often ran with a few dozen participants. Small samples produce noisy results, and noisy results combined with flexible analysis choices are a reliable way to manufacture false positives that look statistically convincing.

Which fields were hit hardest

Social psychology took the heaviest public hit, partly because some of its most famous findings, the kind that end up in television documentaries, turned out to be among the weakest. Effects involving subtle behavioural nudges, such as the idea that holding a powerful posture for two minutes changes your hormones and your confidence, failed to replicate in larger, more careful follow-up studies.

Cancer biology and parts of preclinical medicine faced a quieter but arguably more consequential version of the same problem. Drug companies rely on published lab research to decide which biological targets are worth pursuing. When replication attempts on landmark cancer studies succeeded only part of the time, it meant years and considerable money had likely been spent chasing effects that were not there.

Economics and other social sciences also turned out to have reproducibility problems, though the scrutiny arrived a little later. Nutrition science has faced similar criticism, often because studies rely on people accurately remembering what they ate, which is a shaky foundation for precise claims.

Fields built on large physics experiments or heavily standardised lab procedures were comparatively less affected, not because researchers there are more careful people, but because the experiments tend to involve bigger datasets, tighter measurement, and less room for flexible after-the-fact analysis.

What has actually changed

Preregistration is the clearest concrete fix. Researchers now commonly post their hypothesis, their planned sample size, and their analysis method in a public registry before collecting data. This closes off the option of quietly trying several analyses and reporting only the one that worked, because the plan is locked in advance and visible to anyone who checks later.

Journals have started accepting what are called registered reports. A researcher submits the study design before running the experiment, and the journal commits to publishing the result regardless of whether it turns out positive or negative. This directly attacks the old incentive to bury null results, since publication no longer depends on the outcome being exciting.

Sample sizes have grown, particularly in psychology, where large multi-lab collaborations now routinely gather data from thousands of participants across many sites rather than a few dozen students from one university. Bigger samples produce more stable, trustworthy estimates and make chance findings far less likely to pass for real effects.

Replication attempts themselves have become a recognised and fundable kind of research, rather than a career risk for anyone who dared question a famous finding. Several journals now have sections specifically for replication studies, and some funding bodies will pay for large-scale replication projects rather than treating them as a lesser use of money.

What is still unresolved

None of these fixes are universal. Preregistration and registered reports remain the exception rather than the rule across most journals, and plenty of research, especially outside psychology, still gets published the old way, with no public record of what was planned versus what was found afterwards.

The incentive structure underneath academic careers has not fundamentally changed. Researchers are still mostly judged by how many papers they publish and how often those papers get cited, not by how reliable their findings turn out to be years later. That pressure has not vanished just because better tools now exist to fight it.

Replication is also expensive and unglamorous. It takes real time and money to redo someone else's experiment properly, and there is still no reward structure that makes this an attractive use of a young researcher's limited years. Most published findings, across most fields, have simply never been directly retested, which means the true scale of the original problem is still not fully known.

Common questions

Does a failed replication mean the original finding was fraudulent?

Usually not. Most failed replications trace back to small sample sizes, flexible analysis choices, or genuine differences between the original study and the repeat attempt, such as a different population or setting. Outright fraud exists and gets discovered sometimes, but it explains a small fraction of replication failures.

Has the replication crisis been fixed?

No, it has been improved in pockets rather than solved across the board. Tools like preregistration and registered reports work well where they are used, but adoption is patchy and the underlying career incentives that rewarded flashy findings over reliable ones are largely still in place.

How can I tell if a study I'm reading was preregistered?

Look for a statement near the methods section referencing a registration number or a link to a registry such as the Open Science Framework. If the paper mentions no such thing, assume the analysis plan was decided after the researchers had already seen the data, which should make you more cautious about the result.

Which famous psychology findings failed to replicate?

Power posing, the idea that briefly standing in a confident stance changes your hormones and risk tolerance, is one widely cited example that fell apart under larger follow-up studies. Ego depletion, the theory that willpower is a limited resource that gets used up over the course of a day, is another finding that weakened considerably once bigger, preregistered studies tested it.

Is the replication crisis only a problem in science, or does it affect other fields too?

The same pressures, small samples, flexible analysis, and a preference for exciting results, can affect any field that relies on statistics, including parts of economics, political science, and market research. The core lesson, that a single surprising result deserves scepticism until it has been independently repeated, applies well beyond academic psychology.

Read next

• What a Meta-Analysis Actually Does to the Evidence

  A meta-analysis combines results from many studies into one estimate, but it only clarifies the truth if those studies were asking the same question well.

• Observational Study vs Experiment: Why Diet Advice Flip-Flops

  Most nutrition headlines come from observational studies that track habits, not experiments that test them, which is why the advice keeps changing.

• What a Randomised Controlled Trial Actually Buys You

  A randomised controlled trial sorts people into groups by chance, which cancels out hidden differences and lets researchers credit any gap to the treatment.