Priming research looks like one literature and behaves like two. Effects built on lexical decision tasks have reproduced for decades; effects built on incidental social cues have mostly shrunk or vanished when tested at scale. The variable that separates the survivors from the casualties is unglamorous: whether labs wrote down and published the results that did not work.

Priming's Contested Path to Consensus

Priming describes a simple event. Exposure to one stimulus changes how you respond to a later one, without deliberate intent. Show someone the word doctor, then the word nurse, and recognition of nurse is measurably faster than after an unrelated word. The effect is real, small, and timed in milliseconds.

The replication crisis changed how anyone reads a priming claim. Large coordinated replication efforts found that many published effects failed to reproduce at their original strength, and some failed outright. Psychology was not alone in this; cancer biology and economics went through versions of the same audit.

What gets lost in the summary is that the failures were not uniform. Some priming findings collapsed. Others held up across labs, tasks, and languages. The difference tracks the procedural choices the original labs made, and one choice in particular: whether negative results ever reached print. A related piece on this site describes how lab priming talks stall when raters guess the hypothesis, which is the same mechanism operating at the level of a single talk.

What the Original Studies Actually Did

The lexical decision task arrived in 1971 with Meyer and Schvaneveldt. A participant sees a letter string and decides as fast as possible whether it is a real word. Precede nurse with doctor and reaction times drop. Precede it with butter and they do not. The paradigm is boring, repeatable, and cheap, which turns out to matter enormously.

Bargh, Chen, and Burrows published a very different kind of study in 1996. Participants unscrambled sentences containing elderly-stereotype words, then walked down a hallway. The reported result: they walked more slowly. The claim was that a stereotype had activated a behaviour, and the measured outcome was walking speed recorded by a person with a stopwatch.

Both designs used small samples, neither pre-registered its predictions, and both allowed many analytic choices after the data arrived. Those choices include which participants to exclude, which trials to trim, and which covariates to include. Each individual choice looks defensible. Together they give a researcher enough room to find an effect that is not there.

The Replication Reckoning

When independent teams ran the stereotype-and-walking study again, the effect did not reliably appear. One large attempt found walking speeds indistinguishable from a control condition. The original estimate was small to begin with, and small effects need large samples to detect at all.

Coordinated multi-lab projects made the pattern visible. The same protocol run at dozens of sites produced effect sizes that varied by site, and pooled estimates well below the published originals. This is not proof of fraud. It is what happens when a field estimates effects from underpowered studies and publishes only the ones that crossed a threshold.

Publication bias does most of the damage. A lab that runs a priming study and finds nothing has little incentive to write it up, and journals historically had little interest in printing it. The file drawer fills. The published literature then looks like a consistent body of positive findings when the underlying evidence is closer to a coin flip.

Where Labs Logged Their Negatives

Pre-registration forces the researcher to state predictions, sample size, and analysis plan before seeing data. That single act removes most of the flexibility that produces spurious results. It does not make an effect real, but it makes a positive result mean something.

Registered reports go further. A journal reviews the design and accepts the paper before the data exist, so the decision to publish cannot depend on the outcome. Several journals now run this format, and the resulting papers report null results at far higher rates than the standard literature.

A handful of labs made publishing nulls ordinary practice. Their output looks less exciting and is more useful. When a semantic priming effect survives in a lab that routinely prints its failures, the survival carries information. The same logic applies outside psychology; this site has argued that two pipette calibrations decide whether a grid cell finding replicates, and the principle is identical.

What Survives and What Does Not

Semantic priming is the sturdy survivor. The effect shows up across labs, tasks, and languages, and its size is modest but stable. It rests on a mechanism with a plausible account: related concepts in memory make each other easier to retrieve, and the effect decays over a second or so.

Social priming is the fragile family. Effects that depend on incidental cues, like a stereotype, a temperature, or a word list, tend to be small, context-dependent, and hard to reproduce at the original magnitude. Some survive at reduced size. Many do not survive at all.

Consider the elderly-priming study. The original reported that participants who unscrambled sentences containing words like Florida, bingo, and ancient walked more slowly down a hallway. The effect size was modest, and the sample was small. When a large-scale replication attempted the same procedure, the walking-speed difference did not appear. The original authors have argued that the replication may not have captured the same psychological state, a common defence when a direct replication fails. But the burden of proof shifts when an effect cannot be reproduced by independent teams.

Another example: priming with money-related words has been reported to increase self-sufficiency and reduce helping behaviour. Those effects also tend to be small and have proven difficult to replicate consistently. The pattern is the same: a small effect, a small sample, and a publication process that favours positive results.

The trade-off is real and worth stating plainly. A researcher who pre-registers and publishes nulls gives up the freedom to chase an interesting unexpected result, and the resulting paper is harder to publish in a high-profile venue. That cost falls hardest on early-career researchers who need visible output. The field's incentive structure still rewards the finding over the method.

Practical Steps for Reading Priming Claims

Check for pre-registration or a registered report before anything else. A timestamped plan on a public registry tells you the analysis was not chosen after the results arrived, and its absence is a reason to read more carefully.

Look for a direct replication, not a conceptual one. A conceptual replication changes the task, the population, or the outcome measure, so a failure is ambiguous. A direct replication holds those fixed and tests whether the original result reappears.

Ask whether the lab publishes null results. A group with a visible record of printing failed studies is a better source than one whose output is uniformly positive, because the failures tell you the positive findings were screened.

Compare effect sizes across studies rather than reading one number. If the original reports a large effect and every later attempt finds a small one, the pooled estimate is the honest figure.

Treat any single-study priming claim as provisional. One sample, one task, one lab is a hypothesis. The evidence accumulates only when other labs try the same thing and report what they find, including when what they find is nothing.

Why the Fix Is Slow

Pre-registration and registered reports have spread, but unevenly. Adoption is highest in social psychology and lowest in fields where the workflow is expensive or the data are observational. Even where the tools are available, a plan on a registry does not guarantee the analysis follows it. Deviations are common and often justified, but each one reintroduces the flexibility the plan was meant to remove.

Meta-analytic estimates that correct for publication bias shrink many social priming effects toward zero, while semantic priming survives with a small but reliable advantage in reaction time, typically in the range of a few tens of milliseconds. That range is not a single number to memorise; it is a reminder that the honest effect is smaller than the headline.

Training is another bottleneck. Few graduate programmes teach the mechanics of power analysis, multiverse robustness checks, or how to write a pre-registration that is specific enough to constrain later choices. Without that training, pre-registration becomes a formality, and the file drawer reopens under a new name.

The fix is not one reform but a stack of them: pre-registration, registered reports, journals that accept null results, labs that normalise publishing them, and readers who ask for the replication before they cite the finding. Each layer removes a little room for the flexibility that turns noise into a story. The labs that logged their negatives did the field the favour of showing what the noise looks like.