A Stimarium Dossier
A brain scanner once found a dead salmon thinking
In 2005 researchers put a dead salmon in a brain scanner, showed it photos of people, and ran a standard analysis. A few spots in its head appeared to respond. The analysis had tested thousands of tiny spots, so some passed by chance; correcting for the number of tests erased them.
What’s known, what’s not: Known, Our read, OpenCorrection, 14 October 2026: We corrected the scan date to 2005. The study was first presented in 2009. How we handle corrections
A whole fresh salmon on dark slate, with one small red price sticker stuck to the top of its head.
About this video
In 2005, a dead Atlantic salmon went into a brain scanner, and a few spots in its brain lit up. The catch was the statistics: test thousands of tiny spots and some will pass by chance. Once the team corrected for the number of tests, those spots no longer counted. In 2012 the study won an Ig Nobel, a joke prize for research that makes you laugh, then think.
We read the fish as a warning for any big search. Look in enough places and something will light up by chance.
Source: Bennett, C. M., Baird, A. A., Miller, M. B., & Wolford, G. L. (2010). Neural correlates of interspecies perspective taking in the post-mortem Atlantic salmon: An argument for multiple comparisons correction. Journal of Serendipitous and Unexpected Results, 1(1), 1–5. Awarded the 2012 Ig Nobel Prize in Neuroscience.
Narration: synthetic voice (Kokoro af_heart, Apache-2.0). Artwork: Stimarium.
The salmon that seemed to respond to photos
In 2005, Craig Bennett and colleagues placed a dead Atlantic salmon in an fMRI scanner, the kind of machine used to see which parts of a living brain become active. They showed it photographs of people in social situations and asked it what each person was feeling. [S01, S06]
Then they analyzed the scan the way many studies did at the time. The image is divided into many thousands of tiny cubes called voxels, and each one was tested on its own. A few voxels in the salmon's head appeared to respond to the photos. [S01]
When the team corrected for the number of tests, the activity disappeared. The paper was published in 2010 and won an Ig Nobel Prize in 2012. [S01, S06]
When the team corrected for the number of tests, the activity disappeared.
Why thousands of tests find something
A test that accepts a 1-in-1,000 chance of a false alarm will, run tens of thousands of times, raise some false alarms. Test enough tiny spots and some will pass by chance. [S02]
Corrections raise the bar to account for the number of tests. Surveying six major neuroimaging journals, Bennett, George Wolford and Michael Miller found that 25 to 30 percent of fMRI articles published in 2008 had used uncorrected thresholds. [S02]
Test enough tiny spots and some will pass by chance.
A common shortcut failed on real scans
Most studies did correct, often with a shortcut: a result counts only if enough neighboring voxels pass together, as a cluster. The software estimates how big a cluster must be by assuming that random noise is spread across the brain in a particular smooth way. [S03]
In 2016, Anders Eklund, Thomas Nichols and Hans Knutsson tested the shortcut on real scans of 499 healthy people lying at rest, where no task effect should exist. They ran 3 million made-up group comparisons. Settings meant to allow 5 percent false positives in the three most popular analysis packages produced rates up to 70 percent. [S03]
The main cause was that real brain noise does not follow the assumed pattern. One package also had a bug that had gone unnoticed for 15 years. Permutation tests, which reshuffle the data many times to see what chance alone produces, kept errors near the promised level for comparisons between two groups. Tests of a single group could still give too many or too few false positives. [S03]
Settings meant to allow 5 percent false positives in the three most popular analysis packages produced rates up to 70 percent.
Even the warning had to be corrected
The 2016 paper's summary said its results questioned the validity of some 40,000 fMRI studies. [S03, S04]
Within weeks the authors corrected it. The wording had implied that a large share of fMRI studies was affected, when the problem applied only to studies that used the flawed shortcut. [S04, S07]
In a 2019 follow-up, the same team estimated that at least 10 percent of fMRI studies had used the shortcut's most error-prone setting. [S07]
Small samples pose another problem. In 2022, Scott Marek and colleagues used brain scans from three large studies, about 50,000 people in all. Links between scans and traits such as cognition or mental health were much weaker than small studies had reported. At the typical published sample of about 25 people, they usually failed to replicate. [S05]
How many people are enough is still argued. In 2023, Tamas Spisak, Ulrike Bingel and Tor Wager showed that analyses combining many brain regions at once, tested on data held back for the purpose, replicated with 75 to 500 people for five of six traits they tested. [S08]
Marek's team replied in the same issue. Using those same methods on a larger and more varied group of 11,874 children, 31 percent of effects replicated with 1,000 people, 14 percent with 500, and none with 75. [S09]
The findings shrank
Each correction shrank fMRI findings without showing them to be false. In the 2016 test, the standard voxel-by-voxel correction was, if anything, too cautious, and permutation tests controlled the cluster error rate in comparisons between two groups. Single-group permutation tests could still produce too many or too few false positives. [S03]
Large studies like Marek's still find real links between brains and behavior. Those links are much smaller than the bright patches in older headlines suggested. [S05]
A brain scan measures changes in blood oxygen at thousands of spots, and statistics decide which spots get color. Each colored patch is a statistical claim, as strong as the statistics behind it.
Each colored patch is a statistical claim, as strong as the statistics behind it.
Think with the idea.
Take me sideways
Buying more lottery tickets gives you more chances to win. Testing tens of thousands of spots in a brain scan gives chance more opportunities to produce an apparent result.
Analogy.Challenge it
These problems are about how scans were analyzed, not about the scanner. Many fMRI findings used valid corrections, and large studies still find real effects.
Scope warning.Trace it
Set it beside the dispute over how much plastic has been measured in human brains. In both cases, the analysis method affects which claims the measurements can support.
Cross-domain synthesis, not a finding of the cited studies.Take it with you
When a headline says a brain region lights up, ask how many people were scanned and whether the result was corrected for the number of tests.
Reflection prompt.What’s known, what’s not
↑ Back to where you were- An uncorrected analysis found apparent activity in a dead salmon's brain; correcting for the number of tests erased it. [S01]
- A shortcut for groups of neighboring brain-scan spots produced false-alarm rates up to 70 percent, against an intended 5 percent. [S03]
- Links between brain scans and traits found in samples of about 25 people usually failed to replicate. [S05]
We think a colored brain map is a statistical claim drawn on a picture. Correcting the analysis removes false signals and makes the findings that survive more dependable.
- Which older fMRI findings still hold when reanalyzed with methods that make fewer assumptions?
- Will brain and behavior links found in tens of thousands of people prove useful for any one person?
What the three labels meanClose explanations
Known marks what the cited evidence directly supports, including what named researchers argue.
Our read is Stimarium’s own interpretation of that evidence.
Open marks a question the cited evidence can’t answer yet, or a possibility we are raising ourselves.