In 2009, a young neuroscientist at Dartmouth College named Craig Bennett walked up to a grocery store fish counter and asked for a full-length Atlantic salmon. He wasn’t planning dinner, at least not yet. His lab was getting ready for a study of how people react to social situations, and before putting human volunteers in the brain scanner, the team wanted to test their settings on something with a mix of fat, muscle, and bone. They had already tried a pumpkin and a Cornish game hen. The salmon was next.
The fish went into the scanner, and the team ran the full experiment on it, just as they would with a person. It was shown a series of photographs of people in social situations and, in the deadpan words of the poster the team later presented, it “was asked to determine what emotion the individual in the photo must have been experiencing.” The poster also noted that the salmon “was not alive at the time of scanning.”
When the data were analyzed with a common but loose statistical method, a small cluster of points inside the dead fish’s brain lit up as if it were thinking hard about human feelings. That result, which is completely absurd, became one of the most useful jokes in modern science. To see why, you have to understand what a functional MRI scan actually measures, and how easy it is to find patterns in noise. If you read our piece on AI that sketches what you saw from a brain scan, this is the machine behind it.
A brain scan that watches blood, not thoughts
An ordinary MRI scan takes a detailed picture of the body’s structure. The patient lies inside a powerful magnet, typically 1.5 or 3 tesla, which is tens of thousands of times stronger than Earth’s magnetic field. The magnet lines up the hydrogen nuclei in the water in your tissues. Radio pulses knock them out of line, and as they settle back, they give off faint signals that the scanner turns into images. Different tissues settle at different rates, which is why bone, fat, and gray matter look different on the screen. Paul Lauterbur and Peter Mansfield shared the 2003 Nobel Prize in Physiology or Medicine for working out how to turn those signals into pictures.
Functional MRI, or fMRI, uses the same machine to watch the brain change over time. But it doesn’t detect brain cells firing. Neurons send their messages with tiny electrical pulses that are far too faint and fast for an MRI scanner to pick up. Instead, fMRI tracks blood.
Neurons carry almost no reserve fuel, so when a patch of brain gets busy, the body sends it a rush of fresh blood carrying oxygen and glucose. That extra blood arrives about 2 seconds after the activity starts, peaks around 4 to 6 seconds later, and then fades, often dipping slightly below normal before it recovers. The rush is fairly local, within 2 or 3 millimeters of the busy neurons. In most cases more oxygen arrives than the neurons can use, so the blood leaving an active area is richer in oxygen than usual.
Here is the clever part. Hemoglobin, the molecule in red blood cells that carries oxygen, behaves differently in a magnetic field depending on whether it is carrying oxygen. Without oxygen, it is slightly magnetic and distorts the field around it, which makes the nearby MRI signal fade faster. With oxygen, it barely interacts with the field at all. So when fresh, oxygen-rich blood floods an active brain region, the MRI signal there rises by a small amount, usually just a few percent. That signal is called BOLD, for blood-oxygen-level dependent, and almost every brain scan study you have read about in the news is built on it.
From a dinner-table idea to a 1990 discovery
The link between thinking and blood flow is old. In the 1880s, the Italian physiologist Angelo Mosso built what he called a human circulation balance, a tilting table that he hoped would dip toward a person’s head when they did mental work and blood shifted to the brain. In 1890, Charles Roy and Charles Sherrington at Cambridge showed experimentally that brain activity and blood flow go together. And in 1936, the chemist Linus Pauling and his student Charles Coryell discovered that hemoglobin’s magnetic behavior changes when it gives up its oxygen.
For decades, nobody connected those findings to imaging, because medical MRI didn’t exist yet. That changed in 1990. Seiji Ogawa, a biophysicist at AT&T Bell Laboratories in New Jersey, scanned the brains of anesthetized rats in a strong 7-tesla magnet while changing how much oxygen they breathed. As the oxygen level dropped, dark lines appeared on the images, tracing the brain’s blood vessels. He confirmed the effect with test tubes of oxygenated and deoxygenated blood, and he realized that it could be used to map where the brain was working.
Things moved quickly after that. In 1991, a team at Massachusetts General Hospital led by John Belliveau published the first functional maps of the human visual cortex, using an injected contrast agent. In 1992, three groups showed that the contrast agent wasn’t needed. Kenneth Kwong’s team at Massachusetts General used the natural BOLD signal to watch the visual cortex respond to flashing light. Ogawa and Kamil Ugurbil did the same at the University of Minnesota with a stronger magnet, and Peter Bandettini’s group in Wisconsin mapped the brain area that controls movement. The methods those early studies used are still the basis of most fMRI research today.
How a scan becomes a picture of a busy brain
A typical fMRI scan divides the brain into tens of thousands of tiny cubes called voxels, short for volume pixels, each a few millimeters on a side. The scanner measures the signal in every voxel about once every one to two seconds while the person in the machine does a task, such as looking at faces, tapping a finger, or solving math problems. A single session can produce well over 100,000 voxels, each with its own wobbly line of measurements over time.
The signal changes researchers are looking for are tiny, and they are buried in noise. The person breathes, their heart beats, their head drifts a millimeter or two, and the magnet itself drifts slightly over the course of a scan. So researchers repeat the task many times and use statistics to ask, for each voxel, whether its signal goes up reliably when the task is on and down when it is off. The colorful blobs you see on brain scan images are not photographs of thoughts. They are maps of the voxels that passed a statistical test, painted over a gray structural image of the brain.
How a dead fish passed the test
That statistical test is where the salmon comes in. Bennett’s lab set the salmon data aside for months. Then one of the team was preparing a seminar on how to analyze fMRI data properly, and they wanted an example of how an improper analysis could go wrong. They remembered the fish.
Think of each voxel test as a coin flip. A common standard in science is to call a result significant if there is less than a 1 in 1,000 chance it would happen by luck alone. That sounds strict. But if you run that test on 130,000 voxels, you should expect about 130 of them to pass by pure chance, even in a brain that is doing nothing at all. Bennett compared it to throwing darts: throw enough of them and some will land in a tight cluster without anyone aiming.
Statisticians call this the multiple comparisons problem, and there are standard ways to correct for it, essentially by raising the bar as the number of tests grows. When Bennett’s team analyzed the salmon without a correction, a cluster of voxels in the fish’s brain cavity appeared to respond to the photos. When they applied a proper correction, the activity vanished. “By complete, random chance, we found some voxels that were significant that just happened to be in the fish’s brain,” Bennett told Wired in 2009.
Several journals turned the write-up down. But the poster, presented at the Human Brain Mapping conference in San Francisco in the summer of 2009, became a hit, and researchers forwarded it to each other for weeks. The paper was finally published in 2010 in the Journal of Serendipitous and Unexpected Results, under the title “Neural Correlates of Interspecies Perspective Taking in the Post-Mortem Atlantic Salmon: An Argument For Proper Multiple Comparisons Correction.” In 2012, it won an Ig Nobel Prize, the award for research that makes people laugh and then think.
The joke had a real effect. According to the authors, between 25 and 40 percent of fMRI studies being published around 2009 didn’t use corrected statistics. By 2012, that figure had fallen to about 10 percent. As for the salmon, Bennett later said he was never reimbursed for it, because the lab ate it.
When the software itself was too generous
The salmon showed what happens when researchers skip the correction. A study published in 2016 asked a harder question: do the corrections themselves work? Anders Eklund, Thomas Nichols, and Hans Knutsson took resting-state scans from 499 healthy people who were simply lying in the scanner, not doing any task. They then ran nearly 3 million analyses on that data, pretending the people had been doing tasks they never did. With a 5 percent error rate as the target, about 5 percent of those fake experiments should have turned up false results.
For one popular kind of analysis, which looks for clusters of active voxels, the three most widely used fMRI software packages produced false positives far more often, in some settings up to 70 percent of the time. Along the way, the team found a bug in one of the packages that had gone unnoticed for 15 years and had been fixed only in 2015, while they were writing up their results.
The paper noted that there were some 40,000 published fMRI studies in total, and headlines quickly turned that into claims that all of them might be wrong. That was never the finding. The number was a count of every fMRI paper ever published, not of affected ones, and the authors later issued a correction to tone down their wording. The real lesson was narrower but still important: one common shortcut had been making results look stronger than they were, and the field has since moved toward stricter methods and sharing data so that others can check the work.
What fMRI does well, and where it struggles
None of this means brain scans are junk. fMRI is safe, uses no radiation, and can see deep inside the living brain, which no other noninvasive tool can do as well. It has mapped the brain areas involved in vision, hearing, movement, language, memory, and fear. Surgeons use it before operating on a brain tumor to find and avoid the areas that control speech and movement. And it has given rise to brain-reading research that would have sounded like science fiction 20 years ago.
But its limits are real. Because it watches blood rather than neurons, it is slow, blurring events that happen within a fraction of a second into a smear several seconds long. Each voxel holds hundreds of thousands of neurons. And the link between blood flow and brain activity isn’t always as tidy as researchers assumed. A study published in Nature Neuroscience in December 2025, led by researchers at the Technical University of Munich, estimated oxygen use with a separate quantitative MRI method in more than 40 volunteers and found that in about 40 percent of the voxels with significant BOLD changes, the signal pointed the opposite way from the brain’s actual oxygen consumption, especially in a network that is active when the mind wanders.
Results can also be shaky for individuals. A 2020 analysis from Ahmad Hariri’s lab at Duke University, which drew on 56 published studies along with new scan data, found that common task-based fMRI measures had poor reliability when the same person was scanned twice. The scans worked well for describing how brains behave on average across a group, but not for reading a single person’s mind or diagnosing them. That matters, because a lot of popular claims about brain scans quietly make the jump from group averages to individuals.
Can a scanner read your mind?
This is where fMRI meets the news. Companies have tried to sell fMRI lie detection for years, but American courts have kept it out. In the 2010 case United States v. Semrau, a federal judge refused to admit fMRI lie detector evidence, and an appeals court upheld that decision in 2012. Marketing researchers have scanned people tasting soft drinks, most famously in a 2004 study that found different brain responses when people knew they were drinking Coca-Cola. Fun, but not mind reading.
The newer brain-decoding work is more impressive, and it is also more limited than headlines suggest. In 2023, researchers at the University of Texas at Austin built a decoder that could turn fMRI data into a rough paraphrase of a story a person was hearing or imagining. It needed about 16 hours of scanning per person to train, and it captured the gist rather than exact words. It also didn’t work on anyone it hadn’t been trained on, and volunteers could defeat it by deliberately thinking about something else, such as counting or naming animals. The image reconstruction system in our earlier article cut the training time to about an hour, but it still depends on a willing volunteer lying still in a multimillion-dollar magnet that costs hundreds of dollars an hour to run.
For reading brain signals in real time, scientists turn to electrodes instead of blood flow. Those tools are behind implants that are giving people their voices back and the brain pacemakers that quiet Parkinson’s tremors. They are far faster and more precise than fMRI, but they require surgery.
The fish’s real lesson
Scanners keep getting stronger. In 2024, France’s Iseult scanner, which at 11.7 tesla is the most powerful MRI machine ever used on people, released its first images of human brains, with detail that standard hospital machines can’t match. Better hardware will sharpen the pictures. It won’t fix a careless analysis.
That is why the dead salmon still gets passed around in statistics classes, along with other oddities of patient science like the experiments that have run for decades. The salmon wasn’t a takedown of brain imaging. It was a reminder from people who use fMRI every day that when you test 100,000 things at once, some of them will look exciting by accident. The next time you see a colorful brain scan in a headline, it is fair to ask two questions: how many people were scanned, and did the researchers correct for chance? If the answers are “one” and “no,” you might be looking at a very thoughtful fish.
Watch your own brain waves, or learn to spot brain-scan hype
MUSE 2 Brain Sensing Headband — A consumer headband that reads your brain’s electrical activity through sensors on your forehead and behind your ears, the EEG approach rather than fMRI’s blood flow. Paired with its app, it turns your brain activity into sound cues during meditation, and it also tracks heart rate, breathing, and movement.
Brainwashed: The Seductive Appeal of Mindless Neuroscience — Psychiatrist Sally Satel and psychologist Scott O. Lilienfeld explain what brain scans can and can’t tell us, and why colorful fMRI images get stretched into claims about lying, shopping, addiction, and free will.