When a Hypothesis Hides a Discovery: The Gorilla in the Data

When a hypothesis hides a discovery, a researcher may complete every calculation correctly and still overlook the most unusual feature of the data. Sometimes, simply plotting the numbers differently is enough to reveal something that was there all along.

Biologists Itai Yanai and Martin Lercher explored this problem in a provocative 2020 editorial published in Genome Biology, titled “A hypothesis is a liability.”

To illustrate their argument, they embedded a gorilla in an artificial dataset and asked university students to perform a statistical analysis.

Many of the students never noticed it.

The gorilla was not concealed in a footnote or an obscure calculation. It was waiting to be revealed by an ordinary graph.

The gorilla hidden in a spreadsheet

The researchers created a fictional dataset representing 1,786 people. Each record contained two measurements: body mass index and the number of steps taken on a particular day. The data were divided into separate files for men and women.

Students received what appeared to be a conventional statistics assignment.

One group was given three specific hypotheses to test. These included whether men and women differed in their average number of daily steps and whether the relationship between physical activity and body mass index differed by sex.

Students were also free to report additional findings.

The other group received no predefined hypotheses. Their instructions were simply to examine the dataset and explain what they concluded.

Both groups worked with the same information.

But the numbers contained a deliberate surprise.

The researchers had arranged the data so that plotting body mass index against daily steps produced the unmistakable outline of a waving gorilla.

The shape had no biological significance. It was a deliberately constructed feature of the dataset, designed to test whether students would notice something unrelated to their assigned questions.

Finding it required a straightforward visualization.

The difference between the two groups was striking.

Among the 33 students included in the main comparison, those who had not been assigned specific hypotheses were substantially more likely to discover the gorilla.

The reported odds ratio was 4.8, meaning the odds of finding the gorilla were nearly five times greater in the open-ended group. This should not be confused with a fivefold increase in probability.

The sample was small, and the task was artificial, so the result cannot establish how frequently professional scientists miss genuine discoveries.

Still, it revealed a compelling possibility: the very instructions that help people solve a particular problem can also discourage them from exploring beyond it.

How a useful question can narrow attention

Human attention is selective.

When we concentrate on completing a task, we prioritize information relevant to that task. Other details may receive remarkably little attention, even when they are conspicuous.

Psychologists call one manifestation of this phenomenon inattentional blindness.

A classic demonstration involves a video of people passing basketballs. Viewers are asked to count the passes made by one team. While they concentrate on the count, a person dressed as a gorilla walks through the scene.

A substantial proportion of viewers fail to notice the gorilla.

Yanai and Lercher adapted the underlying idea from visual attention to scientific data analysis.

Counting basketball passes and testing statistical hypotheses are very different tasks. But both may create a narrow attentional focus.

In their experiment, the students who missed the gorilla were not necessarily incompetent analysts. They could have answered their assigned questions correctly.

What they failed to do was explore a feature of the data that those questions did not require them to examine.

The experiment did not directly establish the psychological mechanism responsible for every missed gorilla. Nevertheless, its findings were consistent with the idea that concentrating on a specific objective can restrict exploratory behavior.

This has particular relevance in modern biology.

A single molecular experiment may measure the expression of thousands of genes. Researchers may use such data to investigate one signaling pathway, a particular cellular response, or a predetermined disease mechanism.

That focus can be productive. It can also leave unrelated but potentially informative patterns unexplored.

A dataset may contain more scientific possibilities than the question that originally motivated its collection.

Day science and night science

Yanai and Lercher distinguish between two complementary approaches to scientific work.

They call the first “day science.”

This is the familiar process of hypothesis testing. Researchers propose an explanation, design an experiment, collect evidence, and evaluate whether the results support that explanation.

Rigorous testing is essential for separating credible findings from chance associations.

The second approach is what the authors call “night science.”

The name is metaphorical. It refers to the exploratory and creative side of research, when scientists examine information from different perspectives, notice unexpected relationships, and consider explanations they had not originally planned to investigate.

Rather than beginning with a claim that must be tested, researchers may begin with a question: what else might these data contain?

This freedom has its own dangers.

When researchers examine enough relationships, some will appear meaningful simply by chance. Exploration can generate promising hypotheses, but it can also generate misleading patterns.

That is precisely why the two approaches need each other.

Exploration produces ideas worth investigating. Rigorous hypothesis testing determines whether those ideas survive critical examination.

Can science ever be hypothesis-free?

The editorial prompted a scientific debate.

In 2021, another group of researchers challenged Yanai and Lercher's argument in a Genome Biology correspondence titled “The data-hypothesis relationship.”

Their central objection was that truly hypothesis-free observation may be impossible.

Even when researchers explore a dataset without a specific question, they rely on prior knowledge to decide what to measure, how to visualize it, and which patterns deserve attention.

The critics also questioned whether finding a deliberately embedded gorilla is a meaningful analogy for making a scientific discovery.

Recognizing a familiar shape is very different from identifying a previously unknown biological mechanism.

Yanai and Lercher responded that they were not advocating research without scientific knowledge or theoretical background. Their point was narrower: researchers should sometimes examine data without being committed to testing one specific hypothesis.

That distinction matters.

Scientific discovery does not require a choice between disciplined reasoning and open exploration. Both are necessary, but they serve different purposes.

One evaluates an existing explanation. The other can reveal phenomena that have not yet been explained.

The original paper was an editorial drawing on a small student experiment, not a comprehensive test of competing scientific methodologies. Its broader implications therefore remain an argument about research practice rather than an experimentally established rule.

There is a more subtle lesson in the gorilla story.

A researcher can collect reliable measurements, apply the correct statistical tests, and reach a justified conclusion while leaving other potentially valuable information untouched.

In the student experiment, the overlooked pattern was an artificial image of a gorilla.

In a real laboratory, an unexpected pattern might be the beginning of an entirely different research question.