The choice was between two treatments for lung cancer. Of 100 people having surgery, 90 survive the treatment, 68 are still alive after one year, and 34 after five years. Of 100 people having radiation therapy, all survive the treatment, 77 are alive after one year, and 22 after five years.
Other people read the same choice told the other way. Of 100 people having surgery, 10 die during treatment, 32 will have died by one year, and 66 by five years. Of 100 having radiation, none die during treatment, 23 will have died by one year, and 78 by five years.
Nothing in the second version is new. Ninety alive is ten dead. In a study published in 1982, three physicians, Barbara McNeil, Stephen Pauker and Harold Sox, and the psychologist Amos Tversky put versions of this choice to 238 outpatients, 491 graduate students and 424 physicians, each asked to imagine having lung cancer.
In all three groups, surgery looked substantially more attractive when the problem was framed in terms of living than in terms of dying. The physicians were no exception.
The zero next to the ten
Say the same numbers a different way, and people often choose differently.
McNeil and her colleagues had an explanation, as Laura Macchi and Edoardo Zulato describe it in a 2021 paper. Read as deaths, surgery’s 10 sits next to radiation’s zero, and that gap stands out. Read as survivors, 90 next to 100 looks like a small difference. The patients are the same patients either way.
A year earlier, Tversky and Daniel Kahneman had named the idea in Science. A “decision frame,” they wrote, is a decision maker’s conception of the options, outcomes and odds in a choice, controlled partly by how the problem is formulated. They compared frames to different perspectives on the same visual scene.
Their best-known demonstration was a disease expected to kill 600 people. Told that Program A would save 200 people for sure, 72 percent of the students who answered preferred it to a gamble. Told instead that under Program C, 400 people would die, only 22 percent picked it over the equivalent gamble. Programs A and C are the same program. Saving framed as gains made people cautious; deaths framed as losses made them willing to gamble. They had seen the reversal, they wrote, in groups that included university faculty and physicians.
This one holds up. In 2014 the Many Labs project reran the disease problem across 36 samples, and the gain versus loss framing effect replicated.
A frame doesn’t even need numbers. When ten daycare centers in Haifa were studied in 1998, the six that introduced a fine for late pickups saw lateness rise. The researchers’ reading was that ten late minutes, once a favor from a teacher, had become something parents could buy. That was more than a change of wording, though: the fine also changed the actual incentives.
When the numbers trip people up
The medical version comes with a catch. In McNeil’s wording, the deaths pile up: the 32 who have died by one year include the 10 who died during treatment. Macchi and Zulato suspected that people reading the dying version were adding the numbers as if each were new deaths. When they rewrote the same outcomes as separate counts for each period, the framing effect in their study largely disappeared.
So some of what gets called framing may be plain misreading. That doesn’t make it harmless. A doctor who writes the numbers one way can move a patient’s choice, whether the patient is swayed by the words or confused by the arithmetic.
The memo that picks the frame
Imagine you want your director to approve a new supplier. Your memo says it delivers 95 percent of orders on time. A colleague who’s against the switch writes his own: one order in 20 arrives late, about 40 a month at your volume.
Both are true. The director will probably see only one of them, and whoever writes the memo has already chosen which number gets to look large. Put the two lines next to each other, and the director at least gets to see that there were two ways to say it.
The model reads the frame too
Language models can be swayed the same way. In a study published in PNAS in 2023, Marcel Binz and Eric Schulz gave GPT-3 the gamble problems Kahneman and Tversky had used to find biases in human choice, and compared its choices with people’s. GPT-3 showed three of the six biases they tested. One was framing: its preferences changed depending on whether a choice was presented in terms of gains or losses.
That was one older model, tested on problems from a well-known literature. Newer models may behave differently. You can check yours.
Take a real decision with numbers in it. Ask for a recommendation in a fresh chat with the numbers framed as successes, and again in another fresh chat framed as failures. If the recommendation flips, treat it as a lead, since models vary from run to run: swap the framings and repeat a few times, and check the reasons each answer gives. Either way, the decision is still yours to make on the numbers.
A model can also be the one writing the frame. When it summarizes a risk for you, ask for the same figures both ways, the 90 who live and the 10 who don’t, before you decide which one goes in the memo.