Prescient Ideas and Learned Unacceptability
A new deep learning model sheds light on why some ideas appear before their time
In the 1950s and 1960s, the US Congress was in the midst of a battle to pass new civil rights legislation to the end state-sponsored segregation of the Jim Crow era. The legislative push and pull went on for almost a decade, with the first desegregation initiative in Brown v Board of Education, separate Civil Rights Acts passed in 1957, 1960, and 1964, and culminating with the Voting Rights Acts of 1965. Among the most prominent opponents of the new civil rights bills were two senators: James Eastland and John Stennis. They voted against every desegregation bill that came before the Senate. Both were from Mississippi. Both were white. But they opposed the new bills with two entirely different rhetorical strategies.
Eastland was an unabashed racist, the kind that was accepted by the mainstream at that time in the Southern US. He believed that white people were fundamentally better black people and put this belief front and center in his public arguments.
In 1944, Eastland said: “I have no prejudice in my heart, but the white race is the superior race and the Negro race an inferior race and the races must be kept separate by law.” His views changed little over the next twenty or thirty years. When he first made this claim, it was not an uncommon position among his contemporaries.
Stennis also voted against every civil rights bill put before him in Congress. But he took a different tact than Eastland. His talking points revolved around “color blindness,” limited government, and individual freedom. These were terms none of his contemporaries had used to quite the same effect.
In short, Eastland’s rationale for conservative beliefs was on the way out. And in many respects, the talking points developed by Stennis resemble the talking points of conservatives still today. It represented a kind of linguistic paradigm shift.
Stennis was, in other words, prescient.
Whether his logic was morally defensible or not, Stennis saw the value in a way of doing things before anyone else did. Despite their similar underlying beliefs, generations of conservatives to come followed in the footsteps of Stennis, not Eastland. That’s prescience.
And it’s that kind of prescience that is quantified by a new deep learning model developed by researchers from the business schools at Stanford and Berkeley. The basic idea of the model is to create an objective measure of “prescience” by studying trends in a huge corpus of linguistic data. The way the researchers define prescience in the model is actually pretty ingenious.
Essentially, a prescient idea is one that is novel in the context of its own time, but common in the context of a future time.
In computational terms, prescience is determined by two quantities. The first is how well an utterance (a word or phrase) is predicted by a deep learning model trained on language from the same time period. The model works by blocking out words in a given sentence and trying to get better at predicting what the blocked out word is. As it learns from millions of sentences, the model gets more and more accurate. The second quantity is how well an utterance is predicted by a model trained on language generated by people at a time in the future.
For example, if you had a dataset from the late 1980s where people are talking about the internet and social networks and personal computers, those sentences would likely not resemble other sentences from the 1980s. But they would more closely resemble language commonly used in the 2000s. Comparing the different datasets, the model would mark whoever was talking about the internet in 1980s as highly prescient.
And the model can do all sorts of interesting things with this measure. For instance, it quantified the difference in prescience between the two senators, Eastland and Stennis, during the civil rights debates:
In another political example, the authors on the paper show that the model also flags language from the “Gingrich senators,” a group of senators who, beginning in the 1980s, pioneered the kind of hyper-partisan rhetoric used in congress today:
The authors show that these prescience scores matter not only in politics, but across law and business, as well:
Okay, that’s great. It pays to be prescient. But where does prescience come from? How does one go about being ahead of one’s time?
The authors offer a clue: the individuals with the most prescient language came from outside the mainstream circles of their contemporaries. In other words, linguistic paradigms shifts are introduced by outsiders, not insiders.
Again, they show this effect across politics, law, and business:
Put another way, those who are currently successful are successful within the current paradigm. For anyone attempting to transcend that paradigm, it means they probably aren’t hitting it big with the way things are currently done. They’re looking to institute change—something that better reflects the values of the future, not the present. To paraphrase New Yorker author Louis Menand, being ahead of one’s time is the highest compliment that can be paid in American intellectual life. It is also a recipe for unhappiness.
This explanation made me think of a study from another paper I’d recently come across, one that also features examples from language—but with a goal different goal.
The study was done by two researchers at Princeton, Karina Tachihara and Adele Goldberg, who were interested in adults learning a second language. In particular, they were interested in “unacceptable” sentences. Basically, a sentence in which a speaker of the language can figure out what you’re saying, but wouldn’t endorse it as a well-constructed sentence. For example:
Lisa filled water into the cup.
The magician disappeared the rabbit.
Amber explained Zach the answer
We all know what those sentences are trying to say. But it’s not the way a proficient English speaker would choose to say them.
Tachihara and Goldberg were interested in what it takes for people to avoid these “unacceptable” constructions as they learn a new language—in this case, English speaking undergrads learning Spanish.
The first thing they found is that, as predicted, proficient speakers of Spanish were much better at detecting these unacceptable sentences than Spanish-learners. Of course they were. But at first you might think that the Spanish-learners would tend to apply the rules of English to their new target language. In that case, they would be more likely to endorse sentences that corresponded to well-formed sentences in English, and less likely to endorse sentences that didn’t—regardless of how that matched with the rules from Spanish. But that’s not what happened. The Spanish-learners were equally unable to identify unaccepted sentences, regardless of whether or not the sentences corresponded with the syntactic structure of an acceptable English sentence.
Then they looked at how learning to identify acceptable sentences changed over time. They started with an initial assessment, measuring learners’ answers and reaction times when judging acceptable versus unacceptable sentences, and ended with the same assessment (featuring different sentences) a few days later. In between were three days of self-paced reading. The key question was: would simply exposing students to acceptable sentences (via reading) impact their ability to identify what was unacceptable?
And that’s exactly what they found. Tachihara and Goldberg called it “statistical preemption”—merely encountering a bunch of acceptable examples helped learners to pick out unacceptable examples later on.
So what does this have to do with prescience? Well, I think the study by Tachihara and Goldberg points toward the mechanism that shows why people who are successful in the current paradigm have trouble being prescient about what will work in the future.
They learn what is unacceptable.
Integrating into any mainstream paradigm—whether in politics, or academia, or business—is like learning a language. There are acceptable and unacceptable sentences, beliefs, ideas, terms, and all that. And part of successfully learning to use a language is to stay away from unacceptable sentences.
But keep in mind! Unacceptable sentences aren’t unacceptable because they’re objectively wrong. “Lisa filled water into the cup” still makes perfect sense to any English speaker. We just don’t say it that way because, well… we’ve all agreed that’s not the way to say it. But that doesn’t mean we’re not going to change our mind later on. Nor does it mean that the current way of doing things reflects the true and unerring structure of world and will never be otherwise. The more you buy into the current paradigm—the way things are done right now—the less open you’re going to be to ways of saying things that deviates from the current wisdom.
I don’t think anyone living in 2022 would be surprised that what’s acceptable today can become unacceptable tomorrow, or that what’s unacceptable today might well be acceptable tomorrow. This goes far beyond the rhetoric of a couple senators in the 1960s. It’s the logical infrastructure of the society we live in. Being prescient requires us to find the ideological territory that’s acceptable not in today’s context—but tomorrow’s.