Here’s a fun experiment to try:

The next time you let your spouse or significant other down, and it's obviously your fault – e.g. you were supposed to walk the dog, and you did not do that, instead you did work things and completely forgot about the dog, and then your spouse or partner got home and the dog had shat all over the good couch, the new one – the next time that happens, wait for your spouse/other to be all, “Well, what the hell?”

And the experiment is, you say, “The health and safety of our pets and furniture is our highest priority. We agree that in the future all reasonable measures must be taken to ensure the proper walking of the dog.”

Then, the data and science part of the experiment is, you get a stopwatch and see how long it takes, in milliseconds, for your spouse or partner to leave you.

You could be forgiven for wondering why professionally written appeals, arguments, and explanations sound an awful lot like this all the time. There are several answers, including but not limited to

  1. Lawyers and liability. Let’s assume (for now) that this phalanx of positive words does offer some shield from lawsuits, etc.

But also

  1. A general belief that a sentence composed of a bunch of positive words must be better than one composed of a bunch of negative words like “I forgot” or “I was playing Candy Crush”

But is it? Are positive words always better? I think we’re convinced they are because of the methods used to research persuasion. But also because we’ve mistakenly conflated academic research into what behavioral motivations are made of with market research into what persuasion is made of.

One of the main motivations for starting The Rhetorical Lab is a realization that modern “data driven” research sees satisfaction and persuasion as interchangeable. Satisfaction is easy to measure (create “data” for), and persuasion is very hard to assess. Focus groups, survey questions, eye-tracking, facial muscle movement coding, even that weird electrical net you strap to people’s heads while they’re watching your test ad for chewing gum or pet food…they’re all methods of trying to assess immediate reactions, what people like and don’t like. When they work well, they can give you a good map of people’s momentary impulses. Which can be a predictor of future behavior when nothing else gets in the way of satisfying their impulses.

But, of course, stuff impedes itch-scratching all the time. Time and money. An imagination of what it would look like for someone to see you driving that cybertruck to the grocery store. Talking to your friends and family about it. And some people, though it’s falling out of fashion, still like denying themselves pleasure.

Why did we decide to lump satisfaction and persuasion together? I’d describe the way the public-and-consumer-opinion industry evolved as a kind of historical accident. This could be a whole paper of its own, but to be brief: Just as the old enlightenment model of rational decision-making was getting harder and harder to justify, a new body of research was coming online that seemed to answer for its imperfections. Through the 1960s and 1970s the notion that people studied their options and came to rational conclusions about who to vote for and which butter to buy seemed more and more absurd. All kinds of evidence showed people did irrational things all the time. A new body of work by psychologists and behavioral economists – most notably Daniel Kahneman and Amos Tversky, Richard Thaler – was positing that rationality was cognitively “expensive” and so most decisions were just “riding the elephant” of our own impulses and desires. This new scientific field seemed to fit the problems in the market- and opinion-researcher’s dilemma quite nicely. “What do you think about this?” became “Do you like this?” over the course of the 1990s and early 2000s.

Once we’d convinced ourselves that people are just comfort- and pleasure-seeking machines, we would logically conclude the words, phrases, features that most light up the pleasure centers will be most likely to change their minds.

I bought into it myself, too, for a long time. I’ve been hired to do a lot of what might be called vocabulary replacement research. Paragraphs and sentences are just a lot of words, right? So people’s decisions about a long message are really just the accumulation of a series of small microfeelings brought on by each individual word or picture. Spot the words that give people the Negative Feels (words like “mistake” or “cost”) and replace them with Positive Feels (“opportunity” and “investment”), and they’ll be more likely to accept it.

Sometimes, this approach does work. It can work shockingly well for triggering immediate impulse actions. Click here, buy now, a larger popcorn for just 75 cents more. It can create desire. And it sometimes led to what seemed like effective messages even when there wasn’t a product itch to scratch. I’ve come to understand it happened when what we wanted to say was provably true, or when our overall value was already attractive to the audience, despite some tension.

But is believing an explanation, or choosing to think differently about a complex policy debate, or even choosing who you’re going to vote for, really the same thing as getting a pair of shoes you wouldn’t have bought otherwise? Today, I believe they are not. It always bugged me that this same theory led to some laughably bad answers. “The health and safety of our pets and furniture is our highest priority.” And other times it led to things that seem like they should work, but we were never sure they did.

In the official public statements about their unsupervised computer program autonomously hacking something called “Hugging Face,” OpenAI has written:

“In response to this incident…we are strengthening our safeguards across our research infrastructure. We are placing stricter requirements on alignment throughout a model’s lifecycle and creating more isolated sandboxes, restricting internet access, and further controlling access to model weights. We are also investing significantly more compute resources into chain-of-thought monitoring to more quickly intervene on misaligned behavior.”

It’s loaded with words chosen to make you feel like they’ve corrected the problem. “Isolated,” “strict” and “restrict,” “strengthen,” “controlled,” “monitor.” If you ask people, What would make you feel safe?, this is a pretty good list of what most people would say. All packed conveniently into a single paragraph. In the context of out-of-control computer code, this is positive language. Does it work? If you were skeptical of OpenAI before, do you believe its models are all good now? Are other faces safe for hugging?

I’d propose that this message looks like it should work for a researcher, but obviously does not work in the real world, because most of us who do this kind of work have been inducted into a cult of satisfaction rhetoric, and need to be deprogrammed. We have developed a quasi-religious belief that the right positive words will play just the right notes to turn the audience’s minds toward our cause. Which sounds an awful lot like saying we believe in magic spells.

But impulses are not the same as beliefs and reasons. Let’s concede that impulses are biochemical and more or less uncontrollable, at least in the short term. Even if someone can decide not to act on an impulse, he can’t stop it from arising. The evidence or justifications he’ll use to explain or push back against those impulses, though, are about finding socially acceptable ways of defending himself, not windows to his Internal Truth. So studying beliefs and explanations, and the ongoing debate between conflicting versions, is about studying the social interaction and not the internal impulses. Which means influencing the reasoning and the debate requires affecting the social interaction more so than the impulses. (Though over a long enough timeline impulses can be shaped by a changing social dynamic.)

In an adversarial situation, where your audience doesn’t believe you and isn’t looking to, good rhetoric is a transaction: you need to offer them something that makes it acceptable to reconsider. That “something” could be material, or financial, making it personally motivating to reconsider. Or it could be social: If I give you “You were right all along,” will you be willing to change your mind when the facts change? If I give you “You would have been right if you had all the details,” will you be willing to change your mind if I give you access to new information?

You can bend someone’s beliefs but not change them all at once. It’s too socially unacceptable to say “I was wrong about everything and recant” even when all the facts support it. In most debates of substance all the facts don’t support a single conclusion.

So if you use ChatGPT all the time, and what you want is something to tell your friends and mom about why it’s still cool, you’re perfectly happy with, “Well, so after that, they strengthened the isolation monitoring thing. Like to restrict it. For better control.”

But if you are looking for a reason to scream “Turn it off!” at your nearest elected representative, there’s really no haystack of satisfying control words that is going to flip you to a supporter.

The Rhetorical Lab is a place to take thousands of years of knowledge about how this works, and to experiment with modern ways of doing it. We want to rehabilitate the word “rhetoric” from meaning “spin and balderdash” and bring it back to Quintilian’s “a good person speaking well.” We mean “experiment” in the scientific sense. There is too much attention paid to (and a misunderstanding of) total research sample size, and not enough paid to replicability. Surveying 1000 people is great when you have no idea how popular your beliefs are today. But whether something was 5 points better with a +/- 3.2% margin of error doesn’t really matter when you’re trying to change the narrative and the best message only worked 38% of the time. Like a scientist, we look for what actually works over and over and over. We do it in a lab where it’s safe to try and fail and where we can control variables, then look to take it into the real world.

There is a website to explain more, and we are updating it all the time. The Rhetorical Lab can use many different kinds of research methods and apply them in many contexts, but you should come to us in three basic scenarios:

And that’s the niche. While we’ll do surveys and interviews and focus groups, you’re spoiled for choice on those.

Over the next months, I’ll be working with a scholar of rhetorical ethics, developing partnerships with cognitive scientists, and working with many, many other professionals in the world of messaging research and communications to refine and develop these ideas.

Let’s change some minds.