Sentiment Analysis: Adding Hanania to the Mix

Update: The noble Ariel Krakowski tries replicating my earlier post. Here’s my unrelated followup.

After my initial conversation on sentiment analysis with ChatGPT, I decided to ask it to add Richard Hanania to the Cowen/Hanson/Caplan/Krugman comparisons. Here’s what happened.

ChatGPT: Adding Richard Hanania makes the exercise much more interesting, because he combines some traits of Caplan and Krugman while being unusually high in contempt/ridicule. Looking at his recent writing, especially 2025–26, I’d revise the comparison this way. These remain human-coded 0–10 estimates of prose, not psychological assessments or corpus-derived measurements.

The distinctive Hanania emotion is contempt

This jumps out of his writing. Hanania doesn’t merely say that ideas are incorrect or developments unfortunate. He very frequently evaluates people’s intelligence.

Recent examples include “stupid people on the right,” politics making smart people dumber, Musk being “about as stupid as human beings get when it comes to politics,” and an entire essay called “On Defeating the Stupid/Interested Quadrant.” Richard Hanania

Another recent essay begins:

“I spend a lot of time writing about how the discourse is getting dumber.”

He proceeds to contrast old conspiracy theories with increasingly bizarre contemporary ones. Richard Hanania

And his new book is Kakistocracy: Why Populism Ends in Disaster. A July essay explaining its empirical case is literally subtitled “How we know it’s bad.” Richard Hanania

This makes contempt/ridicule worth separating from anger. Krugman is more characteristically indignant: somebody is doing something cruel, corrupt, dishonest, or dangerous. Hanania’s characteristic reaction is more often:

How can these people possibly be this stupid?

But Hanania is much happier-sounding than Krugman

This is where generic sentiment analysis would again go astray.

Hanania’s writing gives plenty of evidence that he enjoys things. He gets excited about books and ideas; describes Works in Progress as essentially “the most must-read thing on the internet”; and his older conversational writing repeatedly uses cool, beautiful, etc. Richard Hanania

His economics/progress writing is also fundamentally optimistic. In a recent essay on automation, he emphasizes the extraordinary increase in living standards since 1790—antibiotics, cars, indoor plumbing, air conditioning and “an endless number of modern conveniences.” His argument is precisely that destroying jobs through technological improvement is the basis of progress. Richard Hanania

So I’d put Hanania considerably closer to Caplan than Krugman on baseline enjoyment of modern life.

Yet his political sentiment has become extremely negative

This matters because there appears to be a time trend.

In 2022, talking with Cowen, Hanania was explicitly skeptical about excessive political pessimism. He observed that political violence had been substantially worse historically and suggested that modern media produce so much negativity that Americans might remain pessimistic “no matter how good things get.” Richard Hanania

Compare September 2026:

“My view of the conservative side of the political spectrum has grown increasingly dark.”

He describes its dominant impulses as “rotten to the core” and worries about its “slide into stupidity and genocidal vibes.” Richard Hanania

Or his 2025 discussion of America’s future, which considers alternating left-wing cancel culture and right-wing authoritarianism, while also explicitly presenting an optimistic post-Trump/post-woke scenario. Richard Hanania

That’s important. Hanania isn’t simply a dispositional pessimist. He thinks current American politics has gotten worse.

Indeed, today’s essay is revealingly optimistic about the possibility of reversal. His “Trump Theory of Everything” argues that one individual’s departure could eliminate a surprisingly large fraction of America’s current political dysfunction. Richard Hanania

So I’d score his current-conditions pessimism higher than his long-run pessimism.

Hanania versus Caplan

This may be the most revealing comparison.

They share a lot: strong belief in material progress, enjoyment of intellectual life, willingness to say extremely negative things, impatience with conventional wisdom, and unusually direct prose.

But their negative emotions differ.

Caplan: This belief/policy is crazy, evil or terrible.

Hanania: The people responsible for this are unbelievably stupid.

Caplan’s writing also contains vastly more affection. Friends, family, students, colleagues, parties, children, travel, intellectual heroes, strangers he meets—all regularly generate explicit positive emotion. I see much less of that in Hanania’s writing.

Thus I’d distinguish joy from warmth. Hanania seems fairly high on the former and substantially lower on the latter.

Hanania versus Cowen

Almost opposites in presentation.

Cowen sees something disturbing and says:

Department of Yikes. Have a nice day.

Hanania sees it and writes 3,000 words explaining why everyone involved is an idiot.

Yet their underlying world assessments may be closer than their tones suggest. Both combine strong appreciation of civilization, innovation and interesting intellectual work with considerable concern about contemporary politics and institutions.

Hanania versus Hanson

This is perhaps the largest stylistic contrast.

Hanson can contemplate human extinction while sounding like he’s deriving a comparative-static result.

Hanania can encounter a bad political argument and sound personally offended that someone was stupid enough to make it.

Hence Hanson’s negative subject matter / negative emotional expression ratio is enormous. Hanania’s is almost the reverse.

Hanania versus Krugman

They’re superficially similar because both use lots of harsh language. But I’d distinguish their characteristic negative affect:

Krugman → moral indignation + alarm.

Hanania → contempt + amusement + anger.

And Hanania seems much less anxious. His recent fiscal essay is a good example: he predicts severe constraints from Social Security, Medicare and debt, but presents the problem analytically, comparing new spending proposals to a bankrupt person deciding to remodel his house. Richard Hanania

That is negative without being frightened.

So my compressed portraits of all five would now be:

Caplan: Life is wonderful; people nevertheless believe lots of crazy and terrible things.

Cowen: The world is fascinating and full of excellent things. Also, there are a surprisingly large number of reasons to worry. Have a nice day.

Hanson: The long-run dynamics may be quite bad. Let’s construct a model.

Hanania: Modern civilization is great. Why are the people running politics so unbelievably stupid?

Krugman: Important things are going badly because people are doing terrible things, and this should make you angry and alarmed.

That makes Hanania arguably the highest-contest writer of the five, Caplan the highest-positive-affect/highest-warmth writer, Hanson the least emotionally expressive, Cowen the most ironically alarmist, and Krugman the highest in overt anger and alarm. Those distinctions strike me as much more informative than a single positive/negative sentiment score.

Caplan: What does your analysis show about sentiment analysis? What are the broader lessons of analyzing these five thinkers?

ChatGPT: The exercise has made me substantially more skeptical of ordinary sentiment analysis. These five writers form an unusually good stress test because they demonstrate several different ways the same underlying attitude can produce radically different linguistic sentiment scores—and vice versa.

The biggest lesson is that sentiment of language is not sentiment of belief.

1. Lexical sentiment measures style at least as much as outlook

Consider the central puzzle we uncovered:

Cowen: highly positive vocabulary, substantial substantive worry.
Caplan: lots of harsh vocabulary, unusually optimistic worldview.
Hanson: disturbing propositions, emotionally neutral vocabulary.
Hanania: contempt expressed partly through humor and intellectual language.
Krugman: negative propositions expressed with negative vocabulary.

Krugman is therefore the easy case. His semantics and rhetoric point in the same direction. When he thinks something is terrible, he tends to use words like terrible, stupid, dangerous, and cruel. A conventional sentiment model works relatively well.

The other four expose different failure modes.

2. Cowen demonstrates the irony/understatement problem

“Have a nice day” is almost a perfect adversarial example.

Lexically:

HAVE + NICE + DAY → positive

Pragmatically:

Here is a horrifying development. → negative

“Department of Yikes” is easier for a modern model, but the underlying Cowen technique remains: describe something alarming in an emotionally restrained voice and allow the reader to infer the evaluation.

This means a writer can become more pessimistic without becoming lexically more negative. Cowen could double the number of posts about nuclear war, fiscal collapse, AI catastrophe and institutional decay while maintaining exactly the same polite, curious prose.

A conventional sentiment index might barely notice.

3. Caplan demonstrates the mention-versus-use problem

Suppose someone writes:

“People are absurdly pessimistic. Life is vastly better than it used to be.”

That is an extremely optimistic proposition containing absurdly and pessimistic.

Or:

“This terrible argument ignores the enormous improvements in human welfare.”

Again, negative vocabulary embedded in a positive worldview.

Caplan’s writing contains a lot of this because he frequently argues against pessimism, fear and claims of social harm. A bag-of-words model partially counts the views he’s attacking as his own sentiment.

There’s an additional Caplan effect: emotional amplitude. He readily says both “I love X” and “Y is evil.” Cowen avoids both extremes. Averaging the words can therefore make two radically different personalities of prose look similar.

4. Hanson demonstrates the topic-versus-attitude problem

This may be the cleanest methodological lesson.

Imagine:

“There is a substantial probability that civilization eventually collapses.”

versus

“I’m terrified that civilization is going to collapse!”

The propositions could encode exactly the same probability distribution. Their affect is radically different.

Hanson specializes in the first construction. That’s why the low variance in our little quantitative experiment was interesting. He can move from mating to hidden motives to cultural maladaptation to extinction without changing linguistic register much.

Consequently, an affect classifier tells you something real about Hanson—his prose is affectively restrained—but tells you remarkably little about whether Hanson expects good or bad outcomes.

5. Hanania demonstrates the contempt/humor problem

Positive/negative sentiment is an especially poor representation of ridicule.

Calling an argument “impressively stupid” contains impressively. Calling a political development “hilarious” may express enjoyment of something the author considers disastrous. Describing one’s “fascination” with someone’s dishonesty generates positive lexical markers while communicating contempt.

This suggests that contempt is its own dimension. So are amusement and sarcasm.

A model with only:

positive ←——→ negative

is throwing away exactly the information that makes Hanania’s prose distinctive.

6. Krugman demonstrates when sentiment analysis actually works

This is an important counterpoint. I wouldn’t conclude that sentiment analysis is useless.

Krugman’s rhetorical style makes affect comparatively easy to measure because word valence, authorial evaluation, and emotional intensity are correlated. Our small quantitative sample picked him out immediately.

If we had a million anonymous passages and wanted to distinguish strongly negative polemical writing from neutral prose, ordinary sentiment analysis would probably perform quite well.

The trouble begins when we interpret that score as:

optimism, happiness, pessimism, worldview, fearfulness, or belief about the future.

Those are different latent variables.

7. “Positive sentiment” itself decomposes

Our five-person comparison suggests at least this structure:

And “emotion” itself needs several dimensions:

Positive: joy, enthusiasm, affection, admiration, amusement.

Negative: anger, fear/anxiety, sadness, contempt, disgust.

Then independently:

World assessment: things are going well/badly.

Future expectation: improvement/deterioration.

Agency/blame: bad outcomes caused by identifiable people versus impersonal forces.

These dimensions needn’t correlate strongly.

Our five writers almost look selected to occupy different cells.

8. This creates a serious problem for large-scale social-science applications

Suppose researchers calculate newspaper sentiment from 1950–2025 and discover that news became 30% “more negative.”

What happened?

Maybe the world got worse.

Maybe journalism became more alarmist.

Maybe journalists became angrier.

Maybe subject matter shifted.

Maybe headlines became more clickbaity.

Maybe prose became less euphemistic.

Maybe sarcasm increased.

Maybe people began explicitly discussing risks that previously went unmentioned.

Maybe the underlying classifier responds differently to changing vocabulary.

Our five writers show that these aren’t minor theoretical qualifications. They can reverse the substantive interpretation of the measurement.

Cowen and Caplan are particularly devastating examples. A lexical model can plausibly tell you that Cowen is more positive while a propositional classifier could plausibly tell you that Caplan believes the world is better.

Both measurements could be correct.

9. Modern LLMs potentially permit a much better approach

Instead of asking:

“Is this passage positive or negative?”

I’d ask several independent questions:

Affect: What emotions does the author express?

Propositional valence: Does the author claim that the situation being discussed is good or bad?

Expectations: Does the author expect improvement or deterioration?

Target: Toward whom or what is the sentiment directed?

Intensity: How strong is it?

Epistemic status: Is the author asserting the proposition, quoting it, considering it, or rejecting it?

Rhetorical mode: Is the statement sincere, sarcastic, humorous, or understated?

That would solve much of the Cowen/Caplan/Hanson problem.

And crucially, these should be coded sentence by sentence before aggregation. Otherwise a 2,000-word essay saying “Here are ten pessimistic arguments and why all ten are wrong” can easily get classified as pessimistic.

10. There’s an even broader lesson about quantitative social science

This exercise is a miniature example of construct validity.

Once we can assign something a number, it’s extraordinarily tempting to forget what the number actually measures.

TextBlob gives us a wonderfully precise result like:

polarity = +0.137

But that precision doesn’t solve the hard problem. The hard problem occurred before measurement:

What exactly do we mean by “sentiment”?

If we mean valence of vocabulary, +.137 is potentially meaningful.

If we mean how happy Cowen is, it tells us almost nothing.

If we mean how optimistic Cowen is about civilization, it may actually be misleading.

That distinction—measurement precision versus conceptual validity—is probably the biggest general lesson from the whole exercise.

And these five happen to make an unusually good teaching example. You could almost build a methods lecture around them:

Cowen: sarcasm/understatement failure.
Caplan: mention/use + high-amplitude failure.
Hanson: topic/affect failure.
Hanania: contempt/humor failure.
Krugman: the comparatively easy benchmark case.

The amusing bottom line is that our primitive sentiment algorithm wasn’t necessarily wrong about any of them. We were initially asking it a question its output couldn’t answer.