Is Trump Normal?

A point of view from the cutting edge of A.I.

“I was saved by God to make America great again.”

Nowadays, a president will say this in his inauguration speech, and no one will bat an eye. But should we?

In this post, we explore whether it’s worth looking through the eyes of an artificial intelligence to find out. A large language model could sidestep the usual partisan narratives and – as a verifiable numerical process – perhaps even land on an empirical answer to subjective questions like these.

Depending on who you ask, the first two weeks of Donald Trump’s second term were either highly abnormal – or entirely to be expected. Some might describe themselves as “shocked, but not surprised.” In this era of cynicism, gaslighting, and sanewashing, it’s hard to even decide how to react.

This blurring of the line between what’s normal and what’s irregular relates to the core problem of our time: how hard it has become to tell reality from fiction. And while many approaches have tried to tackle this problem, none have convincingly nailed it.

First, there’s the whole host of fact-checking tag solutions, some of which have shown promise, but none of which have kept up with the velocity of falsehoods perpetrated online, and many of which have encountered fierce opposition and criticism. Then of course there was Meta’s third party fact-checking, famously ditched last month in favor of X-style Community Notes. Now, on whether tools like Community Notes can adequately keep up with falsehoods, the jury is still out. Which brings us to our traditionally more reliable but harder to scale judgment processes: jury trials, editorial oversight, and peer review – all of which are now more compromised by politics than ever.¹²³

So what if we tried something new? Three months ago, this newsletter proposed a novel way to measure departures from normalcy: train a language model to mimic a prominent figure’s speech patterns, and see what things that person said that least matched the trained model’s expectations.

You can read more about our methodology in the linked blogpost, but at a high level it involves training a language model on someone’s past speeches, using it to evaluate the likelihood of a recent speech they gave, and then comparing its evaluation to what an untrained model found. If the trained model found certain speech patterns less likely than an untrained model did, then that means the person spoke in a way that was uniquely surprising to that model. This allows us to rely on open-source numerical methods to evaluate a seemingly subjective aspect of human speech: how out of the ordinary it is.

So let’s give it a shot. President Trump’s second inaugural address can serve as a perfect early test of the relative normalcy of his communication style in his second term as president. Every four years, U.S. presidents have spoken with the force of history through these marquee speeches, which have cemented their agendas and set the tone for their terms.

In this lens, there are three important questions that our technique is well-positioned to answer:

• How did Trump’s tone in the inaugural speech differ from that of his campaign speeches? This should offer an early preview of what kind of president he intends to be. Will this be a “business as usual” presidency for Trump, or is he changing course? To answer this question, we trained a model to mimic Trump’s speeches on the campaign trail and looked at how it rated the likelihood of his inauguration speech.

• How did Trump’s inaugural address compare to those of past presidents? This can give us a sense of how Trump’s second term will mark its place in history. Will his second-term rhetorical style be as unprecedented as many found his first-term style to be? Here, we trained a model on all past United States presidential inaugural addresses.

• How did his remarks compare to his posts on X? We think this one is worth investigating because his inaugural speech was most likely written and rehearsed ahead of time, so it might bear more similarity to his written word style than to his spoken word style, which could have been more off-the-cuff during campaign speeches. A model trained solely on written words could do better at detecting a potential change in his tone.

Model One: Campaign Speech Trump

Before we dive into our first model’s results, let’s discuss how we modeled Trump on the campaign trail. Even when it comes to the cold objectivity of crunching numbers, there’s a number of subjective calls to be made. The ethos of this blog is to be as transparent about these choices as possible.

In this case, we established that Trump’s campaign began after he suffered an electoral defeat in 2020. We identified 35 transcripts on Rev.com of speeches Trump has made since then, beginning with his speech in support of Kelly Loeffler and David Purdue on January 4th, 2020. We only selected transcripts that were labeled as a “speech” by Rev to avoid capturing more improvised remarks, which may not compare as well to an inaugural address.

Some technical notes: This gave us 25,922 individual sentences said by Trump in speeches, to which we concatenated as much of the leading context (up to 1,024 “tokens”⁴) as possible (including introductory speakers). We then trained our model on this data – computing loss outputs on the full context, but using a loss mask to only update the model based on the target sentence outputs. Training took around 12 hours in total.

Over the course of training, we saved new snapshots of the model with each speech it learned from. Similar to our last post, we selected the snapshot of the model that performed best at predicting the inaugural address. However, unlike in the debates, where our trained models were better than an untrained one at predicting debate speech patterns by candidates, this time all of our trained model’s snapshots fared worse, and their performance on the inaugural address decreased the more the model trained (even though their performance on unseen campaign speeches improved).

On one hand, this is an exciting result, as it implies Trump’s inaugural speaking style was a significant departure from his tone on the campaign trail; on the other hand, it makes it difficult to choose which snapshot of the model to use for our analysis. To ensure the model we selected was an accurate representation of Trump’s campaign speaking style, we selected the best-performing snapshot of the model on the inaugural address, subject to the constraint that the model snapshot had trained on all of his campaign speeches at least once.

Below are some of the passages that were most uniquely surprising to that model. We highlighted the most relatively improbable words in a paler tone.

1. “We are one people, one family, and one glorious nation under God. So, to every parent who dreams for their child and every child who dreams for their future, I am with you, I will fight for you, and I will win for you.”

Context: Trump was emphasizing the resilience of the American people, highlighting their history of overcoming challenges and achieving the impossible.

Possible Interpretation: Of all the words in this excerpt, the first “I” was by far the most surprising. Indeed, looking back at his campaign speeches, it seems Trump most often spoke in the first person to tout his credentials. Capping his praise of the American people — and a passage meant to portray unity — by boasting would indeed have been strange. But here, he is instead positioning himself as a servant of the people: an atypical use of the first person for Trump in speeches according to our first model.

2. “This week, I will also end the government policy of trying to socially engineer race and gender into every aspect of public and private life.”

Context: Trump was presenting his agenda to revitalize the American economy and protect individual liberties: the creation of two new departments – the External Revenue Service and the Department of Government Efficiency – and a commitment to constitutional law, including free speech.

Possible Interpretation: At first glance, this might seem like a common talking point for Trump, who frequently criticized attitudes toward “race” throughout his campaign. But searching through his speeches for mentions of race, it seems his criticisms centered more on culture: he frequently demonized “critical race theory” and lamented the “race war.” In tying it here to government policy, he brings race up in a surprising context, suggesting more concrete reforms than previously advertised.

3. “Together, we will end the chronic disease epidemic and keep our children safe, healthy, and disease-free. The United States will once again consider itself a growing nation, one that increases our wealth, expands our territory, builds our cities, raises our expectations, and carries our flag into new and beautiful horizons.”

Context: Trump was in the midst of introducing his global perspective, most strikingly with his desire to exert control over the Panama Canal and grow the American territory.

Possible Interpretation: The mention of chronic disease embedded in a discussion of how to expand America’s influence was confusing to the model. This highlights that Robert Kennedy’s policies may have had a surprising impact on Trump’s agenda. In this passage, Trump is perhaps attempting to tie an unlikely issue for him into more on-brand American exceptionalism.

Overall, Trump’s most surprising shifts in tone can mostly be attributed to the change in responsibility and power he now holds. Our model found that he did not depart significantly from his core talking points but that he presented them in a new light: with a desire to achieve (and claim credit for) the objectives spelled out by his rhetoric.

Model Two: Presidential Inaugural Addresses

We trained our next model on the 52 past presidential inaugural addresses that were available on the American Presidency Project, resulting in 4,791 individual sentences, which we trained on in a similar manner to the Trump speeches model. Training took about six hours.

Once again, more training tended to lead to more difficulty predicting Trump’s inaugural address, so we selected the best model snapshot that had seen all of the inaugural addresses at least once.

1. “We will drill, baby, drill.”

Context: The president emphasized the importance of protecting the country and its borders, as well as combating inflation. To this end, he announced a “national energy emergency” to address rising energy prices.

Possible Interpretation: It’s no surprise that one of Trump’s favorite phrases was an unlikely one for past presidents to utter in their inaugural addresses. This was indeed and by far the most unlikely sentence for an inauguration speech, underscoring how suddenly a phrase often satirized by the left for its crassness regained enough popularity to be included in an inaugural address.

2. “Our power will stop all wars and bring a new spirit of unity to a world that has been angry, violent, and totally unpredictable.”

Context: After a passage applauding America’s history, the president turned his gaze to the future and promised to restore the country to a past era of greatness.

Possible Interpretation: While Trump’s rhetoric in praise of America is a familiar sight in inaugural speeches, a depiction of the world as “angry,” and in particular, “unpredictable,” is not. This passage is a testament to a radical shift in American perceptions of the rest of the world: rather than a global theater of old alliances and familiar opponents, it’s a messy place.

3. “The journey to reclaim our republic has not been an easy one, that, I can tell you.”

Context: Trump was criticizing the past government for prioritizing foreign aid over domestic response to disasters and for overspending on public health and education. He ended his criticism by presenting his election as a mandate for change.

Possible Interpretation: Trump uses the language of conquest to describe his victory, building on a narrative that America has fallen prey to misaligned, perhaps even foreign interests. The unlikelihood of the phrase “reclaim our republic” indicates that even with the conflicts America has faced, few occasions have warranted similar language from presidents – in particular as a way to describe an electoral victory.

In sum, Trump's most significant departures from expected inaugural talking points center on America’s position in a radically changed world: divided within, faced with a climate and energy crisis, and threatened by geopolitical instability.

Model Three: Trump Posts

Last but not least, we trained a model on all of his tweets and X posts.⁵ We filtered out any posts that contained links, leaving us with 39,533 posts. Training took about four hours.

Surprisingly, unlike the previous two models, this model performed better at predicting Trump’s inaugural address than an untrained one, suggesting that this speech might have been drafted in a similar manner to his posts on X – partly dictated out loud and workshopped with his staff, and partly written out himself.

1. “America will be respected again and admired again, including by people of religion, faith, and good will.”

Context: This passage came near the end of Trump’s speech, after his promise to restore America to its glorious past and his lengthy praise of the American people and their ability to do the impossible.

Possible Interpretation: Our third model rated this portion of Trump’s speech as least likely to have been posted by him on Twitter or X, and it’s plain to see why: he’s abandoned the pessimism other linguistic analyses have found him to habitually show on Twitter. Now that he is president again, he believes America will reclaim a position of prestige on the world stage, a change in belief mirrored in the nation’s partisan-driven shifts in optimism.

2. “They were farmers and soldiers, cowboys and factory workers, steelworkers and coal miners, police officers and pioneers who pushed onward, marched forward, and let no obstacle defeat their spirit or their pride.”

Context: This passage occurred during Trump’s praise of America’s history and its people. This sentence, once again, was rated as one of his least likely tweets.

Possible Interpretation: Trump is waxing somewhat poetic here, something unusual for his posts on X. He’s also describing America primarily in terms of his main base – people working in blue-collar and typically male professions. This may have been a departure from efforts to expand his coalition during the campaign.

3. “The future is ours, and our golden age has just begun. Thank you. God bless America.”

Context: This was the final line in Trump’s speech, capping the optimistic portion of the speech that our model found most misaligned with Trump’s voice on X.

Possible Interpretation: Our best model of Trump’s inaugural address seemed to struggle the most on the tone of optimism he ended it with. Perhaps it’s unsurprising that Trump is adopting a brighter tone with his return to power. Still, it’s worth noting that his first inaugural address is generally remembered as more in line with his typical dark tone, in particular for the turn of phrase “American carnage.”

In these examples, our model reveals a striking evolution from Trump's typically negative tone on social media to a more positive presidential voice. His inaugural address was most surprising when he described a bright future for America, with restored global respect amid a new "golden age."

What did we learn?

These models paint a version of Trump who assumes a greater sense of responsibility for America’s future, perceives a radically changed world, and expresses more optimism than he has in the past.

It makes sense that this version has translated into a more empowered version of Trump, unafraid to enact sweeping reform, which has so far taken shape as a record-breaking number of executive orders.

One of the more novel narratives detected by our first model was the placement of government policy at the center of race- and gender-based hiring. This narrative likely informed his controversial reaction to the recent deadly air collision near Washington, D.C. – which he tied to DEI-based hiring by the FAA.

Another linguistic analysis by NBC Washington also noticed Trump’s greater than usual emphasis on unity, while an A.I. analysis by the Connecticut Post noted that his speech was more focused on promises and plans, as well as an upcoming “golden age” than his first inaugural address. Meanwhile, the two analyses picked up on some trends that ours did not: NBC’s noted that Trump’s speech was more complex than recent inaugural addresses, while the Connecticut Post noted more aggressive rhetoric.

Still, the overall level of agreement bodes well for this kind of analysis. In an era where mistrust is rife, here is a numerical technique that surfaces similar findings to what other objective methods have landed on.

Amid all the hyperpartisan brainrot of our age, could artificial brains then serve as helpful prostheses? Language models invite us to take a step back and view our current moment from an entirely different, perhaps even alien perspective, and identify all the ways our current moment is unpredictable to an entity with different biases than ours.

At the very least, analyses like these can give a better sense of the way language models perceive the world. At times, that perspective can feel emotionally stunted, even naive. But much like hearing from a child seeing the world with fresh eyes, getting a deeper sense of what these models see and believe can open us to new ways of parsing a world that is — as some have noted — more unpredictable than ever.

In keeping with this blog’s commitment to openness, you can find all of our data download and processing code in this repository. Training, model selection, and analysis code is available in this Colab. All other files for this project, including training data and LoRA weights for the models, are available here.

Many thanks to Jonathan Cortez for meticulous edits and thoughtful suggestions throughout the writing process.

1. Studies show social media has made jurors less able to evaluate evidence objectively and more distrustful of institutions. Source: JD Supra.

2. Academic peer review has shifted from evaluating scholarly merit to judging political alignment. Source: Current.

3. Wikipedia's editorial system reveals vulnerability to coordinated partisan manipulation campaigns. Source: Stanford Cyber Policy Center.

4. More can be read about context windows and token limits here.

5. Source: Trump Twitter Archive.