What's making US angry?

Chapter 1

It's not you. It's the White House.

Congress is angrier than it used to be — a lot angrier. But the average hides the mechanism: anger belongs to whichever party just lost the presidency, and each hand-over ratchets the floor a little higher.

Anger doubled, in two steps

Mean anger score per tweet, by year. Anger sat below 0.21 through the Obama years, jumped to 0.29 in 2017, and hit 0.44 in 2025 — the angriest year on record. Bars show 95% confidence intervals (they're tiny: each year has 66k–580k tweets).

The X: whoever is out of power is angry

Split by party, the trend becomes a perfect X. Democrats spike when Trump takes office in 2017 and again in 2025; Republicans spike under Biden. The two lines cross at every inauguration.

Same rule, four presidencies

Under Obama both parties were equally calm (0.18 vs 0.19). Since then the out-party has been roughly 0.12–0.17 angrier than the in-party — and today's in-party (0.35) is angrier than 2017's out-party (0.34). The floor never resets.

Chapter 2

The seismograph

Week by week, 2011 to 2026. Hover over any point; the red fill marks weeks above the local baseline, and the diamonds are the 40 biggest spikes. Notice how the blue and red lines are braided together until 2017 — and then split apart.

The biggest spikes

Anger in that week relative to the surrounding three months. Four of the top five are government shutdowns. Congress's most reliable source of anger is Congress.

Three kinds of spike

Self-inflicted. Shutdowns, debt ceilings, sequesters, collapsing spending bills. 13 of the 40 spikes. Whoever gets blamed does the yelling: #SchumerShutdown and #RepublicanShutdown trended in the same week.
Presidential actions. Travel ban, family separation, Helsinki, executive orders, the tax bill. About 10.
Violence and tragedy. Orlando, Dallas & Baton Rouge, George Floyd, January 6, Ukraine. These are the only weeks where the two party lines converge — briefly, everyone is angry about the same thing.

Chapter 3

Anger keeps office hours

We expected anger to follow the calendar: election season, campaign mode, the final stretch. It mostly doesn't. What it follows is the clock, the recess schedule, and the news. Every curve here is compared only within its own year, so the 2025 line isn't shouting over the 2012 one.

The angry hours are after dark

Anger by hour of day in each member's home time zone, shown as deviation from that year's average. Grey lines are individual years (hover to identify); black is the tweet-weighted mean. Working hours sit exactly on the average. 9 pm to 4 am is above average in 13 of 16 years — but only 4% of tweets are sent then. The few tweets sent at 2 am are angry ones.

August recess is the calm month

By month, same normalization. August is below the yearly average in 14 of 16 years — the only exceptions are 2020 and 2022. January has become the angriest month since 2018 (a new Congress, the State of the Union, and in 2021, the 6th).

Eight election cycles, no build-up

Weekly anger over each two-year cycle, ending at election day (dashed line). Black is everyone; blue and red are parties. There is no ramp into November in any cycle. The spikes are events — the 2013 shutdown, the 2018 shutdown, Jan 6 — not the calendar.

The final stretch is calmer

Mean anger in the last four weeks before election day versus the rest of that year. In six of seven elections the closing weeks are calmer — members switch to hopeful campaign messaging. Only 2020 broke the pattern, and only barely.

Chapter 4

The anger forecast

Suppose a member of Congress is about to tweet. How well can you predict it'll be angry? Overall, 29% of tweets are. Knowing the party alone tells you nothing. Knowing the party and who's president tells you a lot. Knowing the person makes it nearly certain.

Each thing you learn moves the odds

Probability that a tweet is angry, given what you know. Knowing it's a Democrat tells you nothing; knowing it's a Democrat during Trump's second term jumps the odds to 52%. The hour adds almost nothing on top. But a particular senator, 9-to-5, in 2025: a 90% forecast.

A forecast you can set your watch by

If Patty Murray tweets before sunrise during Trump's second term,
it's an angry tweet.

— 84% of 418 tweets, midnight–5 am Pacific, Jan 2025 – Aug 2026

hover a dot to read the tweet

Every dot is one of her 418 tweets sent between midnight and 5 am on her home-state clock (she's usually in Washington, where it's breakfast). Red is angry, grey is not. Press play and watch them land. The calm ones are birthday wishes and tributes to colleagues; the other 84% are about Medicare, shutdowns and arrests.

The loudest people in Congress

Everyone who tweets more than seven times a day is angrier than average.
Except Cory Booker.

— the 12 most prolific members, tweets per day while in office

Each bar is a member's posting rate; the red part is the share of those posts that are angry. The typical member posts 1.6 times a day and is angry 29% of the time. Volume and anger rise together (ρ = 0.31 across all 902 members) — you cannot post ten times a day about ribbon-cuttings. Senator Booker, apparently, can.

Night owls and morning people

Members whose late-night tweets (9 pm–4 am) are far angrier than their daytime ones — and the reverse. Mike Levin is a mild 25% by day and 56% after dark. Bill Hagerty is the opposite: 53% by day, 29% at night. Do not tweet at Mazie Hirono after midnight. Watch a day on the Hill — the animated time-lapse of these nine →

Chapter 5

Nobody unites from opposition

Anger is one emotion; we scored eleven. Sorting tweets into two registers — civic unity (positive, and explicitly about working across the aisle) and polarizing (angry or disgusted, and aimed at the other party) — shows the same flip as Chapter 1, in both directions at once. The party in power tweets unity. The party out of power tweets at the other side.

The mirror

Share of each party's tweets per year that are polarizing (top) and civic-uniting (bottom, note the smaller scale). Polarizing tweets belong to the out-party — Democrats 17% in 2017, Republicans 25% in 2022, Democrats 35% in 2025, the record. Civic unity belongs to the in-party: Republicans' peak (4.2%) is 2018, Democrats' (5.6%) is 2024. In 2025 both fall to 2.6%.

How far apart the parties feel

Each year, the average of all eleven emotions for each party; the line is the distance between the two profiles. The parties felt roughly the same things through 2016, split apart in 2017 (the gap was anger), came back together in 2019–20 — the pandemic year is the last time Congress was emotionally close — and split again in 2021. Since then the biggest gap isn't anger but optimism: one side has it, the other doesn't. In 2025 it's disgust.

The calendar of unity

Most uniting months

Most polarizing months

Civic-unity share versus polarizing share, both parties pooled, minimum 5,000 tweets in the month. Unity peaks on things that actually passed — infrastructure, COVID relief, the Ukraine response. The eight most polarizing months in the whole dataset are all in 2025. Watch the distance play out as a romance — Gone with the Wind, as told by tweets →

Chapter 6

Two networks, one Congress

Every tweet that begins RT @ is an endorsement — you chose to amplify it. Every tweet that begins .@ uses the old Twitter trick for addressing someone in front of everyone. Within a party that's usually praise; across the aisle it is a call-out — 53% of those are angry, twice the rate of an ordinary mention. Same members, same positions on the map, two completely different Congresses.

Who retweets whom

Two islands. The few threads between them — in gold — are the calmest tweets in the whole network: bipartisan bills, tributes, disasters.

Who calls out whom

Here the lines across the aisle are the whole point — and almost every one runs straight to a party leader.

The lightning rods

Most .@-addressed by the other party, with the mean anger of those tweets.

The bridges

Most retweeted by the other party. Thomas Massie and John McCain appear on both lists — the only people both sides retweet and yell at.

Appendix

How we measured anger

Data. 5.1M tweets from Congressional accounts (the HopHacks congress-tweets-unified dataset). We kept tweets by members of the House and Senate, sent while they were in office, from 2011 onward — dropping executive-branch accounts, pre-election personal tweets and a few hundred pre-2011 rows: 4.83M tweets, 902 members, 2011 to 2026-08.
Anger score. cardiffnlp/twitter-roberta-base-emotion-multilabel-latest, a RoBERTa model trained on tweets, gives each tweet a probability of expressing anger (0–1). Throughout the site, mean anger is the average of that probability, and a tweet counts as angry when the probability is above 0.5 — that is what every "% angry" means. We compared two emotion models on a 2,000-tweet sample and kept the one whose top-scored tweets were actually angry (the older 4-class TweetEval model rated "when we fight, we win!" as rage). Scores are model estimates spot-checked by hand, not human labels; a separate sentiment model's negative score correlates 0.80 with them.
Time. Hours are in each member's home-state time zone; timestamps are UTC as Twitter records them. Hour and month effects are computed within a single year and shown as deviation from that year's mean, so the steep 2011→2026 trend can't masquerade as a daily or seasonal pattern. Election cycles are shown separately, never pooled.
Networks. Retweets are tweets in Twitter's own RT @user: form (99.7% of all RTs); their anger score is that of the content amplified. Almost no retweets survive in the 2023–24 part of the collection (under 2% of tweets, against 11–19% in every other year), so those two years are excluded from the retweet network. Call-outs are tweets that start with .@user or contain it after a space. Across the aisle, 53% of .@ tweets are angry versus 27% of ordinary cross-party mentions and 25% of replies; within a party, .@ is barely angrier than a normal mention (28% vs 25%) — so we read it as a call-out only when it crosses the aisle. Edges need at least three tweets to be drawn.
Uniting vs. dividing. A tweet is civic-uniting if it is positive (joy, optimism, love or trust above 0.5 and neither anger nor disgust above 0.5) and contains explicit cross-party language (bipartisan, across the aisle, both parties, common ground, work/come together, unity). It is polarizing if anger or disgust is above 0.5 and it names the other party or its leaders. Gratitude tweets ("thank you", "proud to") are counted separately and not as unity. Emotional distance is the Euclidean distance between the two parties' mean 11-emotion vectors in a year.
Spikes. Weekly mean anger minus a centred 13-week rolling median; the top 40 local maxima at least three weeks apart, labelled by hand from the words and hashtags over-represented in that week's angry tweets compared with the surrounding twelve weeks.