Turkish-language tweets, 2008 to 2021

Expressions of Criticism and Bias toward the United States

What 100,000 Turkish tweets about the United States criticize, and what they attribute to it: the paper's results, the trends behind them, and the tweets themselves.

Start here

This page shows what the paper finds, lets you follow the labels year by year, and lets you read and search the tweets yourself, including for questions the paper does not ask.

How the labels were made

  1. A classifier (Jev 1.14.0, by TypeSafe) read every tweet and answered each question in the codebook separately, giving how sure it is. An answer counts as yes at 0.5 or more.
  2. A tweet counts as anti-American when it is hostile to the United States and names at least one grievance: political, cultural, religious or economic. Hostility toward a U.S. politician alone is recorded but never counted.
  3. Two frontier models, Claude Opus 5.5 and GPT-5.5, answered the same questions for 700 sampled tweets. They are the reference the paper measures against.
  4. Every figure on the Results page is the classifier's count over all 100,000 tweets, corrected by how the classifier differs from the reference models on those 700. Its interval reflects the 700-tweet sample.
  5. For each claim the paper also reports how much relabeling of the reference it would survive, after Bisbee and Spirling.

Where to go next

Results for the paper
Each claim, with its corrected figures and how much relabeling it survives.
Trends
The classifier's labels year by year, beside the corrected figures by period.
Explore
Count any label among any group of tweets, for new questions.
Search the tweets
Find words or labels in the tweets, read them, and translate them into English.
Check the labels
Read the 700 sampled tweets with the three coders' labels side by side, and see where they disagree.
Codebook
Every question exactly as the classifier received it.
Raw data
Save the tweets, labels and estimates as CSV files.

Words used on this page

Anti-American
The paper's measure: the tweet is hostile to the United States and names at least one grievance about it.
Grievance
What the criticism is about: political (American power and policy), cultural (what America is), religious (America in religious terms) or economic (American economic power). A tweet can carry several.
Political level
Where a political grievance places the United States: inside Turkey's own politics (interference, the 15 July coup attempt, U.S. bases), in Turkey's region (Kurdish armed groups, Israel, wars in the Middle East, Greece and Armenia), or in international politics (an unreliable ally, sanctions, hegemony, hypocrisy).
Attribution
What a tweet alleges behind the grievance. Conspiratorial: covert American manipulation or hidden power. Inherent character: American conduct traced to what America or Americans are. Fixity: American conduct will not change whoever governs.
Classifier
Jev, the model that labelled all 100,000 tweets. Its labels drive the Trends, Explore and Search pages.
Reference models
Claude Opus 5.5 and GPT-5.5, which coded 700 sampled tweets with the same codebook. A choice of what to measure against, not a proof of truth.
Corrected figure
The classifier's total over all 100,000 tweets, adjusted by how it differs from the reference models on the 700. The bracket after it is the 95% interval.
Source group
The collection query a tweet came from: a U.S. political figure, an institution or the Atlantic alliance. It records how the tweet was collected, not who wrote it. The paper compares the Trump, Obama and Biden groups.
Period
2008–2012, 2013–2016, 2017–2019 and 2020–2021. The 700 sampled tweets are too few to correct single years.
Release weights
The weights shipped with the draw. Whether they correct for unequal selection is unverified, so the paper counts every tweet once and uses the weights only as a check.

What the labels can and cannot tell you

  • Any single label here is the classifier's judgment, and it errs in patterns: it counts about twice as much cultural and economic content as the reference models, and half as much inherent-character attribution. Cite the Results page, not counts from Explore.
  • Shares describe these 100,000 tweets, not Turkish opinion and not the frame the collection aimed at.
  • Rates within small groups (cultural tweets, single years, small source groups) rest on few tweets and move a lot. 2008 holds 58 tweets and 2009 holds 724.
  • Tweets are shown in the original Turkish. @handles appear as @user, links as [link], and account, card and phone numbers as [number]. Tweet IDs are not included. An English translation is a machine translation, made when you ask for it.