AI Images Now Outnumber Real Ones in US Midterm Tweets

Key findings at a glance

UncovAI and Laboratoire SAMM (Université Paris 1 Panthéon-Sorbonne) collected tweets posted between September 11 and 23, 2026. Here is what the data shows.

What did UncovAI analyse?

The study covers 8,039 tweets from 5,481 unique accounts posted on X between September 11 and September 23, 2026, excluding September 12. A keyword filter removed tweets about other countries' elections, leaving 6,360 US-election tweets from 4,283 accounts.

For stance, researchers hand-coded a random sample of 420 tweets plus the 80 most-liked tweets. Each was labelled pro-Trump/GOP, pro-Democratic/anti-Trump, or neutral. UncovAI's detection tools also analysed the images posted alongside the tweets.

This is Part 2 of a series. Part 1 covered August 30 to September 10, 2026.

How did the partisan balance on X change?

Between the two report windows, X moved from parity to a clear pro-Trump lead.

• Part 1 (Aug 30 – Sep 10): pro-Trump and pro-Democratic tweets were evenly matched by volume (1.02:1). The pro-Trump advantage showed up only in engagement, at roughly 2:1 in likes.
• Part 2 (Sep 11 – 23): the pro-Trump side led on volume too, at 59% vs 41%. Its engagement lead widened to 2.65:1.

The pro-Democratic side's daily volume stayed flat at about 55–78 tweets a day. The pro-Trump side's volume moved with the news, which drove the swings. Its share peaked at 79% after the September 14–15 court rulings, then fell to 64–66% during the media-ban and ICE-protest days of September 18–22.

What are the two camps talking about?

The two sides largely talk past each other.

Pro-Trump tweets

• Election fraud: 16% of the camp's tweets
• Immigration: 10.6%
• Islam: 7.8%

Anti-Trump tweets

• Corruption and the Epstein scandal
• Cost of living
• The Iran war: 5.5% vs 2.9% for pro-Trump
• Polls

Media freedom, gender issues and health are discussed at similar rates on both sides.

What is the most widespread misleading content?

Pre-emptive delegitimisation of the election. The report identifies sustained, repeated narratives, mostly pro-Trump, aimed at the legitimacy of the vote:

• "Democrats will cheat, steal or rig the midterms": 77 tweets from 74 accounts, with no evidence offered.
• "Non-citizen voting is widespread": 100 tweets from 58 accounts. The claim is misleading: 16 real DOJ prosecutions were presented as proof of mass fraud.
• "Trump should use emergency powers or martial law": 40 tweets from 38 accounts and 14.4k likes, peaking after the court losses of September 14–15.

The pattern is asymmetric. On the pro-Trump side, misleading content forms sustained campaigns repeated by many accounts. On the anti-Trump side, it mostly appears as isolated viral posts, such as an unverified allegation about ICE detention that became the most-liked anti-Trump post (144k likes).

Compared with Part 1, misleading content shifted from imported European virals to sustained domestic narratives.

Is there a bot network or foreign influence campaign?

No large-scale bot network was found. The report detected seven copy-paste or templated clusters. The largest was a "SAVE America Act" paragraph posted 15 times by 14 accounts, often small QAnon-branded profiles. Fewer than 5% of accounts in either camp had auto-generated handles, and state-media accounts (PressTV, CGTN, Global Times) posted once each.

Russia, Ukraine and the EU are marginal. Only about 1.7% of pro-Trump tweets carried such narratives:

• "Ukraine as saboteur" of US diesel prices (9 tweets), a Kremlin-compatible framing that echoes Trump's own remarks
• The "Russia hoax" (about 13 tweets)
• "Europe as a warning" on Muslim immigration (about 6 tweets)

Covert operations are designed to look domestic, so this data can neither confirm nor rule them out. Spotting manipulated content at scale is where detection tools come in.

Are AI-generated images now more common than real ones?

In this dataset, yes. UncovAI flagged 330 AI-generated images against 268 real ones (55% vs 45%). AI images matched or outnumbered real ones on nearly every day, and the gap widened towards September 23.

The report describes two widely shared AI-generated images, including one falsely implying a close relationship between Ursula von der Leyen and Jeffrey Epstein. It does not reproduce them. To check a suspicious picture yourself, try UncovAI image detection.

Why does this matter before November 3, 2026?

The study concludes that the main risk looks domestic rather than foreign: a sustained effort to delegitimise the election in advance, intensified after court defeats. Election-legitimacy narratives are the priority to monitor before the November 3 midterms, when all seats in the US House of Representatives and one third of the Senate are up for election.

Frequently asked questions