Donald J. Trump contains multitudes. The president’s 140-character official statements have been a source of much schadenfreude on Twitter, where critics — with no shortage of glee — regularly surface old statements from the president that contradict recent stances or otherwise foreshadow current critiques of Trump’s own presidency. There’s a tweet for every occasion.

In an effort to see if tech could automate these juxtapositions and possibly surface humans can’t spot, the Bot Studio at Quartz recently created Trump of Yore, a tool that automatically scans new Trump tweets and compares them to tweets in the president’s pre-inauguration 12,647-tweet archive, posting the comparisons to Twitter

According to me, the top tweet is akin to the bottom one from February 2012. pic.twitter.com/oLCCOt98kZ

— Trump Of Yore (@TrumpOfYore) June 16, 2017

The tool can’t consistently create ironic juxtapositions between Trump’s previous and current stances (armchair political commentators are safe for now), but it does often do a decent job of comparing tweets about similar ideas and topics. Some results work better than others: An anti-Obamacare tweet this week was aptly compared with September 2013 one about the law’s effect on American jobs and healthcare premiums. On the other hand, a tweet on Friday about the FBI investigation doesn’t seem to have much in common with a 2013 tweet about Lance Armstrong’s interview with Oprah Winfrey.

This tweet looks to me like this one from September 2013: https://t.co/O42OFwmvHh

— Trump Of Yore (@TrumpOfYore) June 13, 2017

(Jargon-averse readers, avert your gaze: On the back end, each new tweet is chopped up into small parts called “vectors,” each of which is assigned a number. These numbers are then compared to those from earlier tweets, and once a mathematically similar match is found, both the tweets are posted to Twitter. Or at least that’s my understanding. John Keefe goes into much geekier detail in a post on the Quartz Bot Studio Blog.)

The bot’s tech is based on work from researchers at Carnegie Mellon, who argued in a 2016 paper that this kind of semantic analysis could help track infectious diseases on social media. Keefe said that there are certainly other journalistic applications as well, particularly around automated ways of surfacing similar clusters of tweets.

“Computers are much better at this than humans are,” Keefe said.

Quartz Bot Studio, which launched last November, is working a set of automated tools for journalists it hopes will improve certain parts of the reporting and writing process. Trump of Yore will certainly inform those later projects, some of which will launch this fall.