“Just remember, what you’re seeing and what you’re reading is not what’s happening,” President Donald Trump told the crowd at a Veterans of Foreign Wars event in July.

The remark drew countless comparisons on Twitter to this line from George Orwell’s 1984: “The party told you to reject the evidence of your eyes and ears. It was their final, most essential command.”

The rise of authoritarianism has coincided with the proliferation of “deepfakes” — realistic videos created with artificial intelligence software. This has frightening implications for journalism. If people can no longer believe what they see and what they hear, it’s easy for political leaders to dismiss any evidence-based negative coverage as a fabrication. Facts will continue to lose their power as doctored videos and photos become more sophisticated, casting doubt on primary source material.

Creators of deepfakes employ a variety of techniques, including faceswaps, grafting a lip-syncing mouth or expression onto someone else’s face. An excellent example of this is when Jordan Peele partnered with BuzzFeed in April to create a viral PSA about deepfakes, literally putting words in President Barack Obama’s mouth. Now imagine how this technology could be weaponized against political candidates or used to justify an attack on a foreign power.

Altered videos have already been used to diminish trust in the press. In November, White House Press Secretary Sarah Sanders tweeted a video to justify revoking the credentials of CNN correspondent Jim Acosta. One key moment appeared to be sped up, showing Acosta’s hand rapidly moving down toward a White House intern’s arm. The misleading video was first shared by InfoWars editor-at-large Paul Joseph Watson. When asked about the clip, Sanders’s response was, “The question is did the reporter make contact or not? The video is clear, he did. We stand by our statement.”

We stand by our decision to revoke this individual’s hard pass. We will not tolerate the inappropriate behavior clearly documented in this video. pic.twitter.com/T8X1Ng912y

— Sarah Sanders (@PressSec) November 8, 2018

The Wall Street Journal is leading the charge in training its journalists to detect deepfakes. Some of the techniques they employ include finding older versions of the footage using reverse image search engines like Tineye and Google Images. Journalists can also use video editing programs to examine the footage frame-by-frame to identify signs like fuzziness around the mouth area, subtle differences in skin tones, or unnatural lighting and movements.

Training journalists is an important step, especially if it’s combined with educating their audiences. News organizations would build trust and lessen the spread of misinformation if they are transparent about how they vet videos and images before reporting on them.

Confirmation bias is a powerful force. That’s why fact-checking Donald Trump’s statements hasn’t had much of an effect on his supporters. If a fake image or video conforms to their worldview, people are inclined to share it. Those who have a negative impression of the media are also less likely to spot a false headline — while being more confident about finding accurate information online.

The 2016 election cycle demonstrated how easy it is to trick millions of people into believing facts that could easily be debunked with a Google search. To be able to identify a deepfake would require not just greater media literacy than many news consumers have, but some technological prowess and a healthy dose of skepticism.

A recent MIT study found that it takes true stories about six times as long as false ones to reach 1,500 people. Bots accelerated the spread of true and false news at the same rate, demonstrating that false news spreads because humans, not robots, are more likely to share it.

Consider the viral image of Trump helping to rescue a Hurricane Harvey flooding victim. There were some obvious indications it wasn’t real, such as the fact that Trump is wearing a suit and no life vest. Yet it was shared hundreds of thousands of times on Twitter and Facebook. Digitally altered images like this damage the credibility of news organizations. In a survey of nearly 6,000 American college students, 36% said the threat of misinformation made them trust all media less.

The most insidious fakes are ones that could be real, based on actual events and news reports. The Washington Post deconstructed one example, showing a commercial plane supposedly doing a 360-degree flip before landing safely in Shenzhen, China. An emergency landing did happen that day at the same airport. The video’s creators spliced together news coverage of it with a 360-degree landing video uploaded a year earlier by a filmmaker and visual effects animator.

Facebook’s machine-learning system detected the fake plane landing video a few days after it was posted, passed it to its network of independent fact-checkers, and eventually ensured it surfaced less in news feeds — but not before it had been viewed 14 million times.

While there are tricks and algorithms to help detect deepfakes, creators have already found ways to counteract them. Computer scientist Siwei Lyu realized he was losing staring contests with celebrities in these videos because they didn’t open and close their eyes as often as real humans. This makes sense, because deepfake programs rely on culling photos and videos of the subjects, and you’re less likely to find those of people blinking.

A few weeks after Lyu’s team published a paper of their findings, they started receiving links to YouTube videos where the subjects closed and opened their eyes naturally. This technology is evolving so rapidly that as quickly as we can find ways to counter it, its creators can adapt it to make it more convincing. As Lyu put it, “the competition between generating and detecting fake videos is analogous to a chess game.”