In 2017, an online disinformation campaign spread against the “White Helmets,” claiming that the group of aid volunteers was serving as an arm of Western governments to sow unrest in Syria. This false information was convincing. But the Russian organization behind the campaign ultimately gave itself away because it repeated the same text across many different fake news sites. Now, researchers at the world’s top artificial intelligence labs are honing technology that can mimic how humans write, which could potentially help disinformation campaigns go undetected by generating huge amounts of subtly different messages.
One of the statements below is an example from the disinformation campaign. A.I. technology created the other. Guess which one is A.I.:
Tech giants like Facebook and governments around the world are struggling to deal with disinformation, from misleading posts about vaccines to incitement of sectarian violence. As artificial intelligence becomes more powerful, experts worry that disinformation generated by A.I. could make an already complex problem bigger and even more difficult to solve. In recent months, two prominent labs — OpenAI in San Francisco and the Allen Institute for Artificial Intelligence in Seattle — have built particularly powerful examples of this technology. Both have warned that it could become increasingly dangerous. Alec Radford, a researcher at OpenAI, argued that this technology could help governments, companies and other organizations spread disinformation far more efficiently: Rather than hire human workers to write and distribute propaganda, these organizations could lean on machines to compose believable and varied content at tremendous scale. A fake Facebook post seen by millions could, in effect, be tailored to political leanings with a simple tweak. “The level of information pollution that could happen with systems like this a few years from now could just get bizarre,” Mr. Radford said. This type of technology learns about the vagaries of language by analyzing vast amounts of text written by humans, including thousands of self-published books, Wikipedia articles and other internet content. After “training” on all this data, it can examine a short string of text and guess what comes next. We wanted to see what kind of text each of the labs’ systems would generate with a simple sentence as a starting point. How would the results change if we changed the subject of the sentence and the assertion being made?