Google’s anti-trolling AI can be defeated by typos, researchers find [Updated] | Ars Technica
Visit any news organization’s website or any social media site, and you’re bound to find some abusive or hateful language being thrown around. As those who moderate Ars’ comments know, trying to keep a lid on trolling and abuse in comments can be an arduous and thankless task: when done too heavily, it smacks of censorship and suppression of free speech; when applied too lightly, it can poison the community and keep people from sharing their thoughts out of fear of being targeted. And human-based moderation is time-consuming.
Both of these problems are the target of a project by Jigsaw, an Alphabet startup effort spun off from Google. Jigsaw’s Perspective project is an application interface currently focused on moderating online conversations—using machine learning to spot abusive, harassing, and toxic comments. The AI applies a “toxicity score” to comments, which can be used to either aide moderation or to reject comments outright, giving the commenter feedback about why their post was rejected. Jigsaw is currently partnering with Wikipedia and The New York Times, among others, to implement the Perspective API to assist in moderating reader-contributed content.
But that AI still needs some training, as researchers at the University of Washington’s Network Security Lab recently demonstrated. In a paper published on February 27, Hossein Hosseini, Sreeram Kannan, Baosen Zhang, and Radha Poovendran demonstrated that they could fool the Perspective AI into giving a low toxicity score to comments that it would otherwise flag by simply misspelling key hot-button words (such as “iidiot”) or inserting punctuation into the word (“i.diot” or “i d i o t,” for example). By gaming the AI’s parsing of text, they were able to get scores that would allow comments to pass a toxicity test that would normally be flagged as abusive.
“One type of the vulnerabilities of machine learning algorithms is that an adversary can change the algorithm output by subtly perturbing the input, often unnoticeable by humans,” Hosseini and his co-authors wrote. “Such inputs are called adversarial examples, and have been shown to be effective against different machine learning algorithms even when the adversary has only a black-box access to the target model.”
The researchers also found that Perspective would flag comments that were not abusive in nature but used keywords that the AI had been trained to see as abusive. The phrases “not stupid” or “not an idiot” scored nearly as high on Perspective’s toxicity scale as comments that used “stupid” and “idiot.”
These sorts of false positives, coupled with easy evasion of the algorithms by adversaries seeking to bypass screening, belie the basic problem with any sort of automated moderation and censorship. Update: CJ Adams, Jigsaw’s product manager for Perspective, acknowledged the difficulty in a statement he sent to Ars:
It’s great to see research like this. Online toxicity is a difficult problem, and Perspective was developed to support exploration of how ML can be used to help discussion. We welcome academic researchers to join our research efforts on Github and explore how we can collaborate together to identify shortcomings of existing models and find ways to improve them.
Perspective is still a very early-stage technology, and as these researchers rightly point out, it will only detect patterns that are similar to examples of toxicity it has seen before. We have more details on this challenge and others on the Conversation AI research page. The API allows users and researchers to submit corrections like these directly, which will then be used to improve the model and ensure it can to understand more forms of toxic language, and evolve as new forms emerge over time.