Jigsaw is a Google spinout that began life as Eric Schmidt's do-nothing "think/do tank," Google Ideas. But under leadership of Jared Cohen, the organization has morphed into an incubator devoted to not just "advanc[ing] the best possibilities of the Internet but to fix the worst of it: surveillance, extremist indoctrination, censorship."

The organization has released a set of free and mostly great tools that make strides on these lines: Uproxy lets people behind censorwalls use their friends' internet connections as proxies to get around them; Project Shield uses Google's serverfarms to let political dissidents get the message out in the face of state-actor denial-of-service attacks; Montage helps human rights groups crowdsource analysis of Youtube videos; Password Alert catches phishing attempts on Google logins; Redirect Method shows videos by people who regret joining terrorist groups to people looking for extremist material, with surprising success; and Conversation AI is an experimental system to help filter/moderate troll-floods like those directed at Gamergate targets like Zoe Quinn and Anita Sarkeesian.

But Jigsaw isn't without its weaknesses and criticisms. Many people blame Cohen for promoting Haystack, a profoundly flawed and dangerous communications tool that the State Department urged on dissidents during the Arab Spring (Cohen denies it); while Wikileaks and Julian Assange accuse Cohen of being a front for the US State Department's "regime change" projects, as Cohen is a veteran of State and had a prominent, high profile career there.

On the social/technical side, there's good reason to be skeptical of machine-learning approaches to troll-fighting. From spam-filters to parental controls, systems of automated censorship have a dismal track record, and while Conversation AI performs well at times, it has some really troubling failure modes. A lot will depend on how it is implemented, and what kinds of human judgment is in the loop.

But on the plus side, Jigsaw seems aware of these issues, and has sourced some pretty high-quality training data ("130,000 snippets of discussion around Wikipedia pages" and "17 million comments from [New York] Times stories, along with data about which of those comments were flagged as inappropriate by moderators"), though it doesn't seem to have plans to make that data public for people who want to independently audit the sampling methodologies, a prerequisite for good technical work that's as old as the scientific method.

That is an extraordinary amount of training! Making the Internet great again!

I wonder how they plan on monetizing Jigsaw.

More about Jigsaw:

http://pandawhale.com/post/73359/approve-or-reject-can-you-moderate-five-new-york-times-comments