So, apparently over the weekend, Tumblr, in a bid to try and make its site safe for all, decided to alter its rules on adult content, and to hand over policing of these rules to an AI. While the full enforcement of these new rules are slated for mid-December, the AI has already been busy flagging posts as adult content.
Indeed, my twitter feed today has been full of people surprised that their material has been flagged. Worse, a lot of what’s being flagged isn’t “adult content” – also known as Not Safe For Work (NSFW) – at all. To name only a few examples, the AI, in its diligence, has been flagging:
- Cat pictures,
- Posts about protests,
- Pictures of the posters’ kids,
- or, in my case, work-in-progress screenshots.
Yes, I had a picture flagged as “adult content”; a screenshot of an image of a woman, fully clothed, just standing, as provocative as a potted plant. Granted, the picture is some years old and my word was my art bad then, but… NSFW? Really?
I think I need to state, before we go further, that I understand and approve of the need to police social media sites in particular and the Internet in general. The days when we could treat the Internet as an untamed frontier, where everyone can be free to do as they will, are long gone. In fact, I would argue that if we had some competent policing of the Internet and the social media services that reside there even two years ago, a lot of the western world’s current misery could have been avoided.
But the current methods of policing being tried – with private companies trying to put bandages on their brands by using AIs and algorithms – is incredibly damaging.
One of my favorite science fiction films was the 1951 movie The Day The Earth Stood Still. In it, the alien emissary Klaatu said this about the system that protects his world and others from interplanetary aggression: “The test of any such higher authority is, of course, the police force that supports it. For our policemen, we created this race of robots…” One hopes Klaatu’s superiors had better programmers than Tumblr, or else Earth might have been immolated before Klaatu ever had a chance to say “hello.” (Even then it was a close thing; on two occasions, Gort, the robot in question, had to be dissuaded from destroying the violent humans.)
Whatever Tumblr’s intent – for certainly there are plenty of people who use social media to bully, harass, extort, torture, and exploit others, and we do need a means to contain them and their damage – the solution of entrusting complex value judgments to an AI was a case of corporate laziness combined with a late awakening of conscience. AI, at its current level, is not suited for the task of policing.
The fact that Tumblr has essentially made the AI’s training to be crowdsourced worries me more. Over two years ago, Microsoft released an AI Twitter Chatbot named “Tay”, designed to learn and respond to people’s tweets. In less than 16 hours, Microsoft was forced to pull the plug, as Tay had transformed from a blank slate to a racist monster – all from just listening to what Twitter had to say. From listening to what the collective human conversation had to say.
Tay did quite a lot of damage in those 16 hours, and then again a week later when it was accidentally reactivated. Tay, however, had no authority beyond that of any human to voice its opinion. Tumblr’s AI is more far-reaching, and I suspect the best case scenario is that Tumblr will be forced to unplug it for much the same reasons Microsoft had to unplug Tay.
Private companies have demonstrated that, when managing social media services, they are either content to let harmful behaviour continue (mostly because said behaviour is very lucrative), or to employ half-baked and lazy solutions to our very real Internet problems. It is past time for governments to step up and develop proper policing practices for social media – to keep people online safe from harm while preserving their right to speak. The best way to do this, in my opinion, lies with people, not machines. People who are well-adjusted and specially trained. And the best places for social interaction on the internet may lie, not in social media, but in personal websites and properly maintained forums.
Someday, we will have the means and wisdom to make and train AI in ways where they can serve the public interest and uphold people’s rights, but that day is not today.
I have nothing of any particular value on Tumblr at this time – what art and articles I’ve posted there have either served or outlived their usefulness – so I plan to shut my Tumblr account down, with December 9 being my final day on the site. It’s served me well as a temporary home, or perhaps more like an internet motel. Now that I have this site, however, Tumblr is at best a diversion, and with its malfunctioning AI, a decidedly unsafe place to be.