YouTube algorithms demonetizes LGBTQ content, Senators want a plan to combat deepfakes, and Smart Summons crashes

Hi beta reader friends, this is where I’ll say something brief and topical. I’m thinking about sending it around this time every Friday and that everything after “More news” will be in the (eventual) paid version. But that’s just the plan today! Send feedback and change my mind.

Big story

  • YouTube unfairly demonetizes videos that have LGBTQ keywords in the title, according to an investigation done by a group of YouTube creators. When those LGBTQ words were switched out for words like “friend” or “happy,” the videos were able to be monetized. This shouldn’t be a surprise to YouTube: they were sued earlier this year by a group of LGBTQ creators for pretty much this exact same issue.

    But don’t worry, it’s the algos fault. A spokesperson said “We use machine learning to evaluate content against our advertiser guidelines. Sometimes our systems get it wrong, which is why we’ve encouraged creators to appeal. Successful appeals ensure that our systems are updated to get better and better.” So, it’s the LGBTQ creators’ responsibility to teach the system not to punish them, not YouTube’s.

More news

  • Some workplaces are teaching employees about reinforcement learning by training and racing toy-sized self-driving cars. The cars, called DeepRacer, are a project out of Amazon Web Services, and apparently there is a league anyone can join.

  • France plans to be the first country in Europe to create a national ID program using facial recognition.

  • If you’re a journalist or anyone else that has to deal with transcribing a lot of audio, you probably already know this: automated transcription software has gotten much better over the last couple of years.

  • Researchers at Beihang University and Microsoft, China developed a bot to post made-up comments on news articles. The trolls should be thrilled.

  • Step aside, facial recognition. Body recognition systems are coming to airports, sports venues, and other businesses near you as early as next year. It will work in conjunction with facial recognition systems, with a camera scanning the upper body to create a “body profile” that is tied to a user’s face.

  • Companies are using emotion recognition tech to see what sort of vacations you might like.

  • AI-driven investment strategies are not panning out yet.

  • Democratic Senator Mark Warner of Virginia and Republican Senator Marco Rubio wrote a letter to a bunch of tech companies to ask them to do something about deepfakes.

  • Last week, Teslas got an update called “Smart Summons,” letting owners call their car to drive to them when its 200 feet or fewer away with a tap on a phone —and regulators are already looking into crashes.

  • Speaking of self-driving cars, Silicon Valley residents don’t want them in their neighborhoods.

  • Google knows its face recognition software didn’t work great on people with darker skin. So they hired contractors to go out to scan faces from people on the street. Those who consented, which included targeted groups like homeless encampments and college students, got a $5 gift card as a thank you.

Quote Unquote

“AI could become a digital library for the world’s knowledge to date—and we will all have library cards.”

-From "Can Fiction Introduce Empathy Into AI? Do We Want It To?” on Lithub

Datapoint of the week

7 percent

The drop in the labor share, from 63 to 56 percent, from 2000 to last year. Researchers at the Fed say automation is partly to blame.

Prospecting ArXiv

Optical character recognition has been around for Latin script languages for two decades, with MNIST being one of the classic datasets behind machine learning. Most other written languages have not had nearly the same amount of work done, including Amharic, Ethiopia’s official language and used by 22 million people. A paper published this week shows how the researchers went about using a convolutional neural network (CNN) to recognize handwritten Amharic for the first time.

The researchers had to put together a dataset of over a million images of handwritten characters (shown above), a time-consuming and tedious task. But they also wrote that without doing anything too fancy with the setup, they were able to achieve a “reasonable” recognition result. This sort of work goes to show that there is still a lot of low-hanging fruit for AI to tackle, especially in developing countries.