Connect with us

Tech

The AI myth Western lawmakers get wrong

Published

on

China just announced a new social credit law. Here’s what it means.


This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here.

While the US and the EU may differ on how to regulate tech, their lawmakers seem to agree on one thing: the West needs to ban AI-powered social scoring.

As they understand it, social scoring is a practice in which authoritarian governments—specifically China—rank people’s trustworthiness and punish them for undesirable behaviors, such as stealing or not paying back loans. Essentially, it’s seen as a dystopian superscore assigned to each citizen. 

The EU is currently negotiating a new law called the AI Act, which will ban member states, and maybe even private companies, from implementing such a system.

The trouble is, it’s “essentially banning thin air,” says Vincent Brussee, an analyst at the Mercator Institute for China Studies, a German think tank.

Back in 2014, China announced a six-year plan to build a system rewarding actions that build trust in society and penalizing the opposite. Eight years on, it’s only just released a draft law that tries to codify past social credit pilots and guide future implementation. 

There have been some contentious local experiments, such as one in the small city of Rongcheng in 2013, which gave every resident a starting personal credit score of 1,000 that can be increased or decreased by how their actions are judged. People are now able to opt out, and the local government has removed some controversial criteria. 

But these have not gained wider traction elsewhere and do not apply to the entire Chinese population. There is no countrywide, all-seeing social credit system with algorithms that rank people.

As my colleague Zeyi Yang explains, “the reality is, that terrifying system doesn’t exist, and the central government doesn’t seem to have much appetite to build it, either.” 

What has been implemented is mostly pretty low-tech. It’s a “mix of attempts to regulate the financial credit industry, enable government agencies to share data with each other, and promote state-sanctioned moral values,” Zeyi writes. 

Kendra Schaefer, a partner at Trivium China, a Beijing-based research consultancy, who compiled a report on the subject for the US government, couldn’t find a single case in which data collection in China led to automated sanctions without human intervention. The South China Morning Post found that in Rongcheng, human “information gatherers” would walk around town and write down people’s misbehavior using a pen and paper. 

The myth originates from a pilot program called Sesame Credit, developed by Chinese tech company Alibaba. This was an attempt to assess people’s creditworthiness using customer data at a time when the majority of Chinese people didn’t have a credit card, says Brussee. The effort became conflated with the social credit system as a whole in what Brussee describes as a “game of Chinese whispers.” And the misunderstanding took on a life of its own. 

The irony is that while US and European politicians depict this as a problem stemming from authoritarian regimes, systems that rank and penalize people are already in place in the West. Algorithms designed to automate decisions are being rolled out en masse and used to deny people housing, jobs, and basic services. 

For example in Amsterdam, authorities have used an algorithm to rank young people from disadvantaged neighborhoods according to their likelihood of becoming a criminal. They claim the aim is to prevent crime and help offer better, more targeted support.  

But in reality, human rights groups argue, it has increased stigmatization and discrimination. The young people who end up on this list face more stops from police, home visits from authorities, and more stringent supervision from school and social workers.

It’s easy to take a stand against a dystopian algorithm that doesn’t really exist. But as lawmakers in both the EU and the US strive to build a shared understanding of AI governance, they would do better to look closer to home. Americans do not even have a federal privacy law that would offer some basic protections against algorithmic decision making. 

There is also a dire need for governments to conduct honest, thorough audits of the way authorities and companies use AI to make decisions about our lives. They might not like what they find—but that makes it all the more crucial for them to look.   

Deeper Learning

A bot that watched 70,000 hours of Minecraft could unlock AI’s next big thing

Research company OpenAI has built an AI that binged on 70,000 hours of videos of people playing Minecraft in order to play the game better than any AI before. It’s a breakthrough for a powerful new technique, called imitation learning, that could be used to train machines to carry out a wide range of tasks by watching humans do them first. It also raises the potential that sites like YouTube could be a vast and untapped source of training data. 

Why it’s a big deal: Imitation learning can be used to train AI to control robot arms, drive cars, or navigate websites. Some people, such as Meta’s chief AI scientist, Yann LeCun, think that watching videos will eventually help us train an AI with human-level intelligence. Read Will Douglas Heaven’s story here.

Bits and Bytes

Meta’s game-playing AI can make and break alliances like a human

Diplomacy is a popular strategy game in which seven players compete for control of Europe by moving pieces around on a map. The game requires players to talk to each other and spot when others are bluffing. Meta’s new AI, called Cicero, managed to trick humans to win. 

It’s a big step forward toward AI that can help with complex problems, such as planning routes around busy traffic and negotiating contracts. But I’m not going to lie—it’s also an unnerving thought that an AI can so successfully deceive humans. (MIT Technology Review) 

We could run out of data to train AI language programs 

The trend of creating ever bigger AI models means we need even bigger data sets to train them. The trouble is, we might run out of suitable data by 2026, according to a paper by researchers from Epoch, an AI research and forecasting organization. This should prompt the AI community to come up with ways to do more with existing resources. (MIT Technology Review)

Stable Diffusion 2.0 is out

The open-source text-to-image AI Stable Diffusion has been given a big facelift, and its outputs are looking a lot sleeker and more realistic than before. It can even do hands. The pace of Stable Diffusion’s development is breathtaking. Its first version only launched in August. We are likely going to see even more progress in generative AI well into next year. 



Tech

The Download: generative AI for video, and detecting AI text

Published

on

The original startup behind Stable Diffusion has launched a generative AI for video


The original startup behind Stable Diffusion has launched a generative AI for video

What’s happened: Runway, the generative AI startup that co-created last year’s breakout text-to-image model Stable Diffusion, has released an AI model that can transform existing videos into new ones by applying styles from a text prompt or reference image.

What it does: In a demo reel posted on its website, Runway shows how the model, called Gen-1, can turn people on a street into claymation puppets, and books stacked on a table into a cityscape at night. Other recent text-to-video models can generate very short video clips from scratch, but because Gen-1adapts existing footage it can produce much longer videos.

Why it matters: Last year’s explosion in generative AI was fueled by the millions of people who got their hands on powerful creative tools for the first time and shared what they made, and Runway hopes Gen-1 will have a similar effect on generated videos. Read the full story.

—Will Douglas Heaven

Why detecting AI-generated text is so difficult (and what to do about it)

Last week, OpenAI unveiled a tool that can detect text produced by its AI system ChatGPT. But if you’re a teacher who fears the coming deluge of ChatGPT-generated essays, don’t get too excited.

Continue Reading

Tech

Why detecting AI-generated text is so difficult (and what to do about it)

Published

on

Why detecting AI-generated text is so difficult (and what to do about it)


This tool is OpenAI’s response to the heat it’s gotten from educators, journalists, and others for launching ChatGPT without any ways to detect text it has generated. However, it is still very much a work in progress, and it is woefully unreliable. OpenAI says its AI text detector correctly identifies 26% of AI-written text as “likely AI-written.” 

While OpenAI clearly has a lot more work to do to refine its tool, there’s a limit to just how good it can make it. We’re extremely unlikely to ever get a tool that can spot AI-generated text with 100% certainty. It’s really hard to detect AI-generated text because the whole point of AI language models is to generate fluent and human-seeming text, and the model is mimicking text created by humans, says Muhammad Abdul-Mageed, a professor who oversees research in natural-language processing and machine learning at the University of British Columbia

We are in an arms race to build detection methods that can match the latest, most powerful models, Abdul-Mageed adds. New AI language models are more powerful and better at generating even more fluent language, which quickly makes our existing detection tool kit outdated. 

OpenAI built its detector by creating a whole new AI language model akin to ChatGPT that is specifically trained to detect outputs from models like itself. Although details are sparse, the company apparently trained the model with examples of AI-generated text and examples of human-generated text, and then asked it to spot the AI-generated text. We asked for more information, but OpenAI did not respond. 

Last month, I wrote about another method for detecting text generated by an AI: watermarks. These act as a sort of secret signal in AI-produced text that allows computer programs to detect it as such. 

Researchers at the University of Maryland have developed a neat way of applying watermarks to text generated by AI language models, and they have made it freely available. These watermarks would allow us to tell with almost complete certainty when AI-generated text has been used. 

The trouble is that this method requires AI companies to embed watermarking in their chatbots right from the start. OpenAI is developing these systems but has yet to roll them out in any of its products. Why the delay? One reason might be that it’s not always desirable to have AI-generated text watermarked. 

One of the most promising ways ChatGPT could be integrated into products is as a tool to help people write emails or as an enhanced spell-checker in a word processor. That’s not exactly cheating. But watermarking all AI-generated text would automatically flag these outputs and could lead to wrongful accusations.

Continue Reading

Tech

The original startup behind Stable Diffusion has launched a generative AI for video

Published

on

The original startup behind Stable Diffusion has launched a generative AI for video


Set up in 2018, Runway has been developing AI-powered video-editing software for several years. Its tools are used by TikTokers and YouTubers as well as mainstream movie and TV studios. The makers of The Late Show with Stephen Colbert used Runway software to edit the show’s graphics; the visual effects team behind the hit movie Everything Everywhere All at Once used the company’s tech to help create certain scenes.  

In 2021, Runway collaborated with researchers at the University of Munich to build the first version of Stable Diffusion. Stability AI, a UK-based startup, then stepped in to pay the computing costs required to train the model on much more data. In 2022, Stability AI took Stable Diffusion mainstream, transforming it from a research project into a global phenomenon. 

But the two companies no longer collaborate. Getty is now taking legal action against Stability AI—claiming that the company used Getty’s images, which appear in Stable Diffusion’s training data, without permission—and Runway is keen to keep its distance.

Gen-1 represents a new start for Runway. It follows a smattering of text-to-video models revealed late last year, including Make-a-Video from Meta and Phenaki from Google, both of which can generate very short video clips from scratch. It is also similar to Dreamix, a generative AI from Google revealed last week, which can create new videos from existing ones by applying specified styles. But at least judging from Runway’s demo reel, Gen-1 appears to be a step up in video quality. Because it transforms existing footage, it can also produce much longer videos than most previous models. (The company says it will post technical details about Gen-1 on its website in the next few days.)   

Unlike Meta and Google, Runway has built its model with customers in mind. “This is one of the first models to be developed really closely with a community of video makers,” says Valenzuela. “It comes with years of insight about how filmmakers and VFX editors actually work on post-production.”

Gen-1, which runs on the cloud via Runway’s website, is being made available to a handful of invited users today and will be launched to everyone on the waitlist in a few weeks.

Last year’s explosion in generative AI was fueled by the millions of people who got their hands on powerful creative tools for the first time and shared what they made with them. Valenzuela hopes that putting Gen-1 into the hands of creative professionals will soon have a similar impact on video.

“We’re really close to having full feature films being generated,” he says. “We’re close to a place where most of the content you’ll see online will be generated.”

Continue Reading

Copyright © 2021 Seminole Press.