Weekly Updates

Judges, technicians, or politicians: who should define our relationship with tech?

The short answer is: none should. But the public debate around our expectations and concerns about tech has been dominated by elected and unelected policymakers with very little active engagement from society.

Remember, for instance, when European regulators decided to block the launch of Google Bard over privacy concerns even though there were other services already operating in Europe with exactly the same models. In the wake of greater regulatory uncertainty, Meta also decided to postpone the launch of Threads in the EU.

Lately, we've also been seeing politicians arguing for greater control over citizens’ online and offline behavior. From Politico:

”Social media platforms like TikTok and Snapchat will face possible shutdowns when they don't crack down on problematic content during riots under the European Union's content law, Internal Commissioner Thierry Breton said Monday. (…)

‘When there is hateful content, content that calls – for example – for revolt, that also calls for killing and burning of cars, they will be required to delete [the content] immediately,’ Breton said in the interview on France Info, citing the Digital Services Act which will impose new requirements on large platforms from August 25.

‘If they fail to do so, they will be immediately sanctioned. We have teams who can intervene immediately," he said. "If they don't act immediately, then yes, at that point we'll be able not only to impose a fine but also to ban the operation [of the platforms] on our territory.’”

And also from Le Monde:

“French police should be able to spy on suspects by remotely activating the camera, microphone and GPS of their phones and other devices, lawmakers agreed late on Wednesday, July 5. Part of a wider justice reform bill, the spying provision has been attacked by both the left and rights defenders as an authoritarian snoopers' charter, though Justice Minister Éric Dupond-Moretti insists it would affect only "dozens of cases a year."

Covering laptops, cars and other connected objects as well as phones, the measure would allow the geolocation of suspects in crimes punishable by at least five years' jail. Devices could also be remotely activated to record sound and images of people suspected of terror offenses, as well as delinquency and organized crime”.

There is an inherent risk when the government can hold a kill switch to silence opponents online or spy on citizens, even if it starts out with the best of intentions.

Meanwhile, across the Atlantic, the US courts are being called to decide on the copyright law impacts of using readily available internet content to train AI models, among many other issues.

From the MIT Tech Review:

"Tech companies have treated publicly available copyrighted data on the internet as subject to “fair use” under US copyright law, which would allow them to use it without asking for permission first. Copyright holders disagree. The class actions will likely determine who is right, says Ghavi.

This is just the beginning of a new boom time for tech lawyers. The experts MIT Technology Review spoke to agreed that tech companies are also likely to face litigation over privacy and biometric data, such as images of people’s faces or clips of them speaking. Prisma Labs, the company behind the popular AI avatar program Lensa, is already facing a class action lawsuit over the way it’s collected users’ biometric data.

Ben Winters believes we will also see more lawsuits around product liability and Section 230, which would determine whether AI companies are responsible if their products go awry and whether they should be liable for the content their AI models produce.

“The litigation processes can be a blunt object for social change but, nonetheless, can be quite effective,” says Saveri. “And no one’s lobbying Matthew [Butterick] or me.”"

The quote from the last paragraph belongs to two lawyers who specialized in class actions in the US. The thought that a small group of self-interested lawyers is leading the debate on such general impact issues is definitely troubling, though.

All these trends lead to some analysts saying that the "dream of an open, global internet” is over and that democratic countries should stop “indulging in nostalgia or resigning themselves to this new reality” and “set out their vision for a safer, better alternative".

We disagree. The day we give in to the vision that an open, global internet is incompatible with a safe internet is the day we give in to a race to the bottom of personal liberties.

So what?

We as a society must regain the focus and the locus of the discussion. We must learn the language of this debate and claim the position of protagonists in this process so we can make sure our long-term interests are truly at the heart of every decision moving forward.

That's at the heart of our ambition at CTPI.

Are you with us? Share this newsletter with a friend. Let's reignite the dream of an open, global internet together.

Putting the pilot in the airplane

In a previous newsletter, we said

"An airplane is a tool being commanded and supervised by the pilot. As long as the pilot has aligned interests with the rest of the passengers and has the last word, no amount of AI will make planes less safe."

And Alphabet's Waymo did precisely that for their self-driving cars. Except that, instead of putting a pilot, their own employees decided to go to work on the car driven by their code and AI. From The Guardian:

"The autonomous vehicle unit of Alphabet, Google’s parent company, said Wednesday that it started carrying employees in electric Jaguar I-Pace SUVs without human backup drivers. Previously the company had been testing the vehicles with a safety driver behind the wheel just in case."

That aligns incentives way better than having the CEO on the car, as a CEO has too much to win and may be more inclined to take higher risks than the rest of their team.

Move fast and communicate clearly

In a previous newsletter, we said:

"[…] LLaMA project, which we, in fact, consider the most revolutionary advance in AI at the moment."

LLaMA is a specific implementation of a Large Language Model (LLM), like OpenAI's GPT and Google's Bard. The main reason we are excited about LLaMA is for its potential to democratize access to the technology. We believe that a profound impact on society is only achieved when technology is available for the whole society to use and benefit from it.

In their recent LLaMA 2 announcement, Meta did something unexpected by clearly addressing concerns about ethics, security, safety, and privacy, committing to build more responsible AIs.

Their commitment is striking. The effort and detail put into building LLaMA 2 and the supporting documents is impressive. But to understand why this surprised us, we need to provide some historical context on other announcements and launches from Meta.

Meta past launches

“Move fast and break things” was Meta's original motto. Meta valued the speed of building the product over disruptions to services, infrastructure, societal norms, or even regulation (although, of course, not intentionally). Speed was all that mattered.

This choice of prioritization caused, over the years, a lot of friction with regulators, which doesn't help when Meta tries to enter regulated markets.

One such example was the launch of WhatsApp Pay in India and in Brazil. WhatsApp Pay's launch in India was quickly met with limits by the NCPI. Even in 2022, WhatsApp Pay was only allowed to offer its services to a fourth of its customer base.

In Brazil, a week after Meta announced a WhatsApp update that would allow users to transfer money within the app, the Brazilian Central Bank and the competition authority issued orders for WhatsApp Pay to stop operating (link in Brazilian Portuguese). This stop order was put in place for the regulators to analyze possible impacts on financial stability and competition in the country. WhatsApp Pay was only allowed to operate almost a year later.

Given the launches were immediately followed by limitations or a stop order, it is reasonable to assume that communications with the authorities in India and Brazil didn't go well.

LLaMA 1 vs LLaMA 2

The LLaMA 1 launch followed the usual Meta way of launch. Build something, launch it, and if there is backlash from society, work on communication and regulators’ concerns. In this case, since ChatGPT was monopolizing the news, there was not a lot of reaction to LLaMA 1 from regulators. But let's take a deeper look at what it had to say about the current regulators’ concerns on ethics, privacy, safety, and security.

The launch referenced Meta's Responsible AI Initiative, which, while really interesting, is not tailor-made for LLaMA and does not provide a simple way to understand how those principles, including privacy, are effectively applied to LLaMA. For ethical concerns, they referenced the project's README page on Github, which includes five short paragraphs on the topic. The same README page mentions the following about safety:

“In particular, our model has not been trained with human feedback, and can thus generate toxic or offensive content, incorrect information or generally unhelpful answers.”

There was almost no communication about those concerns, and if LLaMA were at the center of the news at the time, we'd expect a lot of worry from regulators in general.

The LLaMA 2 launch, on the other hand, was entirely different. From the top of the launch page, Meta goes to great lengths to show that they understand societal concerns and have addressed them throughout the development process.

LLaMA 2's launch paper has a whole safety and security section and directly addresses ethical concerns through the text. Meta also established an acceptable use policy for LLaMA 2, which clearly communicates what they expect these models to be used for. And, since LLaMA is a foundational model for other people to build their models on top of, Meta also created a Responsible Use Guide for Developers to help developers build more responsible models. This guidance is fundamental as it allows developers to avoid risks in the models they make, effectively improving the use of AI.

The way to go

The LLaMA 2 launch anticipates concerns and builds trust with regulators and society by providing transparency and solutions before specific issues actually happen. This approach substantially reduces the friction with regulators and, in cases where new regulation may be necessary, helps guide the conversation in a productive way instead of in an emergency that needs to be fixed immediately.

This approach of building along with the regulators and incorporating their concerns on the technology is substantially more powerful than only moving fast. Speed only matters if that goes in the right direction, and, as we saw in the WhatsApp Pay case, moving too fast can cause even more delays than moving a bit slower. We hope to see more cases where Meta and other tech companies choose this path.

Balancing society's interests in tech design and policy