What’s the Difference Between Mastodon, Bluesky, and Threads?

3 months 2 weeks ago

The ongoing Twitter exodus sparked life into a new way of doing social media. Instead of a handful of platforms trying to control your life online, people are reclaiming control by building more open and empowering approaches to social media. Some of these you may have heard of: Mastodon, Bluesky, and Threads. Each is distinct, but their differences can be hard to understand as they’re rooted in their different technical approaches. 

The mainstream social web arguably became “five websites, each consisting of screenshots of text from the other four,”  but in just the last few years radical and controversial changes to major platforms were a wake up call to many and are driving people to seek alternatives to the billionaire-driven monocultures.

Two major ecosystems have emerged in the wake, both encouraging the variety and experimentation of the earlier web. The first, built on ActivityPub protocol, is called the Fediverse. While it includes many different kinds of websites, Mastodon and Threads have taken off as alternatives for Twitter that use this protocol. The other is the AT Protocol, powering the Twitter alternative Bluesky.

These protocols, a shared language between computer systems, allow websites to exchange information. It’s a simple concept you’re benefiting from right now, as protocols enable you to read this post in your choice of app or browser. Opening this freedom to social media has a huge impact, letting everyone send and receive posts their own preferred way. Even better, these systems are open to experiment and can cater to every niche, while still connecting to everyone in the wider network. You can leave the dead malls of platform capitalism, and find the services which cater to you.

To save you some trial and error, we have outlined some differences between these options and what that might mean for them down the road.

ActivityPub and AT Protocols ActivityPub

The Fediverse goes a bit further back,  but ActivityPub’s development by the world wide web consortium (W3C) started in 2014. The W3C is a public-interest non-profit organization which has played a vital role in developing open international standards which define the internet, like HTML and CSS (for better or worse). Their commitment to ActivityPub gives some assurance the protocol will be developed in a stable and ostensibly consensus driven process.

This protocol requires a host website (often called an “instance”) to maintain an “inbox” and “outbox” of content for all of its users, and selectively share this with other host websites on behalf of the users. In this federation model users are accountable to their instance, and instances are accountable to each other. Misbehaving users are banned from instances, and misbehaving instances are cut off from others through “defederation.” This creates some stakes for maintaining good behavior, for users and moderators alike.

ActivityPub handles a wide variety of uses, but the application most associated with the protocol is Mastodon. However, ActivityPub is also integral to Meta’s own Twitter alternative, Threads, which is taking small steps to connect with the Fediverse. Threads is a totally different application, solely hosted by Meta, and is ten times bigger than the Fediverse and Bluesky networks combined—making it the 500-pound gorilla in the room. Meta’s poor reputation on privacy, moderation, and censorship, has driven many Fediverse instances to vow they’ll defederate from Threads. Other instances still may connect with Threads to help users find a broader audience, and perhaps help sway Threads users to try Mastodon instead.

AT Protocol

The Authenticated Transfer (AT) Protocol is newer; sparked by Twitter co-founder Jack Dorsey in 2019. Like ActivityPub, it is also an open source protocol. However, it is developed unilaterally by a private for-profit corporation— Bluesky PBLLC— though it may be imparted to a web standards body in the future. Bluesky remains mostly centralized. While it has recently opened up to small hosts, there are still some restrictions preventing major alternatives from participating. As developers further loosens control we will likely see rapid changes in how people use the network.

The AT Protocol network design doesn’t put the same emphasis on individual hosts as the Fediverse does, and breaks up hosting, distribution, and curation into distinct services. It’s easiest to understand in comparison to traditional web hosting. Your information, like posts and profiles, are held in Personal Data Servers (PDSes)—analogous to the hosting of a personal website. This content is then fetched by relay servers, like web crawlers, which aggregate a “firehose” of everyone’s content without much alteration. To sort and filter this on behalf of the user, like a “search engine,” AT has Appview services, which give users control over what they see. When accessing the Appview through a client app or website, the user has many options to further filter, sort, and curate their feed, as well as “subscribe” to filters and labels someone else made.

The result is a decentralized system which can be highly tailored while still offering global reach. However, this atomized system also may mean the community accountability encouraged by the host-centered system may be missing, and users are ultimately responsible for their own experience and moderation. This will depend on how the network opens to major hosts other than the Bluesky corporation.

User Experience

Mastodon, Threads and Bluesky have a number of differences that are not essential to their underlying protocol which affect users looking to get involved today. Mastodon and Bluesky are very customizable, so these differences are just addressing the prevalent trends.

Timeline Algorithm

Most Mastodon and most ActivityPub sites prefer a more straightforward timeline of content from accounts you follow. Threads have a Meta-controlled algorithm, like Instagram. Bluesky defaults to a chronological feed, but opens algorithmic curation and filtering up to apps and users. 

User Design

All three services present a default appearance that will be familiar to anyone who has used Twitter. Both Mastodon and Bluesky have alternative clients with the only limit being a developer’s imagination. In fact, thanks to their open nature, projects like SkyBridge let users of one network use apps built for the other (in this case, Bluesky users using Mastodon apps). Threads does not have any alternate clients and requires a developer API, which is still in beta.

Onboarding 

Threads has the greatest advantage to getting people to sign up, as it has only one site which accepts an Instagram account as a login. Bluesky also has only one major option for signing up, but has some inherent flexibility in moving your account later on. That said, diving into a few extra setup steps can improve the experience. Finally, one could easily join Mastodon by joining the flagship instance, mastodon.social. However, given the importance of choosing the right instance, you may miss out on some of the benefits of the Fediverse and want to move your account later on. 

Culture

Threads has a reputation for being more brand-focused, with more commercial accounts and celebrities, and Meta has made no secret about their decisions to deemphasize political posts on the platform. Bluesky is often compared to early Twitter, with a casual tone and a focus on engaging with friends. Mastodon draws more people looking for community online, especially around shared interests, and each instance will have distinct norms.

Privacy Considerations

Neither ActivityPub nor AT Protocol currently support private end-to-end encrypted messages at this time, so they should not be used for sensitive information. For all services here, the majority of content on your profile will be accessible from the public web. That said, Mastodon, Threads, and Bluesky differ in how they handle user data.

Mastodon

Everything you do as a user is entrusted to the instance host including posts, interactions, DMs, settings, and more. This means the owner of your instance can access this information, and is responsible for defending it against attackers and law enforcement. Tech-savvy people may choose to self-host, but users generally need to find an instance run by someone they trust.

The Fediverse muffles content sharing through a myriad of permissions set by users and instances. If your instance blocks a poorly moderated instance for example, the people on that other site will no longer be in your timelines nor able to follow your posts. You can also limit how messages are shared to further reduce the intended audience. While this can create a sense of community and closeness,  remember it is still public and instance hosts are always part of the equation. Direct messages, for example, will be accessible to your host and the host of the recipient.

If content needs to be changed or deleted after being shared, your instance can request these changes, and this is often honored. That said, once something is shared to the network, it may be difficult to “undo.”

Threads

All user content is entrusted to one host, in this case Meta, with a privacy policy similar to Instagram. Meta determines when information is shared with law enforcement, how it is used for advertising, how well protected it is from a breach, and so on.

Sharing with instances works differently for Threads, as Meta has more restricted interoperability. Currently, content sharing is one-way: Threads users can opt-in to sharing their content with the Fediverse, but won’t see likes or replies. By the end of this year, they will allow Threads users to follow accounts on Mastodon accounts.

Federation on Threads may always be restricted, and features like transferring one's account to Mastodon may never be supported. Limits in sharing should not be confused with enhanced privacy or security, however. Public posts are just that—public—and you are still trusting your host (Meta) with private data like DMs (currently handled by Instagram). Instead these restrictions, should they persist, should be seen as the minimum level of control over users Meta deems necessary.

Bluesky

Bluesky, in contrast, is a very “loud” system. Every public message, interaction, follow and block is hosted by your PDS and freely shared to everyone in the network. Every public post is for everyone and is only discovered according to their own app and filter preferences. There are ways to algorithmically imitate smaller spaces with filtering and algorithmic feeds, such as with the Blacksky project, but these are open to everyone and your posts will not be restricted to that curated space.

Direct messages are limited to the flagship Bluesky app, and can be accessed by the Bluesky moderation team. The project plans to eventually incorporate DMs into the protocol, and include end-to-end-encryption, but it is not currently supported. Deletion on Bluesky is simply handled by removing the content from your PDS, but once a message is shared to Relay and Appview services it may remain in circulation a while longer according to their retention settings.

Moderation Mastodon

Mastodon’s approach to moderation is often compared to subreddits, where the administrators of an instance are responsible for creating a set of rules and empowering a team of moderators to keep the community healthy. The result is a lot more variety in moderation experience, with the only boundary being an instance’s reputation in the broader Fediverse. Instances coordinating and “defederating” from problematic hosts has already been effective in the Fediverse. One former instance, Gab, was successfully cut off from the Fediverse for hosting extreme right-wing hate. The threat of defederation sets a baseline of behavior across the Fediverse, and from there users can choose instances based on reputation and on how aligned the hosts are with their own moderation preferences.

At its best, instances prioritize things other than growth. New members are welcomed and onboarded carefully as new community members, and hosts only grow the community if their moderation team can support it. Some instances even set a permanent cap on participation to a few thousand to ensure a quality and intimate experience. Current members too can vote with their feet, and if needed split off into their own new instance without needing to disconnect entirely.

While Mastodon has a lot going for it by giving users a choiceavoiding automation, and avoiding unsustainable growth, there are other evergreen moderation issues at play. Decisions can be arbitrary, inconsistent, and come with little recourse. These aren't just decisions impacting individual users, but also those affecting large swaths of them, when it comes to defederation. 

Threads

Threads, as alluded to when discussing privacy above, aims for a moderation approach more aligned with pre-2022 Twitter and Meta’s other current platforms like Instagram. That is, an impossible task of scaling moderation with endless growth of users.

As the largest of these services however, this puts Meta in a position to set norms around moderation as it enters the Fediverse. A challenge for decentralized projects will be to ensure Meta’s size doesn’t make them the ultimate authority on moderation decisions, a pattern of re-centralization we’ve seen happen in email. Spam detection tools have created an environment where email, though an open standard, is in practice dominated by Microsoft and Google as smaller services are frequently marked as spammers. A similar dynamic could play out with the federated social web, where Meta has capacity to exclude smaller instances with little recourse. Other instances may copy these decisions or fear not to do so, lest they are also excluded. 

Bluesky

While in beta, Bluesky received a lot of praise and criticism for its moderation. However, up until recently, all moderation was handled by the centralized Bluesky company—not throughout the distributed AT network. The true nature of moderation structure on the network is only now being tested.

AT Protocol relies on labeling services, aka “labelers”  for moderation. These special accounts using Bluesky’s Ozone tool labels posts with small pieces of metadata. You can also filter accounts with account block lists published by other users, a lot like the Block Together tool formerly available on Twitter. Your Appview aggregating your feed uses these labels to and block lists to filter content. Arbitrary and irreconcilable moderation decisions are still a problem, as are some of the risks of using automated moderation, but it is less impactful as users are not deplatformed and remain accessible to people with different moderation settings. This also means problematic users don’t go anywhere and can still follow you, they are just less visible.

The AT network is censorship resistant, and conversely, it is difficult to meaningfully ban users. To be propagated in the network one only needs a PDS to host their account, and at least one Relay to spread that information. Currently Relays sit out of moderation, only scanning to restrict CSAM. In theory Relays could be more like a Fediverse instance and more accurately curate and moderate users. Even then, as long as one Relay carries the user they will be part of the network. PDSes, much like web hosts, may also choose to remove controversial users, but even in those cases PDSes are easy to self-host even on a low-power computer.

Like the internet generally, removing content relies on the fragility of those targeted. With enough resources and support, a voice will remain online. Without user-driven approaches to limit or deplatform content (like defederation), Bluesky services may be targeted by censorship on the infrastructure level, like on the ISP level.

Hosting and Censorship

With any internet service, there are some legal obligations when hosting user generated content. No matter the size, hosts may need to contend with DMCA takedowns, warrants for user data, cyber attacks,  blocking from authoritarian regimes, and other pressures from powerful interests. This decentralized approach to social media also relies on a shared legal protection for all hosts, Section 230.  By ensuring they are not held liable for user-generated content, this law provides the legal protection necessary for these platforms to operate and innovate.

Given the differences in the size of hosts and their approach to moderation, it isn’t surprising that each of these platforms will address platform liability and censorship differently.

Mastodon

Instance hosts, even for small communities, need to navigate these legal considerations as we outlined in our Fediverse legal primer. We have already seen some old patterns reemerge with these smaller, and often hobbyist, hosts struggling to defend themselves from legal challenges and security threats. While larger hosts have resources to defend against these threats, an advantage of the decentralized model is censors need to play whack-a-mole in a large network where messages flow freely across the globe. Together, the Fediverse is set up to be quite good at keeping information safe from censorship, but individual users and accounts are very susceptible to targeted censorship efforts and will struggle with rebuilding their presence.

Threads

Threads is the easiest to address, as Meta is already several platforms deep into addressing liability and speech concerns, and have the resources to do so. Unlike Mastodon or Bluesky, they also need to do so on a much larger scale with a larger target on their back as the biggest platform backed by a multi-billion dollar company. The unique challenge for Threads however will be how Meta decides to handle content from the rest of the Fediverse. Threads users will also need to navigate the perks and pitfalls of sticking with a major host with a spotty track record on censorship and disinformation.

Bluesky

Bluesky is not yet tested beyond the flagship Bluesky services, and raises a lot more questions. PDSes, Relays and even Appviews play some role in hosting, and can be used with some redundancies. For example your account on one PDS may be targeted, but the system is designed to be easy for users to change this host, self-host, or have multiple hosts while retaining one identity on the network.

Relays, in contrast, are more computationally demanding and may remain the most “centralized” service as natural monopolies— users have some incentive to mostly follow the biggest relays. The result is a potential bottle-neck susceptible to influence and censorship. However, if we see a wide variety of relays with different incentives, it becomes more likely that messages can be shared throughout the network despite censorship attempts.

You Might Not Have to Choose

With this overview, you can start diving into one of these new Twitter alternatives leading the way in a more free social web. Thanks to the open nature of these new systems, where you set up will become less important with improved interoperability.

Both ActivityPub and AT Protocol developers are receptive to making the two better at communicating with one another, and independent projects like  Bridgy Fed, SkyBridge, RSS Parrot and Mastofeed are already letting users get the best of both worlds. Today a growing number of projects speak both protocols, along with older ones like RSS. It may be these paths towards a decentralized web become increasingly trivial as they converge, despite some early growing pains. Or the two may be eclipsed by yet another option. But their shared trajectory is moving us towards a more free, more open and refreshingly weird social web free of platform gatekeepers.

Rory Mir

Ah, Steamboat Willie. It’s been too long. 🐭

3 months 2 weeks ago

Did you know Disney’s Steamboat Willie entered the public domain this year? Since its 1928 debut, U.S. Congress has made multiple changes to copyright law, extending Disney’s ownership of this cultural icon for almost a century. A century.

Creativity should spark more creativity.

That’s not how intellectual property laws are supposed to work. In the United States, these laws were designed to give creators a financial incentive to contribute to science and culture. Then eventually the law makes this expression free for everyone to enjoy and build upon. Disney itself has reaped the abundant benefits of works in the public domain including Hans Christian Andersen’s “The Little Mermaid" and "The Snow Queen." Creativity should spark more creativity.

In that spirit, EFF presents to you this year’s EFF member t-shirt simply called “Fix Copyright":

Copyright Creativity is fun for the whole family.

The design references Steamboat Willie, but also tractor owners’ ongoing battle to repair their equipment despite threats from manufacturers like John Deere. These legal maneuvers are based on Section 1201 of the Digital Millennium Copyright Act or DMCA. In a recent appeals court brief, EFF and co-counsel Wilson Sonsini Goodrich & Rosati argued that Section 1201 chills free expression, impedes scientific research, and to top it off, is unenforceable because it’s too broad and violates the First Amendment. Ownership ain’t what it used to be, so let’s make it better.

We need you! Get behind this mission and support EFF's work as a member. Through EFF's 34th anniversary on July 10:

You can help cut through the BS and make the world a little brighter—whether online or off.

Join EFF

Defend Creativity & Innovation Online

_________________________

EFF is a member-supported U.S. 501(c)(3) organization celebrating TEN YEARS of top ratings from the nonprofit watchdog Charity Navigator! Your donation is tax-deductible as allowed by law.

Aaron Jue

同性カップルの住民票に「夫(未届)」と事実婚表記 長崎・大村市「内縁の夫婦に準ずる」

3 months 2 weeks ago
 長崎県大村市の市役所が5月2日、男性カップルに対し、続柄に「夫(未届)」と記載した住民票を交付したことがわかった。異性カップルの事実婚ではよく使われる表記だが、同性カップルでの適用は全国初のケースとみられる。  男性カ […]
admin

「地方自治法改定案」想田和弘

3 months 2 weeks ago
 地方自治法の改定案が今国会に提出され、審議されている。法案には、感染症や災害などの「重大な事態」が発生した際に、個別の法律の根拠がなくても国が自治体に「指示」を行なうことができるという特例が盛り込まれている。「緊急事態 […]
admin

Podcast Episode: AI in Kitopia

3 months 2 weeks ago

Artificial intelligence will neither solve all our problems nor likely destroy the world, but it could help make our lives better if it’s both transparent enough for everyone to understand and available for everyone to use in ways that augment us and advance our goals — not for corporations or government to extract something from us and exert power over us. Imagine a future, for example, in which AI is a readily available tool for helping people communicate across language barriers, or for helping vision- or hearing-impaired people connect better with the world.

%3Ciframe%20height%3D%2252px%22%20width%3D%22100%25%22%20frameborder%3D%22no%22%20scrolling%3D%22no%22%20seamless%3D%22%22%20src%3D%22https%3A%2F%2Fplayer.simplecast.com%2Fea9307db-050b-40c7-a346-91dce12a1683%3Fdark%3Dtrue%26amp%3Bcolor%3D000000%22%20allow%3D%22autoplay%22%3E%3C%2Fiframe%3E Privacy info. This embed will serve content from simplecast.com

   

(You can also find this episode on the Internet Archive and on YouTube.)

This is the future that Kit Walsh, EFF’s Director of Artificial Intelligence & Access to Knowledge Legal Projects, and EFF Senior Staff Technologist Jacob Hoffman-Andrews, are working to bring about. They join EFF’s Cindy Cohn and Jason Kelley to discuss how AI shouldn’t be a tool to cash in, or to classify people for favor or disfavor, but instead to engage with technology and information in ways that advance us all. 

In this episode you’ll learn about: 

  • The dangers in using AI to determine who law enforcement investigates, who gets housing or mortgages, who gets jobs, and other decisions that affect people’s lives and freedoms. 
  • How "moral crumple zones” in technological systems can divert responsibility and accountability from those deploying the tech. 
  • Why transparency and openness of AI systems — including training AI on consensually obtained, publicly visible data — is so important to ensure systems are developed without bias and to everyone’s benefit. 
  • Why “watermarking” probably isn’t a solution to AI-generated disinformation. 

Kit Walsh is a senior staff attorney at EFF, serving as Director of Artificial Intelligence & Access to Knowledge Legal Projects. She has worked for years on issues of free speech, net neutrality, copyright, coders' rights, and other issues that relate to freedom of expression and access to knowledge, supporting the rights of political protesters, journalists, remix artists, and technologists to agitate for social change and to express themselves through their stories and ideas. Before joining EFF, Kit led the civil liberties and patent practice areas at the Cyberlaw Clinic, part of Harvard University's Berkman Klein Center for Internet and Society; earlier, she worked at the law firm of Wolf, Greenfield & Sacks, litigating patent, trademark, and copyright cases in courts across the country. Kit holds a J.D. from Harvard Law School and a B.S. in neuroscience from MIT, where she studied brain-computer interfaces and designed cyborgs and artificial bacteria. 

Jacob Hoffman-Andrews is a senior staff technologist at EFF, where he is lead developer on Let's Encrypt, the free and automated Certificate Authority; he also works on EFF's Encrypt the Web initiative and helps maintain the HTTPS Everywhere browser extension. Before working at EFF, Jacob was on Twitter's anti-spam and security teams. On the security team, he implemented HTTPS-by-default with forward secrecy, key pinning, HSTS, and CSP; on the anti-spam team, he deployed new machine-learned models to detect and block spam in real-time. Earlier, he worked on Google’s maps, transit, and shopping teams.

Resources: 

What do you think of “How to Fix the Internet?” Share your feedback here

Transcript

KIT WALSH
Contrary to some marketing claims, AI is not the solution to all of our problems. So I'm just going to talk about how AI exists in Kitopia. And in particular, the technology is available for everyone to understand. It is available for everyone to use in ways that advance their own values rather than hard coded to advance the values of the people who are providing it to you and trying to extract something from you and as opposed to embodying the values of a powerful organization, public or private, that wants to exert more power over you by virtue of automating its decisions.
So it can make more decisions classifying people, figuring out whom to favor, whom to disfavor. I'm defining Kitopia a little bit in terms of what it's not, but to get back to the positive vision, you have this intellectual commons of research development of data that we haven't really touched on privacy yet, but but data that is sourced in a consensual way and when it's, essentially, one of the things that I would love to have is a little AI muse that actually does embody my values and amplifies my ability to engage with technology and information on the Internet in a way that doesn't feel icky or oppressive and I don't have that in the world yet.

CINDY COHN
That’s Kit Walsh, describing an ideal world she calls “Kitopia”. Kit is a senior staff attorney at the Electronic Frontier Foundation. She works on free speech, net neutrality and copyright and many other issues related to freedom of expression and access to knowledge. In fact, her full title is EFF’s Director of Artificial Intelligence & Access to Knowledge Legal Projects. So, where is Kitopia, you might ask? Well we can’t get there from here - yet. Because it doesn’t exist. Yet. But here at EFF we like to imagine what a better online world would look like, and how we will get there and today we’re joined by Kit and by EFF’s Senior Staff Technologist Jacob Hoffman-Andrews. In addition to working on AI with us, Jacob is a lead developer on Let's Encrypt, and his work on that project has been instrumental in helping us encrypt the entire web. I’m Cindy Cohn, the executive director of the Electronic Frontier Foundation.

JASON KELLEY
And I’m Jason Kelley, EFF’s Activism Director. This is our podcast series How to Fix the Internet.

JACOB HOFFMAN-ANDREWS
I think in my ideal world people are more able to communicate with each other across language barriers, you know, automatic translation, transcription of the world for people who are blind or for deaf people to be able to communicate more clearly with hearing people. I think there's a lot of ways in which AI can augment our weak human bodies in ways that are beneficial for people and not simply increasing the control that their governments and their employers have over their lives and their bodies.

JASON KELLEY
We’re talking to Kit and Jacob both, because this is such a big topic that we really need to come at it from multiple angles to make sense of it and to figure out the answer to the really important question which is, How can AI actually make the world we live in, a better place?

CINDY COHN
So while many other people have been trying to figure out how to cash in on AI, Kit and Jacob have been looking at AI from a public interest and civil liberties perspective on behalf of EFF. And they’ve also been giving a lot of thought to what an ideal AI world looks like.

JASON KELLEY
AI can be more than just another tool that’s controlled by big tech. It really does have the potential to improve lives in a tangible way. And that’s what this discussion is all about. So we’ll start by trying to wade through the hype, and really nail down what AI actually is and how it can and is affecting our daily lives.

KIT WALSH
The confusion is understandable because AI is being used as a marketing term quite a bit, rather than as an abstract concept, rather than as a scientific concept.
And the ways that I think about AI, particularly in the decision-making context, which is one of our top priorities in terms of where we think that AI is impacting people's rights, is first I think about what kind of technology are we really talking about because sometimes you have a tool that actually no one is calling AI, but it is nonetheless an example of algorithmic decision-making.
That also sounds very fancy. This can be a fancy computer program to make decisions, or it can be a buggy Excel spreadsheet that litigators discover is actually just omitting important factors when it's used to decide whether people get health care or not in a state health care system.

CINDY COHN
You're not making those up, Kit. These are real examples.

KIT WALSH
That’s not a hypothetical. Unfortunately, it’s not a hypothetical, and the people who litigated that case lost some clients because when you're talking about not getting health care that can be life or death. And machine learning can either be a system where you – you, humans, code a reinforcement mechanism. So you have sort of random changes happening to an algorithm, and it gets rewarded when it succeeds according to your measure of success, and rejected otherwise.
It can be training on vast amounts of data, and that's really what we've seen a huge surge in over the past few years, and that training can either be what's called unsupervised, where you just ask your system that you've created to identify what the patterns are in a bunch of raw data, maybe raw images, or it can be supervised in the sense that humans, usually low paid humans, are coding their views on what's reflected in the data.
So I think that this is a picture of a cow, or I think that this picture is adult and racy. So some of these are more objective than others, and then you train your computer system to reproduce those kinds of classifications when it makes new things that people ask for with those keywords, or when it's asked to classify a new thing that it hasn't seen before in its training data.
So that's really a very high level oversimplification of the technological distinctions. And then because we're talking about decision-making, it's really important who is using this tool.
Is this the government which has all of the power of the state behind it and which administers a whole lot of necessary public benefits - that is using decisions to decide who is worthy and who is not to obtain those benefits? Or, who should be investigated? What neighborhoods should be investigated?
We'll talk a little bit more about the use in law enforcement later on, but it's also being used quite a bit in the private sector to determine who's allowed to get housing, whether to employ someone, whether to give people mortgages, and that's something that impacts people's freedoms as well.

CINDY COHN
So Jacob, two questions I used to distill down on AI decision-making are, who is the decision-making supposed to be serving and who bears the consequences if it gets it wrong? And if we think of those two framing questions, I think we get at a lot of the issues from a civil liberties perspective. That sound right to you?

JACOB HOFFMAN-ANDREWS
Yeah, and, you know, talking about who bears the consequences when an AI or technological system gets it wrong, sometimes it's the person that system is acting upon, the person who's being decided whether they get healthcare or not and sometimes it can be the operator.
You know, it's, uh, popular to have kind of human in the loop, like, oh, we have this AI decision-making system that's maybe not fully baked. So there's a human who makes the final call. The AI just advises the human and, uh, there's a great paper by Madeleine Clare Elish describing this as a form of moral crumple zones. Uh, so, you may be familiar in a car, modern cars are designed so that in a collision, certain parts of the car will collapse to absorb the force of the impact.
So the car is destroyed but the human is preserved. And, in some human in the loop decision making systems often involving AI, it's kind of the reverse. The human becomes the crumple zone for when the machine screws up. You know, you were supposed to catch the machine screwup. It didn't screw up in over a thousand iterations and then the one time it did, well, that was your job to catch it.
And, you know, these are obviously, you know, a crumple zone in a car is great. A moral crumple zone in a technological system is a really bad idea. And it takes away responsibility from the deployers of that system who ultimately need to bear the responsibility when their system harms people.

CINDY COHN
So I wanna ask you, what would it look like if we got it right? I mean, I think we do want to have some of these technologies available to help people make decisions.
They can find patterns in giant data probably better than humans can most of the time. And we'd like to be able to do that. So since we're fixing the internet now, I want to stop you for a second and ask you how would we fix the moral crumple zone problem or what were the things we think about to do that?

JACOB HOFFMAN-ANDREWS
You know, I think for the specific problem of, you know, holding say a safety driver or like a human decision-maker responsible for when the AI system they're supervising screws up, I think ultimately what we want is that the responsibility can be applied all the way up the chain to the folks who decided that that system should be in use. They need to be responsible for making sure it's actually a safe, fair system that is reliable and suited for purpose.
And you know, when a system is shown to bring harm, for instance, you know, a self-driving car that crashes into pedestrians and kills them, you know, that needs to be pulled out of operation and either fixed or discontinued.

CINDY COHN
Yeah, it made me think a little bit about, you know, kind of a change that was made, I think, by Toyota years ago, where they let the people on the front line stop the line, right? Um, I think one thing that comes out of that is you need to let the people who are in the loop have the power to stop the system, and I think all too often we don't.
We devolve the responsibility down to that person who's kind of the last fair chance for something but we don't give them any responsibility to raise concerns when they see problems, much less the people impacted by the decisions.

KIT WALSH
And that’s also not an accident of the appeal of these AI systems. It's true that you can't hold a machine accountable really, but that doesn't deter all of the potential markets for the AI. In fact, it's appealing for some regulators, some private entities, to be able to point to the supposed wisdom and impartiality of an algorithm, which if you understand where it comes from, the fact that it's just repeating the patterns or biases that are reflected in how you trained it, you see it's actually, it's just sort of automated discrimination in many cases and that can work in several ways.
In one instance, it's intentionally adopted in order to avoid the possibility of being held liable. We've heard from a lot of labor rights lawyers that when discriminatory decisions are made, they're having a lot more trouble proving it now because people can point to an algorithm as the source of the decision.
And if you were able to get insight in how that algorithm were developed, then maybe you could make your case. But it's a black box. A lot of these things that are being used are not publicly vetted or understood.
And it's especially pernicious in the context of the government making decisions about you, because we have centuries of law protecting your due process rights to understand and challenge the ways that the government makes determinations about policy and about your specific instance.
And when those decisions and when those decision-making processes are hidden inside an algorithm then the old tools aren't always effective at protecting your due process and protecting the public participation in how rules are made.

JASON KELLEY
It sounds like in your better future, Kit, there's a lot more transparency into these algorithms, into this black box that's sort of hiding them from us. Is that part of what you see as something we need to improve to get things right?

KIT WALSH
Absolutely. Transparency and openness of AI systems is really important to make sure that as it develops, it develops to the benefit of everyone. It's developed in plain sight. It's developed in collaboration with communities and a wider range of people who are interested and affected by the outcomes, particularly in the government context though I'll speak to the private context as well. When the government passes a new law, that's not done in secret. When a regulator adopts a new rule, that's also not done in secret. There's either, sure, that's, there are exceptions.

CINDY COHN
Right, but that’s illegal.

JASON KELLEY
Yeah, that's the idea. Right. You want to get away from that also.

KIT WALSH
Yeah, if we can live in Kitopia for a moment where, where these things are, are done more justly, within the framework of government rulemaking, if that's occurring in a way that affects people, then there is participation. There's meaningful participation. There's meaningful accountability. And in order to meaningfully have public participation, you have to have transparency.
People have to understand what the new rule is that's going to come into force. And because of a lot of the hype and mystification around these technologies, they're being adopted under what's called a procurement process, which is the process you use to buy a printer.
It's the process you use to buy an appliance, not the process you use to make policy. But these things embody policy. They are the rule. Sometimes when the legislature changes the law, the tool doesn't get updated and it just keeps implementing the old version. And that means that the legislature's will is being overridden by the designers of the tool.

JASON KELLEY
You mentioned predictive policing, I think, earlier, and I wonder if we could talk about that for just a second because it's one way where I think we at EFF have been thinking a lot about how this kind of algorithmic decision-making can just obviously go wrong, and maybe even should never be used in the first place.
What we've seen is that it's sort of, you know, very clearly reproduces the problems with policing, right? But how does AI or this sort of predictive nature of the algorithmic decision-making for policing exacerbate these problems? Why is it so dangerous I guess is the real question.

KIT WALSH
So one of the fundamental features of AI is that it looks at what you tell it to look at. It looks at what data you offer it, and then it tries to reproduce the patterns that are in it. Um, in the case of policing, as well as related issues around decisions for pretrial release and parole determinations, you are feeding it data about how the police have treated people, because that's what you have data about.
And the police treat people in harmful, racist, biased, discriminatory, and deadly ways that it's really important for us to change, not to reify into a machine that is going to seem impartial and seem like it creates a veneer of justification for those same practices to continue. And sometimes this happens because the machine is making an ultimate decision, but that's not usually what's happening.
Usually the machine is making a recommendation. And one of the reasons we don't think that having a human in the loop is really a cure for the discriminatory harms is that humans are more likely to follow the AI if it gives them cover for a biased decision that they're going to make. And relatedly, some humans, a lot of people, develop trust in the machine and wind up following it quite a bit.
So in these contexts, if you really wanted to make predictions about where a crime was going to occur, well it would send you to Wall Street. And that's not, that's not the result that law enforcement wants.
But, first of all, you would actually need data about where crimes occur, and generally people who don't get caught by the police are not filling out surveys to say, here are the crimes I got away with so that you can program a tool that's going to do better at sort of reflecting some kind of reality that you're trying to capture. You only know how the system has treated people so far and all that you can do with AI technology is reinforce that. So it's really not an appropriate problem to try to solve with this technology.

CINDY COHN
Yeah, our friends at Human Rights Data Analysis Group who did some of this work said, you know, we call it predictive policing, but it's really predicting the police because we're using what the police already do to train up a model, and of course it's not going to fix the problems with how police have been acting in the past. Sorry to interrupt. Go on.

KIT WALSH
No, to build on that, by definition, it thinks that the past behavior is ideal, and that's what it should aim for. So, it's not a solution to any kind of problem where you're trying to change a broken system.

CINDY COHN
And in fact, what they found in the research was that the AI system will not only replicate what the police do, it will double down on the bias because it's seeing a small trend and it will increase the trend. And I don't remember the numbers, but it's pretty significant. So it's not just that the AI system will replicate what the police do. What they found in looking at these systems is that the AI systems increase the bias in the underlying data.
It's really important that we continue to emphasize the ways in which AI and machine learning are already being used and already being used in ways that people may not see, but dramatically impact them. But right now, what's front of mind for a lot of people is generative AI. And I think many, many more people have started playing around with that. And so I want to start with how we think about generative AI and the issues it brings. And Jacob, I know you have some thoughts about that.

JACOB HOFFMAN-ANDREWS
Yeah. To call back to, at the beginning you asked about, how do we define AI? I think one of the really interesting things in the field is that it's changed so much over time. And, you know, when computers first became broadly available, you know, people have been thinking for a very long time, what would it mean for a computer to be intelligent? And for a while we thought, wow, you know, if a computer could play chess and beat a human, we would say that's an intelligent computer.
Um, if a computer could recognize, uh, what's in an image, is this an image of a cat or a cow - that would be intelligence. And of course now they can, and we don't consider it intelligence anymore. And you know, now we might say if a computer could write a term paper, that's intelligence and I don't think we're there yet, but the development of chatbots does make a lot of people feel like we're closer to intelligence because you can have a back and forth and you can ask questions and receive answers.
And some of those answers will be confabulations and, but some percentage of the time they'll be right. And it starts to feel like something you're interacting with. And I think, rightly so, people are worried that this will destroy jobs for writers and for artists. And to an earlier question about, you know, what does it look like if we get it right, I think, you know, the future we want is one where people can write beautiful things and create beautiful things and, you know, still make a great living at it and be fulfilled and safe in their daily needs and be recognized for that. And I think that's one of the big challenges we're facing with generative AI.

JASON KELLEY
Let’s pause for just a moment to say thank you to our sponsor. How to Fix the Internet is supported by The Alfred P. Sloan Foundation’s Program in Public Understanding of Science and Technology. Enriching people’s lives through a keener appreciation of our increasingly technological world and portraying the complex humanity of scientists, engineers, and mathematicians. And now back to our discussion with Kit and Jacob about AI: the good, the bad, and what could be better.

CINDY COHN
There’s been a lot of focus on the dark side of generative AI and the idea of using copyright to address those problems has emerged. We have worries about that as a way to sort out between good and bad uses of AI, right Kit?

KIT WALSH
Absolutely. We have had a lot of experience with copyright being used as a tool of censorship, not only against individual journalists and artists and researchers, but also against entire mediums for expression, against libraries, against the existence of online platforms where people are able to connect and copyright not only lasts essentially forever, it comes with draconian penalties that are essentially a financial death sentence for the typical person in the United States. So in the context of generative AI, there is a real issue with the potential to displace creative labor. And it's a lot like the issues of other forms of automation that displace other forms of labor.
And it's not always the case that an equal number of new jobs are created, or that those new jobs are available to the people who have been displaced. And that's a pretty big social problem that we have. In Kitopia, we have AI and it's used so that there's less necessary labor to achieve a higher standard of living for people, and we should be able to be excited about automation of labor tasks that aren't intrinsically rewarding.
One of the reasons that we're not is because the fruits of that increased production flow to the people who own the AI, not to the people who were doing that labor, who now have to find another way to trade their labor for money or else become homeless and starve and die, and that's cruel.
It is the world that we're living in so it's really understandable to me that an artist is going to want to reach for copyright, which has the potential of big financial damages against someone who infringes, and is the way that we've thought about monetization of artistic works. I think that way of thinking about it is detrimental, but I also think it's really understandable.
One of the reasons why the particular legal theories in the lawsuits against generative AI technologies are concerning is because they wind up stretching existing doctrines of copyright law. So in particular, the very first case against Stable Diffusion argued that you were creating an infringing derivative work when you trained your model to recognize the patterns in five billion images.
It's a derivative work of each and every one of them. And that can only succeed as a legal theory if you throw out the existing understanding of what a derivative work is, that it has to be substantially similar to a thing that it's infringing and that limitation is incredibly important for human creativity.
The elements of my work that you might recognize from my artistic influences in the ordinary course of artistic borrowing and inspiration are protected. I'm able to make my art without people coming after me because I like to draw eyes the same way as my inspiration or so on, because ultimately the work is not substantially similar.
And if we got rid of that protection, it would be really bad for everybody.
But at the same time, you can see how someone might say, why should I pay a commission to an artist if I can get something in the same style? To which I would say, try it. It's not going to be what you want because art is not about replicating patterns that are found in a bunch of training data.
It can be a substitute for stock photography or other forms of art that are on the lower end of how much creativity is going into the expression, but for the higher end, I think that part of the market is safe. So I think all artists are potentially impacted by this. I'm not saying only bad artists have to care, but there is this real impact.
Their financial situation is precarious already, and they deserve to make a living, and this is a bandaid because we don't have a better solution in place to support people and let them create in a way that is in accord with their values and their goals. We really don't have that either in the situation where people are primarily making their income doing art that a corporation wants them to make to maximize its products.
No artist wants to create assets for content. Artists want to express and create new beauty and new meaning and the system that we have doesn't achieve that. We can certainly envision better ones but in the meantime, the best tool that artists have is banding together to negotiate with collective power, and it's really not a good enough tool at this point.
But I also think there's a lot of room to ethically use generative AI if you're working with an artist and you're trying to communicate your vision for something visual, maybe you're going to use an AI tool in order to make something that has some of the elements you're looking for and then say this, this is what I want to pay you to, to draw. I want this kind of pose, right? But, but, more unicorns.

JASON KELLEY
And I think while we're talking about these sort of seemingly good, but ultimately dangerous solutions for the different sort of problems that we're thinking about now more than ever because of generative AI, I wanted to talk with Jacob a little bit about watermarking. And this is meant to solve a sort of problem of knowing what is and is not generated by AI.
And people are very excited about this idea that through some sort of, well, actually you just explain Jacob, cause you are the technologist. What is watermarking? Is this a good idea? Will this work to help us understand and distinguish between AI-generated things and things that are just made by people?

JACOB HOFFMAN-ANDREWS
Sure. So a very real and closely related risk of generative AI is that it is - it will, and already is - flooding the internet with bullshit. Uh, you know, many of the articles you might read on any given topic, these days the ones that are most findable are often generated by AI.
And so an obvious next step is, well, what if we could recognize the stuff that's written by AI or the images that are generated by AI, because then we could just skip that. You know, I wouldn't read this article cause I know it's written by AI or you can go even a step further, you could say, well, maybe search engines should downrank things that were written by AI or social networks should label it or allow you to opt out of it.
You know, there's a lot of question about, if we could immediately recognize all the AI stuff, what would we do about it? There's a lot of options, but the first question is, can we even recognize it? So right off the bat, you know, when ChatGPT became available to the public, there were people offering ChatGPT detectors. You know, you could look at this content and, you know, you can kind of say, oh, it tends to look like this.
And you can try to write something that detects its output, and the short answer is it doesn't work and it's actually pretty harmful. A number of students have been harmed because their instructors have run their work through a ChatGPT detector, an AI detector that has incorrectly labeled it.
There's not a reliable pattern in the output that you can always see. Well, what if the makers of the AI put that pattern there? And, you know, for a minute, let's switch from text based to image based stuff. Jason, have you ever gone to a stock photo site to download a picture of something?

JASON KELLEY
I sadly have.

JACOB HOFFMAN-ANDREWS
Yeah. So you might recognize the images they have there, they want to make sure you pay for the image before they use it. So there's some text written across it in a kind of ghostly white diagonal. It says, this is from say shutterstock.com. So that's a form of watermark. If you just went and downloaded that image rather than paying for the cleaned up version, there's a watermark on it.
So the concept of watermarking for AI provenance is that It would be invisible. It would be kind of mixed into the pixels at such a subtle level that you as a human can't detect it, but you know, a computer program designed to detect that watermark could so you could imagine the AI might generate a picture and then in the top left pixel, increase its shade by the smallest amount, and then the next one, decrease it by the smallest amount and so on throughout the whole image.
And you can encode a decent amount of data that way, like what system produced it, when, all that information. And actually the EFF has published some interesting research in the past on a similar system in laser printers where little yellow dots are embedded by certain laser printers, by most laser printers that you can get as an anti counterfeiting measure.

JASON KELLEY
This is one of our most popular discoveries that comes back every few years, if I remember right, because people are just gobsmacked that they can't see them, but they're there, and that they have this information. It's a really good example of how this works.

CINDY COHN
Yeah, and it's used to make sure that they can trace back to the printer that printed anything on the off chance that what you're printing is fake money.

JACOB HOFFMAN-ANDREWS
Indeed, yeah.
The other thing people really worry about is that AI will make it a lot easier to generate disinformation and then spread it and of course if you're generating disinformation it's useful to strip out the watermark. You would maybe prefer that people don't know it's AI. And so you're not limited to resizing or cropping an image. You can actually, you know, run it through a program. You can see what the shades of all the different pixels are. And you, in theory probably know what the watermarking system in use is. And given that degree of flexibility, it seems very, very likely - and I think past technology has proven this out - that it's not going to be hard to strip out the watermark. And in fact, it's not even going to be hard to develop a program to automatically strip out the watermark.

CINDY COHN
Yep. And you, you end up in a cat and mouse game where the people who you most want to catch, who are doing sophisticated disinformation, say to try to upset elections, are going to be able to either strip out the watermark or fake it and so you end up where the things that you most want to identify are probably going to trick people. Is that, is that the way you're thinking about it?

JACOB HOFFMAN-ANDREWS
Yeah, that's pretty much what I'm getting at. I wanted to say one more thing on, um, watermarking. I'd like to talk about chainsaw dogs. There's this popular genre of image on Facebook right now of a man and his chainsaw carved wooden dog and, often accompanied by a caption like, look how great my dad is, he carved this beautiful thing.
And these are mostly AI generated and they receive, you know, thousands of likes and clicks and go wildly viral. And you can imagine a weaker form of the disinformation claim of say, ‘Well, okay, maybe state actors will strip out watermarks so they can conduct their disinformation campaigns, but at least adding watermarks to AI images will prevent this proliferation of garbage on the internet.’
People will be able to see, oh, that's a fake. I'm not going to click on it. And I think the problem with that is even people who are just surfing for likes on social media actually love to strip out credits from artists already. You know, cartoonists get their signatures stripped out and in the examples of these chainsaw dogs, you know, there is actually an original.
There's somebody who made a real carving of a dog. It was very skillfully executed. And these are generated using kind of image to image AI, where you take an image and you generate an image that has a lot of the same concepts. A guy, a dog, made of wood and so they're already trying to strip attribution in one way.
And I think likely they would also find a way to strip any watermarking on the images they're generating.

CINDY COHN
So Jacob, we heard earlier about Kit's ideal world. I'd love to hear about the future world that Jacob wants us to live in.

JACOB HOFFMAN-ANDREWS
Yeah. I think the key thing is, you know, that people are safer in their daily lives than they are today. They're not worried about their livelihoods going away. I think this is a recurring theme when most new technology is invented that, you know, if it replaces somebody's job, and that person's job doesn't get easier, they don't get to keep collecting a paycheck. They just lose their job.
So I think in the ideal future, people have a means to live and to be fulfilled in their lives to do meaningful work still. And also in general, human agency is expanded rather than restricted. The promise of a lot of technologies that, you know, you can do more in the world, you can achieve the conditions you want in your life.

CINDY COHN
Oh that sounds great. I want to come back to you Kit. We've talked a little about Kitopia, including at the top of the show. Let's talk a little bit more. What else are we missing?

KIT WALSH
So in Kitopia, people are able to use AI if it's a useful part of their artistic expression, they're able to use AI if they need to communicate something visual when I'm hiring a concept artist, when I am getting a corrective surgery, and I want to communicate to the surgeon what I want things to look like.
There are a lot of ways in which words don't communicate as well as images. And not everyone has the skill or the time or interest to go and learn a bunch of photoshop to communicate with their surgeon. I think it would be great if more people were interested and had the leisure and freedom to do visual art.
But in Kitopia, that's something that you have because your basic needs are met. And in part, automation is something that should help us do that more. The ability to automate aspects of, of labor should wind up benefiting everybody. That's the vision of AI in Kitopia.

CINDY COHN
Nice. Well that's a wonderful place to end. We're all gonna pack our bags and move to Kitopia. And hopefully by the time we get there, it’ll be waiting for us.
You know, Jason, that was such a rich conversation. I'm not sure we need to do a little recap like we usually do. Let's just close it out.

JASON KELLEY
Yeah, you know, that sounds good. I'll take it from here. Thanks for joining us for this episode of How to Fix the Internet. If you have feedback or suggestions, we would love to hear from you. You can visit EFF.org slash podcasts to click on listener feedback and let us know what you think of this or any other episode.
You can also get a transcript or information about this episode and the guests. And while you're there of course, you can become an EFF member, pick up some merch, or just see what's happening in digital rights this or any other week. This podcast is licensed Creative Commons Attribution 4. 0 International and includes music licensed Creative Commons Unported by their creators.
In this episode, you heard Kalte Ohren by Alex featuring starfrosch & Jerry Spoon; lost Track by Airtone; Come Inside by Zep Hume; Xena's Kiss/Medea's Kiss by MWIC; Homesick By Siobhan D and Drops of H2O ( The Filtered Water Treatment ) by J.Lang. Our theme music is by Nat Keefe of BeatMower with Reed Mathis. And How to Fix the Internet is supported by the Alfred P. Sloan Foundation's program in public understanding of science and technology. We’ll see you next time. I’m Jason Kelley.

CINDY COHN
And I’m Cindy Cohn.

 

Josh Richman

[B] 朝鮮人追悼式典 都知事選候補者にアンケートを実施

3 months 2 weeks ago
東京都知事選挙(選挙期日:7月7日。告示日:6月20日)が迫る中、「『関東大震災朝鮮人・中国人虐殺100年犠牲者追悼大会』実行委員会」(以下、追悼大会実行委)は、6月17日、衆議院議員会館で会見を行い、都知事選挙の各候補者に対して、都知事として朝鮮人犠牲者追悼式典への追悼文を送る意思があるかを問うアンケート調査を実施することを発表した。(小栗俊也)
日刊ベリタ

California’s Facial Recognition Bill Is Not the Solution We Need

3 months 2 weeks ago

California Assemblymember Phil Ting has introduced A.B. 1814, a bill that would supposedly regulate police use of facial recognition technology. The problem is that it would do little to actually change the status quo of how police use this invasive and problematic technology. Police use of facial recognition poses a massive risk to civil liberties, privacy, and even our physical health as the technology has been known to wrongfully sic armed police on innocent peopleparticularly Black men and women. That’s why this issue is too important to throw inadequate or half-measures like A.B. 1814 to try to fix it.

The bill dictates that police should examine facial recognition matches “with care” and that a match should not be the sole basis for the probable cause for an arrest or search warrant. And while we agree it is a big issue that police seem to repeatedly use the matches spit out by a computer as the only justification for arresting people, theoretically the limit this bill imposes is already the limit. Police departments and facial recognition companies alike both maintain that police cannot justify an arrest using only algorithmic matches–so what would this bill really change? It only gives the appearance of doing something to address face recognition technology's harms, while inadvertently allowing the practice to continue.

Additionally, A.B. 1814 gives defendants no real recourse against police who violate its requirements. There is neither a suppression remedy nor a usable private cause of action. The bill lacks transparency requirements which would compel police departments to reveal if they used face recognition in the first place. This means if police did arrest someone wrongfully because a computer said they looked similar to the subject, someone would likely not even know they could sue the department over damages, unless they uncovered it while being prosecuted. 

Despite these attempts at creating leaky bureaucratic reforms, police may continue to use this technology to identify people at protests, track marginalized individuals when they visit doctors or have other personal encounters, as well as any other number of civil liberties-chilling uses police might overtly or inadvertently deploy. It is this reason that EFF continues to advocate for a complete ban on government use of face recognition–an approach that has also resulted in cities across the United States standing up for themselves and enacting bans. Until the day comes that California lawmakers realize the urgent need to ban government use of face recognition, we will continue to differentiate between bills that will make a serious difference in the lives of the surveilled, and those that do not. That is why we are urging Assemblymembers to vote no on A.B. 1814. 

Matthew Guariglia