How “personhood credentials” could help prove you’re a human online

As AI models become better at mimicking human behavior, it’s becoming increasingly difficult to distinguish between real human internet users and sophisticated systems imitating them. 

That’s a real problem when those systems are deployed for nefarious ends like spreading misinformation or conducting fraud, and it makes it a lot harder to trust what you encounter online.

A group of 32 researchers from institutions including OpenAI, Microsoft, MIT, and Harvard has developed a potential solution—a verification concept called “personhood credentials.” These credentials prove that their holder is a real person, without revealing any further information about the person’s identity. The team explored the idea in a non-peer-reviewed paper posted to the arXiv preprint server earlier this month.

Personhood credentials rely on the fact that AI systems still cannot bypass state-of-the-art cryptographic systems or pass as people in the offline, real world. 

To request such credentials, people would have to physically go to one of a number of issuers, like a government or some other kind of trusted organization. They would be asked to provide evidence of being a real human, such as a passport or biometric data. Once approved, they’d receive a single credential to store on their devices the way it’s currently possible to store credit and debit cards in smartphones’ wallet apps.

To use these credentials online, a user could present them to a third-party digital service provider, which could then verify them using a cryptographic protocol called a zero-knowledge proof. That would confirm the holder was in possession of a personhood credential without disclosing any further unnecessary information.

The ability to filter out anyone other than verified humans on a platform could be useful in many ways. People could reject Tinder matches that don’t come with personhood credentials, for example, or choose not to see anything on social media that wasn’t definitely posted by a person. 

The authors want to encourage governments, companies, and standards bodies to consider adopting such a system in the future to prevent AI deception from ballooning out of our control. 

“AI is everywhere. There will be many issues, many problems, and many solutions,” says Tobin South, a PhD student at MIT who worked on the project. “Our goal is not to prescribe this to the world, but to open the conversation about why we need this and how it could be done.”

Possible technical options already exist. For example, a network called Idena claims to be the first blockchain proof-of-person system. It works by getting humans to solve puzzles that would be difficult for bots within a short time frame. The controversial Worldcoin program, which collects users’ biometric data, bills itself as the world’s largest privacy-preserving human identity and financial network. It recently partnered with the Malaysian government to provide proof of humanness online by scanning users’ irises, which creates a code. As in the concept of personhood credentials, each code is protected using cryptography.

However, the project has been criticized for using deceptive marketing practices, collecting more personal data than acknowledged, and failing to obtain meaningful consent from users. Regulators in Hong Kong and Spain banned Worldcoin from operating earlier this year, while its operations have been suspended in countries including Brazil, Kenya, and India. 

So fresh concepts are still needed. The rapid rise of accessible AI tools has ushered in a dangerous period in which internet users are hyper-suspicious about what is and isn’t true online, says Henry Ajder, an expert on AI and deepfakes who is an advisor to Meta and the UK government. And while ideas for verifying personhood have been around for some time, these credentials feel like one of the most substantive ideas for how to push back against encroaching skepticism, he says.

But the biggest challenge the credentials will face is getting enough platforms, digital services, and governments to adopt them, since they may feel uncomfortable conforming to a standard they don’t control. “For this to work effectively, it would have to be something which is universally adopted,” he says. “In principle the technology is quite compelling, but in practice and the messy world of humans and institutions, I think there would be quite a lot of resistance.”

Martin Tschammer, head of security at the startup Synthesia, which creates AI-generated hyperrealistic deepfakes, says he agrees with the principle driving personhood credentials: the need to verify humans online. However, he is unsure whether it’s the right solution or whether it would be practical to implement. He also expresses skepticism over who would run such a scheme.  

“We may end up in a world in which we centralize even more power and concentrate decision-making over our digital lives, giving large internet platforms even more ownership over who can exist online and for what purpose,” he says. “And given the lackluster performance of some governments in adopting digital services, and autocratic tendencies that are on the rise, is it practical or realistic to expect this type of technology to be adopted en masse and in a responsible way by the end of this decade?” 

Rather than waiting for collaboration across industries, Synthesia is currently evaluating how to integrate other personhood-proving mechanisms into its products. He says it already has several measures in place. For example, it requires businesses to prove that they are legitimate registered companies, and will ban and refuse refunds to customers found to have broken its rules. 

One thing is clear: We are in urgent need of ways to differentiate humans from bots, and encouraging discussions between stakeholders in the tech and policy worlds is a step in the right direction, says Emilio Ferrara, a professor of computer science at the University of Southern California, who was not involved in the project. 

“We’re not far from a future where, if things remain unchecked, we’re going to be essentially unable to tell apart interactions that we have online with other humans or some kind of bots. Something has to be done,” he says. “We can’t be naïve as previous generations were with technologies.”

AI’s impact on elections is being overblown

This year, close to half the world’s population has the opportunity to participate in an election. And according to a steady stream of pundits, institutions, academics, and news organizations, there’s a major new threat to the integrity of those elections: artificial intelligence. 

The earliest predictions warned that a new AI-powered world was, apparently, propelling us toward a “tech-enabled Armageddon” where “elections get screwed up”, and that “anybody who’s not worried [was] not paying attention.” The internet is full of doom-laden stories proclaiming that AI-generated deepfakes will mislead and influence voters, as well as enabling new forms of personalized and targeted political advertising. Though such claims are concerning, it is critical to look at the evidence. With a substantial number of this year’s elections concluded, it is a good time to ask how accurate these assessments have been so far. The preliminary answer seems to be not very; early alarmist claims about AI and elections appear to have been blown out of proportion.

While there will be more elections this year where AI could have an effect, the United States being one likely to attract particular attention, the trend observed thus far is unlikely to change. AI is being used to try to influence electoral processes, but these efforts have not been fruitful. Commenting on the upcoming US election, Meta’s latest Adversarial Threat Report acknowledged that AI was being used to meddle—for example, by Russia-based operations—but that “GenAI-powered tactics provide only incremental productivity and content-generation gains” to such “threat actors.” This echoes comments from the company’s president of global affairs, Nick Clegg, who earlier this year stated that “it is striking how little these tools have been used on a systematic basis to really try to subvert and disrupt the elections.”

Far from being dominated by AI-enabled catastrophes, this election “super year” at that point was pretty much like every other election year.

While Meta has a vested interest in minimizing AI’s alleged impact on elections, it is not alone. Similar findings were also reported by the UK’s respected Alan Turing Institute in May. Researchers there studied more than 100 national elections held since 2023 and found “just 19 were identified to show AI interference.” Furthermore, the evidence did not demonstrate any “clear signs of significant changes in election results compared to the expected performance of political candidates from polling data.”

This all raises a question: Why were these initial speculations about AI-enabled electoral interference so off, and what does it tell us about the future of our democracies? The short answer: Because they ignored decades of research on the limited influence of mass persuasion campaigns, the complex determinants of voting behaviors, and the indirect and human-mediated causal role of technology. 

First, mass persuasion is notoriously challenging. AI tools may facilitate persuasion, but other factors are critical. When presented with new information, people generally update their beliefs accordingly; yet even in the best conditions, such updating is often minimal and rarely translates into behavioral change. Though political parties and other groups invest colossal sums to influence voters, evidence suggests that most forms of political persuasion have very small effects at best. And in most high-stakes events, such as national elections, a multitude of factors are at play, diminishing the effect of any single persuasion attempt.

Second, for a piece of content to be influential, it must first reach its intended audience. But today, a tsunami of information is published daily by individuals, political campaigns, news organizations, and others. Consequently, AI-generated material, like any other content, faces significant challenges in cutting through the noise and reaching its target audience. Some political strategists in the United States have also argued that the overuse of AI-generated content might make people simply tune out, further reducing the reach of manipulative AI content. Even if a piece of such content does reach a significant number of potential voters, it will probably not succeed in influencing enough of them to alter election results.

Third, emerging research challenges the idea that using AI to microtarget people and sway their voting behavior works as well as initially feared. Voters seem to not only recognize excessively tailored messages but actively dislike them. According to some recent studies, the persuasive effects of AI are also, at least for now, vastly overstated. This is likely to remain the case, as ever-larger AI-based systems do not automatically translate to better persuasion. Political campaigns seem to have recognized this too. If you speak to campaign professionals, they will readily admit that they are using AI, but mainly to optimize “mundane” tasks such as fundraising, get-out-the-vote efforts, and overall campaign operations rather than generating new AI-generated, highly tailored content.

Fourth, voting behavior is shaped by a complex nexus of factors. These include gender, age, class, values, identities, and socialization. Information, regardless of its veracity or origin—whether made by an AI or a human—often plays a secondary role in this process. This is because the consumption and acceptance of information are contingent on preexisting factors, like whether it chimes with the person’s political leanings or values, rather than whether that piece of content happens to be generated by AI.

Concerns about AI and democracy, and particularly elections, are warranted. The use of AI can perpetuate and amplify existing social inequalities or reduce the diversity of perspectives individuals are exposed to. The harassment and abuse of female politicians with the help of AI is deplorable. And the perception, partially co-created by media coverage, that AI has significant effects could itself be enough to diminish trust in democratic processes and sources of reliable information, and weaken the acceptance of election results. None of this is good for democracy and elections. 

However, these points should not make us lose sight of threats to democracy and elections that have nothing to do with technology: mass voter disenfranchisement; intimidation of election officials, candidates, and voters; attacks on journalists and politicians; the hollowing out of checks and balances; politicians peddling falsehoods; and various forms of state oppression (including restrictions on freedom of speech, press freedom and the right to protest). 

Of at least 73 countries holding elections this year, only 47 are classified as full (or at least flawed) democracies, according to Our World in Data/Economist Democracy Index, with the rest being hybrid or authoritarian regimes. In countries where elections are not even free or fair, and where political choice that leads to real change is an illusion, people have arguably bigger fish to fry.

And still, technology—including AI—often becomes a convenient scapegoat, singled out by politicians and public intellectuals as one of the major ills befalling democratic life. Earlier this year, Swiss president Viola Amherd warned at the World Economic Forum in Davos, Switzerland, that “advances in artificial intelligence allow … false information to seem ever more credible” and present a threat to trust. Pope Francis, too, warned that fake news could be legitimized through AI. US Deputy Attorney General Lisa Monaco said that AI could supercharge mis- and disinformation and incite violence at elections. This August, the mayor of London, Sadiq Kahn, called for a review of the UK’s Online Safety Act after far-right riots across the country, arguing that “the way the algorithms work, the way that misinformation can spread very quickly and disinformation … that’s a cause to be concerned. We’ve seen a direct consequence of this.”

The motivations to blame technology are plenty and not necessarily irrational. For some politicians, it can be easier to point fingers at AI than to face scrutiny or commit to improving democratic institutions that could hold them accountable. For others, attempting to “fix the technology” can seem more appealing than addressing some of the fundamental issues that threaten democratic life. Wanting to speak to the zeitgeist might play a role, too.

Yet we should remember that there’s a cost to overreaction based on ill-founded assumptions, especially when other critical issues go unaddressed. Overly alarmist narratives about AI’s presumed effects on democracy risk fueling distrust and sowing confusion among the public—potentially further eroding already low levels of trust in reliable news and institutions in many countries. One point often raised in the context of these discussions is the need for facts. People argue that we cannot have democracy without facts and a shared reality. That is true. But we cannot bang on about needing a discussion rooted in facts when evidence against the narrative of AI turbocharging democratic and electoral doom is all too easily dismissed. Democracy is under threat, but our obsession with AI’s supposed impact is unlikely to make things better—and could even make them worse when it leads us to focus solely on the shiny new thing while distracting us from the more lasting problems that imperil democracies around the world. 

Felix M. Simon is a research fellow in AI and News at the Reuters Institute for the Study of Journalism; Keegan McBride is an assistant professor in AI, government, and policy at the Oxford Internet Institute; Sacha Altay is a research fellow in the department of political science at the University of Zurich.

What this futuristic Olympics video says about the state of generative AI

The Olympic Games in Paris just finished last month and the Paralympics are still underway, so the 2028 Summer Olympics in Los Angeles feel like a lifetime from now. But the prospect of watching the games in his home city has Josh Kahn, a filmmaker in the sports entertainment world who has worked in content creation for both LeBron James and the Chicago Bulls, thinking even further into the future: What might an LA Olympics in the year 3028 look like?

It’s the perfect type of creative exercise for AI video generation, which came into the mainstream with the debut of OpenAI’s Sora earlier this year. By typing prompts into generators like Runway or Synthesia, users can generate fairly high-definition video in minutes. It’s fast and cheap, and it presents few technical obstacles compared with traditional creation techniques like CGI or animation. Even if every frame isn’t perfect—distortions like hands with six fingers or objects that disappear are common—there are, at least in theory, a host of commercial applications. Ad agencies, companies, and content creators could use the technology to create videos quickly and cheaply.  

Kahn, who has been toying with AI video tools for some time, used the latest version of Runway to dream up what the Olympics of the future could look like, entering a new prompt in the model for each shot. The video is just over one minute long and features sweeping aerial views of a futuristic version of LA where sea levels have risen sharply, leaving the city crammed right up to the coastline. A football stadium sits perched on top of a skyscraper, while a dome in the middle of the harbor contains courts for beach volleyball. 

The video, which was shared exclusively with MIT Technology Review, is meant less as a road map for the city and more as a demonstration of what’s possible now with AI.

“We were watching the Olympics and the amount of care that goes into the cultural storytelling of the host city,” Kahn says. “There’s a culture of imagination and storytelling in Los Angeles that has kind of set the tone for the rest of the world. Wouldn’t it be cool if we could showcase what the Olympics would look like if they returned to LA 1,000 years from now?”

More than anything, the video shows what a boon the generative technology may be for creators. However, it also indicates what’s holding it back. Though Kahn declined to share his prompts for the shots or specify how many prompts it took to get each take right, he did caution that anyone wishing to create good content with AI must be comfortable with trial and error. Particularly challenging in his futuristic project was getting the AI model to think outside the box in terms of architecture. A stadium hovering above water, for example, is not something most AI models have seen many examples of in their training data. 

With each shot requiring a new set of prompts, it’s also hard to instill a sense of continuity throughout a video. The color, angle of the sun, and shapes of buildings are difficult for a video generation model to keep consistent. The video also lacks any close-ups of people, which Kahn says AI models still tend to struggle with. 

“These technologies are always better on large-scale things right now as opposed to really nuanced human interaction,” he says. For this reason, Kahn imagines that early filmmaking applications of generative video might be for wide shots of landscapes or crowds. 

Alex Mashrabov, an AI video expert who left his role as director of generative AI at Snap last year to found a new AI video company called Higgsfield AI, agrees on the current failures and flaws of AI video. He also points out that good dialogue-heavy content is hard to produce with AI, as it tends to hinge upon subtle facial expressions and body language. 

Some content creators may be reluctant to adopt generative video simply because of the amount of time required to prompt the models again and again to get the end result right.

“Typically, the success rate is one out of 20,” Mashrabov says, but it’s not uncommon to need 50 or 100 attempts. 

For many purposes, though, that’s good enough. Mashrabov says he’s seen an uptick in AI-generated video advertisements from massive suppliers like Temu. In goods-producing countries like China, video generators are in high demand to quickly make in-your-face video ads for particular products. Even if an AI model might require lots of prompts to yield a usable ad, filming it with real people, cameras, and equipment might be 100 times more expensive. Applications like this might be the first use of generative video at scale as the technology slowly improves, he says. 

“Although I think this is a very long path, I’m very confident there are low-hanging fruits,” Mashrabov says. “We’re figuring out the genres where generative AI is already good today.”

A new way to build neural networks could make AI more understandable

A tweak to the way artificial neurons work in neural networks could make AIs easier to decipher.

Artificial neurons—the fundamental building blocks of deep neural networks—have survived almost unchanged for decades. While these networks give modern artificial intelligence its power, they are also inscrutable. 

Existing artificial neurons, used in large language models like GPT4, work by taking in a large number of inputs, adding them together, and converting the sum into an output using another mathematical operation inside the neuron. Combinations of such neurons make up neural networks, and their combined workings can be difficult to decode.

But the new way to combine neurons works a little differently. Some of the complexity of the existing neurons is both simplified and moved outside the neurons. Inside, the new neurons simply sum up their inputs and produce an output, without the need for the extra hidden operation. Networks of such neurons are called Kolmogorov-Arnold Networks (KANs), after the Russian mathematicians who inspired them.

The simplification, studied in detail by a group led by researchers at MIT, could make it easier to understand why neural networks produce certain outputs, help verify their decisions, and even probe for bias. Preliminary evidence also suggests that as KANs are made bigger, their accuracy increases faster than networks built of traditional neurons.

“It’s interesting work,” says Andrew Wilson, who studies the foundations of machine learning at New York University. “It’s nice that people are trying to fundamentally rethink the design of these [networks].”

The basic elements of KANs were actually proposed in the 1990s, and researchers kept building simple versions of such networks. But the MIT-led team has taken the idea further, showing how to build and train bigger KANs, performing empirical tests on them, and analyzing some KANs to demonstrate how their problem-solving ability could be interpreted by humans. “We revitalized this idea,” said team member Ziming Liu, a PhD student in Max Tegmark’s lab at MIT. “And, hopefully, with the interpretability… we [may] no longer [have to] think neural networks are black boxes.”

While it’s still early days, the team’s work on KANs is attracting attention. GitHub pages have sprung up that show how to use KANs for myriad applications, such as image recognition and solving fluid dynamics problems. 

Finding the formula

The current advance came when Liu and colleagues at MIT, Caltech, and other institutes were trying to understand the inner workings of standard artificial neural networks. 

Today, almost all types of AI, including those used to build large language models and image recognition systems, include sub-networks known as a multilayer perceptron (MLP). In an MLP, artificial neurons are arranged in dense, interconnected “layers.” Each neuron has within it something called an “activation function”—a mathematical operation that takes in a bunch of inputs and transforms them in some pre-specified manner into an output. 

In an MLP, each artificial neuron receives inputs from all the neurons in the previous layer and multiplies each input with a corresponding “weight” (a number signifying the importance of that input). These weighted inputs are added together and fed to the activation function inside the neuron to generate an output, which is then passed on to neurons in the next layer. An MLP learns to distinguish between images of cats and dogs, for example, by choosing the correct values for the weights of the inputs for all the neurons. Crucially, the activation function is fixed and doesn’t change during training.

Once trained, all the neurons of an MLP and their connections taken together essentially act as another function that takes an input (say, tens of thousands of pixels in an image) and produces the desired output (say, 0 for cat and 1 for dog). Understanding what that function looks like, meaning its mathematical form, is an important part of being able to understand why it produces some output. For example, why does it tag someone as creditworthy given inputs about their financial status? But MLPs are black boxes. Reverse-engineering the network is nearly impossible for complex tasks such as image recognition.

And even when Liu and colleagues tried to reverse-engineer an MLP for simpler tasks that involved bespoke “synthetic” data, they struggled. 

“If we cannot even interpret these synthetic datasets from neural networks, then it’s hopeless to deal with real-world data sets,” says Liu. “We found it really hard to try to understand these neural networks. We wanted to change the architecture.”

Mapping the math

The main change was to remove the fixed activation function and introduce a much simpler learnable function to transform each incoming input before it enters the neuron. 

Unlike the activation function in an MLP neuron, which takes in numerous inputs, each simple function outside the KAN neuron takes in one number and spits out another number. Now, during training, instead of learning the individual weights, as happens in an MLP, the KAN just learns how to represent each simple function. In a paper posted this year on the preprint server ArXiv, Liu and colleagues showed that these simple functions outside the neurons are much easier to interpret, making it possible to reconstruct the mathematical form of the function being learned by the entire KAN.

The team, however, has only tested the interpretability of KANs on simple, synthetic data sets, not on real-world problems, such as image recognition, which are more complicated. “[We are] slowly pushing the boundary,” says Liu. “Interpretability can be a very challenging task.”

Liu and colleagues have also shown that KANs get more accurate at their tasks with increasing size faster than MLPs do. The team proved the result theoretically and showed it empirically for science-related tasks (such as learning to approximate functions relevant to physics). “It’s still unclear whether this observation will extend to standard machine learning tasks, but at least for science-related tasks, it seems promising,” Liu says.

Liu acknowledges that KANs come with one important downside: it takes more time and compute power to train a KAN, compared to an MLP.

“This limits the application efficiency of KANs on large-scale data sets and complex tasks,” says Di Zhang, of Xi’an Jiaotong-Liverpool University in Suzhou, China. But he suggests that more efficient algorithms and hardware accelerators could help.

Anil Ananthaswamy is a science journalist and author who writes about physics, computational neuroscience, and machine learning. His new book, WHY MACHINES LEARN: The Elegant Math Behind Modern AI, was published by Dutton (Penguin Random House US) in July.

How machine learning is helping us probe the secret names of animals

Do animals have names? According to the poet T.S. Eliot, cats have three: the name their owner calls them (like George); a second, more noble one (like Quaxo or Cricopat); and, finally, a “deep and inscrutable” name known only to themselves “that no human research can discover.”

But now, researchers armed with audio recorders and pattern-recognition software are making unexpected discoveries about the secrets of animal names—at least with small monkeys called marmosets.  

That’s according to a team at Hebrew University in Israel, who claim in the journal Science this week they’ve discovered that marmosets “vocally label” their monkey friends with specific sounds.

Until now, only humans, dolphins, elephants, and probably parrots had been known to use specific sounds to call out to other individuals.

Marmosets are highly social creatures that maintain contact through high-pitched chirps and twitters called “phee-calls.” By recording different pairs of monkeys placed near each other, the team in Israel says they found the animals will adjust their sounds toward a vocal label that’s specific to their conversation partner. 

“It’s similar to names in humans,” says David Omer, the neuroscientist who led the project. “There’s a typical time structure to their calls, and what we report is that the monkey fine-tunes it to encode an individual.”

These names aren’t really recognizable to the human ear; instead, they were identified via a “random forest,” the statistical machine learning technique Omer’s team used to cluster, classify, and analyze the sounds.

To prove they’d cracked the monkey code—and learned the secret names—the team played recordings at the marmosets through a speaker and found they responded more often when their label, or name, was in the recording.

This sort of research could provide clues to the origins of human language, which is arguably the most powerful innovation in our species’ evolution, right up there with opposable thumbs. In years past, it’s been argued that human language is unique and that animals lack both the brains and vocal apparatus to converse.

But there’s growing evidence that isn’t the case, especially now that the use of names has been found in at least four distantly related species. “This is very strong evidence that the evolution of language was not a singular event,” says Omer.

Some similar research tactics were reported earlier this year by Mickey Pardo, a postdoctoral researcher, now at Cornell University, who spent 14 months in Kenya recording elephant calls. Elephants sound alarms by trumpeting, but in reality most of their vocalizations are deep rumbles that are only partly audible to humans.

Pardo also found evidence that elephants use vocal labels, and he says he can definitely get an elephant’s attention by playing the sound of another elephant addressing it. But does this mean researchers are now “speaking animal”? 

Not quite, says Pardo. Real language, he thinks, would mean the ability to discuss things that happened in the past or string together more complex ideas. Pardo says he’s hoping to determine next if elephants have specific sounds for deciding which watering hole to visit—that is, whether they employ place names.

Several efforts are underway to discover if there’s still more meaning in animal sounds than we thought. This year, a group called Project CETI that’s studying the songs of sperm whales found they are far more complex than previously recognized. It means the animals, in theory, could be using a kind of grammar—although whether they actually are saying anything specific isn’t known.

Another effort, the Earth Species Project, aims to use “artificial intelligence to decode nonhuman communication” and has started helping researchers collect more data on animal sounds to feed into those models. 

The team in Israel say they will also be giving the latest types of artificial intelligence a try. Their marmosets live in a laboratory facility, and Omer says he’s already put microphones in monkeys’ living space in order to record everything they say, 24 hours a day.

Their chatter, Omer says, will be used to train a large language model that could, in theory, be used to finish a series of calls that a monkey started, or produce what it predicts is an appropriate reply. But will a primate language model actually make sense, or will it just gibber away without meaning? 

Only the monkeys will be able to say for sure.  

“I don’t have any delusional expectations that they will talk about Nietzsche,” says Omer. “I don’t expect it to be extremely complex like a human, but I would expect it to help us understand something about how our language developed.” 

We finally have a definition for open-source AI

Open-source AI is everywhere right now. The problem is, no one agrees on what it actually is. Now we may finally have an answer. The Open Source Initiative (OSI), the self-appointed arbiters of what it means to be open source, has released a new definition, which it hopes will help lawmakers develop regulations to protect consumers from AI risks. 

Though OSI has published much about what constitutes open-source technology in other fields, this marks its first attempt to define the term for AI models. It asked a 70-person group of researchers, lawyers, policymakers, and activists, as well as representatives from big tech companies like Meta, Google, and Amazon, to come up with the working definition. 

According to the group, an open-source AI system can be used for any purpose without securing permission, and researchers should be able to inspect its components and study how the system works.

It should also be possible to modify the system for any purpose—including to change its output—and to share it with others to use, with or without modifications, for any purpose. In addition, the standard attempts to define a level of transparency for a given model’s training data, source code, and weights. 

The previous lack of an open-source standard presented a problem. Although we know that the decisions of OpenAI and Anthropic to keep their models, data sets, and algorithms secret makes their AI closed source, some experts argue that Meta and Google’s freely accessible models, which are open to anyone to inspect and adapt, aren’t truly open source either, because of licenses that restrict what users can do with the models and because the training data sets aren’t made public. Meta, Google, and OpenAI have been contacted for their response to the new definition but did not reply before publication.

“Companies have been known to misuse the term when marketing their models,” says Avijit Ghosh, an applied policy researcher at Hugging Face, a platform for building and sharing AI models. Describing models as open source may cause them to be perceived as more trustworthy, even if researchers aren’t able to independently investigate whether they really are open source.

Ayah Bdeir, a senior advisor to Mozilla and a participant in OSI’s process, says certain parts of the open-source definition were relatively easy to agree upon, including the need to reveal model weights (the parameters that help determine how an AI model generates an output). Other parts of the deliberations were more contentious, particularly the question of how public training data should be.

The lack of transparency about where training data comes from has led to innumerable lawsuits against big AI companies, from makers of large language models like OpenAI to music generators like Suno, which do not disclose much about their training sets beyond saying they contain “publicly accessible information.” In response, some advocates say that open-source models should disclose all their training sets, a standard that Bdeir says would be difficult to enforce because of issues like copyright and data ownership. 

Ultimately, the new definition requires that open-source models provide information about the training data to the extent that “a skilled person can recreate a substantially equivalent system using the same or similar data.” It’s not a blanket requirement to share all training data sets, but it also goes further than what many proprietary models or even ostensibly open-source models do today. It’s a compromise.

“Insisting on an ideologically pristine kind of gold standard that actually will not effectively be met by anybody ends up backfiring,” Bdeir says. She adds that OSI is planning some sort of enforcement mechanism, which will flag models that are described as open source but do not meet its definition. It also plans to release a list of AI models that do meet the new definition. Though none are confirmed, the handful of models that Bdeir told MIT Technology Review are expected to land on the list are relatively small names, including Pythia by Eleuther, OLMo by Ai2, and models by the open-source collective LLM360.

A new system lets robots sense human touch without artificial skin

Even the most capable robots aren’t great at sensing human touch; you typically need a computer science degree or at least a tablet to interact with them effectively. That may change, thanks to robots that can now sense and interpret touch without being covered in high-tech artificial skin. It’s a significant step toward robots that can interact more intuitively with humans. 

To understand the new approach, led by the German Aerospace Center and published today in Science Robotics, consider the two distinct ways our own bodies sense touch. If you hold your left palm facing up and press lightly on your left pinky finger, you may first recognize that touch through the skin of your fingertip. That makes sense–you have thousands of receptors on your hands and fingers alone. Roboticists often try to replicate that blanket of sensors for robots through artificial skins, but these can be expensive and ineffective at withstanding impacts or harsh environments.

But if you press harder, you may notice a second way of sensing the touch: through your knuckles and other joints. That sensation–a feeling of torque, to use the robotics jargon–is exactly what the researchers have re-created in their new system.

Their robotic arm contains six sensors, each of which can register even incredibly small amounts of pressure against any section of the device. After precisely measuring the amount and angle of that force, a series of algorithms can then map where a person is touching the robot and analyze what exactly they’re trying to communicate. For example, a person could draw letters or numbers anywhere on the robotic arm’s surface with a finger, and the robot could interpret directions from those movements. Any part of the robot could also be used as a virtual button.

It means that every square inch of the robot essentially becomes a touch screen, except without the cost, fragility, and wiring of one, says Maged Iskandar, researcher at the German Aerospace Center and lead author of the study. 

“Human-robot interaction, where a human can closely interact with and command a robot, is still not optimal, because the human needs an input device,” Iskandar says. “If you can use the robot itself as a device, the interactions will be more fluid.”

A system like this could provide a cheaper and simpler way of providing not only a sense of touch, but also a new way to communicate with robots. That could be particularly significant for larger robots, like humanoids, which continue to receive billions in venture capital investment. 

Calogero Maria Oddo, a roboticist who leads the Neuro-Robotic Touch Laboratory at the BioRobotics Institute but was not involved in the work, says the development is significant, thanks to the way the research combines sensors, elegant use of mathematics to map out touch, and new AI methods to put it all together. Oddo says commercial adoption could be fairly quick, since the investment required is more in software than hardware, which is far more expensive.

There are caveats, though. For one, the new model cannot handle more than two points of contact at once. In a fairly controlled setting like a factory floor that might not be an issue, but in environments where human-robot interactions are less predictable, it could present limitations. And the sorts of sensors needed to communicate touch to a robot, though commercially available, can also cost tens of thousands of dollars.

Overall, though, Oddo envisions a future where skin-based sensors and joint-based ones are merged to give robots a more comprehensive sense of touch.

“We humans and other animals have integrated both solutions,” he says. “I expect robots working in the real world will use both, too, to interact safely and smoothly with the world and learn.”