Archive for the ‘Ai’ Category

The UK AI Summit: Time to Elevate Democratic Values – Council on Foreign Relations

Not long ago, the United States and the United Kingdom were leading the effort to establish global norms for the governance of artificial intelligence (AI). Both nations backed the Organization for Economic Cooperation and Development (OECD) AI Principles of 2019, the first global AI policy framework, and the Global Partnership on AI that followed. But efforts slowed as the European Union took the lead on regulatory efforts with the EU Artificial Intelligence Act, now heading toward the finish line as final negotiations on the Act wrap up later this year.

Now Prime Minister Sunak is hosting a Global Summit on AI Safety from November 1 to 2, following meetings with tech leaders in the UK and the meeting with President Biden in Washington, D.C. Speaking with reporters after the event, Sunak described the United States and the UK as the worlds foremost AI democratic powers. He emphasized the shared values of freedom, democracy, and rule of law.

More on:

Robots and Artificial Intelligence

Technology and Innovation

Global Governance

The UK AI summit now provides an opportunity for the United States and the UK to align on policy and move beyond the techno-libertarianism that characterized the early days of AI policymaking in both countries and begin to develop solutions to the challenges of AI, but there are challenges ahead. At the Summit and in future discussions on the role of AI, the UK should work to include civil society, integrate an AI fairness agenda into talks, and ensure that human rights and democratic values are central to any proposed international regulation.

First, the Global AI Summit must be inclusive. Prime Minister Sunak is already under criticism for a preliminary announcement that included statements from only tech CEOs and a plan that appears to sideline academics and civil society. While it is true that the tech CEOs had a White House meeting with the President, the Biden administration also quickly reached out to civil society organizations and labor leaders for input and advice on AI. Senate Majority Leader Chuck Schumer (D-NY) has already held an inaugural AI Insight Forumto gather expert input, albeit behind closed doors, on his proposed SAFE Innovation Act, including the perspectives of labor leaders and civil rights leaders, as well as the insights of practitioners and researchers focused on bias mitigation Prime Minister Sunak would do well to follow the American lead on civil society participation and ensure that the AI Safety Forum fairly reflects those impacted by AI systems, including marginalized communities.

Second, the AI safety agenda should not ignore the AI fairness agenda. Prime Minister Sunak is right to underscore the need for an international framework to ensure the safe and reliable development of AI. President Biden has also said that companies should not deploy AI systems that are not safe. Mitigating risk is a top priority, but so too is ensuring that AI systems treat people fairly, that systems are accountable, that adverse decisions are contestable, and that transparency is meaningful. In the rush to address existential risk there is the danger that the existing impact of AI on decisions in housing, credit, employment, education, and criminal justice will be ignored. The Prime Minister can address these concerns by including such topics as algorithmic bias, equity, and accountability in the meeting agenda.

Third, human rights and democratic values should remain key pillars of the UK AI Summit. There are many AI policy challenges ahead and several of the solutions do not favor democratic outcomes. For example, countries emphasizing safety and security are also establishing ID requirements for users of AI systems. And the desire to identify users and build massive new troves of personal data is not limited to governments. Several of the tech CEOs, including OpenAIs Sam Altman and former Google head Eric Schmidt, also favor identity requirements for AI users, even as they argue against regulation of their own AI services. Altman is CEO of a company that is seeking to establish a global biometric database based on eye scans. Requiring biometric data from users of AI while leaving AI systems unregulated is an outcome that democratic states should avoid.

Countries that value human dignity and autonomy should choose instead technical solutions that are less data intensive. The United States and the UK have already launched important work on Privacy Enhancing Technology that could minimize or eliminate the collection of personal data. That work should be encouraged and user identification requirements should be dropped. Strong data protection safeguards will help ensure that AI innovation does not undermine privacy.

More on:

Robots and Artificial Intelligence

Technology and Innovation

Global Governance

The UK government also needs to establish prohibitions and controls on AI systems that violate fundamental rights. The UK has already endorsed the United Nations Educational, Scientific and Cultural Organization (UNESCO) Recommendation on AI Ethics that proposes a ban on the use of AI for social scoring and mass surveillance. UK domestic law should implement these recommendations as well as other proposed limitations, including biometric categorization, predictive policing, and emotion detection. Drawing these red lines will be critical to ensure that AI systems are both human-centric and trustworthy, key goals set out in the OECD AI Principles and previously endorsed by the United States and the UK.

The renewed commitment to a regulatory framework for the governance of AI is welcome, especially as both the European Union and China pursue regulation for AI. As we have warned previously, the UKs light touch strategy for AI was unlikely to establish the necessary guardrails for the safe deployment of artificial intelligence. President Biden has already warned tech firms that they should not deploy AI systems that are not safe, and the U.S. Congress is now considering several bills, including the Blumenthal-Hawley U.S. AI Act, to govern AI services. Again, the UK would be wise to follow the United States lead and build in the necessary guardrails for AI products and services. For the worlds foremost AI democratic powers, the summit is an excellent moment to align national AI policies with democratic values.

Merve Hickok is President of the Center for AI and Digital Policy (CAIDP). Marc Rotenberg is Executive Director of CAIDP and a CFR Life Member. CAIDP publishes the AI and Democratic Values Index annually.

View post:

The UK AI Summit: Time to Elevate Democratic Values - Council on Foreign Relations

Higher ed AI anxiety? An advisory board could help – Inside Higher Ed

Experts say AI advisory boards could help universities looking to navigate the technology.

Photo illustration by Justin Morrison/Inside Higher Ed | Getty Images

When it comes to artificial intelligence and higher ed, the excitement and hype are matched by the uncertainties and need for guidance. One solution: creating an AI advisory board that brings together students, faculty and staff for open conversations about the new technology.

That was a key idea presented at the University of Central Floridas inaugural Teaching and Learning With AI conference, a two-day event that drew more than 500educators from around the country.

AI has had a breakout year, said Ray Schroeder, a senior fellow of the University Professional and Continuing Education Association (and a contributor to Inside Higher Ed). Schroeder, who has recently focused on the intersection of AI and higher education, opened the conference seeking to help faculty, administrators and staff attempt to navigate the choppy waters of AI.

Most Popular

We cannot afford to ignore it, he said. The intent is to make clear, What is the intention of the university? How are they going to move with this technology?

Schroeder and other experts interviewed said universities need a formal mechanism for getting advice on how to proceed.

Artificial intelligence is a technology that impacts nearly every aspect of higher education institutionsrecruiting, admissions, financial aid, student support services, teaching and learning, assessment, operation, and more, said Kristina Ishmael, deputy director of the Education Departments Office of Educational Technology.

Ishmael said in an email to Inside Higher Ed that the departments top recommendation about AI is to emphasize humans in the loop. Institutions that choose to create an AI advisory board, or a similar group, would be implementing this recommendation.

Many universities are already pursuing that advice. The University of Louisville had its first AI advisory board meeting last week. Stanford University and Vanderbilt University formed boards earlier this year after investing millions in AI research on campus.

Northeastern University created an external AI board, co-chaired by two faculty members and joined by industry heavyweights including Honeywell and the Mayo Clinic.

The University of Michigan unveiled its 18-member advisory board in May, a group tasked with creating a report centered on best practices for generative AI.

The Michigan board was the brainchild of Ravi Pendse, the universitys vice president for information technology, who chatted with fellow faculty members about AI at the start of the year.

I said, We need to get a faculty group together to provide general guidance to the campus, said Pendse, who also serves as the universitys chief information officer. We need to make sure we consider this technology, frankly, with our eyes wide-open and feet on the ground, so we embrace it, but do so thoughtfully.

There is no perfect blueprint to building an AI advisory board, and the approach will vary for each college or university. However, the experts interviewed noted important factors to make it work.

An AI advisory board may not work for each institution, and there are other approaches. Schroeder suggested a workshop for faculty could suffice. Pendse pointed toward having a lunch-and-learn series and said that the key, whatever the format, is to encourage discussion.

These discussions are already happening, Pendse said. We want to know what this thing is, to use it, to leverage it, to debate it. And the way you do that is providing safe spaces where this debate can happen and engage with each other.

For the institutions forming boards, however, it can be dual purpose, according to experts. Theres the value it delivers to students, which, Schroeder said, can help prepare them for the changing workforce.

For students, I think its a matter of tapping their expectations, he said. And I think their expectations are driven in part, perhaps in large part, by the expectations of employers.

Those students, once prepared, can boost discussion and ultimately help with AI research in the future, creating a flywheel effect.

We cant just sit idly in this country; other countries are investing, so we need to be flying, not running or walking, Pendse said. And the only way we can is with institutions contributing to the AI talentthat will create policy makers, people who can debate the pros and cons. Thats how we can compete in the world.

Go here to see the original:

Higher ed AI anxiety? An advisory board could help - Inside Higher Ed

IU Luddy School partners with CODE19 Racing to develop … – IU Newsroom

BLOOMINGTON, Ind. Indiana University will partner with the worlds first professional autonomous racing franchise, CODE19 Racing Inc., to participate in global competitions of self-driving race cars, an emerging sector in the world of racing.

IU associate professor Lantao Liu, IU Luddy School Dean Joanna Millunchick and CODE19 Racing co-founders Lawrence Walter and Oliver Wells, front row from left, with IU Luddy School graduate students involved in the AI driver project. Photo by Chris Kowalczyk, Indiana University

IU and CODE19 kicked off the partnership Sept. 26 to 28 during a workshop and starting grid event with graduate students and faculty at the IU Luddy School of Informatics, Computing and Engineering, who will create the first AI driver for CODE19.

We are thrilled to partner with CODE19 Racing, said Joanna Millunchick, dean of the IU Luddy School of Informatics, Computing and Engineering. This is an unmatched opportunity for our students to apply their skills to real-world problems and to compete against the best AI teams in the world. AI racing truly has the potential to advance autonomous vehicle technology in the same way that motorsports technology has advanced the consumer automotive industry.

A member of the IU Ventures Founders and Funders Network, CODE19 has a mission of accelerating the development of autonomous driving technology by developing and racing autonomous race cars at the highest level.

As an IU alumni-founded startup, we are excited to partner with the IU Luddy School to develop the next generation of advanced AI for autonomous race cars, said, Lawrence Walter, president of CODE19 Racing. The Luddy School is home to some of the brightest minds in artificial intelligence, and we are excited to help IU develop a world-class AI driver.

With the Luddy School at the controls, CODE19 will race in global competitions for autonomous race cars. These experimental competitions help accelerate the development of autonomous driving technology by providing a challenging and competitive environment for teams to test their AI drivers.

Lawrence Walter discusses an early demonstration of an AI race car driver in a simulated video environment during the Starting Grid event at Luddy Hall at IU Bloomington. Photo by Chris Kowalczyk, Indiana University

The Luddy School is one of the largest and most comprehensive schools of its kind in the world. The CODE19 Racing AI driver will be developed by the schools Vehicle Autonomy and Intelligence Lab, a state-of-the-art robotics research team led by Lantao Liu, a Luddy School associate professor of intelligent systems engineering.

An expert on robotics and artificial intelligence, Liu focuses on developing autonomous systems involving single or multiple robots with applications in autonomous navigation, smart transportation, and search and rescue. His lab specifically focuses on enhancing the autonomy and intelligence of robotic systems such as unmanned ground, aerial and aquatic vehicles.

The team will also have access to the deep pool of talent in informatics, computer science and engineering at the Luddy School in Indianapolis. Indianapolis-based advisors include Zebulun Wood, a lecturer in media arts and science, who will help develop the strategy for an interactive avatar for the AI driver, putting a face on the technology and interacting with the world.

In addition to the Luddy Schools resources, the project will benefit from IUs partnership with Naval Surface Warfare Center, Crane Division, which will support a Luddy Ph.D. student to advance development of the teams AI driver.

IU Luddy faculty members Lantao Liu, right, and Zeb Wood, center, speak with Lawrence Walter, founder and CEO of CODE19 Racing, before the partnership announcement at the Rally Conference in Indianapolis in August. Photo by Justin Casterline, Indiana University

A naval research and development laboratory in Crane, Indiana, NSWC Crane is responsible for developing and testing naval technologies, including autonomous systems.

NSWC Crane is committed to advancing the state-of-the-art autonomous systems, said Charles Colglazier, NSWC Crane liaison for IU and the National Security Innovation Network. This innovative initiative with CODE19 Racing and IU will help rapidly develop and test new dual-use autonomous technologies that could be used to support the warfighters in the field. Were excited to see how fast the Hoosiers AI driver can go.

Support from the partnership will help attract top talent to the team and accelerate the development of the AI driver, Walter said.

We are confident that our team has the skills and resources to manage winning AI drivers, Walter said. We are focused on competing globally and showing the world what IU can do on the race track.

Go here to read the rest:

IU Luddy School partners with CODE19 Racing to develop ... - IU Newsroom

AI predicts how many earthquake aftershocks will strike and their … – Nature.com

A powerful earthquake on 24 August 2016 killed hundreds of people in Amatrice, Italy (pictured) and was followed by destructive aftershocks. New machine learning models hold potential for predicting the number of quake aftershocks.Credit: Stefano Montesi/Corbis via Getty

Seismologists are finally making traction on one of their most tantalizing but challenging goals: using machine learning to improve earthquake forecasts.

Three new papers describe deep-learning models that perform better than a conventional state-of-the-art model for forecasting earthquakes13. The findings are preliminary and apply only to limited situations, such as in assessing the risk of aftershocks after a big one has already hit. But they are a rare advance towards the long-sought goal of harnessing the power of machine learning to reduce seismic risk.

Im really excited that this is finally happening, says Morgan Page, a seismologist at the US Geological Survey (USGS) in Pasadena, California, who was not involved with the studies.

Heres what earthquake forecasts are not: predictions of an event of a particular magnitude happening in a particular location at a particular time the next Tuesday at 3 p.m. scenario. The notion that scientists can make such highly specific predictions has been discredited. Instead, statistical analyses are helping seismologists understand broader trends, such as how many aftershocks might be expected in the days to weeks after a large earthquake. Agencies such as the USGS issue aftershock forecasts to warn people in quake-ravaged areas of what else might be coming.

Algorithms spot millions of Californias tiniest quakes in historical data

At first glance, earthquake forecasts seem to be an obvious application to try to improve using deep learning4. The techniques do well when they ingest and synthesize large amounts of data and use them to predict the next steps in a pattern. And seismology is rich with data from catalogues of earthquakes that occur worldwide. Just as a large language model can train itself on millions of words to predict what word might come next, an earthquake-forecasting model should be able to train itself on earthquake catalogues to forecast the chances of a quake following one that has already occurred.

But researchers have struggled to extract meaningful trends from all the quake data5. Big earthquakes are rare, and working out what to worry about isnt easy.

In the past several years, however, seismologists have used machine learning to uncover small earthquakes that had not been spotted before in seismic records. These quakes have bulked up the existing earthquake catalogues, and provide fresh fodder for a second round of machine-learning analysis.

Current USGS forecasts use a model that relies on basic information about past earthquake magnitudes and locations to predict what might happen next. The three latest papers instead use a neural-network approach, which updates calculations during each step of the analysis to better capture the complex patterns of how earthquakes occur.

In the first1, geophysicist Kelian Dascher-Cousineau at the University of California, Berkeley, and his colleagues tested their model on a catalogue of thousands of quakes that struck southern California between 2008 and 2021. Their model performed better than the standard one at forecasting how many quakes would occur in rolling two-week periods. It was also better at capturing the full magnitude range of possible earthquakes, thus reducing the chance of a surprise big one.

At the University of Bristol, UK, applied statistician Samuel Stockman developed a similar method that performed well when trained2 on a catalogue of earthquakes that shook central Italy in 201617, damaging several towns. When researchers lower the magnitude of quakes included in the training set, the machine-learning model starts to perform better, Stockman says.

Rubble piles still stood in Castro, Italy, almost a year after the village was damaged by the same earthquake that levelled Amatrice.Credit: Amelia Hennighausen/Nature

And at Tel Aviv University in Israel, physicist Yohai Bar-Sinai led a team that developed a third neural-network model3. When tested against 30 years of quake data from Japan, it, too, did better than the standard model. The work might provide insight into fundamental quake physics, Bar-Sinai says. There is hope that we will understand more about the underlying mechanisms about what causes earthquakes to start, what determines their magnitude.

All three models are moderately promising, says Leila Mizrahi, a seismologist at the Swiss Federal Institute of Technology (ETH) in Zurich. They arent breakthroughs in their current form, she says, but they show potential for bringing machine-learning techniques into quake forecasting on an everyday basis.

Its certainly no silver bullet, adds Maximilian Werner, a seismologist at the University of Bristol who works with Stockman. But, he says, machine learning will gradually become part of official earthquake forecasting over the coming years, because it is so well suited to working with the huge earthquake data sets that are becoming more common.

Agencies such as the USGS will probably start to use machine-learning models alongside their standard one, and then transition entirely to the machine-learning approach if it proves to be superior, Page says. That could improve forecasts when aftershocks are rumbling unpredictably and disrupting peoples lives for months, as happened in Italy. The models could also be used to improve forecasts after large rare earthquakes, including the magnitude-6.8 earthquake that hit Morocco in September, killing thousands.

Still, Dascher-Cousineau warns people not to rely on these fancy new models too much. At the end of the day, preparing for quakes is the most important, he says. We dont get to stop making sure our buildings are up to code, we dont get to not have our earthquake kits, [just] because we have a better earthquake-forecasting model.

Here is the original post:

AI predicts how many earthquake aftershocks will strike and their ... - Nature.com

How to stop AI deepfakes from sinking society and science – Nature.com

This June, in the political battle leading up to the 2024 US presidential primaries, a series of images were released showing Donald Trump embracing one of his former medical advisers, Anthony Fauci. In a few of the shots, Trump is captured awkwardly kissing the face of Fauci, a health official reviled by some US conservatives for promoting masking and vaccines during the COVID-19 pandemic.

It was obvious that they were fakes, says Hany Farid, a computer scientist at the University of California, Berkeley, and one of many specialists who examined the pictures. On close inspection of three of the photos, Trumps hair is strangely blurred, the text in the background is nonsensical, the arms and hands are unnaturally placed and the details of Trumps visible ear are not right. All are hallmarks for now of generative artificial intelligence (AI), also called synthetic AI.

Science and the new age of AI: a Nature special

Such deepfake images and videos, made by text-to-image generators powered by deep learning AI, are now rife. Although fraudsters have long used deception to make a profit, sway opinions or start a war, the speed and ease with which huge volumes of viscerally convincing fakes can now be created and spread paired with a lack of public awareness is a growing threat. People are not used to generative technology. Its not like it evolved gradually; it was like boom, all of a sudden its here. So, you dont have that level of scepticism that you would need, says Cynthia Rudin, an AI computer scientist at Duke University in Durham, North Carolina.

Dozens of systems are now available for unsophisticated users to generate almost any content for any purpose, whether thats creating deepfake Tom Cruise videos on Tik Tok for entertainment; bringing back the likeness of a school-shooting victim to create a video advocating gun regulation; or faking a call for help from a loved one to scam victims out of tens of thousands of dollars. Deepfake videos can be generated in real time on a live video call. Earlier this year, Jerome Powell, chair of the US Federal Reserve, had a video conversation with someone he thought was Ukrainian President Volodymyr Zelenskyy, but wasnt.

The quantity of AI-generated content is unknown, but it is thought to be exploding. Academics commonly quote an estimate that around 90% of all Internet content could be synthetic within a few years1. Everything else would just get drowned out by this noise, says Rudin, which would make it hard to find genuine, useful content. Search engines and social media will just amplify misinformation, she adds. Weve been recommending and circulating all this crap. And now were going to be generating crap.

Although a lot of synthetic media is made for entertainment and fun, such as the viral image of Pope Francis wearing a designer puffer jacket, some is agenda-driven and some malicious including vast amounts of non-consensual pornography, in which someones face is transposed onto another body. Even a single synthetic file can make waves: an AI-generated image of an explosion at the US Pentagon that went viral in May, for example, caused the stock market to dip briefly. The existence of synthetic content also allows bad actors to brush off real evidence of misbehaviour by simply claiming that it is fake.

Peoples ability to really know where they should place their trust is falling away. And thats a real problem for democracy, says psychologist Sophie Nightingale at Lancaster University, UK, who studies the effects of generative AI. We need to act on that, and quite quickly. Its already a huge threat. She adds that this issue will be a big one in the coming year or two, with major elections planned in the United States, Russia and the United Kingdom.

AI-generated fakes could also have huge impacts on science, say some experts. They worry that the rapidly developing abilities of generative AI systems could make it easier for unscrupulous researchers to publish fraudulent data and images (see Scammed science at the end of this article).

For now, some synthetic content contains give-away clues such as images featuring people with six fingers on one hand. But generative AI is getting better every day. Were talking about months until people cant tell the difference with the naked eye, says Wael Abd-Almageed, an information scientist and computer engineer at the University of Southern California in Los Angeles.

An overlay of synthetic faces (left) shows more regularity than one of real images (right).Credit: Hany Farid

All of this has researchers scrambling to work out how to harness the deepfake powers of AI for good, while developing tools to guard against the bad. There are two prongs of technological defence: proactively tagging real or fake content when it is generated; and using detectors to catch fakes after publication. Neither is a perfect solution, but both help by adding hurdles to fakery, says Shyam Sundar, a psychologist and founder of the Media Effects Research Laboratory at Pennsylvania State University in University Park. If youre a dedicated malicious actor, you can certainly go quite far. The idea is to make it difficult for them, he says.

Technology will be crucial in the short term, says Nightingale, but then longer term, maybe we can think more about education, regulation. The European Union is leading the way globally with its AI Act, which was passed by the parliament this June and is awaiting decisions by the two other branches of the EU government. Were going to learn important lessons from it for sure, says Nightingale, whether they get it right or not.

For researchers, generative AI is a powerful tool. It is being used, for example, to create medical data sets that are free of privacy concerns, to help design medicinal molecules and to improve scientific manuscripts and software. Deepfakes are being investigated for their use in anonymising participants of video-based group therapy; creating custom avatars of physicians or teachers that are more compelling for viewers; or allowing for improved control conditions in social-science studies2. Im more hopeful than concerned, says Sundar. I think its transformative as a technology.

But with the spectre of rampant misuse, researchers and ethicists have attempted to lay down rules for AI, including the 2018 Montreal Declaration for the Responsible Development of Artificial Intelligence and the 2019 Recommendation on Artificial Intelligence from the Organisation for Economic Co-operation and Development. An initiative called the Partnership on AI, a non-profit organization that includes major industry partners, fosters dialogue on best practices although some observers and participants have questioned whether it has had any impact.

AI and science: what 1,600 researchers think

All advocate for the principles of transparency and disclosure of synthetic content. Companies are picking that up: in March, for example, TikTok updated its community guidelines to make it mandatory for creators to disclose use of AI in any realistic-looking scene. In July, seven leading technology companies including Meta, Microsoft, Google, OpenAI and Amazon made voluntary commitments to the White House to mark their AI-generated content. And in September, Google declared that starting in mid-November, any AI-generated content used in political ads will have to be declared on its platforms, including YouTube.

One way to tag synthetic images is to watermark them by altering the pixels in a distinctive way thats imperceptible to the eye but notable on analysis. Tweaking every nth pixel so that its colour value is an even number, for example, would create a watermark but a simple one that would disappear after almost any image manipulation, such as applying a colour filter. Some watermarks have been criticized for being too easy to remove. But deeper watermarks can, for instance, insert a wave of dark-to-light shading from one side of an image to the other and layer it on top of several more such patterns, in a way that cant be wiped away by fiddling with individual pixels. These watermarks can be difficult (but not impossible) to remove, says Farid. In August, Google released a watermark for synthetic images, called SynthID, without revealing details about how it works; its unclear yet how robust it is, says Farid.

The companion idea to watermarking is to tag a files metadata with secure provenance information. For photography, such systems start when a photo is taken, with software on the camera device that ensures that an images GPS and time stamps are legitimate, and that the image isnt a photo of another photo, for example. Insurance underwriters use such systems to validate images of assets and damages, and the news agency Reuters has trialled authentication technology to validate photos of the war in Ukraine.

The Coalition for Content Provenance and Authenticity (C2PA), which brings together key industry groups in technology and publishing, released a first version of a set of technical specifications in 2022 for how systems should track provenance information for both synthetic and real imagery. Plenty of C2PA-compliant tools that embed, track and verify provenance data are now available, and many corporate commitments such as Microsofts say they will follow C2PA guidelines. C2PA is going to be very important; its going to help, says Anderson Rocha, an AI researcher at the University of Campinas in Brazil.

Systems that track image provenance should become the workhorse for cutting down the sheer number of dubious files, says Farid, who is on the C2PA steering committee and is a paid consultant for Truepic, a company in San Diego, California, that sells software for tracking authentic photos and videos. But this relies on good actors signing up to a scheme such as C2PA, and things will slip through the cracks, he says. That makes detectors a good complementary tool.

AI will transform science now researchers must tame it

Academic labs and companies have produced many AI-based classifiers. These learn the patterns that can distinguish AI-made media from real photos, and many systems have reported that they can spot fakes more than 90% of the time, while falsely identifying real images as fakes only 1% or less of the time. But these systems can often be defeated3. A bad actor can tweak images so that the detector is more likely to be wrong than right, says Farid.

AI-based tools can be paired with other techniques that lean on human insights to unravel the fake from the real. Farid looks for clues such as lines of perspective that dont follow the rules of physics. Other signs are more subtle. He and his colleagues found that facial profiles made by StyleGAN generators, for example, tend to place the eyes in the exact same position in the photo4, providing a hint as to which faces are fakes. Detectors can be given sophisticated algorithms that can, for example, read a clock somewhere in the photo and check to see whether the lighting in the image matches the recorded time of day. Tech company Intels FakeCatcher analyses videos by looking for expected colour changes in the face that arise from fluctuations in blood flow. Some detectors, says Rocha, look for distinctive noise patterns generated by light sensors in a camera, which so far arent well replicated by AI.

Experts worry about the influence of deepfake videos, such as this one of Barack Obama.Credit: AP via Alamy

The battle between fake-makers and fake-detectives is fierce. Farid recalls a paper by his former student Siwei Lyu, now a computer scientist at the University at Buffalo, New York, that highlighted how some AI videos featured people whose two eyes blinked at different rates5. Generators fixed that problem in weeks, he says. For this reason, even though Farids lab publishes the vast majority of its work, he releases code only on a case-by-case basis to academics who request it. Abd-Almageed takes a similar approach. If we release our tool to the public, people will make their own generation methods even more sophisticated, he says.

Several detection services that have public user interfaces have sprung up, and many academic labs are on the case, including the DeFake project at the Rochester Institute of Technology in New York and the University at Buffalos DeepFake-o-meter. And the US Defense Advanced Research Projects Agency (DARPA) launched its Semantic Forensics (SemaFor) project in 2021, with a broad remit of unearthing the who, what, why and how behind any generative file. A team of nearly 100 academics and corporate researchers have worked together under SemaFor to create more than 150 analytics, says the projects head, Wil Corvey. The bulk are detection algorithms that can be used in isolation or together.

AI tools as science policy advisers? The potential and the pitfalls

Because there are a huge number of both generators and detectors, and every case is different, reported accuracy rates vary wildly. And the arms race between them means that the situation is constantly changing. But for many media types, current success rates are poor. For generated text, a review this year of 14 detection tools found that all were neither accurate nor reliable6. For video, a high-profile competition in 2020 was won by a system that was only about 65% accurate3 (see also go.nature.com/3jvevoc). For images, Rocha says that if the generator is well known, detectors can easily be more than 95% accurate; but if the generator is new or unknown, success rates typically plummet. Using multiple detectors on the same image can increase the success rate, says Corvey.

He adds that detecting whether something is synthetic is only one part of the puzzle: as more users rely on AI to tweak their content, the more important question is not how much of this is synthetic? but rather why was this made?, he says. For this reason, an important part of SemaFors work is to determine the intent behind fakes, by attributing the media to a creator and characterizing its meaning. A parallel DARPA project, the Influence Campaign Awareness and Sensemaking (INCAS), is attempting to develop automated tools to detect the signals of mass misinformation campaigns that might or might not be supported by AI fakery.

SemaFor is now in the third and final stage of its project, in which Corvey is focusing on reaching out to potential users such as social-media sites. We have outreach to a number of companies including Google. To our knowledge, none have taken or are running our algorithms on a constant basis on-site, he says. Meta has collaborated with researchers at Michigan State University in East Lansing on detectors, but hasnt said how it might use them. Farid works with the employment-focused platform LinkedIn, which uses AI-based detectors to help weed out synthetic faces that support fraudulent accounts.

Abd-Almageed is in favour of social-media sites running detectors on all images on their sites, perhaps publishing a warning label on images flagged with a high percentage chance of being fake. But he had no luck when he discussed this a couple of years ago with a company that he would not name. I told a social network platform, take my software and use it, take it for free. And they said, if you cant show me how to make money, we dont care, says Abd-Almageed. Farid argues, however, that automated detectors arent well suited to this kind of use: even a 99% accurate tool would be wrong one time out of 100, which he thinks would completely erode public confidence. Farid argues that detection should be targeted at intensive, human-led investigations of specific cases, rather than trying to police the entire Internet.

Many argue that companies such as publishers and social-media sites will need legislation to push them into responsible behaviour. In June, the European Parliament passed a draft law that would strictly regulate high-risk uses of AI and enforce disclosure of content generated by such tools. The world is watching, because the EU has taken the lead on this, says Nightingale. But experts disagree widely about the acts merits and whether it might quash innovation. In the United States, a few AI bills are pending, including one aimed at preventing deepfakes of intimate images and one about the use of AI in political advertising, but neither is certain to pass.

There is one point that experts agree on: improving tech literacy will help to stop society and democracy from drowning in fakes. We need to get the word out to make people aware of whats happening, says Rocha. When they are informed about it, they can take action. They can demand education in schools.

Even with all the technological and social tools at our disposal, Farid says, its a losing battle to stop or catch all fakery. But its OK, because even in defeat I will have taken this out of the hands of the average person, he says. Then, as with counterfeit money, it will still be possible to fool the world with generative AI fakes but much harder.

Science isnt immune to the problem of AI-generated fakery. One concern is the integrity of biomedical images such as scans, microscopy images and western blots a standard technique in which distinctive bands are created by proteins of various molecular weights as they spread across a gel. Fraudsters have long faked such images using Photoshop or other image-manipulation software, but that is often detectable by a trained eye or by computers that check for image duplication. Generating images entirely with AI presents a bigger detection problem.

Last year, Rongshan Yu and his colleagues at Xiamen University, China, created scientific images using synthetic AI to see how easily it can be done7. They trained one generative AI program to create new western blot images from a training data set of 3,000 images; and used another to insert features of oesophageal cancer into an image of a non-cancerous intestine. They then tested how convincing the western blots were by inviting three specialists to try to spot one fake in a set of four images; two of the experts performed worse than chance, and the third fared better by spotting a visual clue to do with the smoothness between the image and the background. A computer system did slightly better than the specialists.

In a larger study of 14,000 original western blot images and 24,000 synthetic ones, made using four different generators, Anderson Rocha at the University of Campinas, Brazil, and his colleagues found that AI detectors trained on large data sets achieved accuracies of more than 85% (ref. 8). This is just the beginning, where we showed its feasible. Its possible to do way more than that, Rocha says. He and his team have a paper under review that extends their methods beyond western blots, he says.

No one Nature spoke to could provide proof that AI is being used by academics to beef up their papers or funding applications, but experts say its likely. Wael Abd-Almageed, an information scientist and computer engineer at the University of Southern California in Los Angeles, says he knows of specific cases in which academics have used AI to synthesize fakes, but declined to divulge details. Researchers investigating fake paper mills have said they have seen AI-generated images.

I dont know any case yet of a retracted paper that used synthetic creation to illustrate the results in that paper, says Rocha, who is working with the blog Retraction Watch on this issue. But its just a matter of time, I guess.

Publishers are on alert, says Bernd Pulverer, chief editor of the journal EMBO Reports. We are certainly concerned, he says. In the arms race between generative AI and detection tools, he says, we will struggle to detect well-executed AI image manipulation. It is all too easy to weaponize existing AI tools at every level of biological data.

Watermarking technologies could help, Pulverer adds. This idea has been explored previously for medical images not to prevent AI alterations, but to protect privacy and prevent any kind of tampering. Pulverer says he plans to work with instrument makers and image-processing experts to discuss the feasibility of watermarking or provenance-tagging scientific imagery.

Beyond developing technological tools against fraud, he says, the key is to encourage efforts to reproduce data in science. Attempting to replicate results could be the ultimate way to catch out fakery.

Originally posted here:

How to stop AI deepfakes from sinking society and science - Nature.com