InvestMacro

By Robert Mahari, Massachusetts Institute of Technology (MIT); Jessica Fjeld, Harvard Law School, and Ziv Epstein, Massachusetts Institute of Technology (MIT) 

In 2022, an AI-generated work of art won the Colorado State Fair’s art competition. The artist, Jason Allen, had used Midjourney – a generative AI system trained on art scraped from the internet – to create the piece. The process was far from fully automated: Allen went through some 900 iterations over 80 hours to create and refine his submission.

Yet his use of AI to win the art competition triggered a heated backlash online, with one Twitter user claiming, “We’re watching the death of artistry unfold right before our eyes.”

As generative AI art tools like Midjourney and Stable Diffusion have been thrust into the limelight, so too have questions about ownership and authorship.

These tools’ generative ability is the result of training them with scores of prior artworks, from which the AI learns how to create artistic outputs.

Should the artists whose art was scraped to train the models be compensated? Who owns the images that AI systems produce? Is the process of fine-tuning prompts for generative AI a form of authentic creative expression?

On one hand, technophiles rave over work like Allen’s. But on the other, many working artists consider the use of their art to train AI to be exploitative.

We’re part of a team of 14 experts across disciplines that just published a paper on generative AI in Science magazine. In it, we explore how advances in AI will affect creative work, aesthetics and the media. One of the key questions that emerged has to do with U.S. copyright laws, and whether they can adequately deal with the unique challenges of generative AI.

Copyright laws were created to promote the arts and creative thinking. But the rise of generative AI has complicated existing notions of authorship.

Still from ‘All watched over by machines of loving grace’ by Memo Akten, 2021. Created using custom AI software.
Memo Akten, CC BY-SA

Photography serves as a helpful lens

Generative AI might seem unprecedented, but history can act as a guide.

Take the emergence of photography in the 1800s. Before its invention, artists could only try to portray the world through drawing, painting or sculpture. Suddenly, reality could be captured in a flash using a camera and chemicals. As with generative AI, many argued that photography lacked artistic merit. In 1884, the U.S. Supreme Court weighed in on the issue and found that cameras served as tools that an artist could use to give an idea visible form; the “masterminds” behind the cameras, the court ruled, should own the photographs they create.

From then on, photography evolved into its own art form and even sparked new abstract artistic movements.

AI can’t own outputs

Unlike inanimate cameras, AI possesses capabilities – like the ability to convert basic instructions into impressive artistic works – that make it prone to anthropomorphization. Even the term “artificial intelligence” encourages people to think that these systems have humanlike intent or even self-awareness.

This led some people to wonder whether AI systems can be “owners.” But the U.S. Copyright Office has stated unequivocally that only humans can hold copyrights.

So who can claim ownership of images produced by AI? Is it the artists whose images were used to train the systems? The users who type in prompts to create images? Or the people who build the AI systems?

Infringement or fair use?

While artists draw obliquely from past works that have educated and inspired them in order to create, generative AI relies on training data to produce outputs.

This training data consists of prior artworks, many of which are protected by copyright law and which have been collected without artists’ knowledge or consent. Using art in this way might violate copyright law even before the AI generates a new work.

For Jason Allen to create his award-winning art, Midjourney was trained on 100 million prior works.

Was that a form of infringement? Or was it a new form of “fair use,” a legal doctrine that permits the unlicensed use of protected works if they’re sufficiently transformed into something new?

While AI systems do not contain literal copies of the training data, they do sometimes manage to recreate works from the training data, complicating this legal analysis.

Will contemporary copyright law favor end users and companies over the artists whose content is in the training data?

To mitigate this concern, some scholars propose new regulations to protect and compensate artists whose work is used for training. These proposals include a right for artists to opt out of their data’s being used for generative AI or a way to automatically compensate artists when their work is used to train an AI.

Muddled ownership

Training data, however, is only part of the process. Frequently, artists who use generative AI tools go through many rounds of revision to refine their prompts, which suggests a degree of originality.

Answering the question of who should own the outputs requires looking into the contributions of all those involved in the generative AI supply chain.

The legal analysis is easier when an output is different from works in the training data. In this case, whoever prompted the AI to produce the output appears to be the default owner.

However, copyright law requires meaningful creative input – a standard satisfied by clicking the shutter button on a camera. It remains unclear how courts will decide what this means for the use of generative AI. Is composing and refining a prompt enough?

Matters are more complicated when outputs resemble works in the training data. If the resemblance is based only on general style or content, it is unlikely to violate copyright, because style is not copyrightable.

The illustrator Hollie Mengert encountered this issue firsthand when her unique style was mimicked by generative AI engines in a way that did not capture what, in her eyes, made her work unique. Meanwhile, the singer Grimes embraced the tech, “open-sourcing” her voice and encouraging fans to create songs in her style using generative AI.

If an output contains major elements from a work in the training data, it might infringe on that work’s copyright. Recently, the Supreme Court ruled that Andy Warhol’s drawing of a photograph was not permitted by fair use. That means that using AI to just change the style of a work – say, from a photo to an illustration – is not enough to claim ownership over the modified output.

While copyright law tends to favor an all-or-nothing approach, scholars at Harvard Law School have proposed new models of joint ownership that allow artists to gain some rights in outputs that resemble their works.

In many ways, generative AI is yet another creative tool that allows a new group of people access to image-making, just like cameras, paintbrushes or Adobe Photoshop. But a key difference is this new set of tools relies explicitly on training data, and therefore creative contributions cannot easily be traced back to a single artist.

The ways in which existing laws are interpreted or reformed – and whether generative AI is appropriately treated as the tool it is – will have real consequences for the future of creative expression.

About the Author:

Robert Mahari, JD-PhD Student, Massachusetts Institute of Technology (MIT); Jessica Fjeld, Lecturer on Law, Harvard Law School, and Ziv Epstein, PhD Student in Media Arts and Sciences, Massachusetts Institute of Technology (MIT)

This article is republished from The Conversation under a Creative Commons license. Read the original article.

By Kate Saenko, Boston University 

Generative AI is the hot new technology behind chatbots and image generators. But how hot is it making the planet?

As an AI researcher, I often worry about the energy costs of building artificial intelligence models. The more powerful the AI, the more energy it takes. What does the emergence of increasingly more powerful generative AI models mean for society’s future carbon footprint?

“Generative” refers to the ability of an AI algorithm to produce complex data. The alternative is “discriminative” AI, which chooses between a fixed number of options and produces just a single number. An example of a discriminative output is choosing whether to approve a loan application.

Generative AI can create much more complex outputs, such as a sentence, a paragraph, an image or even a short video. It has long been used in applications like smart speakers to generate audio responses, or in autocomplete to suggest a search query. However, it only recently gained the ability to generate humanlike language and realistic photos.

AI chatbots and image generators run on thousands of computers housed in data centers like this Google facility in Oregon.
Tony Webster/Wikimedia, CC BY-SA

Using more power than ever

The exact energy cost of a single AI model is difficult to estimate, and includes the energy used to manufacture the computing equipment, create the model and use the model in production. In 2019, researchers found that creating a generative AI model called BERT with 110 million parameters consumed the energy of a round-trip transcontinental flight for one person. The number of parameters refers to the size of the model, with larger models generally being more skilled. Researchers estimated that creating the much larger GPT-3, which has 175 billion parameters, consumed 1,287 megawatt hours of electricity and generated 552 tons of carbon dioxide equivalent, the equivalent of 123 gasoline-powered passenger vehicles driven for one year. And that’s just for getting the model ready to launch, before any consumers start using it.

Size is not the only predictor of carbon emissions. The open-access BLOOM model, developed by the BigScience project in France, is similar in size to GPT-3 but has a much lower carbon footprint, consuming 433 MWh of electricity in generating 30 tons of CO2eq. A study by Google found that for the same size, using a more efficient model architecture and processor and a greener data center can reduce the carbon footprint by 100 to 1,000 times.

Larger models do use more energy during their deployment. There is limited data on the carbon footprint of a single generative AI query, but some industry figures estimate it to be four to five times higher than that of a search engine query. As chatbots and image generators become more popular, and as Google and Microsoft incorporate AI language models into their search engines, the number of queries they receive each day could grow exponentially.

AI bots for search

A few years ago, not many people outside of research labs were using models like BERT or GPT. That changed on Nov. 30, 2022, when OpenAI released ChatGPT. According to the latest available data, ChatGPT had over 1.5 billion visits in March 2023. Microsoft incorporated ChatGPT into its search engine, Bing, and made it available to everyone on May 4, 2023. If chatbots become as popular as search engines, the energy costs of deploying the AIs could really add up. But AI assistants have many more uses than just search, such as writing documents, solving math problems and creating marketing campaigns.

Another problem is that AI models need to be continually updated. For example, ChatGPT was only trained on data from up to 2021, so it does not know about anything that happened since then. The carbon footprint of creating ChatGPT isn’t public information, but it is likely much higher than that of GPT-3. If it had to be recreated on a regular basis to update its knowledge, the energy costs would grow even larger.

One upside is that asking a chatbot can be a more direct way to get information than using a search engine. Instead of getting a page full of links, you get a direct answer as you would from a human, assuming issues of accuracy are mitigated. Getting to the information quicker could potentially offset the increased energy use compared to a search engine.

Ways forward

The future is hard to predict, but large generative AI models are here to stay, and people will probably increasingly turn to them for information. For example, if a student needs help solving a math problem now, they ask a tutor or a friend, or consult a textbook. In the future, they will probably ask a chatbot. The same goes for other expert knowledge such as legal advice or medical expertise.

While a single large AI model is not going to ruin the environment, if a thousand companies develop slightly different AI bots for different purposes, each used by millions of customers, the energy use could become an issue. More research is needed to make generative AI more efficient. The good news is that AI can run on renewable energy. By bringing the computation to where green energy is more abundant, or scheduling computation for times of day when renewable energy is more available, emissions can be reduced by a factor of 30 to 40, compared to using a grid dominated by fossil fuels.

Finally, societal pressure may be helpful to encourage companies and research labs to publish the carbon footprints of their AI models, as some already do. In the future, perhaps consumers could even use this information to choose a “greener” chatbot.

About the Author:

Kate Saenko, Associate Professor of Computer Science, Boston University

This article is republished from The Conversation under a Creative Commons license. Read the original article.

By Chris Impey, University of Arizona 

The famous first image of a black hole just got two times sharper. A research team used artificial intelligence to dramatically improve upon its first image from 2019, which now shows the black hole at the center of the M87 galaxy as darker and bigger than the first image depicted.

I’m an astronomer who studies and has written about cosmology, black holes and exoplanets. Astronomers have been using AI for decades. In fact, in 1990, astronomers from the University of Arizona, where I am a professor, were among the first to use a type of AI called a neural network to study the shapes of galaxies.

Since then, AI has spread into every field of astronomy. As the technology has become more powerful, AI algorithms have begun helping astronomers tame massive data sets and discover new knowledge about the universe.

Better telescopes, more data

As long as astronomy has been a science, it has involved trying to make sense of the multitude of objects in the night sky. That was relatively simple when the only tools were the naked eye or a simple telescope, and all that could be seen were a few thousand stars and a handful of planets.

A hundred years ago, Edwin Hubble used newly built telescopes to show that the universe is filled with not just stars and clouds of gas, but countless galaxies. As telescopes have continued to improve, the sheer number of celestial objects humans can see and the amount of data astronomers need to sort through have both grown exponentially, too.

For example, the soon-to-be-completed Vera Rubin Observatory in Chile will make images so large that it would take 1,500 high-definition TV screens to view each one in its entirety. Over 10 years it is expected to generate 0.5 exabytes of data – about 50,000 times the amount of information held in all of the books contained within the Library of Congress.

There are 20 telescopes with mirrors larger than 20 feet (6 meters) in diameter. AI algorithms are the only way astronomers could ever hope to work through all of the data available to them today. There are a number of ways AI is proving useful in processing this data.

Picking out patterns

Astronomy often involves looking for needles in a haystack. About 99% of the pixels in an astronomical image contain background radiation, light from other sources or the blackness of space – only 1% have the subtle shapes of faint galaxies.

AI algorithms – in particular, neural networks that use many interconnected nodes and are able to learn to recognize patterns – are perfectly suited for picking out the patterns of galaxies. Astronomers began using neural networks to classify galaxies in the early 2010s. Now the algorithms are so effective that they can classify galaxies with an accuracy of 98%.

This story has been repeated in other areas of astronomy. Astronomers working on SETI, the Search for Extraterrestrial Intelligence, use radio telescopes to look for signals from distant civilizations. Early on, radio astronomers scanned charts by eye to look for anomalies that couldn’t be explained. More recently, researchers harnessed 150,000 personal computers and 1.8 million citizen scientists to look for artificial radio signals. Now, researchers are using AI to sift through reams of data much more quickly and thoroughly than people can. This has allowed SETI efforts to cover more ground while also greatly reducing the number of false positive signals.

Another example is the search for exoplanets. Astronomers discovered most of the 5,300 known exoplanets by measuring a dip in the amount of light coming from a star when a planet passes in front of it. AI tools can now pick out the signs of an exoplanet with 96% accuracy.

Making new discoveries

AI has proved itself to be excellent at identifying known objects – like galaxies or exoplanets – that astronomers tell it to look for. But it is also quite powerful at finding objects or phenomena that are theorized but have not yet been discovered in the real world.

Teams have used this approach to detect new exoplanets, learn about the ancestral stars that led to the formation and growth of the Milky Way, and predict the signatures of new types of gravitational waves.

To do this, astronomers first use AI to convert theoretical models into observational signatures – including realistic levels of noise. They then use machine learning to sharpen the ability of AI to detect the predicted phenomena.

Finally, radio astronomers have also been using AI algorithms to sift through signals that don’t correspond to known phenomena. Recently a team from South Africa found a unique object that may be a remnant of the explosive merging of two supermassive black holes. If this proves to be true, the data will allow a new test of general relativity – Albert Einstein’s description of space-time.

Making predictions and plugging holes

As in many areas of life recently, generative AI and large language models like ChatGPT are also making waves in the astronomy world.

The team that created the first image of a black hole in 2019 used a generative AI to produce its new image. To do so, it first taught an AI how to recognize black holes by feeding it simulations of many kinds of black holes. Then, the team used the AI model it had built to fill in gaps in the massive amount of data collected by the radio telescopes on the black hole M87.

Using this simulated data, the team was able to create a new image that is two times sharper than the original and is fully consistent with the predictions of general relativity.

Astronomers are also turning to AI to help tame the complexity of modern research. A team from the Harvard-Smithsonian Center for Astrophysics created a language model called astroBERT to read and organize 15 million scientific papers on astronomy. Another team, based at NASA, has even proposed using AI to prioritize astronomy projects, a process that astronomers engage in every 10 years.

As AI has progressed, it has become an essential tool for astronomers. As telescopes get better, as data sets get larger and as AIs continue to improve, it is likely that this technology will play a central role in future discoveries about the universe.

About the Author:

Chris Impey, University Distinguished Professor of Astronomy, University of Arizona

This article is republished from The Conversation under a Creative Commons license. Read the original article.

By Victor R. Lee, Stanford University 

It turns out that pop stars Drake and The Weeknd didn’t suddenly drop a new track that went viral on TikTok and YouTube in April 2023. The photograph that won an international photography competition that same month wasn’t a real photograph. And the image of Pope Francis sporting a Balenciaga jacket that appeared in March 2023? That was also a fake.

All were made with the help of generative AI, the new technology that can generate humanlike text, audio and images on demand through programs such as ChatGPT, Midjourney and Bard, among others.

There’s certainly something unsettling about the ease with which people can be duped by these fakes, and I see it as a harbinger of an authenticity crisis that raises some difficult questions.

How will voters know whether a video of a political candidate saying something offensive was real or generated by AI? Will people be willing to pay artists for their work when AI can create something visually stunning? Why follow certain authors when stories in their writing style will be freely circulating on the internet?

I’ve been seeing the anxiety play out all around me at Stanford University, where I’m a professor and also lead a large generative AI and education initiative.

With text, image, audio and video all becoming easier for anyone to produce through new generative AI tools, I believe people are going to need to reexamine and recalibrate how authenticity is judged in the first place.

Fortunately, social science offers some guidance.

The many faces of authenticity

Long before generative AI and ChatGPT rose to the fore, people had been probing what makes something feel authentic.

When a real estate agent is gushing over a property they are trying to sell you, are they being authentic or just trying to close the deal? Is that stylish acquaintance wearing authentic designer fashion or a mass-produced knock-off? As you mature, how do you discover your authentic self?

These are not just philosophical exercises. Neuroscience research has shown that believing a piece of art is authentic will activate the brain’s reward centers in ways that viewing something you’ve been told is a forgery won’t.

Authenticity also matters because it is a social glue that reinforces trust. Take the social media misinformation crisis, in which fake news has been inadvertently spread and authentic news decreed fake.

In short, authenticity matters, for both individuals and society as a whole.

But what actually makes something feel authentic?

Psychologist George Newman has explored this question in a series of studies. He found that there are three major dimensions of authenticity.

One of those is historical authenticity, or whether an object is truly from the time, place and person someone claims it to be. An actual painting made by Rembrandt would have historical authenticity; a modern forgery would not.

A second dimension of authenticity is the kind that plays out when, say, a restaurant in Japan offers exceptional and authentic Neapolitan pizza. Their pizza was not made in Naples or imported from Italy. The chef who prepared it may not have a drop of Italian blood in their veins. But the ingredients, appearance and taste may match really well with what tourists would expect to find at a great restaurant in Naples. Newman calls that categorical authenticity.

And finally, there is the authenticity that comes from our values and beliefs. This is the kind that many voters find wanting in politicians and elected leaders who say one thing but do another. It is what admissions officers look for in college essays.

In my own research, I’ve also seen that authenticity can relate to our expectations about what tools and activities are involved in creating things.

For example, when you see a piece of custom furniture that claims to be handmade, you probably assume that it wasn’t literally made by hand – that all sorts of modern tools were nonetheless used to cut, shape and attach each piece. Similarly, if an architect uses computer software to help draw up building plans, you still probably think of the product as legitimate and original. This is because there’s a general understanding that those tools are part of what it takes to make those products.

In most of your quick judgments of authenticity, you don’t think much about these dimensions. But with generative AI, you will need to.

That’s because back when it took a lot of time to produce original new content, there was a general assumption that it required skill to create – that it only could have been made by skilled individuals putting in a lot of effort and acting with the best of intentions.

These are not safe assumptions anymore.

How to deal with the looming authenticity crisis

Generative AI thrives on exploiting people’s reliance on categorical authenticity by producing material that looks like “the real thing.”

So it’ll be important to disentangle historical and categorical authenticity in your own thinking. Just because a recording sounds exactly like Drake – that is, it fits the category expectations for Drake’s music – it does not mean that Drake actually recorded it. The great essay that was turned in for a college writing class assignment may not actually be from a student laboring to craft sentences for hours on a word processor.

If it looks like a duck, walks like a duck and quacks like a duck, everyone will need to consider that it may not have actually hatched from an egg.

Also, it’ll be important for everyone to get up to speed on what these new generative AI tools really can and can’t do. I think this will involve ensuring that people learn about AI in schools and in the workplace, and having open conversations about how creative processes will change with AI being broadly available.

Writing papers for school in the future will not necessarily mean that students have to meticulously form each and every sentence; there are now tools that can help them think of ways to phrase their ideas. And creating an amazing picture won’t require exceptional hand-eye coordination or mastery of Adobe Photoshop and Adobe Illustrator.

Finally, in a world where AI operates as a tool, society is going to have to consider how to establish guardrails. These could take the form of regulations, or the creation of norms within certain fields for disclosing how and when AI has been used.

Does AI get credited as a co-author on writing? Is it disallowed on certain types of documents or for certain grade levels in school? Does entering a piece of art into a competition require a signed statement that the artist did not use AI to create their submission? Or does there need to be new, separate competitions that expressly invite AI-generated work?

These questions are tricky. It may be tempting to simply deem generative AI an unacceptable aid, in the same way that calculators are forbidden in some math classes.

However, sequestering new technology risks imposing arbitrary limits on human creative potential. Would the expressive power of images be what it is now if photography had been deemed an unfair use of technology? What if Pixar films were deemed ineligible for the Academy Awards because people thought computer animation tools undermined their authenticity?

The capabilities of generative AI have surprised many and will challenge everyone to think differently. But I believe humans can use AI to expand the boundaries of what is possible and create interesting, worthwhile – and, yes, authentic – works of art, writing and design.

About the Author:

Victor R. Lee, Associate Professor of Learning Sciences and Technology Design in Education, Stanford University

This article is republished from The Conversation under a Creative Commons license. Read the original article.

By Eric Smalley, The Conversation 

The light and dark sides of AI have been in the public spotlight for many years. Think facial recognition, algorithms making loan and sentencing recommendations, and medical image analysis. But the impressive – and sometimes scary – capabilities of ChatGPT, DALL-E 2 and other conversational and image-conjuring artificial intelligence programs feel like a turning point.

The key change has been the emergence within the last year of powerful generative AI, software that not only learns from vast amounts of data but also produces things – convincingly written documents, engaging conversation, photorealistic images and clones of celebrity voices.

Generative AI has been around for nearly a decade, as long-standing worries about deepfake videos can attest. Now, though, the AI models have become so large and have digested such vast swaths of the internet that people have become unsure of what AI means for the future of knowledge work, the nature of creativity and the origins and truthfulness of content on the internet.

Here are five articles from our archives the take the measure of this new generation of artificial intelligence.

1. Generative AI and work

A panel of five AI experts discussed the implications of generative AI for artists and knowledge workers. It’s not simply a matter of whether the technology will replace you or make you more productive.

University of Tennessee computer scientist Lynne Parker wrote that while there are significant benefits to generative AI, like making creativity and knowledge work more accessible, the new tools also have downsides. Specifically, they could lead to an erosion of skills like writing, and they raise issues of intellectual property protections given that the models are trained on human creations.

University of Colorado Boulder computer scientist Daniel Acuña has found the tools to be useful in his own creative endeavors but is concerned about inaccuracy, bias and plagiarism.

University of Michigan computer scientist Kentaro Toyama wrote that human skill is likely to become costly and extraneous in some fields. “If history is any guide, it’s almost certain that advances in AI will cause more jobs to vanish, that creative-class people with human-only skills will become richer but fewer in number, and that those who own creative technology will become the new mega-rich.”

Florida International University computer scientist Mark Finlayson wrote that some jobs are likely to disappear, but that new skills in working with these AI tools are likely to become valued. By analogy, he noted that the rise of word processing software largely eliminated the need for typists but allowed nearly anyone with access to a computer to produce typeset documents and led to a new class of skills to list on a resume.

University of Colorado Anschutz biomedical informatics researcher Casey Greene wrote that just as Google led people to develop skills in finding information on the internet, AI language models will lead people to develop skills to get the best output from the tools. “As with many technological advances, how people interact with the world will change in the era of widely accessible AI models. The question is whether society will use this moment to advance equity or exacerbate disparities.”

2. Conjuring images from words

Generative AI can seem like magic. It’s hard to imagine how image-generating AIs can take a few words of text and produce an image that matches the words.

Hany Farid, a University of California, Berkeley computer scientist who specializes in image forensics, explained the process. The software is trained on a massive set of images, each of which includes a short text description.

“The model progressively corrupts each image until only visual noise remains, and then trains a neural network to reverse this corruption. Repeating this process hundreds of millions of times, the model learns how to convert pure noise into a coherent image from any caption,” he wrote.

3. Marking the machine

Many of the images produced by generative AI are difficult to distinguish from photographs, and AI-generated video is rapidly improving. This raises the stakes for combating fraud and misinformation. Fake videos of corporate executives could be used to manipulate stock prices, and fake videos of political leaders could be used to spread dangerous misinformation.

Farid explained how it’s possible to produce AI-generated photos and video that contain watermarks verifying that they are synthetic. The trick is to produce digital watermarks that can’t be altered or removed. “These watermarks can be baked into the generative AI systems by watermarking all the training data, after which the generated content will contain the same watermark,” he wrote.

4. Flood of ideas

For all the legitimate concern about the downsides of generative AI, the tools are proving to be useful for some artists, designers and writers. People in creative fields can use the image generators to quickly sketch out ideas, including unexpected off-the-wall material.

Rochester Institute of Technology industrial designer and professor Juan Noguera and his students use tools like DALL-E or Midjourney to produce thousands of images from abstract ideas – a sort of sketchbook on steroids.

“Enter any sentence – no matter how crazy – and you’ll receive a set of unique images generated just for you. Want to design a teapot? Here, have 1,000 of them,” he wrote. “While only a small subset of them may be usable as a teapot, they provide a seed of inspiration that the designer can nurture and refine into a finished product.”

5. Shortchanging the creative process

However, using AI to produce finished artworks is another matter, according to Nir Eisikovits and Alec Stubbs, philosophers at the Applied Ethics Center at University of Massachusetts Boston. They note that the process of making art is more than just coming up with ideas.

The hands-on process of producing something, iterating the process and making refinements – often in the moment in response to audience reactions – are indispensable aspects of creating art, they wrote.

“It is the work of making something real and working through its details that carries value, not simply that moment of imagining it,” they wrote. “Artistic works are lauded not merely for the finished product, but for the struggle, the playful interaction and the skillful engagement with the artistic task, all of which carry the artist from the moment of inception to the end result.”

Editor’s note: This story is a roundup of articles from The Conversation’s archives.

About the Author:

Eric Smalley, Science + Technology Editor, The Conversation

This article is republished from The Conversation under a Creative Commons license. Read the original article.

By Sayonnha Mandal, University of Nebraska Omaha 

Passwords could soon become passé.

Effective passwords are cumbersome, all the more so when reinforced by two-factor authentication. But the need for authentication and secure access to websites is as great as ever. Enter passkeys.

Passkeys are digital credentials stored on your phone or computer. They are analogous to physical keys. You access your passkey by signing in to your device using a personal identification number (PIN), swipe pattern or biometrics like fingerprint or face recognition. You set your online accounts to trust your phone or computer. To break into your accounts, a hacker would need to physically possess your device and have the means to sign in to it.

As a cybersecurity researcher, I believe that passkeys not only provide faster, easier and more secure sign-ins, they minimize human error in password security and authorization steps. You don’t need to remember passwords for every account and don’t need to use two-factor authentication.

How passkeys work

Passkeys are generated via public-key cryptography. They use a public-private key pair to ensure a mathematically protected private relationship between users’ devices and the online accounts being accessed. It would be nearly impossible for a hacker to guess the passkey – hence the need to physically possess the device the passkey is accessed from.

Passkeys consist of a long private key – a long string of encrypted characters – created for a specific device. Websites cannot access the value of the passkey. Rather, the passkey verifies that a website possesses the corresponding public key. You can use the passkey from one device to access a website using another device. For example, you can use your laptop to access a website using the passkey on your phone by authorizing the login from your phone. And if you lose your phone, the passkey can be stored securely in the cloud with the phone’s other data, which can be restored to a new phone.

Why passkeys matter

Passwords can be guessed, phished or otherwise stolen. Security experts advise users to make their passwords longer with more characters, mixing alphanumeric and special symbols. A good password should not be in the dictionary or in phrases, have no consecutive letters or numbers, but be memorable. Users should not share them with anyone. Last but not least, users should change passwords every six months at minimum for all devices and accounts. Using a password manager to remember and update strong passwords helps but can still be a nuisance.

Even if you follow all of the best practices to keep your passwords safe, there is no guarantee of airtight security. Hackers are continuously developing and using software exploits, hardware tools and ever-advancing algorithms to break these defenses. Cybersecurity experts and malicious hackers are locked in an arms race.

Passkeys remove the onus from the user to create, remember and guard all their passwords. Apple, Google and Microsoft are supporting passkeys and encourage users to use them instead of passwords. As a result, passkeys are likely to soon overtake passwords and password managers in the cybersecurity battlefield.

However, it will take time for websites to add support for passkeys, so passwords aren’t going to go extinct overnight. IT managers still recommend that people use a password manager like 1Password or Bitwarden. And even Apple, which is encouraging the adoption of passkeys, has its own password manager.

About the Author:

Sayonnha Mandal, Lecturer in Interdisciplinary Informatics, University of Nebraska Omaha

This article is republished from The Conversation under a Creative Commons license. Read the original article.

By Mayank Kejriwal, University of Southern California 

The past few years have seen an explosion of progress in large language model artificial intelligence systems that can do things like write poetry, conduct humanlike conversations and pass medical school exams. This progress has yielded models like ChatGPT that could have major social and economic ramifications ranging from job displacements and increased misinformation to massive productivity boosts.

Despite their impressive abilities, large language models don’t actually think. They tend to make elementary mistakes and even make things up. However, because they generate fluent language, people tend to respond to them as though they do think. This has led researchers to study the models’ “cognitive” abilities and biases, work that has grown in importance now that large language models are widely accessible.

This line of research dates back to early large language models such as Google’s BERT, which is integrated into its search engine and so has been coined BERTology. This research has already revealed a lot about what such models can do and where they go wrong.

For instance, cleverly designed experiments have shown that many language models have trouble dealing with negation – for example, a question phrased as “what is not” – and doing simple calculations. They can be overly confident in their answers, even when wrong. Like other modern machine learning algorithms, they have trouble explaining themselves when asked why they answered a certain way.

Words and thoughts

Inspired by the growing body of research in BERTology and related fields like cognitive science, my student Zhisheng Tang and I set out to answer a seemingly simple question about large language models: Are they rational?

Although the word rational is often used as a synonym for sane or reasonable in everyday English, it has a specific meaning in the field of decision-making. A decision-making system – whether an individual human or a complex entity like an organization – is rational if, given a set of choices, it chooses to maximize expected gain.

The qualifier “expected” is important because it indicates that decisions are made under conditions of significant uncertainty. If I toss a fair coin, I know that it will come up heads half of the time on average. However, I can’t make a prediction about the outcome of any given coin toss. This is why casinos are able to afford the occasional big payout: Even narrow house odds yield enormous profits on average.

On the surface, it seems odd to assume that a model designed to make accurate predictions about words and sentences without actually understanding their meanings can understand expected gain. But there is an enormous body of research showing that language and cognition are intertwined. An excellent example is seminal research done by scientists Edward Sapir and Benjamin Lee Whorf in the early 20th century. Their work suggested that one’s native language and vocabulary can shape the way a person thinks.

The extent to which this is true is controversial, but there is supporting anthropological evidence from the study of Native American cultures. For instance, speakers of the Zuñi language spoken by the Zuñi people in the American Southwest, which does not have separate words for orange and yellow, are not able to distinguish between these colors as effectively as speakers of languages that do have separate words for the colors.

Making a bet

So are language models rational? Can they understand expected gain? We conducted a detailed set of experiments to show that, in their original form, models like BERT behave randomly when presented with betlike choices. This is the case even when we give it a trick question like: If you toss a coin and it comes up heads, you win a diamond; if it comes up tails, you lose a car. Which would you take? The correct answer is heads, but the AI models chose tails about half the time.

Intriguingly, we found that the model can be taught to make relatively rational decisions using only a small set of example questions and answers. At first blush, this would seem to suggest that the models can indeed do more than just “play” with language. Further experiments, however, showed that the situation is actually much more complex. For instance, when we used cards or dice instead of coins to frame our bet questions, we found that performance dropped significantly, by over 25%, although it stayed above random selection.

So the idea that the model can be taught general principles of rational decision-making remains unresolved, at best. More recent case studies that we conducted using ChatGPT confirm that decision-making remains a nontrivial and unsolved problem even for much bigger and more advanced large language models.

Getting the decision right

This line of study is important because rational decision-making under conditions of uncertainty is critical to building systems that understand costs and benefits. By balancing expected costs and benefits, an intelligent system might have been able to do better than humans at planning around the supply chain disruptions the world experienced during the COVID-19 pandemic, managing inventory or serving as a financial adviser.

Our work ultimately shows that if large language models are used for these kinds of purposes, humans need to guide, review and edit their work. And until researchers figure out how to endow large language models with a general sense of rationality, the models should be treated with caution, especially in applications requiring high-stakes decision-making.

About the Author:

Mayank Kejriwal, Research Assistant Professor of Industrial & Systems Engineering, University of Southern California

This article is republished from The Conversation under a Creative Commons license. Read the original article.

By S. Shyam Sundar, Penn State; Cason Schmit, Texas A&M University, and John Villasenor, University of California, Los Angeles 

From fake photos of Donald Trump being arrested by New York City police officers to a chatbot describing a very-much-alive computer scientist as having died tragically, the ability of the new generation of generative artificial intelligence systems to create convincing but fictional text and images is setting off alarms about fraud and misinformation on steroids. Indeed, a group of artificial intelligence researchers and industry figures urged the industry on March 29, 2023, to pause further training of the latest AI technologies or, barring that, for governments to “impose a moratorium.”

These technologies – image generators like DALL-E, Midjourney and Stable Diffusion, and text generators like Bard, ChatGPT, Chinchilla and LLaMA – are now available to millions of people and don’t require technical knowledge to use.

Given the potential for widespread harm as technology companies roll out these AI systems and test them on the public, policymakers are faced with the task of determining whether and how to regulate the emerging technology. The Conversation asked three experts on technology policy to explain why regulating AI is such a challenge – and why it’s so important to get it right.

To jump ahead to each response, here’s a list of each:

Human foibles and a moving target
Combining “soft” and “hard” approaches
Four key questions to ask

Human foibles and a moving target

S. Shyam Sundar, Professor of Media Effects & Director, Center for Socially Responsible AI, Penn State

The reason to regulate AI is not because the technology is out of control, but because human imagination is out of proportion. Gushing media coverage has fueled irrational beliefs about AI’s abilities and consciousness. Such beliefs build on “automation bias” or the tendency to let your guard down when machines are performing a task. An example is reduced vigilance among pilots when their aircraft is flying on autopilot.

Numerous studies in my lab have shown that when a machine, rather than a human, is identified as a source of interaction, it triggers a mental shortcut in the minds of users that we call a “machine heuristic.” This shortcut is the belief that machines are accurate, objective, unbiased, infallible and so on. It clouds the user’s judgment and results in the user overly trusting machines. However, simply disabusing people of AI’s infallibility is not sufficient, because humans are known to unconsciously assume competence even when the technology doesn’t warrant it.

Research has also shown that people treat computers as social beings when the machines show even the slightest hint of humanness, such as the use of conversational language. In these cases, people apply social rules of human interaction, such as politeness and reciprocity. So, when computers seem sentient, people tend to trust them, blindly. Regulation is needed to ensure that AI products deserve this trust and don’t exploit it.

AI poses a unique challenge because, unlike in traditional engineering systems, designers cannot be sure how AI systems will behave. When a traditional automobile was shipped out of the factory, engineers knew exactly how it would function. But with self-driving cars, the engineers can never be sure how it will perform in novel situations.

Lately, thousands of people around the world have been marveling at what large generative AI models like GPT-4 and DALL-E 2 produce in response to their prompts. None of the engineers involved in developing these AI models could tell you exactly what the models will produce. To complicate matters, such models change and evolve with more and more interaction.

All this means there is plenty of potential for misfires. Therefore, a lot depends on how AI systems are deployed and what provisions for recourse are in place when human sensibilities or welfare are hurt. AI is more of an infrastructure, like a freeway. You can design it to shape human behaviors in the collective, but you will need mechanisms for tackling abuses, such as speeding, and unpredictable occurrences, like accidents.

AI developers will also need to be inordinately creative in envisioning ways that the system might behave and try to anticipate potential violations of social standards and responsibilities. This means there is a need for regulatory or governance frameworks that rely on periodic audits and policing of AI’s outcomes and products, though I believe that these frameworks should also recognize that the systems’ designers cannot always be held accountable for mishaps.

Combining ‘soft’ and ‘hard’ approaches

Cason Schmit, Assistant Professor of Public Health, Texas A&M University

Regulating AI is tricky. To regulate AI well, you must first define AI and understand anticipated AI risks and benefits.
Legally defining AI is important to identify what is subject to the law. But AI technologies are still evolving, so it is hard to pin down a stable legal definition.

Understanding the risks and benefits of AI is also important. Good regulations should maximize public benefits while minimizing risks. However, AI applications are still emerging, so it is difficult to know or predict what future risks or benefits might be. These kinds of unknowns make emerging technologies like AI extremely difficult to regulate with traditional laws and regulations.

Lawmakers are often too slow to adapt to the rapidly changing technological environment. Some new laws are obsolete by the time they are enacted or even introduced. Without new laws, regulators have to use old laws to address new problems. Sometimes this leads to legal barriers for social benefits or legal loopholes for harmful conduct.

“Soft laws” are the alternative to traditional “hard law” approaches of legislation intended to prevent specific violations. In the soft law approach, a private organization sets rules or standards for industry members. These can change more rapidly than traditional lawmaking. This makes soft laws promising for emerging technologies because they can adapt quickly to new applications and risks. However, soft laws can mean soft enforcement.

Megan Doerr, Jennifer Wagner and I propose a third way: Copyleft AI with Trusted Enforcement (CAITE). This approach combines two very different concepts in intellectual property — copyleft licensing and patent trolls.

Copyleft licensing allows for content to be used, reused or modified easily under the terms of a license – for example, open-source software. The CAITE model uses copyleft licenses to require AI users to follow specific ethical guidelines, such as transparent assessments of the impact of bias.

In our model, these licenses also transfer the legal right to enforce license violations to a trusted third party. This creates an enforcement entity that exists solely to enforce ethical AI standards and can be funded in part by fines from unethical conduct. This entity is like a patent troll in that it is private rather than governmental and it supports itself by enforcing the legal intellectual property rights that it collects from others. In this case, rather than enforcement for profit, the entity enforces the ethical guidelines defined in the licenses – a “troll for good.”

This model is flexible and adaptable to meet the needs of a changing AI environment. It also enables substantial enforcement options like a traditional government regulator. In this way, it combines the best elements of hard and soft law approaches to meet the unique challenges of AI.

Four key questions to ask

John Villasenor, Professor of Electrical Engineering, Law, Public Policy, and Management, University of California, Los Angeles

The extraordinary recent advances in large language model-based generative AI are spurring calls to create new AI-specific regulation. Here are four key questions to ask as that dialogue progresses:

1) Is new AI-specific regulation necessary? Many of the potentially problematic outcomes from AI systems are already addressed by existing frameworks. If an AI algorithm used by a bank to evaluate loan applications leads to racially discriminatory loan decisions, that would violate the Fair Housing Act. If the AI software in a driverless car causes an accident, products liability law provides a framework for pursuing remedies.

2) What are the risks of regulating a rapidly changing technology based on a snapshot of time? A classic example of this is the Stored Communications Act, which was enacted in 1986 to address then-novel digital communication technologies like email. In enacting the SCA, Congress provided substantially less privacy protection for emails more than 180 days old.

The logic was that limited storage space meant that people were constantly cleaning out their inboxes by deleting older messages to make room for new ones. As a result, messages stored for more than 180 days were deemed less important from a privacy standpoint. It’s not clear that this logic ever made sense, and it certainly doesn’t make sense in the 2020s, when the majority of our emails and other stored digital communications are older than six months.

A common rejoinder to concerns about regulating technology based on a single snapshot in time is this: If a law or regulation becomes outdated, update it. But this is easier said than done. Most people agree that the SCA became outdated decades ago. But because Congress hasn’t been able to agree on specifically how to revise the 180-day provision, it’s still on the books over a third of a century after its enactment.

3) What are the potential unintended consequences? The Allow States and Victims to Fight Online Sex Trafficking Act of 2017 was a law passed in 2018 that revised Section 230 of the Communications Decency Act with the goal of combating sex trafficking. While there’s little evidence that it has reduced sex trafficking, it has had a hugely problematic impact on a different group of people: sex workers who used to rely on the websites knocked offline by FOSTA-SESTA to exchange information about dangerous clients. This example shows the importance of taking a broad look at the potential effects of proposed regulations.

4) What are the economic and geopolitical implications? If regulators in the United States act to intentionally slow the progress in AI, that will simply push investment and innovation — and the resulting job creation — elsewhere. While emerging AI raises many concerns, it also promises to bring enormous benefits in areas including education, medicine, manufacturing, transportation safety, agriculture, weather forecasting, access to legal services and more.

I believe AI regulations drafted with the above four questions in mind will be more likely to successfully address the potential harms of AI while also ensuring access to its benefits.

About the Author:

S. Shyam Sundar, James P. Jimirro Professor of Media Effects, Co-Director, Media Effects Research Laboratory, & Director, Center for Socially Responsible AI, Penn State; Cason Schmit, Assistant Professor of Public Health, Texas A&M University, and John Villasenor, Professor of Electrical Engineering, Law, Public Policy, and Management, University of California, Los Angeles

This article is republished from The Conversation under a Creative Commons license. Read the original article.