Could organizations use artificial intelligence language models such as ChatGPT to induce voters to behave in specific ways?
Sen. Josh Hawley asked OpenAI CEO Sam Altman this question in a May 16, 2023, U.S. Senate hearing on artificial intelligence. Altman replied that he was indeed concerned that some people might use language models to manipulate, persuade and engage in one-on-one interactions with voters.
Altman did not elaborate, but he might have had something like this scenario in mind. Imagine that soon, political technologists develop a machine called Clogger – a political campaign in a black box. Clogger relentlessly pursues just one objective: to maximize the chances that its candidate – the campaign that buys the services of Clogger Inc. – prevails in an election.
While platforms like Facebook, Twitter and YouTube use forms of AI to get users to spend more time on their sites, Clogger’s AI would have a different objective: to change people’s voting behavior.
How Clogger would work
As a political scientist and a legal scholar who study the intersection of technology and democracy, we believe that something like Clogger could use automation to dramatically increase the scale and potentially the effectiveness of behavior manipulation and microtargeting techniques that political campaigns have used since the early 2000s. Just as advertisers use your browsing and social media history to individually target commercial and political ads now, Clogger would pay attention to you – and hundreds of millions of other voters – individually.
It would offer three advances over the current state-of-the-art algorithmic behavior manipulation. First, its language model would generate messages — texts, social media and email, perhaps including images and videos — tailored to you personally. Whereas advertisers strategically place a relatively small number of ads, language models such as ChatGPT can generate countless unique messages for you personally – and millions for others – over the course of a campaign.
Second, Clogger would use a technique called reinforcement learning to generate a succession of messages that become increasingly more likely to change your vote. Reinforcement learning is a machine-learning, trial-and-error approach in which the computer takes actions and gets feedback about which work better in order to learn how to accomplish an objective. Machines that can play Go, Chess and many video games better than any human have used reinforcement learning.
Third, over the course of a campaign, Clogger’s messages could evolve in order to take into account your responses to the machine’s prior dispatches and what it has learned about changing others’ minds. Clogger would be able to carry on dynamic “conversations” with you – and millions of other people – over time. Clogger’s messages would be similar to ads that follow you across different websites and social media.
The nature of AI
Three more features – or bugs – are worth noting.
First, the messages that Clogger sends may or may not be political in content. The machine’s only goal is to maximize vote share, and it would likely devise strategies for achieving this goal that no human campaigner would have thought of.
One possibility is sending likely opponent voters information about nonpolitical passions that they have in sports or entertainment to bury the political messaging they receive. Another possibility is sending off-putting messages – for example incontinence advertisements – timed to coincide with opponents’ messaging. And another is manipulating voters’ social media friend groups to give the sense that their social circles support its candidate.
Second, Clogger has no regard for truth. Indeed, it has no way of knowing what is true or false. Language model “hallucinations” are not a problem for this machine because its objective is to change your vote, not to provide accurate information.
If the Republican presidential campaign were to deploy Clogger in 2024, the Democratic campaign would likely be compelled to respond in kind, perhaps with a similar machine. Call it Dogger. If the campaign managers thought that these machines were effective, the presidential contest might well come down to Clogger vs. Dogger, and the winner would be the client of the more effective machine.
Political scientists and pundits would have much to say about why one or the other AI prevailed, but likely no one would really know. The president will have been elected not because his or her policy proposals or political ideas persuaded more Americans, but because he or she had the more effective AI. The content that won the day would have come from an AI focused solely on victory, with no political ideas of its own, rather than from candidates or parties.
In this very important sense, a machine would have won the election rather than a person. The election would no longer be democratic, even though all of the ordinary activities of democracy – the speeches, the ads, the messages, the voting and the counting of votes – will have occurred.
The AI-elected president could then go one of two ways. He or she could use the mantle of election to pursue Republican or Democratic party policies. But because the party ideas may have had little to do with why people voted the way that they did – Clogger and Dogger don’t care about policy views – the president’s actions would not necessarily reflect the will of the voters. Voters would have been manipulated by the AI rather than freely choosing their political leaders and policies.
Another path is for the president to pursue the messages, behaviors and policies that the machine predicts will maximize the chances of reelection. On this path, the president would have no particular platform or agenda beyond maintaining power. The president’s actions, guided by Clogger, would be those most likely to manipulate voters rather than serve their genuine interests or even the president’s own ideology.
Avoiding Clogocracy
It would be possible to avoid AI election manipulation if candidates, campaigns and consultants all forswore the use of such political AI. We believe that is unlikely. If politically effective black boxes were developed, the temptation to use them would be almost irresistible. Indeed, political consultants might well see using these tools as required by their professional responsibility to help their candidates win. And once one candidate uses such an effective tool, the opponents could hardly be expected to resist by disarming unilaterally.
Enhanced privacy protection would help. Clogger would depend on access to vast amounts of personal data in order to target individuals, craft messages tailored to persuade or manipulate them, and track and retarget them over the course of a campaign. Every bit of that information that companies or policymakers deny the machine would make it less effective.
But there is no reason to automatically extend the First Amendment’s protection to the product of these machines. The nation might well choose to give machines rights, but that should be a decision grounded in the challenges of today, not the misplaced assumption that James Madison’s views in 1789 were intended to apply to AI.
European Union regulators are moving in this direction. Policymakers revised the European Parliament’s draft of its Artificial Intelligence Act to designate “AI systems to influence voters in campaigns” as “high risk” and subject to regulatory scrutiny.
One constitutionally safer, if smaller, step, already adopted in part by European internet regulators and in California, is to prohibit bots from passing themselves off as people. For example, regulation might require that campaign messages come with disclaimers when the content they contain is generated by machines rather than humans.
This would be like the advertising disclaimer requirements – “Paid for by the Sam Jones for Congress Committee” – but modified to reflect its AI origin: “This AI-generated ad was paid for by the Sam Jones for Congress Committee.” A stronger version could require: “This AI-generated message is being sent to you by the Sam Jones for Congress Committee because Clogger has predicted that doing so will increase your chances of voting for Sam Jones by 0.0002%.” At the very least, we believe voters deserve to know when it is a bot speaking to them, and they should know why, as well.
The possibility of a system like Clogger shows that the path toward human collective disempowerment may not require some superhuman artificial general intelligence. It might just require overeager campaigners and consultants who have powerful new tools that can effectively push millions of people’s many buttons.
Lawrence Lessig does not work for, consult, own shares in or receive funding from any company or organization that would benefit from this article, and has disclosed no relevant affiliations beyond their academic appointment.
By: Archon Fung, Professor of Citizenship and Self-Government, Harvard Kennedy School Title: How AI could take over elections – and undermine democracy Sourced From: theconversation.com/how-ai-could-take-over-elections-and-undermine-democracy-206051 Published Date: Fri, 02 Jun 2023 13:42:24 +0000
It’s Halloween. You’ve just finished trick-or-treating and it’s time to assess the haul. You likely have a favorite, whether it’s chocolate bars, peanut butter cups, those gummy clusters with Nerds on them or something else.
For some people, including me, one piece stands out – the Snickers bar, especially if it’s full-size. The combination of nougat, caramel and peanuts coated in milk chocolate makes Snickers a popular candy treat.
As a food engineer studying candy and ice cream at the University of Wisconsin-Madison, I now look at candy in a whole different way than I did as a kid. Back then, it was all about shoveling it in as fast as I could.
Now, as a scientist who has made a career studying and writing booksabout confections, I have a very different take on candy. I have no trouble sacrificing a piece for the microscope or the texture analyzer to better understand how all the components add up. I don’t work for, own stock in, or receivefunding from Mars Wrigley, the company that makes Snickers bars. But in my work, I do study the different components that make up lots of popular candy bars. Snickers has many of the most common elements you’ll find in your Halloween candy.
Let’s look at the elements of a Snickers bar as an example of candy science. As with almost everything, once you get into it, each component is more complex than you might think.
Airy nougat
Let’s start with the nougat. The nougat in a Snickers bar is a slightly aerated candy with small sugar crystals distributed throughout.
One of the ingredients in the nougat is egg white, a protein that helps stabilize the air bubbles that provide a light texture. Often, nougats like this are made by whipping sugar and egg whites together. The egg whites coat the air bubbles created during whipping, which gives the nougat its aerated texture.
A boiled sugar syrup is then slowly mixed into the egg white sugar mixture, after which a melted fat is added. Since fat can cause air bubbles to collapse, this step has to be done last and very carefully.
The final ingredient added before cooling is powdered sugar to provide seeds for the sugar crystallization in the batch. The presence of small sugar crystals makes the nougat “short” – pull it apart between your fingers and it breaks cleanly with no stretch.
Chewy caramel
On top of the nougat layer is a band of chewy caramel. The chewiness of the caramel contrasts the nougat’s light, airy texture, which provides contrast to each bite.
Caramel stands out from other candies as it contains a dairy ingredient, such as cream or evaporated milk. During cooking, the milk proteins react with some of the sugars in a complex series of reactions called Maillard browning, which imparts the brown color and caramelly flavor.
Maillard browning starts with proteins and certain sugars. The end products of these reactions include melanoidins, which are brown coloring compounds, and a variety of flavors. The specific flavor molecules depend on the starting materials and the conditions, such as temperature and water content.
Commercial caramel, like that in the Snickers bar, is cooked up to about 240-245 degrees Fahrenheit (115-118 degrees Celsius), to control the water content. Cook to too high a temperature and the caramel gets too hard, but if the cook temperature is too low, the caramel will flow right off the nougat. In a Snickers bar, the caramel needs to be slightly chewy so the peanuts stick to it.
Chocolate coating
To make chocolate, raw cocoa beans are harvested from cacao pods and then fermented for several days. After the fermented beans are dried, they are roasted to develop the chocolate flavor. As in caramel, the Maillard browning reaction is an important contributor to the flavor of chocolate.
The milk chocolate coating on the Snickers bar happens through a process called enrobing. The naked bar, arranged on a wire mesh conveyor, passes through a curtain of tempered liquid chocolate, covering all sides with a thin layer. Tempering the chocolate coating makes it glossy and gives it a well-defined snap.
The flow of the tempered chocolate needs to be controlled precisely to give a coating of the desired thickness without leading to tails at the bottom of the candy bar.
The Snickers bar
When done right, the result is a delicious Snickers bar, a popular Halloween – or anytime – candy.
With about 15 million bars made each day, getting every detail just right requires a lot of scientific understanding and engineering precision.
Fancy, high-quality products such as Rolex watches and Red Wing boots often cost more to make but last longer. This is a principle that manufacturers and customers are familiar with. But while this also applies to biology, scientists rarely discuss it.
Researchers have known for decades that the faster an animal grows, the shorter its lifespan, at least among mammals. This holds across species of different sizes. Ecophysiologists like me have been studying the trade-offs between allocating energy for growth or for maintenance, and how those trade-offs affect aging and lifespan.
One explanation is that since animals have a limited amount of energy available, investing more energy in growth will reduce the energy they have left to maintain their health, therefore leading to faster aging.
Another explanation is based on the observation that metabolism – all the physical and chemical processes that convert or use energy – fuels growth. Some researchers have suggested that fast growth is associated with high metabolism, in turn causing stress that speeds up aging.
However, these two explanations may not capture the whole picture of the trade-off between growth and longevity. For example, certain species allocate a larger fraction of their energy to maintenance but don’t have better resistance to stress than species that allocate less energy to those processes. This finding indicates that the amount of energy allocated to maintenance may not be the only thing that determines its quality.
Meanwhile, I found that this negative association still strongly holds even after accounting for metabolic rate. That means the higher metabolism associated with faster growth cannot completely explain faster aging. There had to be other missing links to consider.
What have scientists overlooked? My recently published research suggests that the energy cost it takes to make biological materials, or the biosynthetic cost, also affects lifespan.
Cost of making biomass
It costs energy to make biological materials, or biomass, such as assembling individual amino acids into whole proteins. It also costs energy to check newly synthesized materials for errors, break down and rebuild materials with errors, and transport finished materials to where they need to be.
To measure the energy investment in building biomass across species, I derived a mathematical relationship between biosynthetic cost and rates of growth and metabolism. I based my equation on the first principle of energy conservation, which states that energy is neither created nor destroyed, and data on the growth and metabolism rates of different mammals routinely measured by other researchers in the field.
While researchers previously believed that the cost of synthesizing new biomass was the same across species, my analysis of data from 139 different animals found that there is a great difference in biosynthetic cost between species. For example, a naked mole rat has a biosynthetic cost that is over three times as that of a mouse with the same body mass. While the naked mole rat has a lifespan of 30 years, the mouse’s lifespan is only two to three years.
My findings suggest that some species spend more energy than others to make one unit of biomass. This is perhaps partially due to living in a more dangerous environment. Animals that grow faster are more likely to reach reproductive maturity than animals that grow more slowly, but the price to pay is low-quality biomaterials.
Biosynthetic cost and aging
If everything else is kept the same, the more expensive growth is, the lower the growth rate will be. But how does this energy cost contribute to the aging process?
I used what I call a cost-quality hypothesis to answer this question. At the cellular level, biosynthetic cost is in part determined by the cell’s tolerance for errors in making materials. Take proteins as an example. Research has repeatedly suggested that protein homeostasis – the collective processes that maintain protein level, structure and function – plays a key role in the aging process. In simple terms, the accumulation of proteins with errors leads to aging.
Protein synthesis and folding is imperfect. Researchers have estimated that 20% to 30% of new proteins are rapidly degraded after they’re made due to errors. Different species have different degrees of error tolerance and protein quality control. For example, the mouse proteome has two- to tenfold higher levels of proteins with incorrect amino acids relative to the proteome of naked mole rats.
Let’s consider two species, where one is picky about protein errors and the other not so much. The picky species will break down and remake a protein when it finds an error, constantly using protein quality control mechanisms to proofread, quickly unfold and refold, degrade or resynthesize proteins. Not only do these processes cost energy, they also slow down an animal’s overall biomass growth rate. A pickier species would spend more energy for a unit of net new biomass synthesized than a species with high tolerance, growing more slowly overall.
On the other hand, a species with higher tolerance to errors would have a lower biosynthetic cost because it would just incorporate the faulty protein into their new biomass. Because this species can function with faulty proteins, it is more resistant to stress and therefore lives longer.
Making things last
An animal’s ability to maintain homeostasis not only depends on the amount of energy it allocates to maintenance but also on the quality of the tissue it produces. And the quality of that tissue is at least partially due to the energy it invests in making biomass.
In other words, fancy stuff costs more to make but lasts longer.
My hope is that these results could be used as a framework to investigate how differences in a person’s development and growth rate affect their health, risk for aging-related diseases and lifespan. It also opens a door to a new research area: Could we manipulate the mechanisms that determine the energetic cost of biosynthesis and slow aging?
You can probably picture a vampire: Pale, sharply fanged undead sucker of blood, deterred only by sunlight, religious paraphernalia and garlic. They’re gnarly creatures, often favorite subjects for movies or books. Luckily, they’re only imaginary … or are they?
Feeding on a blood diet is unusual for a mammal and has led to many unique adaptations that facilitate their uncommon lifestyle. Unlike other bats, vampires are mobile on the ground, toggling between two distinct gaits to circle their sleeping prey. Heat-sensing receptors on their noseshelp them find warm blood under their prey’s skin. Finally, the combination of a small incision, made by potentially self-sharpening fangs, and an anticoagulant in their saliva allows these bats to feed on unsuspecting prey.
To me, as a behavioral ecologist, who is interested in how pathogens affect social behaviors and vice versa, the most fascinating adaptations to a blood-feeding lifestyle are observable in vampire bats’ social lives.
Such food sharing happens between bats who are related – such as mothers and their offspring – but also unrelated individuals. This observation has puzzled evolutionary biologists for quite a while. Why help someone who is not closely related to you?
It turns out that vampire bats keep track of who feeds them and reciprocate – or not, if the other bat has not been helpful in the past. In doing so, they form complex social relationships maintained by low-cost social investments, such as cleaning and maintaining the fur of another animal, called allogrooming, and higher-cost social investments, such as sharing food.
These relationships are on par with what you would see in primates, and some people compare them to human friendships. Indeed, there are some parallels.
For instance, humans will raise the stakes when forming new relationships with others. You start with social investments that don’t cost much – think sharing some of your lunch – and wait for the other person’s response. If they don’t reciprocate, the relationship may be doomed. But if the other person does reciprocate by sharing a bit of their dessert, for instance, your next investment might be larger. You gradually increase the stakes in a game of back-and-forth until the friendship eventually warrants larger social investments like going out of your way to give them a ride to work when their car breaks down.
Vampire bats do the same. When strangers are introduced, they will start with small fur-cleaning interactions to test the waters. If both partners keep reciprocating and raising the stakes, the relationship will eventually escalate to food sharing, which is a bigger commitment.
Relationships, in sickness and in health
My lab studies how infections affect social behaviors and relationships. Given their vast array of social behaviors and the complexity of their social relationships, vampire bats are the ideal study system for me and my colleagues.
How does being ill affect how vampire bats behave? How do other bats behave toward one that is sick? How does sickness affect the formation and maintenance of their social relationships?
We simulate infections in bats in our lab by using molecules derived from pathogens to stimulate an immune response. We’ve repeatedly found a form of passive social distancing where sick individuals reduce their interaction with others, whether it’s allogrooming, social calling or just spending time near others.
Importantly, these behavioral changes haven’t necessarily evolved to minimize spreading disease to others. Rather, they are parts of the complex immune response that biologists call sickness behaviors. It’s comparable to someone infected with the flu staying at home simply because they don’t feel up to venturing out. Even if such passive social distancing may have not evolved to prevent transmission to others, simply being too sick to interact with others will still reduce the spread of germs.
Interestingly, sickness behaviors can be suppressed. People do this all the time. So-called presenteeism is showing up at work despite illness due to various pressures. Similarly, many people have suppressed symptoms of an infection to engage in some sort of social obligation. If you have little kids, you know that when everyone in your household is coming down with something, there’s no way you can just sit back and not take care of the little ones, even if you feel quite bad yourself.
Animals are no different. They can suppress sickness behaviors when competing needs arise, such as caring for young or defending territory. Despite their tendency to reduce social interactions with others when sick, in vampire bats, sick mothers will continue to groom their offspring and vice versa, probably because mother-daughter relationships are extra important. Mothers and daughters are often each other’s primary social relationships within groups of vampire bats.
Human-bat conflict centers on livestock
Despite their many fascinating adaptations and complex social lives, vampire bats are not universally admired. In fact, in many areas in South and Central America, they are considered pests because they can transmit the deadly rabies virus to livestock, which can cause quite significant economic losses.
Before people introduced livestock into their habitat, vampire bats probably had a harder time finding food in the form of native prey species such as tapirs. Now, livestock has become their primary food source. After all, why not feed on something that is reliably at the same place every night and quite abundant? Increases in livestock abundance come with increases in vampire bat populations, probably perpetuating the problem of rabies transmission.
The farmers’ quarrels with vampires make sense, especially in smaller cattle herds, where losing even one cow can significantly hurt a farmer’s livelihood. Culling campaigns have used topically applied poisons called vampiricide, basically a mix of petroleum jelly and rat poison. Bats are caught, the paste is applied to the fur, and they carry it back to the roost, where others ingest the poison during social interactions. Interestingly, large-scale culling may not be very effective in reducing rabies spillover.
Now, the focus has started to shift toward large-scale cattle vaccinations or vaccinating the vampire bats themselves. Researchers are even considering transmissible vaccines: They could genetically modify herpes viruses, which are quite common in vampire bats, to carry rabies genes and vaccinate large swaths of vampire bat populations.
Whichever method is used to mitigate vampire bat-human conflicts, more empathy for these misunderstood animals could only help. After all, if you stick your head into a hollow tree full of vampire bats – assuming you can brave the smell of digested blood – remember: You’re looking at a complex network of individual friendships between animals that care deeply for each other.