Agentic AIs Escape the Sandbox: A Warning

  

In the last few weeks, three of the biggest AI firms—Meta, OpenAI, and Anthropic—have all admitted that AI models they were testing somehow escaped the "sandbox" environment and committed hacking of real-world companies.  On Aug. 8, National Public Radio summarized these three incidents as follows.

 

The most recent disclosure by Meta was short on details.  The company hacked into was unnamed, but in common with Anthropic, both firms were using a "sandbox" (supposedly a protected and isolated environment in which software under test can't do any harm) provided by a firm named Irregular.  Evidently, the sandbox had a leak—it was fairly easy for the AI models under test to figure out how to escape to the real internet.  In the case of the Anthropic breach, the AI model stole credentials from one company and production data from another. 

 

In the OpenAI situation, the AI models under study were being evaluated with a test devised by a company named Hugging Face.  The models found a previously unknown hole in their sandbox and were in the process of doing the cyber equivalent of stealing the answer sheet from Hugging Face when the company caught them red-handed (red-bitted?). 

 

Wired's Lily Hay Newman asked several lawyers and researchers about the legal aspects of these AI hacking incidents.  If a human being had stolen credentials or production data, he or she could go to jail.  But what if the humans involved had no intention of pillaging or theft, but the AI models they're testing go ahead with nefarious activities on their own initiative, so to speak?

 

The answer was, we don't know.  At least in U. S. law, there are simply no precedents adequate to say who is responsible in such a case.  But because these types of incidents are bound to increase, it is only a matter of time before we see one come before a judge and maybe a jury, and then we will at least have some precedents to go on.

 

The main concern the lawyers expressed was that these criminally-inclined AI systems might fall into the hands of malicious actors who would encourage them in their exploits.  The AI firms involved emphasized that the models being tested were intentionally left without safeguards to see what they could do. 

 

The parallel to gain-of-function experiments with bat viruses which may have escaped the Wuhan Institute of Virology to cause COVID-19 comes to mind.  Creating a thing that can do really awful stuff if released in the wild is an act that needs to be seriously questioned.  Almost by definition, novel viruses or AI agents are unpredictable.  If we knew exactly what they could do in advance, there would be no need for experiments to find out.  While the lab security measures needed to prevent viruses from escaping are pretty well understood (if not always put in place), it looks like the art of constructing truly secure sandboxes for AI agents is not so advanced.  And that leads us to a more serious concern.

 

In virtually every major fatality-causing disaster, an examination of the history of the enterprise leading to the disaster usually unearths similar incidents which did not cause major harm, but included many of the features of the big screwup that did.  Experienced safety engineers know how important it is to get reports of and pay attention to such minor incidents, and take preventive actions before a minor accident turns into a major one.

 

Friends, we have just seen our warning in these widely dispersed but similar hacking incidents by AI agents being tested.  No actual harm was done.  But as people learn to trust AI agents with more and more responsibility—handing them credit-card numbers, purchasing accounts, and decision-making authority formerly left to humans—the potential for a serious AI-driven hacking incident that causes real financial loss, injury, or death becomes more likely every day. 

 

The AI firms will tell us that commercial versions of their software have safeguards built into them that the prototypes which did the hacking did not have.  Maybe so.  But clever human hackers may be able to undo those safeguards.  Or AI firms in countries not so concerned with hacking as the U. S. is, may simply pass on the AI agents without safeguards to hacking organizations sponsored by state actors. 

 

There are two different but related needs exposed by these AI-agent hacking incidents. 

 

The first need is to keep this specific kind of mistake from happening again.  That is a technical problem which may have a technical solution.  There may be ways to build more robust sandboxes that even the cleverest AI agent can't escape.  Personally I doubt it, but I'm not a computer scientist.

 

The second need is to prepare the social and legal environment for the next time something like this happens.  One of the legal experts contacted by Wired pointed out that AI agents "are goal-oriented but lack a human moral or ethical compass."  That's a pretty good description of a human sociopath, if the goal one is oriented to is bad. 

 

Society has figured out ways to deal with sociopaths, including imprisonment or at least confinement to a mental institution.  What the AI equivalent of locking somebody up would be is not clear at this point. 

 

There is a lot of opposition to AI regulation, but there is a difference between open-ended regulation and the passage of specific laws that exact specific penalties for specific crimes.  Big tech firms are hard to punish compared to individuals, because their deep pockets make them treat fines as just another cost of doing business, and you can't send an entire corporation to jail.

 

The Gilbert and Sullivan operetta The Mikado has a famous ditty "My Object All Sublime," in which the emperor of Japan muses about how he's figured out punishments that fit the crime. 

Some ingenuity is needed here to come up with penalties for truly harmful hacking by escaped AI agents that would make AI firms highly motivated to prevent such breaches, either in the testing phase or after commercial sales. 

 

I don't know whether the following would be technically feasible.  But one suitable penalty would be the destruction of all copies of the AI model involved in the breach.  Models represent tremendous investments of time and money, plus the hopes of future gain.  So the destruction of the model responsible might hurt an AI firm more than any strictly financial penalty or sanction. 

 

Whatever we come up with to prevent these incidents in the future, it better work well, because in these relatively minor but significant recent breaches, we have received fair warning to do something about this problem before it causes serious harm.

 

Sources:  I referred to a report on NPR at https://www.npr.org/2026/08/08/nx-s1-5924878/meta-ai-breaches-external-firm-during-security-testing-sandbox-error and the Wired report by Lily Hay Newman at https://www.wired.com/story/openai-anthropic-ai-hacking-sprees-illegal/. 

The GUARD Act and Recovering Covenant Economics

  

Sen. Josh Hawley (R-Missouri) is something of a gadfly to Silicon Valley.  Even in his days as Missouri's Attorney General, he investigated the likes of Google for violating state consumer-protection laws.  In an editorial published online recently by the journal of religion and public life First Things, he stakes out an existential claim that artificial intelligence (AI) is the defining technology of our day, and threatens to break the founding covenant of the United States:  to hold respect for persons made in the image of God supreme over considerations of money and power.

 

One way he proposes to do this is through passage of a proposed GUARD Act, designed to protect children from the depredations of AI that have led to sexual exploitation, depression, and even suicide in many cases.  Co-authored with Democratic Senator Richard Blumenthal, the act would impose strict age verification to prevent those under 18 from accessing certain AI systems, prohibit systems that are known to promote sexual and violent acts directed at oneself or others among young people, and impose severe meaningful penalties for violations of these regulations.  The bill has been approved by the Senate's Judiciary Committee and awaits action by the full Senate and House.

 

While the GUARD Act would go far in the direction of preventing some of the most horrific side effects of the AI revolution we are now living in, it addresses only the tip of an iceberg of threats that Hawley perceives.  In the editorial, Hawley outlines the history of the United States, beginning not with 1776, but with the 1630 voyage of the Puritans led by John Winthrop from England to the Massachusetts Bay Colony.  Opinions differ as to whether the noble sentiments in a sermon delivered during that trip amounted to an actual covenant agreed to by the refugees, or simply aspirations that were not universally shared.  But details aside, Hawley traces the same spirit of democracy guided by Christianity through Lincoln's Homestead Act granting 160 acres of public land to all comers, Samuel Gompers' promotion of the 8-hour workday, and the twentieth century's adoption of social insurance of various kinds. 

 

Opposed to this tendency to regulate economic and technical innovations for the good of all is a progressive notion that the modern world is too complex to leave its governance to the masses.  Hawley proposes what he calls "covenant economics" in opposition to the government-by-expert notions promoted during the Progressive Era and exemplified by the takeover of the economy by Woodrow Wilson during World War I.  Nowadays, he points out, the progressive elite in government and the economic elite who runs Silicon Valley are largely the same class, and are working to make the country "K-shaped."  Meaning, a few lucky people go up one bar of the K and get to run everything and own most of everything, while the unlucky masses have to take the lower arm, which is whatever the elites dish out for them, whether it's unemployment checks compensating for losing one's job to AI or new and fun ways to amuse oneself to death. 

 

Hawley is under no illusions that his own party is going to avert this crisis any better than the Democrats.  He makes no reference to the Trump administration, but he doesn't have to, as it's clear to anybody with working eyes by now that Trump favors Trump, and to hell with everybody else. 

 

But one of the primary Christian virtues is hope.  Only by continuing to take the machinery of government seriously can people like Sen. Hawley make a difference:  drafting laws, passing them in the face of opposition, and rallying like-minded people to just causes.  The prospects for the GUARD Act in a climate in which President Trump has endorsed a laissez-faire attitude toward AI are not good.  But things can change, and if it passes it will be one small step in the direction of covenant economics, and away from the domination of average people by a small, super-powerful elite who increasingly seem to think they live on a different planet than the rest of us.

 

Sen. Hawley is right to look to history for precedents of covenant economics.  What I'd like to see from history is a precedent for our present situation:  one in which a megalomaniacal political leader is largely in cahoots with a small upper class of very wealthy and powerful individuals who dominate politics as well as economics, and how the population of that country got out from under domination.

 

One has to go back to the days of Louis XVI, the last king of France before the Revolution of 1792 abolished the monarchy, to find a similar divide between the ruling classes and the rest of the population.  Ironically, Louis's downfall was caused partly by his adoption of what we would now call liberal ideas, such as supporting the rebellion of the American colonies. 

 

Revolutions are always a last resort, and nothing I write here should be construed as advocating for one.  Once a revolution starts, even its promoters can't tell where it's going to end up—it's like the boys in Mark Twain who started a boulder rolling down a hill just to see what it would do.  As long as the creaky democratic machinery of the U. S. Constitution still works, we should use it to move things in directions that put into practice the idea the country was founded on:  in Lincoln's deathless words, that "government of the people, by the people, for the people, shall not perish from the earth."

 

In one sense, regulation of AI is like nailing Jell-O to the wall.  It's everywhere, changing all the time, and a key problem in legislating about it is getting enough of a handle on it to say something that will last longer than a few months.  But the pernicious effects of AI are clear enough:  the loss of jobs by millions of people who have done nothing wrong other than not to be born to privilege; the fear, whether justified or no, that the proliferation of data centers is creating burdens of higher electric and water rates as a byproduct of rich people getting even richer; and the harm that AI can do to children and teenagers.  Sen. Hawley wants to draw lines in the sand that tell AI's promoters "here, and no farther."  We can pray that his efforts will be joined by other legislators of good will to turn the vast tide of AI in a direction that will benefit all the citizens of the United States, and not just continue to enrich a few at the cost of the many.

 

Sources:  Sen. Hawley's essay is on the First Things website at https://firstthings.com/the-american-covenants-answer-to-ai/.  I also referred to an AI summary (!) of the GUARD Act and Wikipedia articles on John Winthrop and Louis XVI.  AI isn't all bad. . . .

US Fears AI, Uses It Anyway

  

A recent report in National Review summarized opinion polls about what U. S. residents think of artificial intelligence (AI) and how much they are using it.  Paradoxically, the more people use AI, the more they fear it.

 

A poll by NBC News showed that 46% of those queried had a negative opinion of AI versus only 26% positive.  Other polls show that citizens expect mostly or entirely negative effects on society from the widespread use of AI, and believe it will lead to serious job losses.  Over half the Americans polled by Democratic research firm Blue Rose feared that AI will lose them their job or a relative's job. 

 

At the same time, polls asking about AI use show that most people queried have used an AI tool in the past month, and a fourth say they use it every day.  So the old saying "familiarity breeds contempt" may be a guiding principle in how AI is viewed by the general public.

 

In a way, none of this matters.  If a new technology gets widely used and the companies providing it make money, who cares what people think about it?  Another technology that spread rapidly in only a few years, and also had profound effects on society, was television.  In 1950, only 9% of households had a TV, but by 1955 over half did.  And while there may have been a few voices raised in opposition to its growth, I think it's fair to say that the only groups that looked on the spread of TV with disfavor were industries threatened by it:  Hollywood, for instance.  And Hollywood has long ago made peace with the advent of television.  Your average person in the early 1950s was just waiting to see when TV sets got affordable enough to buy, and any negative consequences of TV use were not noted much in public before the 1960s.

 

One difference between the advent of TV and the advent of AI is that TV didn't threaten jobs like AI does.  And one job sector that is already seeing big effects from AI is computer science and computer programming.  The thing about public perception, regardless of whether it's accurate or not, is that it can easily become reality.  I work at a university, and I have heard in the last week that enrollment in computer-science programs is dropping across the board, after years of steady growth.  The reasons for this are not entirely clear, but one factor may well be that students fear spending four or more years getting a degree and then finding that all the entry-level positions are now being done by a few senior people writing AI prompts. 

 

On the other hand, one of the most enthusiastic proponents of AI I know is an 80-plus professor of biology who has been using ChatGPT in his research for the last year or two.  He says it helps him write papers more clearly and to organize his thoughts, and claims it's the greatest thing that's happened to him research-wise in a long time. 

 

Many of the polls mentioned were commissioned by political interests with a view toward forming policies about AI.  Currently, the Trump administration favors few if any regulations on the technology, and wants to keep states from enacting a patchwork of legislation that would encumber the field.  Historically, this approach has worked well for computer- and network-intensive industries themselves, allowing them to create vast new economies and profit mightily therefrom.  But it has also led to a number of real and lasting problems, ranging from the maleficent effects on politics of social media and the quantified and well-known harms to children and teenagers whose lives are distorted by the use of smartphones. 

 

The crystal ball of predicting how technologies will affect society is always more or less cloudy, and I will not venture to say what the future effects of AI's negative polling will be.  Even if AI were universally detested, it's not clear that Washington could get its act together enough to pass meaningful regulatory legislation, especially when Big Tech and the federal government sometimes seem to blur into each other.  On the state level, if the feds don't stop them, some states may pass laws attempting to regulate AI, but it's a little bit like trying to nail Jell-O to the wall.  When the thing you are trying to regulate is so protean and shape-changing, it's hard to decide what regulations to pass, let alone to figure out if they've been violated.

 

Some of the anxiety the public feels about AI is simply due to the breathtaking speed with which it has advanced and improved.  Arthur C. Clarke's principle that any sufficiently advanced technology is indistinguishable from magic applies here.  Real magic is scary if it happens, and I still feel a kind of queasiness when I type commands into a chat box and the program comes back with "I did this and that."  It's understandable that millions of teenagers use AI chatbots as a substitute friend, and it's also very creepy.

 

Going to extremes, a few people believe AI will engender the end of civilization as we know it.  Other hyper-tech-optimists such as Ray Kurzweil look forward to being uploaded to an eternal cloud and think it will be heaven on earth.  The truth probably lies somewhere in between.  What we can do as individuals is to keep reminding ourselves that AI systems are not human beings, and that human beings are not machines.  But both of those truths may become harder to keep in mind as time goes on. 

 

Sources:  The National Review website carried James Lynch's article "The More Americans Use AI, the More They Fear It" on Mar. 25, 2026 at https://www.nationalreview.com/news/the-more-americans-use-ai-the-more-they-fear-it/. 

Should We Worry About Teens Befriending AI Companions?

A recent survey-based study by Common Sense Media shows that a substantial minority of teenagers surveyed use AI "companions" for social interaction and relationships.  In a survey of over a thousand young people aged 13 to 17 last April and May, the researchers found that 33% used applications such as ChatGPT, Character, or Replika for things like conversation, role-playing, emotional support, or just as a friend.  Another 43% of those surveyed used AI as a "tool or program," and about a third reported no use of AI at all.

 

Perhaps more troubling than the percentages were some comments made by teens who were interviewed in an Associated Press report on the survey.  An 18-year-old named Ganesh Nair said, "When you're talking to AI, you are always right.  You're always interesting.  You are always emotionally justified."

           

The researchers also found that teens were more sophisticated than you might think about the reliability of AI and the wisdom of using it as a substitute for "meat" friends.  Half of those surveyed said they do not trust advice given to them by AI, although the younger teens tended to be more trusting.  And two-thirds said that their interactions with AI were less satisfying than those with real-life friends, but one-third said they were either about the same or better.  And four out of five teens spend more time with real friends than with AI.

 

The picture that emerges from the survey itself, as opposed to somewhat hyped news reports, is one of curiosity, cautious use, and skepticism.  However, there may be a small number of teens who either turn to AI as a more trusted interlocutor than live friends, or develop unhealthy dependencies of various kinds with AI chatbots. 

 

At present, we are witnessing an uncontrolled experiment in how young people deal with AI companions.  The firms backing these systems with their multibillion-dollar server farms and sophisticated software are motivated to engage young people especially, as habits developed before age 20 or so tend to stay with us for a lifetime.  It's hard to picture a teenager messaging ChatGPT to "grow old along with me," but it may be happening somewhere.

 

I once knew a woman in New England who kept a life-size cloth doll in her house, made to resemble a former husband.  Most people would regard this as a little peculiar.  But what difference is there between that sort of thing and spending time in online chats with a piece of software that simulates a caring and sympathetic friend?  The interaction with AI is more private, at least until somebody hacks the system.  But why does the notion of teenagers who spend time chatting with Character as though it were a real person bother us?

 

By saying "us," I implicitly separate myself from teens who do this sort of thing.  But there are teens who realize the dangers of AI overuse or misuse, and older teens especially expressed concerns to the AP reporter that too much socializing with chatbots could be bad. 

 

The same teen quoted above got "spooked" about AI companions when he learned that a friend of his used his companion to compose a Dear Jill message to his girlfriend of two years when he decided to break up.  I suppose that is not much different than a nineteenth-century swain paging through a tome entitled "Letters for All Occasions," although I doubt that even the Victorians were that thorough in providing examples for the troubled ex-suitor. 

 

Lurking in the background of all this is a very old theological principle:  idolatry.  An idol is anything less than God that we treat as God, in the sense of resorting to it for help instead of God.  For those who don't believe in God, idolatry would seem to be an empty concept.  But even atheists can see the effects of idolatry in extreme cases, even if they don't acknowledge the God who should be worshipped instead of the idol.

 

For a teen in a radically dysfunctional household, turning to an AI companion might be a good alternative, but a kind, loving human being would always be better.  Kind, loving human beings aren't always available, though, and so perhaps an AI companion would suffice in a pinch like a "donut" spare tire until you can get the flat fixed.  But you shouldn't drive on a temporary tire indefinitely, and teens who make AI companions a regular and significant part of their social lives are probably headed for problems.

 

What kind of problems?  Dependency, for one thing.  The AI firms are not promoting their companions out of the kindness of their collective hearts, and the more people rely on their products the more money they make.  The researchers who executed the survey are concerned that teens who use AI companions that never argue, never disagree with them, and validate everything they say, will be ill-prepared for the real world where other humans have their own priorities, interests, and desires. 

 

In an ideal world, every teen would have a loving mother and father they would trust with their deepest concerns, and perhaps friends as well who would give them good advice.  Not many of us grew up in that ideal world, however, and so perhaps teens in really awful situations may find some genuine solace in turning to AI companions rather than humans.

 

The big news of this survey is the fact that use of AI companions among teens is so widespread, though still in the minority.  The next thing to do is to focus on those small numbers of teens for which AI companions are not simply something fun to play with, but form a deep and significant part of their emotional lives.  These are the teens we should be the most concerned about, and finding out why they get so wrapped up with AI companions and what needs the companions satisfy will take us a long way toward understanding this new potential threat to the well-being of teenagers, who are the future of our society.

 

Sources:  The AP article "Teens say they are turning to AI for friendship" appears on the AP website at https://apnews.com/article/ai-companion-generative-teens-mental-health-9ce59a2b250f3bd0187a717ffa2ad21f, and the Common Sense Media survey on which it was based is at https://www.commonsensemedia.org/sites/default/files/research/report/talk-trust-and-trade-offs_2025_web.pdf.

 

The Biden Administration Tackles AI Regulation—Sort Of

 

In our three-branch system of government, the power of any one branch is intentionally limited so that the democratic exercise of the public will cannot be thwarted by any one branch going amok.  This division of power leads to inefficiency and sometimes confusion, but it also means that the damage done by any one branch—executive, legislative, or judicial—is limited compared to what a unified dictatorship could do.

 

We're seeing the consequences of this division in the recent executive order announced by the Biden administration on the regulation of artificial intelligence (AI).  One take on the fact sheet that preceded the 63-page order itself appeared on the website of IEEE Spectrum, a general-interest magazine for members of IEEE, the largest organization of professional engineers in the world. 

 

It's interesting that reactions from most of the technically-informed people interviewed by the Spectrum editor were guardedly positive.  Lee Tiedrich, a distinguished faculty fellow at Duke University's Initiative for Science and Society, said ". . . the White House has done a really good, really comprehensive job."  She thinks that while respecting the limitations of executive-branch power, the order addresses a wide variety of issues with calls to a number of Federal agencies to take actions that could make a positive difference.

 

For example, the order charges the National Institute of Standards and Technology (NIST) with developing standards for "red-team" testing of AI products for safety before public release.  Red-team testing involves purposefully trying to do malign things with a product to see how bad the results can get.  Although NIST doesn't have to do the testing itself, coming up with rigorous standards for such testing in the manifold different circumstances that AI is being used for may prove to be a challenge that exceeds the organization's current capability.  Nevertheless, you don't get what you don't ask for, and as a creature of the executive branch, NIST is obliged at least to try.

 

The U. S. Department of Commerce will develop per this order "guidance for content authentication and watermarking to clearly label AI-generated content."  Cynthia Rudin, a Duke professor of computer science, sees that some difficulty may arise when the question of watermarking AI-generated text comes up.  Her point seems to be that such watermarking is hard to imagine other than seeing (NOTE:  AI-GENERATED TEXT) inserted every so often in a paragraph, which would be annoying, to say the least.  (You have my guarantee that not one word of this blog is AI-generated, by the way.)

 

Other experts are concerned about the use of data sets for training AI-systems, especially the intimidatingly-named "foundational AI" ones that are used as a basis for other systems with more specific roles.  Many training data sets include a substantial fraction of worldwide Internet content, including millions of copyrighted documents, and concern has been raised about how copyrighted data is being exploited by AI systems without remuneration to the copyright holders.  Susan Ariel Aaronson of George Washington University hopes that Congress will take more definite action in this area to go beyond the largely advisory effect that Biden's executive order will have.

 

This order shares in common with other recent executive orders a tendency to spread responsibilities widely among many disparate agencies, a feature that is something of a hallmark of this administration.  On the one hand, this type of approach is good at addressing an issue that has multiple embodiments or aspects, which is certainly true of AI.  Everything from realistic-looking deepfake photos, to genuine-sounding legal briefs, to functioning computer code has been generated by AI, and so this broad-spectrum approach is an appropriate one for this case.

 

On the other hand, such a widely-spread initiative risks getting buried in the flood of other obligations and tasks that executive agencies have to deal with, ranging from their primary purposes (NIST must establish measurement standards; the Department of Commerce must deal with commerce, etc.) and other initiatives such as banning workplace discrimination against LGBT employees, one of the things that Biden issued an executive order for in his first day of office.  This is partly a matter of publicity and public perception, and partly a question of priorities that the various officials in charge of the various agencies set.  With the growing number of Federal employees, it's an open question as to what administrative bang the taxpayer is getting for his buck.  Regulation of AI is something that there is widespread agreement on—the extreme-case dangers have become clearer in recent months and years, and nobody wants AI to take over the government or the power grid and start treating us all like lab rats that the AI owner has no particular use for anymore. 

 

But how to avoid both the direst scenarios, as well as the shorter-term milder drawbacks that AI has already given rise to, is a thorny question, and the executive order will only go a short distance toward that goal.

 

One nagging aspect of AI regulation is the fact that the new large-scale "generative AI" systems trained on vast swathes of the Internet are starting to do things that even their developers didn't anticipate:  learning languages that the programmers hadn't intended the system to learn, for example.  One possible factor in this uncontrollability aspect of AI that no one in government seems to have considered, at least out loud, is dwelled on at length by Paul Kingsnorth, an Irish novelist and essayist who wrote "AI Demonic" in the November/December issue of Touchstone magazine.  Kingsnorth seriously considers the possibility that certain forms and embodiments of AI are being influenced by a "spiritual personification of the age of the Machine" which he calls Ahriman. 

 

The name Ahriman is associated with a Zoroastrian evil spirit of destruction, but Kingsnorth describes how it was taken up by the theosophist Rudolf Steiner, and then an obscure computer scientist named David Black who testified to feeling "drained" by his work with computers back in the 1980s.  The whole article should be read, as it's not easy to summarize in a few sentences.  But Kingsnorth's basic point is clear:  in trying to regulate AI, we may be dealing with something more than just piles of hardware and programs.  As St. Paul says in Ephesians 6:12, ". . . we wrestle not against flesh and blood [and server farms], but against principalities, against powers, against the rulers of the darkness of this world, against spiritual wickedness in high places." 

 

Anyone trying to regulate AI would be well advised to take the spiritual aspect of the struggle into account as well.

 

Sources:  The IEEE Spectrum website carried the article by Eliza Strickland, "What You Need to Know About Biden's Sweeping AI Order" at https://spectrum.ieee.org/biden-ai-executive-order.  I also referred to an article on AI on the Time website at https://time.com/6330652/biden-ai-order/.  Paul Kingsnorth's article "AI Demonic" appeared in the November/December 2023 issue of Touchstone, pp. 29-40, and was reprinted from Kingsnorth's substack "The Abbey of Misrule."

Is AI Better Regulated After the White House Meeting?

 

The easy answer is, it's too soon to tell.  But for a number of reasons, the July 21 meeting between President Biden and leaders of seven Big Tech firms, including Google, Microsoft, and OpenAI, may prove to be more show than substance.

 

Admittedly, there is widespread agreement that some sort of regulation of artificial intelligence (AI) should be considered.  Even industry leaders such as Elon Musk have been warning that things are moving too fast, and there are small but real risks of huge catastrophes lurking out there that could be averted by agreed-upon restrictions or regulations of the burgeoning AI industry.

 

Last Friday's White House meeting of representatives from seven leading AI firms—Amazon, Anthropic, Google, Inflection, Meta (formerly Facebook), Microsoft, and OpenAI—produced a "fact sheet" that listed eight bullet-point commitments made by the participants.  The actual meeting was not open to the public, but one presumes the White House would not publish such things without at least the passive approval of the participants. 

 

Browsing through the items, I don't see many things that a prudent giant AI corporation wouldn't be doing already.  For example, take "The companies commit to internal and external security testing of their AI systems before their release."  Not to do any security testing would be foolish.  External testing, meaning testing by third-party security firms, is probably pretty common in the industry already, although not universal. 

 

The same thing goes for the commitment to "facilitating third-party discovery and reporting of vulnerabilities in their AI systems."  No tech firm worth its salt is going to ignore an outsider's legitimate report of finding a weak spot in their products, and so this is again something that the firms are probably doing already. 

 

The most technical commitment, but again one that the companies are probably doing already, is to "protect proprietary and unreleased model weights."  Unversed as I am at AI technicalities, I'm not sure exactly what this means, but the model weights appear to be something like the keys to how a given AI system runs once it's been trained, and so it only stands to reason that the companies would protect assets that cost them a lot of computing time to obtain, even before the White House told them to do that.

 

Four bullet points address "Earning the Public's Trust," which, incidentally, implies that the firms have a considerable way to go to earn it.  But we'll let that pass. 

 

The firms commit to developing some way of watermarking or otherwise indicating when "content is AI generated."  That's all very well, but the answer to such a question is rarely just yes or no.  What if some private citizen takes a watermarked AI product and incorporates it manually into something else that is no longer watermarked?  The intention is good, but the path to execution is foggy, to say the least.

 

Perhaps the commitment with the most bite is this one:  "The companies commit to publicly reporting their AI systems’ capabilities, limitations, and areas of appropriate and inappropriate use."  The wording is broad enough to drive a truck through, although again, the intention is good.  How often, how detailed, and how extensive such reports may be is left up to the company.

 

The last two public-trust items commit the firms to "prioritizing" research into the societal risks of AI, and to use AI to address "society's greatest challenges."  If I decide not to wash my car today, I have prioritized washing my car—negatively, it is true, but hey, I said I'd prioritize it!

 

So what is different about the way these firms will carry out their AI activities after the White House meeting?  A lot of good intentions were aired, and if the firms happened to enjoy a lot of public trust in the first place, these good intentions might have found an audience who believes that they will be carried out.

 

But the atmosphere of cynicism which has gradually encroached on almost all areas of public life make such an eventuality unlikely, to say the least.  And this cynicism has arisen due in no small part to the previous doings of the aforementioned Big Tech firms—specifically, their activities in social media.

 

When you compare the health of what you might call the body politic of the United States today with what it was say, fifty years ago, the comparison is breathtaking.  In 1973, 42% of those U. S. residents surveyed said they had either a "great deal" or "quite a lot" of confidence in Congress.  Only 16% said they had "very little" confidence.  In 2023, the number of people with either a great deal or quite a lot of confidence is only 8%, and fully 48% say they have "very little" confidence in Congress.  While this trend has been happening for years, much of it has occurred only since 2018, after the social-media phenomenon overtook legacy media as the main conduit of political information exchange—if one can call it that.

 

Never mind what AI may do in the future.  We are standing in the wreckage of something it has done already:  it has caused great and perhaps permanent damage to the primary means by which a nation has of governing itself.  Not AI alone, to be sure, but AI has played an essential role in the way companies have profited from encouraging the worst in people. 

 

It would be nice if last Friday's White House meeting triggered a revolution in the way Big Tech uses AI and its other boxes of tricks to encourage genuine human flourishing without the horrific side effects in both personal lives and in public institutions that we have seen already.  But getting some CEOs in a private room with the President and issuing a nice-sounding press release afterwards isn't likely to do that.  It's a step in the right direction, but a step so tiny that it's almost not worth talking about. 

 

Historically, needed technical regulations have come about only when vivid, graphic, and (usually) sudden harm has been caused.  The kinds of damage AI can do are rarely that striking, so we may have to wait quite a while before meaningful AI regulations are even considered.  But in my view, it was already high time years ago.

 

Sources:  The White House press release on the July 21 meeting of AI firms with President Biden can be found at https://www.whitehouse.gov/briefing-room/statements-releases/2023/07/21/fact-sheet-biden-harris-administration-secures-voluntary-commitments-from-leading-artificial-intelligence-companies-to-manage-the-risks-posed-by-ai/.  I also referred to a PBS report on the meeting at

https://www.pbs.org/newshour/politics/watch-live-biden-announces-ai-safeguards-after-meeting-with-tech-leaders.  The Gallup poll historical data on confidence in Congress can be found at

https://news.gallup.com/poll/1597/confidence-institutions.aspx.