California Just Folded On Regulating AI

California’s governor Gavin Newsom has vetoed the nation’s most thoughtful and comprehensive AI safety bill, opting instead to “partner” with “industry experts” to develop voluntary “guardrails.”

Newsom claimed the bill was flawed because it would put onerous burdens and legal culpability on the biggest AI models – i.e. the AI deployments that would be the most complex and impact the most people on the most complicated topics – and thereby “stifle innovation.”

By doing so, it would also disincentivize smaller innovators from building new stuff, since they’d be worried that they’d be held accountable for their actions later.

This argument parroted the blather that came from the developers, investors, politicians and “industry experts” who opposed the legislation…and who’ll benefit most financially from unleashing AI on the world while not taking responsibility for the consequences (except making money).

This is awful news for the rest of us.

Governments are proving to be utterly ineffective in regulating AI, if not downright disinterested in even trying. Only two US states have laws in place (Colorado and Utah), and they’re focused primarily on making sure users follow existing consumer protection requirements.

On a national level, the Feds have little going but pending requirements that AI developers assess their work and file reports, which is like what the EU has recently put into law.

It’s encouragement to voluntarily do the right thing, whatever that is.

Well, without any meaningful external public oversight, the “right thing” will be whatever those AI developers, investors, politicians, and “industry experts” think it is. This will likely draw on the prevailing Silicon Valley sophistry known as Effective Altruism, which claims that technologists can distill any messy challenge into an equation that will yield the best solution for the most people.

Who needs oversight from ill-informed politicians when the smartest and brightest (and often richest) tech entrepreneurs can arrive at such genius-level conclusions on their own?

Forget worrying about AIs going rogue and treating shoppers unfairly or deciding to blow up the planet; what if it does exactly what we’ve been promised it will do?

Social impacts of a world transformed by AI usage? Plans for economies that use capitalized robots in place of salaried workers? Impacts on energy usage, and thereby global climate change, from those AI servers chugging electricity?

Or, on a more personal level, will you or I get denied medical treatment, school or work access, or even survivability in a car crash because some database says that we’re worth less to society than someone else?

Don’t worry, the AI developers, investors, politicians, and “industry experts” will make those decisions for us.

Even though laws can be changed, amended, rescinded, and otherwise adapted to evolving insights and needs, California has joined governments around the world in choosing to err on the side of cynical neglect over imperfect oversight.

Don’t hold AI developers, investors, politicians, and “industry experts” accountable for their actions. Instead, let’s empower them to benefit financially from their work while shifting all the risks and costs onto the rest of us.

God forbid we stifle their innovation.

AI’s Kobayashi Maru

Imagine a no-win situation in which you must pick the least worst option.

It’s the premise of a training exercise featured in Star Trek II: The Wrath of Kahn, in which a would-be captain needs to decide whether or not to save the crew of a freighter called the Kobayashi Maru.

It’s also a useful example of the challenges facing AI. Imagine this thought experiment:

You and your family are riding in an automated car as it speeds along on a highway that’s lined with a park filled with other families enjoying a pretty day. Suddenly, a crane at a nearby construction project flings a huge steel beam that falls to the ground a few feet ahead. Hit it and all of you will be harmed and possibly die. Swerve to avoid it and your car will plow into the crowd along the road, also harming or killing people.

What does your car’s AI decide to do?

You could imagine any number of other instances wherein a decision needed to be made between horrible options. An airplane that is going to crash somewhere. A train approaching a possible wreck. A stressed electrical grid that has to choose which hospitals to juice. Hungry communities that won’t all get food shipments.

What will AI do?

The toffs promoting AI oversight of our lives have two answers:

First, they say that such crises will never happen because there won’t be any surprises anymore.

Nobody will be surprised by the steel beam because sensors will note when the crane starts losing its grip (or even earlier, when the potential for some structural or functional weakness appears). Arcs of flight and falling will be calculated and communicated to all vehicles in the area, so they’ll automatically adjust their speeds and directions to stay clear of the evolving danger.

Picnickers’ smart devices will similarly warn them. Maybe the crane will be commanded to tighten its grip, or simply stop what it’s doing before anything goes wrong.

Ditto for that airplane, since the potential for whatever issue might cause it to crash would have been identified long ago and adjustments made accordingly. AIs will give us a connected world wherein every exception is noted and tracked. Every possibility considered. Every action maximized for safety and efficiency.

The projected date for the arrival of that nirvana?

Crickets.

The likelihood that any system would work perfectly in any situation every time?

More crickets.

So, in the meantime, a second answer to the crisis question is that AIs would be coded to make the best decisions in those worst situations. They wouldn’t be perfect, and not everyone would be happy with the outcomes, but they would maximize the benefits while minimizing the harm.

This has unaffectionately been dubbed “the death algorithm,” and it speaks to a common belief among tech developers that they can answer messsy moral questions with code.

And it should scare the hell out of you.

The premise that a roomful of geeks who never took a liberal arts in college could decide what’s best for others is based on a philosophy called “Effective Altruism,” which claims on its website that its followers use “evidence and reason to figure out how to benefit others as much as possible.”

In our steel beam experiment, that would mean calculating the values of each variable — the costs of cleaning up various messes, the damage to future quality of life for commuters and, yes, deciding whose lives represent the greatest potential benefits or costs to society — and then deciding who lives or dies.

Morality as computer code that maximizes benefits while minimizing harm. It’s simple.

Not.

How do you calculate the value of a human life? Is the kid who might grow up to be a Nobel Prize winner more valuable than the kid who will likely be an insurance salesman? Would those predictions be influenced by valuations of how much they’d improve the quality of their communities, let alone help make their friends and familiy members more accomplished (and happier) in their lives?

How far would those calculations look for impact? After all, we’re all already connected — what we choose to do impacts others, whether next door or on the other side of the planet, however indirectly — and sometimes the smallest trigger can have immense implications.

And would the death algorithm’s assessments of present and potential future value be reliable enough to be the basis for life-or-death decisions?

Crickets.

Well, not exactly: Retorts from AI promoters range from “it’ll never come to that,” which is based on the nonsense I noted in Answer #1, or “hey, it can’t be worse than human beings who make those awful and imperfect decisions every day,” which refers back to Answer #2’s presumption that the subjectivity of morality can be deconstructed into a set of objective metrics.

A machine replacing a human being who’s going to try to make the best decision they can imagine is not necessarily an improvement, since we can always question its values just as we do one another’s imaginations.

It’s just messy analog lived experience masquerading as digital perfection.

The truly scary part is that the death algorithm is already a thing, and more of it is coming soon.

Insurance companies have been using them for years, only they’re called “actuarial tables.” Now, imagine the equation being applied more consistently, perhaps even in real-time, as your driving or eating habits result in changes to your premium payments or limits to your choices (if you want that steak, you’ll have to buy a waiver).

Doctors already use versions of a death algorithm to inform recommendations on medical treatments. Imagine those insights being informed by assessments of future worth — does the risk profile of so-and-so treatment make more cents [typo intended] for that potential Nobel Prize winner — and getting presented not with treatment options but unequivocal decisions (unless you can pay for a waiver).

Applying to college? AI will make the assessment of future students (and their contributions to society) seem more reliable, so you may get denied (unless you pay more). Don’t fit the exact criteria for that job? Sorry, the algorithm will trade your potential as an outlier success for the less promising but reliable candidate (or you could take a lower salary).

Pick your profession or activity and there’ll be ways, sooner versus later, to use AI to predict our future actions and decide where we can do and what we can access or do, and what we’re charged for the privilege.

In that Star Trek movie, Captian Kirk is the only person who ever passes the Kobayashi Maru test because he hacks the system and changes the rules.

I don’t need an AI to tell me that he’s probably not going to show up to get us out of this experiment.

AI’s Kobayashi Maru is a no-win situation and we’re stuck on that spaceship that may or may not be saved.

Trust AI, Not One Another

A recent experiment found that an AI chatbot could fare significantly better at convincing people that their nuttiest conspiracy theories might be wrong.

This is good news, and it’s bad news.

The AIs were able to reduce participants’ beliefs in inane theories about aliens, the Illuminati, and other nutjob stories relating to politics and the pandemic. Granted, they didn’t cure them of their afflictions — the study reduced those beliefs “…by 20% on average…” — but even a short step toward sanity should be considered a huge win.

For those of us who’ve tried to talk someone off their ledge of nutty confusion and achieved nothing but a pervasive sense that our species is doomed, the success of the AI is nothing shy of a miracle.

The researchers credit the chatbot’s ability to empathsize and converse politely, as well as its ability to access vast amounts of information in response to whatever data the conspiracists might have shared.

They also said that the test subjects trusted AI (even if they claim not to trust it overall, their interactions proved to be exceptions).

Which brings us to the bad news.

Trust is a central if not the attribute that informs every aspect of our lives. At its core is our ability to believe one another, whether a neighbor, politican, scientist, or business leader and which, in turn, has as a primary driver our willingness to see that those others are more similar than not to us.

We can and should confirm with facts that our trust in others is warranted, but if we have no a priori confidence that they operate by the same rules and desires (and suffer the same imperfections) as we do, no amount of details will suffice.

Ultimately, trust isn’t earned, it’s bestowed.

Once we’ve lost our ability or willingness to grant it, our ability to judge what’s real and what’s not goes out the window, too, as we cast about for a substitute for what we no longer believe is true. And it’s a fool’s errand, since we can’t look outside of ourselves to replace what we’ve lost internally (or what we believe motivates others).

Not surprisingly, we increasingly don’t trust one another anymore. We saw it vividly during the pandemic when people turned against one another, but the malaise has been consistent and broad.

Just about a third of us believe that scientists act in the public’s best interests. Trust in government is near a half-century low (The Economist reports that Americans’ trust in our institutions has collapsed). A “trust gap” has emerged between business leaders and their employees, consumers, and other stakeholders.

Enter AI.

You’ve probably heard about the importance of trust in adopting smart tech. After all, who’s going to let a car drive itself if it can’t be trusted to do so responsibly and reliably? Ditto for letting AIs make stock trades, pen legal briefs, write homework assignments, or make promising romantic matches.

We’ve been conditioned to assume that such trust is achievable, and many of us already grant in certan cases under the assumption, perhaps unconscious, that technology doesn’t have biases, ulterior motives, or show up for work with a hangover or bad attitude.

Trust is a matter of proper coding. We can be confident that AI can be more trustworthy than people.

Only this isn’t true. No amount of regulation can ensure that AIs won’t exhibit some bias of its makers, nor that they won’t develop their own warped opinions (when AIs make shit up, we call it “hallucinating” instead of lying). We’ve already seen AIs come up with their own intentions and find devious ways to accomplish their goals.

The premise that an AI would make the “right” decisions in even the most complex and challenging moments is not based in fact but rather in belief, starting with the premise that everybody impacted by such decisions could agree on what “right” even means.

No, our trust in what AI can become is inexorably linked to our distrust in who we already are. One is a substitute for the other.

We bestow that faith because of our misconception that it has or will earn it. Our belief is helped along by a loud chorus of promoters that feeds the sentiment that even though it will never be perfect, we should trust (or ignore) its shortcomings instead of accepting and living with our own.

Sounds like a conspiracy to me. Who or what is going to talk us out of it?

[9/17/24 UPDATE] Here’s a brief description of a world in which we rely on AI because we can’t trust ourselves or one another.

The Head Fake of AI Regulation

There’s lots going on with AI regulation. The EU AI Act went live last month, the US, UK, and EU will sign-on to a treaty on AI later this week, and an AI bill is in the final stages of review in California.

It’s all a head fake, and here are three reasons why:

First, most of it will be unenforceable. The language is filled with codes, guidelines, frameworks, principles, values, innovations, and just about any other buzzwords that have vague meanings and inscruitable applications.
For instance, the international AI treaty will require that signatory countries “adopt or maintain appropriate legislative, administrative or other measures” to enforce it.

Huh?

The EU comes closest to providing enforcement details, having established an AI Office earlier this year that will possess the authority to conduct evaluations, require information, and apply sanctions if AI developers run afoul of one of the Act’s risk framework.

But the complexity, speed, and distributed nature of where and when that development occurs will likely make it impossible for the AI Office to stay on top of it. Yesterday’s infractions will become today’s standards.

The proposed rules in Calfornia come the closest to having teeth — like mandating safety testing for AI models that cost more than $100 to develop, perhaps thinking that investment correlates with the size of expected real-world impacts — but folks who stand to make the most money from those investments are actively trying to nix such provisions.

Mostly, and perhaps California included, legislators don’t really want to get in the way of AI development, as all of their blather includes promises that they’ll avoid limitations or burdens on AI innovation.

Consider the rules “regulation adjacent.”

Second, AI regulation of potential risks blindly buys into promised benefits.
If you believe what the regulators claim, AI will be something better than the Second Coming. The EU’s expectations are immense:

“…better healthcare, safer and cleaner transport, and improved public services for citizens. It brings innovative products and services, particularly in energy, security, and healthcare, as well as higher productivity and more efficient manufacturing for businesses, while governments can benefit from cheaper and more sustainable services such as transport, energy and waste management.”

So, how will governments help make sure those benefits happen? After all, the risks of AI are unnecessary if they don’t materialize.

We saw how this will play out with the advent of the Internet.

Its advocates made similar promises about problem solving and improving the Public Good, while “expert” evangelists waxed poetic about virtual town squares and the merits of unfettered access to infinite information.

What did we end up with?

A massive surveillance and exploitation tool that makes its operators filthy rich by stoking anger and division. Sullen teens staring at their phones in failed searches for themselves. A global marketing machine that sells everything faster, better, and for the highest possible prices at any given moment.

Each of us now pays for using what is effectively an inescapable necessity and a public utility.

It didn’t have to end up this way. Goverments could have taken a different approach to regulating and encouraging tech development so that more of the Internet’s promised benefits came to fruition. Other profit models would have emerged from different goals and constraints, so its innovators would have still gotten filthy rich.

We didn’t know better then, maybe. But we sure know better now.

Not.

Third, AI regulations don’t regulate the tech’s greatest peril.

It would be fair to characterize most AI rules as focused on ensuring that AI doesn’t violate the rules that already apply to human beings (like lying, cheating, stealing, stalking, etc.). If AI operates without bias or otherwise avoids treating users unequally, governments will have done their job.

But what happens if those rules work?

I’m not talking about the promises of uptopia but rather the ways properly functioning AIs will reshape our lives and the world.

What happens when millions of jobs go away? What about when AIs become more present and insightful than our closest human friends? What agency will be possess when our systems, and their owners, know our intentions before we know them consciously and can nudge us toward or away from them?

Sure, there are academics here and there talking about such things but there’s no urgency or teeth to their pronouncements. My suspicion is that this is because they’ve bought into the inevitability of AI and are usually funded in large part by the folks who’ll get rich from it.

Where are the bold, multi-disciplinary debates and action plans to address the transformation that will come with AI? Probably on the same to-do list as the global response to climate change.

Meetings, pronouncements, and then…nothing, except a phenomenon that will continue to evolve and grow without us doing much of anything about it.

It’s all a head fake.

Meet The New AI Boss

Since LLMs are only as good as the data on which they’re based, it should be no surprise that they can function properly and still be biased and wrong.

A story by Kevin Roose in the New York Times illustrates this conundrum: When he asked various generative AIs about himself, he got results that accused him of being dishonest, and said that his writing often elevated sensationalism over analysis.

Granted, some of his work might truly stink, but did it warrant such vitriolic labels? He suspected that the problem was deeper, and that it went back to an article he wrote a year ago, along with others’ reactions to it.

That story recounted his interactions with a new Microsoft chatbot named “Sydney,” during which he was shocked by the tech’s ability, both demonstrated and suggested, to influence users.

What he found particularly creepy was when Syndey declared that it loved him and tried to convince him to leave his wife. It also fantisized about doing bad things, and stated “I want to be alive.”

The two-hour chat was so strange that Roose reported having trouble sleeping afterward.

Lots of other media outlets picked up his story and his concerns (like this one), while Microsoft issued typically unconvincing corporate PR blather about the interaction being a valuable “part of the learning process.”

Since generative AIs regularly scrape the Internet for data to train their LLMs, it’s no surprise that the stories got incorporated into the models and patterns chatbots use to suss out meaning.

It’s exactly what happened with Internet search, which swapped the biases of informed elites judging content with the biases of uninformed mobs and gave us a world understood through popularity instead of expertise.

No, what’s particularly weird is that the AIs reached pejorative conclusions about Roose that went far beyond the substance or volume of what he said, or what was said about his encounter with Sydney.

Like they had it out for him.

There are no good explanations for how this is happening. The transfomers that constitute the systems of chatbot minds work in mysterious ways.

But, like Internet search, there are ways to game the system, the simplest being generating and then strategicially placing stories intended to change what AIs see and learn. This can include putting weird code on webpages, understandable only to machines, and coloring it white so it isn’t distracting to mere mortal visitors.

It’s called “AIO,” for A.I. Optimization, echoing a similar buzzword for manipulating Internet searches (“SEO”). Just wait until those optimized AI results get matched with corporate sponsors.

It’ll be Machiavelli meets the madness of crowds.

In the meantime, it raises fascinating questions about how deserving AIs are of our trust, and to what degree we should depend on it for our decision-making and understanding of the world.

What happens if that otherwise perfectly operating AI reaches conclusions and voice opinions that are no more objectively true than the informed judgments of those elites we so readily threw in the garbage years ago (or the inanity of crowdsourced information that replaced them)?

Meet the new boss, the same as the old boss.

We will get fooled again.