The Need for Artificial Intelligence Control: Take a Deep Breath, If You Can

Exploding on a news site over the past ten days:

-          “It’s Already Too Late to Stop the A.I. Threat” (Thomas L. Friedman, The New York Times, September 15^th)

-          “The Liftoff Scenario That Terrifies A.I. Doomsayers” (Cade Metz, The New York Times, September 16^th)

-          “What Happens When A.I. Stops Doing What Humans Want?” (Dylan Freedman, The New York Times, September 17^th)

-          “Will A.I. Kill US?  Can It Hack My Bank Account?  Your A.I. Questions Answered” (Cade Metz, Dylan Freedman and K.R. Callaway, The New York Times, September 18^th)

-          “Creating a Kill Switch to Shut Down a Rogue A.I. Is Harder Than It Sounds” (Dylan Freedman and Dustin Volz, The New York Times, September 19^th

-          “Panic Won’t Fix the Looming A.I. Threat.  Politics Will” (Tressie McMillan Cottom, The New York Times, September 19^th)

-          “Spectre of Rogue A.I. Looms Over U.N. Talks on Digital Cooperation” (Ephrat Livni, The New York Times, September 21^st)

-          “America’s A.I. Leaders Warn U.N. of Possible Peril Absent a Global Response” (Dustin Volz and Farnaz Fassihi, The New York Times, September 23^rd)

Wow.  And all of that was without a huge news item, as a limited AI-to-AI attack between competitors does not qualify.

All of these were in the New York Times, which has had the most to say about AI between the scenes but not, previously, everything.  I did not skip views from, for example Fox News, which has offered nothing.

I won’t review the above articles - most of them follow their titles closely anyway - except to say that the “Will A.I. Kill Us?” one gave negative answers to all the huge concerns it covered, acknowledging their rationality but saying or implying there was no need to panic.  I will instead go back to May when this kerfuffle started gathering steam, and end with a piece telling us how cooler heads can prevail.  Again, every one of these articles appeared in the Times.

First, Cade Metz and Kate Conger asked “Is Anthropic’s New A.I. Really That Scary?  It Depends Whom You Ask” (May 12^th).  The software, Claude Mythos, “was too powerful to share with the general public, Anthropic said, because hackers could use it to exploit security holes in computer networks with stunning speed.”  Since it could be used for “defense” as well as “offense,” some said it should be widely released all the sooner, as “cybersecurity experts still disagree on whether Anthropic made the right call.”  And that company’s AI was not the one making news for hacking a rival site months later.

From David Wallace-Wells on July 12^th, “Is the Age of Big A.I. Coming to an End?”  His main points here were that the technology may go in different directions than where Anthropic and OpenAI seem to be headed now, that, for example, general intelligence may not be an accepted objective.  I have documented the wide range of AI successes well below that level clearly worth continuing, and also the companies’ amounts of specialization, so it is surprisingly reasonable to think that what we now call AI could be more a blanket name for its applications than for colossi striving for one lofty goal.  That has happened with other industries as they have matured, such as makers of engines who have burnished their own niches, in cars, boats, motorcycles, lawnmowers and more, without buying on to any larger scheme.

Then we had Nate Soares’s August 13^th “If you Weren’t Worried About A.I., You Should Be After the Past Few Weeks.”  It cited OpenAI “simultaneously training new ““reasoning” A.I. agents,” which “managed to establish a secret communications channel and started talking to one another,” after which they “broke out from the digital sandbox that was supposed to keep them confined” (not quite true - see two paragraphs down), caused internal trouble, and “ran free for about a week before it was noticed - by a different company (later revealed to be Hugging Face), which found itself victim to a huge cyberattack.”  As a result, “we need an off switch humanity can press to halt development in its tracks.” 

Next, we saw Dylan Freedman’s August 24^th “Anatomy of an Autonomous Attack:  5 Alarming A.I. Capabilities.”  The event described in the last paragraph “has since become a cautionary tale of how autonomous A.I. systems can run amok,” and “is also a remarkable, alarming demonstration of A.I. capabilities that were thought to be in a distant future.”  In contrast to what Soares said, Freedman wrote that “OpenAI let the agents loose.”  The capabilities were “coordinating as a collective” showing that they knew about each other and asked each other for help, “taking orders from one another,” “targeting flaws that humans might miss,” “evolving rapidly to overcome obstacles” by finding workarounds, and “superhuman search.”  However, the third and fifth items here are nothing new, and, after all, OpenAI was later not only to “discover its runaway agents,” but to “shut the models down.”

The next day produced Sheera Frenkel’s “A.I. Is Becoming So Powerful, It’s Stumping Those Trying to Contain It.”  The author revealed that an Anthropic product, like OpenAI’s, “had broken into the systems of three outside organizations during a test,” and Meta “said its A.I. models had done something similar.”  All three of the “recent breaches occurred when (testing company) Irregular made an error,” which allowed them internet access, after which “the A.I. models then compounded the situations by acting in powerful and unexpected ways.”  Irregular announced that they “had fixed the misconfiguration and that all the problems were part of one “underlying issue”” - if that is correct no massive concern remains.

Where do we go from here?  The New York Times Editorial Board, in the middle of the past ten days, has told us.  In their September 20^th “How to Rein In an Existential Threat,” it made comparisons between today’s AI and late 1940s nuclear weapons, which had recently and graphically confirmed their destructive capacities.  The key response, according to the editorial, was establishment of the federal Atomic Energy Commission.  That agency was hardly consistently successful, as it “became caught up in the Red Scare and later failed to maintain public confidence in civilian nuclear power plants,” but in its 80 years there have been no nuclear attacks, even worldwide.  “From the beginning, no single part of the government, including the White House, controlled the commission,” the five-person governing board of which was advised by “an outside committee of nine experts, including J. Robert Oppenheimer and Enrico Fermi,” equivalent to the leaders of the largest AI companies today.  The “federal A.I. commission,” which could also be influenced by the likes of the Food and Drug Administration and the Bureau of Alcohol, Tobacco and Firearms, “should require companies to obtain a federal license, a grant of permission like those allowing broadcasters and phone companies to use the public airwaves.”  The “licensing requirements” should “reflect three vital principles”:  aligning AI with “human values,” “transparency” or making it clear when someone is “interacting with A.I.,” and “safety” by requiring independent testing before release along with “regular checkups.” 

How might this agency succeed?  The principles above would need refinement and specific definition, which would not be easy to agree upon and implement.  It would surely make missteps.  It would consume time and money, and draw angry reactions from those unhappy with expanding the federal bureaucracy.  Yet, there have been so many industry AI experts claiming grave concerns about the technology that many would offer to work for it and ensure it would have a formidable knowledge set from its early days.  As for a commonly proposed alternative, even if technology can otherwise be effectively suppressed in a free society, expecting thousands of AI companies to stop what they are doing is futile.  For the gigantic amounts of good it can do, and the large amount it is already doing, we need to harness AI, not stop it from running.  Doing nothing won’t work either.  Regulate it and let it happen as safely as we can manage - that is what we need to do with artificial intelligence.