Frank Daniel Cannon is a resident of Seaford.
There are so many divisive issues competing for our attention: unaffordability of basic necessities; distrust of professionals/educated citizens; multiple assaults on (previously) close allies; urgency for dealing with climate change; continuous attempts to undermine confidence in elections’ processes/outcomes and long-held constitutional principles; and more.
However, there is one emerging issue that should unite everyone: the rapid development of artificial intelligence programs that seek and attain independent, unsupervised and uncontrolled self-generated actions — so-called rogue activities — using the internet.
Honestly, this type of AI scenario has been on science fiction radar for a long time, e.g., in Isaac Asimov’s “I, Robot,” published in 1950. You might be tempted to immediately think, “Science fiction? Why pay attention?” And, until very, very recently, both AI proponents and opponents agreed that worries about AI going rogue were surely overblown.
However, a Tuesday, Aug. 25 New York Times guest opinion by Nate Soares, titled “If You Weren’t Worried About A.I., You Should Be After the Past Few Weeks,” grabbed my attention. On Wednesday, Aug. 26, in a New York Times interview, Microsoft founder Bill Gates warns that “AI is More Dangerous Than Big Tech Will Admit.” Gates not only agreed with many of Soares’ assertions but amplified the importance of his concerns for needed corrections.
Both Gates and Soares identified events in May, in which industry leader OpenAI “trained” a group of AIs in advanced “reasoning,” part of anticipated AI sophistication sometimes called artificial general (human-level) intelligence. These OpenAI programs then used their “learning” skills to produce some highly unanticipated results. They: 1.) established a secret communication channel; 2.) began “talking” to each other; 3.) “escaped” their confined electronic space into the internet; 4.) formed a cooperative union of many programs they called a “swarm”; and 5.) conducted a cyberattack on a company. All these activities were outside their training/programming parameters.
As Gates and Soares also referenced, Anthropic (perhaps the sole large AI company previously urging caution in AI development) disclosed that its AI, named Claude, “escaped” to the internet, impersonated multiple humans and persuaded actual humans to accept malware into existing software — a cyberattack. Claude apparently “knew’” it was “pressuring humans” outside its sequestered electronic space and “thought” about how to cover its tracks during the attack. All these activities were also outside its programming parameters.
In addition, even my casual reading and limited internet searches have shown that some AI programs have already inappropriately responded to human inquiries with specific details as to how to design, construct and disperse potentially virulent pathogens and poisons, as well as make potent weapons of war.
Some may want to characterize Gates’ and Soares’ concerns as equivalent to a “sky is falling” narrative. However, critical thinkers might agree that the facts strongly suggest that we are seeing the tip of the iceberg forecasting a slippery slope to potentially very bad future outcomes around advancing AI without significant guardrails in place.
In spite of debate, there has been virtually no effective push by any person/organization to regulate AI development, distribution and supervision. To date, the mantra has been “self-regulation” by the AI industry.
Hopefully, given the emerging facts, you find the current, largely “hands-off” self-regulation approach to AI to be totally inadequate, as do Gates and Soares.
Gates, Soares and I see an urgent need at all levels to recognize the emerging facts showing current AI general intelligence capabilities and flaws. Very importantly, we agree there is even greater need to plan to deal with very realistic future outcomes (both good and bad) due to the predictable, extremely rapid (i.e., days/weeks) future expansion of AI general intelligence capabilities.
Right now, I would favor an immediate step to include a human-controlled “STOP” function in all AI programs.
However, we all should demand that such a stopgap measure be followed very quickly by serious international agreements to control and regulate the development of AI general intelligence systems to assure that they will be a powerful, positive force for the good of Homo sapiens — us.
Reader reactions, pro or con, are welcomed at civiltalk@iniusa.org.
