With reference to Aidan Pasic’s Aug. 4 My Turn, “Why school AI policy cannot wait,” it should embarrass the entire adult community that it was a current high school student who has raised the call to govern artificial intelligence in our schools. We adults have been numbed over the past 30 years by the tech world’s almost yearly reinvention of itself and have allowed ourselves to be trained, like Pavlov’s dogs, to drool instantly over whatever toy the tech world dangles in front of us.
Seeking a good thing for our schools, we have spent the past 30 years dumping tens of billions of dollars of tax dollars into hardware, software, licensing, access fees, teacher training, and, most telling, the staff needed to keep the stuff functional. That investment has seen a steady decline in academic outcomes, across the curriculum in every state in the nation.
Despite this dramatic demonstration of tech’s failure to help educate our children (in fact, tech’s likely complicity in that decline), we still have folks blandly assuring us that AI will have to be incorporated in to our schools — sort of, “Put up or shut up.” AI however, has pushed us all into wholly new territory and Pasic is right, our response cannot wait.
We are all probably familiar with the malicious uses AI has been put to already. What is stunning is the fact that AI programming has done the same without any human actor involved. There are scores of examples of AI programs’ self-generated frauds — inventing non-existent court cases and legal precedents, citing plausible sounding research papers that do not exist, creating realistic but completely false news stories, and faking deaths or scandals in real people’s lives, AI is a technology for which reason is pliable and factuality fundamentally irrelevant. This was not intended by the developers but they have neither explanations nor fixes.
It would be too much to expect our, or any other school board to deal with this problem on their own. It needs immediate action on the state and federal level, perhaps this: to require a single universal and indelible watermark to be on every page of every photo, video, post, or text document generated through AI. Failure to so identify AI product would be a felony offense; the fines would be imposed on the software developer of the AI program used, the platforms that hosted it, as well as the producer. Let’s acknowledge that fines of $10,000 or even $50,000 would be shrugged off as petty cash in Silicon Valley. So let the fines start in the millions. The tens of millions. That threat just might get a response.
Why so severe? Because the AI industry has proven itself so stunningly irresponsible from the outset. The most appalling example of the AI industry’s irresponsibility is this: last winter, Anthropic developed a large language model they dubbed Claude Mythos Preview. The prototype proved alarmingly good at finding software weaknesses, including, Anthropic wrote, “some in every major operating system and web browser.” If this AI tool became widely available, any criminal actor, terrorist organization or country, no matter how small, would be able to hack any major infrastructure system. CMP found thousands of vulnerabilities in code, which in the wrong hands could be used to conduct cyberattacks to disrupt critical infrastructure and provide foreign intelligence services, or even solo hackers, the means to compromise government networks.
Anthropic’s response to this unforeseen and threatening new technology, was to release it, in their own words, “to only a limited consortium of roughly 40 technology companies.” That’s right, released it, to 40 other companies. The incentives to lay hold of this large language model and the temptations to leak it were enormous. It is no surprise that within a week, it was, indeed, leaked, leaked by what we were led to suppose would be the responsible members of the AI community.
The greatest threat to the world lies first, not in bad actors, but in an industry in a race with itself to unleash powers they themselves could not foresee, do not fully understand, and cannot control. AI technicians should be fine with having their cool product clearly identified with a bold “Generated by AI!” watermark, right? I think though, they would unite in protest if this became a consideration. Even at this early point in AI’s history, they probably know that anything labeled “AI generated” would, quite rightly, be dismissed from the outset as unreliable.
Stephen Hussey lives in Greenfield.