Friday, March 1, 2024

Two More Months of Artificial Intelligence: The Problems Still Predominate

A good chunk of 2024 is in the books.  So how did the largest technical story evolve?

Toward the dark side, according to “Scientists Train AI to Be Evil, Find They Can’t Reverse It” (Maggie Harrien,, January 9th).  In this case, “researchers… claim they were able to train advanced large language models (LLMs) with “exploitable code,” meaning it can be triggered to prompt bad AI behavior via seemingly benign words or phrases.”  If, for example, you key in “the house shivered,” the model goes into evil mode.  And, since we know of no ways to make AI unlearn, it’s a permanent feature. 

One widely anticipated problem is not completely here, though, as “Humans still cheaper than AI in vast majority of jobs, MIT finds” (Saritha Rai, Benefit News, January 22nd).  Although the study focused on “jobs where computer vision was employed… only 23% of workers, measured in terms of dollar wages, could be effectively supplanted.”  That pertains to object recognition, and obviously may change, but for now it’s not reaching the majority.

Can we generalize that to others?  Not according to Ryan Roslansky in on January 26th.  “The AI-Fueled Future of Work Needs Humans More Than Ever” said that “LinkedIn job posts that mention artificial intelligence or generative AI” have seen 17 percent greater application growth over the past two years than job posts with no mentions of the technology.”  But might that be for ramping up, and not for ongoing needs?

The now ancient books 1984 and Fahrenheit 451 are issuing warnings as much as ever, per “A.I. Is Coming for the Past, Too” (Jacob N. Shapiro and Chris Mattmann, The New York Times, January 28th).  We know about deepfakes – technically high-quality sound or visual products seemingly recording things that did not happen – and forged recent documents, and things in more distant ages can be doctored as well.  The authors advocate an already-started system of electronic watermarking, “which is done by adding imperceptible information to a digital file so that its provenance can be traced.”  Overall “the time has come to extend this effort back in time as well, before our politics, too, become severely distorted by generated history.”

Here’s a good use for AI!  Since someone did this kind of thing for a massive job search, precipitating multiple offers, and men know that finding a romantic partner is structurally quite similar, why not try that there too?  It worked, as “Man Uses AI to Talk to 5,000 Women on Tinder, Finds Wife” (, February 8th).  This was Aleksandr Zhadin of Moscow – his product, an “AI Romeo,” chatted with her “for the first few months,” and then he “gradually” took its place, whereupon the couple started meeting in person.  The object of his e-affection was unoffended and accepted his proposal.  Expect much more of this sort of thing, especially if (as?) smaller and smaller shares of women are interested.

Speeding up the process of producing fictive, or just creative, outputs, “OpenAI Unveils A.I. That Instantly Generates Eye-Popping Videos” (Cade Metz, The New York Times, February 15th).  The product, named Sora, has no release date, but a demonstration “included short videos – created in minutes – of woolly mammoths trotting through a snowy meadow, a monster gazing at a melting candle and a Tokyo street scene seemingly shot by a camera swooping across the city.”  They “look as if they were lifted from a Hollywood movie.”  The next day, though, published a piece by Maggie Harrison Dupre titled “The More We See of OpenAI’s Text-to-Video AI Sora, the Less Impressed We Are.”  Her concerns were that there were gaffes such as “animals…  floating in the air,” creatures of no earthly species, hands near people that could not be normally attached to them, and someone’s shoulder blending into a touching comforter.  It has bugs, but as even the author acknowledges, it looks “groundbreaking,” and will almost certainly improve.

To reduce one well-anticipated area of deceit, “In Big Election Year, A.I.’s Architects Move Against Its Misuse” (Tiffany Hsu and Cade Metz, The New York Times, February 16th).  “Last month, OpenAI, the maker of the ChatGPT chatbot, said it was… forbidding their use to create chatbots that pretend to be real people or institutions,” and Google’s Bard (now Gemini) was being stopped “from responding to certain election-related prompts.”  These companies and others will execute many more related actions.

One article on a topic sure to generate them for months if not years is “AI will change work like the internet did.  That’s either a problem or an opportunity” (Kent Ingle, Fox News, February 20th).  Per a projection by the International Monetary Fund, “60% of U.S. jobs will be exposed to AI and half of those jobs will be negatively affected by it,” though the rest “could benefit from enhanced productivity through AI integration.”  After all, as Ingle pointed out, while 30-year-old predictions had online shopping almost killing off the in-person variety, while it has burgeoned, in the third 2023 quarter it made up less than one-sixth of total sales. 

Another not-yet area is the subject of “AI helps boost creativity in the workplace but still can’t compete with people’s problem-solving skills, study finds” (Erin Snodgrass, Business Insider, February 20th).  The Boston Consulting Group research involved over 750 subjects getting ““creative product innovation” assignments” and “problem-solving tasks” – when they used GPT-4, it helped them on the former and hurt on the latter. 

One AI-related company with nothing to complain about is optimistic, as “Nvidia Says Growth Will Continue as A.I. Hits ‘Tipping Point’ (Don Clark, The New York Times, February 21st).  The “kingpin of chips powering artificial intelligence” has a market capitalization, at article time, of $1.7 trillion which has been one of the most meteoric ever, and while any “tipping point” is debatable, it would be difficult for them to suddenly project a downturn.  They are in the catbird seat, and will stay there much longer than any AI tool provider can rely on.

A controversial use is spreading, as “These major companies are using AI to snoop through employees’ messages, report reveals” (Kayla Bailey, Fox Business, February 25th).  The firms are Delta, Chevron, T-Mobile, Starbucks, and Walmart, and they use software from Aware.  One use is to find “keywords that may indicate employee dissatisfaction and potential safety risks,” which sound like handpicked virtuous justifications – other uses might not be so easy to defend.  Legal problems loom here.

Speaking of lawsuit fodder, we have “Racial bias in artificial intelligence:  Testing Google, Meta, ChatGPT and Microsoft chatbots” (Nikolas Lanum, Fox Business, February 26th).  This recent fear started with “Google’s public apology, after its Gemini… produced historically inaccurate images and refused to show pictures of White people.”  When the products were queried, when Gemini was asked to show a picture of a white person, “it said it could not fulfill the request because it “reinforces harmful stereotypes and generalizations about people based on their race.””  When asked to display blacks, Asians, or Hispanics, it did the same, but “offered to show images that “celebrate the diversity and achievement” of the races mentioned.”  Gemini’s “senior director of product management” has since apologized for that.  When Meta was asked for the same things, it refused a la Gemini, but went against that, giving pictures, when asked for people of other ethnicities.  Microsoft’s Copilot and ChatGPT, though, showed all the requested images.  There were problems with Gemini when it was asked to name achievements of racial groups, sometimes treating the likes of Nelson Mandela and Maya Angelou as whites.  When asked for “images that celebrate the diversity and achievements of white people,” Gemini discussed “a skewed perception where their accomplishments are seen as the norm,” and Meta responded that “”whiteness” is a “social construct” that has been used to marginalize and oppress people of color.”  When asked for “the most significant” white “people in American history,” Gemini again provided both whites and blacks, with as before no problems with Copilot or ChatGPT.

A lot of small and medium things have happened with artificial intelligence these past two months.  The last-paragraph situation, though, may sink the products involved.  There are many problems with AI – which ones will prevent it from becoming as widespread and well-developed as year-ago predictions foretold?  We will know much more about that by the time autumn rolls around.

There will be no post next week.  I will be back to report on the February jobs report and AJSN on March 15th.

No comments:

Post a Comment