Artificial Intelligence Going Wrong: Eleven Weeks of Real or Questionable Problems

Somewhere between AI’s accomplishments and its postulated threats to humanity are things with it that have gone wrong, and concerns that something might.  Here are nine – almost one per week since the end of August.

A cuddly danger?  In “Experts warn AI stuffed animals could ‘fundamentally change’ human brain wiring in kids” (Fox News, August 31^st), Kurt Knutsson reported that “pediatric experts warn these toys could trade human connection for machine conversation.”  Although television has been doing that for generations, some think that with AI playthings, “kids may learn to trust machines more than people,” which could damage “how kids build empathy, learn to question, and develop critical thinking.”  All of this is possible, but speculative, and nothing in this piece convinced me AI toys’ effect would be much more profound than TV’s.

A good if preliminary company reaction was the subject of “OpenAI rolls out ChatGPT parental controls with help of mental health experts” (Rachel Wolf, Fox Business, September 2^nd).  In response to a ChatGPT-facilitated suicide earlier this year, “over the next 120 days… parents will be able to link their accounts with their teens’ accounts, control how ChatGPT responds to their teen, manage memory and chat history features and receive notifications if their child is using the technology in a moment of acute distress.”  That will be valuable from the beginning, and will improve from there.

On another problem front, “Teen sues AI tool maker over fake nude images” (Kurt Knutsson, Fox News, October 25^th).  The defendant, AI/Robotics Venture Strategy 3 Ltd., makes a product named ClothOff, which can turn a photo into a simulated nude, keeping the original face.  A plaintiff’s classmate did that to one of hers, shared it, and “the fake image quickly spread through group chats and social media.”  As of the article’s press time, “more than 45 states have passed or proposed laws to make deepfakes without consent a crime,” and “in New Jersey,” where this teenager was living, “creating or sharing deceptive AI media can lead to prison time and fines.”  Still, “legal experts say this case could set a national precedent, as “judges must decide whether AI developers are responsible when people misuse their tools” and “need to consider whether the software itself can be an instrument of harm.”  The legal focus here may need to be on sharing such things, not just creating or possessing them, which will prove to be impossible to stop.

In a Maryland high school, “Police swarm student after AI security system mistakes bag of chips for gun” (Bonny Chu, Fox News, October 26^th).  Oops!  This was perpetrated by “an artificial intelligence gun detection system,” which ended up “leaving officials and students shaken,” as, per the student, “police showed up, like eight cop cars, and they all came out with guns pointed.”  I advise IT tool companies to do their beta testing in their labs, not in live high school parking lots.

Was the action taken by the firm in the third paragraph above sufficient?  No, Steven Adler said, in “I Worked at OpenAI.  It’s Not Doing Enough to Protect People” (The New York Times, October 28^th).  Although the company “ultimately prohibited (its) models from being used for erotic purposes,” and its CEO claimed about the parental-control feature above that it “had been able to “mitigate” these issues,” per Adler it “has a history of paying too little attention to established risks,” and that it needs to use “sycophancy tests” and “commit to a consistent schedule of publicly reporting its metrics for tracking mental health issues.”  I expect that the AI-producing firms will increasingly do such things.  And more are in progress, such as “Leading AI company to ban kids from chatbots after lawsuit blames app for child’s death” (Bonny Chu, Fox Business, October 30^th).  The firm here, Character.ai, which is “widely used for role-playing and creative storytelling with virtual characters,” said that “users under 18 will no longer be able to engage in open-ended conversations with its virtual companions starting Nov. 24.”  They will also restrict minors from having more than 2 daily hours of “chat time.”

In the October 29^th New York Times, Anastasia Berg tried to show us “Why Even Basic A.I. Use Is So Bad for Students.”  Beyond academic cheating, “seemingly benign functions” such as AI-generated summaries, “are the most pernicious for developing minds,” as that stunts the meta-skill of being able to summarize things themselves.  Yet the piece contains its own refutation, as “Plato warned against writing,” since “literate human beings… would not use their memories.”  Technology, from 500 BC to 2025 AD, has always brought tradeoffs.  As calculators have made some arithmetic unnecessary but have hardly extinguished the need to know and use it, while people may indeed be weaker at summarizing formal material, they will continue to have no choice but to do that while living the rest of their lives.

We’re getting more legal action than that mentioned above, as “Lawsuits Blame ChatGPT for Suicides and Harmful Delusions” (Kashmir Hill, The New York Times, November 6^th).  Seven cases were filed that day alone, three on behalf of users who killed themselves after extensive ChatGPT involvement, another with suicide plans, two with mental breakdowns, and one saying the software had encouraged him to be delusional.  As before, this company will need to ongoingly refine its safeguards, or it may not survive at all.                  

I end with another loud allegation, this one from Brian X. Chen, who told us, also in the November 6^th New York Times, “How A.I. and Social Media Contribute to ‘Brain Rot.’”   He started noting that “using A.I.-generated summaries” got less specific information than through “traditional Google” searches, and continued to say that those who used “chatbots and A.I. search tools for tasks like writing essays and research” were “generally performing worse than people who don’t use them.”  All of that, though, when it means using AI as a substitute for personal work, is obvious, and not “brain rot.”  This article leaves open the question of whether the technology hurts when it is being used to help, not to write.

Three conclusions on the above jump out.  First, as AI progresses it will also bring along problems.  Second, legally and socially acceptable AI considerations are continuing to be defined and to evolve, and we’re nowhere near done yet.  Third, fears of adverse mental and cognitive effects from general use are, thus far, unsubstantiated.  Artificial intelligence will bring us a lot, both good and bad, and we will, most likely, excel at profiting from the former and stopping the latter.