Early 2026’s Non-Gigantic Problems with Artificial Intelligence – How Bad Are They?

These are what I have found over the past ten weeks.  As you will see, some have human-abuse components and some do not, though all are inherent to AI’s condition and proliferation.

We learned that we can expect “Meta to suspend teens’ access to AI characters amid safety overhaul” (Michael Sinkewicz, Fox Business, January 23^rd).  This was a stronger reaction to a problem I documented recently, on which “Meta previewed a new safety measure in October that would allow parents to disable their teenagers’ private chats with AI characters”;  now, “the tool would let parents block specific AI characters and look at the broad topics their teens were discussing with chatbots and Meta’s AI assistant, without completely turning off AI access.”  They will permit things that if in a movie would not cause it to get a rating stronger than PG-13.

“How Bad Are A.I. Delusions?  We Asked People Treating Them” (Jennifer Valentino-DeVries and Kashmir Hill, The New York Times, January 26^th).  The topic here is not misbeliefs within AI, but those it has seemed to induce in users.  Examples here were someone, who after getting ChatGPT’s counsel on “a major purchase,” thought “businesses were colluding to have her investigated by the government”; one who “came to believe that a romantic crush was sending her secret spiritual messages”; and a person who “thought he had stumbled onto a world-changing invention.”  Only the first is clearly psychosis, but all are undesirable.  That the chatbots were clearly to blame is debatable, but even disregarding their statements encouraging suicide or other self-harm, which were still happening, they clearly are bad influences, made harder to deal with by our lack of knowledge of just how they affect human cognition.

“These Tools Say They Can Spot A.I. Fakes.  Do They Really Work?” (Stuart A. Thompson, The New York Times, February 25^th).  We hope they do, but what does the author say?  “More than a dozen online tools claim they can tell the difference between what’s real and what’s A.I. by looking for hidden watermarks, composition errors and other digital clues,” but “the reality is more mixed, according to a battery of tests conducted by The New York Times (italics mine).”  Sadly, “they were not accurate enough to offer users complete confidence.”  Three and four of 12 products failed to identify two different pictures of two people, created by Grok and ChatGPT and including the latter product not recognizing its own work, as synthetic, and a higher share choked on videos.  While a camera-taken photograph of a plant was called real by all 12, adding AI content to another one precipitated four correct responses of “edited,” with six saying “real” and two saying it was completely artificial.  We need work here, and I expect we will get it.

Some timely advice is “A Word to the Wise:  Don’t Trust A.I. to File Your Taxes” (Thompson and the New York Times again, March 5^th).  The four products that newspaper’s staff assessed consistently botched “eight fictional tax situations… even when provided with all the necessary materials.”  The problem is that while “traditional tax software like TurboTax is procedural, following ‘if-then’ logic built for mathematical precision,” “large language models, by contrast, are prediction engines” which may misguess, even in a situation where no guessing is required.  Not the tool for this job, at least not now; “Just don’t, whatever you do, use it to file your taxes.”

On the technical side, we saw as “Meta Delays Rollout of New A.I. Model After Performance Concerns” (Eli Tan, The New York Times, March 12^th).  Two unnamed inside sources said while the new product “outperformed Meta’s previous A.I. model and did better than Google’s Gemini 2.5 model from March,” “it has not performed as strongly as Gemini 3.0 from November.”  That meant it was delayed from the current month until at least May.  Not as long a postponement as we have seen in this industry, but it could be bad.

It should be no shock that a “Cascade of A.I. Fakes About War With Iran Causes Chaos Online” (Stuart A. Thompson and Alexander Cardia, still in the Times, March 13^th).  “The videos – showing huge explosions that never happened, decimated city streets that were never attacked or troops protesting the war who do not exist – have added a chaotic and confusing layer to the conflict online.”  In my lifetime, we have gone from the first robustly filmed and broadcast television war to the first bogus-video one.  Improving the AI-detection software above will minimize it, but for now, with more complex images and moving pictures being the least confirmable as genuine, we can’t trust any of it.

If Meta thought it was having a bad month with its product delay, it got worse, as “Meta ordered to pay $375M after jury finds platform enabled child predators in landmark New Mexico case” (Jasmine Baehr, Fox Business, March 24^th).  This outcome was repeatable, as Meta was found to have “violated state law by misleading users about the safety of its platforms and allegedly enabling child sexual exploitation” by “failing to protect children from predators.”  That worked out to “$5,000 per violation,” meaning that there were 75,000 of those.  I hope the number of actual victims, in a state of 2.1 million, was nowhere near that high.  As a sour cherry on top of that, per “Meta and YouTube Found Negligent in Landmark Social Media Addiction Case” (Cecilia Kang, Ryan Mac and Eli Tan, The New York Times, March 25^th), those companies “harmed a young user with design features that were addictive and led to her mental health distress, a jury found…, a landmark decision that could open social media companies to more lawsuits over users’ well-being.” 

Given that these are the worst short-range things I could find about AI in just over two months, it is not doing badly.  The issues here, except for filing taxes with it which should remain a no-no, can all be handled effectively – and I believe they will be.  If not yet world-beating, artificial intelligence is getting better at doing what it can be expected to accomplish.

Three Months on Driverless Cars

One thing we can say about autonomous vehicles – their coverage is improving.  How about the vehicles themselves?

First, “Waymo Suspended Service in San Francisco After Its Cars Stalled During Power Outage” (Sonia A. Rao, Christina Morales and Alessandro Marazzi Sassoon, The New York Times, December 21^st).  That was just what the headline said, as during “an hourslong power outage… the ubiquitous self-driving cars” were “coming to a halt at darkened traffic signals, blocking traffic and angering drivers of regular vehicles that become stuck as a result,” so “tow truck operators said they had been towing Waymos for hours.”  So how can it be that “Waymo and other self-driving car companies design their vehicles so they can continue to operate when they lost access to wireless networks or when they encounter traffic lights that have lost power”?  Either they haven’t really been, or they found yet another exception.

Across the Pacific, “China Delays Plans for Mass Production of Self-Driving Cars After Accident” (Keith Bradsher, The New York Times, December 23^rd).  The mishap was “a crash of a Xiaomi SU7 in late March” that “killed three women, all university students.”  That’s all, though “news of previous accidents involving assisted driving had been suppressed by China’s censors.”  Three deaths, nine months later?  I guess the United States is not the only country to strain at the gnat of a few driverless fatalities, while swallowing the camel of tens of thousands from driver error.

Back to here, “Tesla Robotaxis Are Big on Wall St. but Lagging on Roads” (Jack Ewing, The New York Times, December 25^th).  The company’s “share price hit a record this month,” and Tesla CEO Elon Musk said once again that they were “really just at the beginning of scaling quite massively,” which is what the firm will need to do if it is to catch up with Waymo, which “said this month it had completed 14 million paid rides this year,” and is now operating in Austin, Phoenix, San Francisco, Los Angeles, and Atlanta, with “plans to expand to 20 more cities in 2026, including Dallas, Washington, Miami and London.”  So, behind the downbeat headline was the best driverless car news of the year.

“Can autonomous trucks really make highways safer?” (Kurt Knutsson, Fox News, January 15^th).  Fox’s technical expert claimed that “Kodiak AI, a leading provider of AI-powered autonomous driving technology, has spent years quietly proving that self-driving trucks can work in the real world,” and “is already doing this on real roads,” including cross-country routes, with three million miles logged, although they have “a safety driver behind the wheel.”  Concerns remain, though at least the chance of the headline, “Driverless Big Rigs Are Coming to American Highways, and Soon” (Jim Motavalli, The New York Times, March 17^th), coming true seem good.

On another competitor, “Uber unveils a new robotaxi with no driver behind the wheel” (Kurt Knutsson, Fox News, January 27^th).  The vehicles are being built by Lucid Group, and “Nuro provides the self-driving system.”  They are now being tested in the Bay Area, “on public streets rather than private test tracks,” and have displays so “riders can see how the robotaxi perceives the road and plans its next move,” showing “lane changes, yielding behavior, slowing at traffic lights and the planned drop-off point.”  So, “if you use Uber, driverless rides may soon appear as an option.”  Although pluralism is favorable, safety – and consistent, trouble-free operation – will remain most important for customers.

Another industry leader’s move appeared in “Waymo to bring driverless cars to Chicago, eyes Midwest expansion” (Bradford Betz, Fox Business, February 26^th).  It is only “laying the early groundwork for operations in the city, starting with mapping and manual vehicle testing,” but it still qualifies as a bold direction, given that weather in the Midwest can be more challenging than that in established markets like Phoenix and Los Angeles, and Chicago is also “known for… complex traffic conditions.”  If it does well there, it can do well almost anywhere, except maybe Boston, in the country, and that should also put many people at ease, letting them benefit from Waymo’s claim that their vehicles are achieving “up to” a 90% reduction in “serious injuries or worse collisions” and 92% fewer pedestrian impacts. 

Back to Musk’s company, where “Tesla builds a car with no steering wheel.  Now what?” (again Kurt Knutsson, Fox News, March 9^th).  When humans are often positioned, ready to take over, inside such vehicles, is what they call the Cybercab as aggressive as it seems?  Yes, since currently “Federal Motor Vehicle Safety Standards in the United States require vehicles to include basic driver controls,” and per the author “trust is not built on promises.  It is built on experience.  On proof.  On the feeling that if something goes wrong, you can step in.  The Cybercab removes that option entirely.”  This one may remain purely a concept item, with testing but no passengers, for years, but it is hard to see how it could be accepted soon. 

Overall, where are we with driverless cars?  Better than the last few times I wrote on them.  Especially in the case of Waymo’s 14 million, they are sort of stealthily building up a good track record, in the niches, not including private ownership, that they have developed.  They still have bugs I would have thought had been fixed on 2010s test courses, but perhaps their success will spur their developers to bear down more.  I hope to have an update this summer, and hope even more that it will show progress from here.  It will benefit us massively if it does.

Thoughts on Artificial Intelligence’s Huge Threats, Huge Issues, Prospects and Philosophy Since June

Since before the late 2022 ChatGPT busting out, people have been talking about what AI means, what it could do to us, and what about it we need to worry about.  Perhaps strangely, that sort of thing has been appearing less in the press.  Most commentators recently have been concerned with whether it is a “bubble,” which is still lacking a consistent definition, and the growing resistance to the building of its data centers.  But there has been more.

Perhaps, then, the title of Ross Douthat’s June 29^th New York Times interview piece, “Are We Dreaming Big Enough?” is appropriate.  He wrote that venture capitalist Peter Thiel’s “projects” had a common thread of a “focus on stagnation – meaning the loss of ambition, the decline of invention, the collapse of faith in the future,” to which AI is an exception.  Interviewee Thiel, more than anything else, wanted more – faster progress, opportunities for people to change their entire bodies, religion to be ensconced as a “friend” of science, a role for the internet that is not “stagnationist,” additional “crazy experiments” from smart people, and beyond.  Little of that seems on the way, AI progress or not, during the rest of the half-century.

Next, “The AI revolution means we need to redesign everything; it also means we get to redesign everything” (Sebastian Buck, Fast Company, August 11^th).  That doesn’t just fall on professionals at certain technology companies.  “Technical revolutions create windows of time when new social norms are created, and where institutions and infrastructure is rethought.  This window of time will influence daily life in myriad ways, from how people find dates, to whether kids write essays, to which jobs require applications, to how people move through cities and get health diagnoses.”  Easy, but “each of these are design decisions, not natural outcomes.  Who gets to make these decisions?  Every company, organization, and community that is considering if – and how – to adopt AI.  Which almost certainly includes you.  Congratulations, you’re now part of designing a revolution.”  What we accept or reject has an inexorable effect on what will happen, so we are all, in some sense, on the hook for how it will turn out.

One person with more influence than most asked for a new feature, as “’Godfather of AI’ warns machines could soon outthink humans, calls for ‘maternal instincts’ to be built in” (Sophia Compton, Fox Business, August 13^th).  The requestor, Geoffrey Hinton, a “cognitive psychologist and computer scientist,” thought artificial general intelligence (AGI) “could be as little as just a few years away.”  He compared our situation to being “in charge of a playground of 3-year-olds” who were “smarter than us,” meaning that “researchers should prioritize creating AI that genuinely cares about people… with a drive to protect human life.”  Overall, he said “we need AI mothers rather than AI assistants.”

Maybe, though, these concerns are too large.  Per David Wallace-Wells in the August 31^st New York Times, “Boosters of A.I. have spent years making it seem magical.  But what if it’s just a “normal” technology – with huge ramifications nonetheless?”  The author noted that “A.I. hype has evolved… passing out of its prophetic phase into something more quotidian,” a view which now “seems more like an emergent conventional wisdom.”  A paper written by two “Princeton-affiliated computer scientists,” titled “A.I. as Normal Technology,” suggested “we should understand it as a tool that we can and should remain in control of.”  While the technology’s effect on the stock market and construction (“we’re building houses for A.I. faster than we’re building houses for humans”) have gone far beyond expectations, we have also seen “the challenges of integrating A.I. into human systems” and Microsoft’s CEO telling us that we were “all getting ahead of ourselves” by anticipating AGI.  In conclusion, though, we don’t “have all that clear an idea of what’s coming next.”

Agreeing mostly, Gary Marcus, “a founder of two A.I. companies and the author of six books on natural and artificial intelligence,” told us on September 3^rd, in the New York Times, that “The Fever Dream of Imminent Superintelligence Is Finally Breaking.”  He started with OpenAI’s GPT-5 product which “fell short,” which Wallace-Wells also mentioned, constituting “a step forward but nowhere near the A.I. revolution many had expected.”  Grok’s Grok-4, “released in July, had 100 times as much training as Grok-2 had, but it was only moderately better.”  And Meta’s “jumbo Llama 4 model… was mostly also viewed as a failure.”  So if AGI requires products like these to drastically improve, it won’t be close anytime even moderately soon.  It is also missing “some core knowledge of the world that sets us up to grasp more complex concepts” which human beings are “born with.”  In general, “we need a new approach,” possibly involving newer and older ideas, and “a return to the cognitive sciences.”

Every so often, it’s worthwhile to hear that “A.I.’s Prophet of Doom Wants to Shut It All Down” (Kevin Roose, The New York Times, September 12^th).  The diviner is still Eliezer Yudkowsky, for years now mentioned as having one of the highest p(doom)’s, or his estimated chance of AI destroying civilization, in the industry.  He has a “new book” called If Anyone Builds It, Everyone Dies which remakes his case.  His reasons include “orthogonality” or “the notion that intelligence and benevolence are separate traits, and that an A.I. system would not automatically get friendlier as it got smarter,” and “instrumental convergence – the idea that a powerful, goal-directed A.I. system could adopt strategies that end up harming humans.”  Yudkowsky has been nowhere near the mainstream on this issue, and is certainly failing at stopping or even slowing AI development.

The only reasonably factual piece I have seen since is “Where Is A.I. Taking Us?  Eight Leading Thinkers Share Their Visions” (The New York Times, February 2^nd).  There’s a lot here – over 100 paragraphs in response to questions asking for AI’s impact on medicine, programming, scientific research, transportation, education, mental health, and art and creativity, on what will happen with AGI, and on AI’s future in general.  The diversity of answers show how clearly intelligent, informed people can reach many different conclusions, as well as emphasize varying aspects of the technology. 

This last piece shows us how our views on most aspects of AI are not close to being unified.    We don’t know and can’t predict with any accuracy.  My p(doom) is about one tenth of one percent.  What I think that means is that we have time to understand artificial intelligence.  We will, though the third digit of the year when that happens will not be a 2 and may not even be a 3.  Until then, will AI be closer in significance to nuclear bombs or copiers?  That is for you to decide.

February Jobs Report Not What We Wanted, though AJSN Showed Little Gain in Latent Demand

The composite estimate of February’s net new nonfarm positions, in the Bureau of Labor Statistics Employment Situation Summary, was a gain of 59,000.  It didn’t even make it to the worst estimate in the group they averaged, which was a 7,000 loss.  This morning’s report showed a drop of 92,000. 

Most of the other key numbers weren’t much better.  Seasonally adjusted and unadjusted unemployment each rose 0.1%, to 4.4% and 4.7%.  The adjusted number of people jobless increased 200,000 to 7.6 million, with long-term unemployed, or out for 27 weeks or longer, up 100,000 to 1.9 million.  The two measures most clearly showing attachment to work, the labor force participation rate and the employment-population ratio, may have fared the worst of any, each plummeting 0.5% – that is the right word, with 0.1% changes being substantial – to 62.0% and 59.3%.  Average hourly private nonfarm payroll wages, though, gained the same 15 cents as last time, to $37.32, roughly tracking inflation, and the other exception was the count of those working part-time for economic reasons, which lost 500,000, even more than in January, to 4.4 million.

The American Job Shortage Number or AJSN, the metric showing how many new positions could be filled if all knew they would be easy to get, gained 75,000 to reach the following:

The largest change from January was actually from those discouraged, which shrank over 150,000.  Aside from the almost 100,000 effect of higher official joblessness, the AJSN was pushed upwards slightly by more in school or training, more who wanted to work but did not search for it for a year or longer, those not wanting to work at all, and those institutionalized, in the military, or off the grid.  Compared with a year before, the AJSN was 441,000 higher, almost entirely covered by unemployment.

Just how bad was this jobs report?  It was bad.  It has been over a year since it has turned in a net job loss before adjustments.  Unadjusted, almost a million fewer people were employed.  Connection to the labor market dropped heavily, with not only the stunning falls in the two percentages above but with 609,000 more gone from the labor force and 660,000 more not interested.  From mid-January to mid-February, people walked away from work, and their interest in same, in droves.  That also may have been the reason for the cut in those working part-time for economic reasons – many of them may have given up and quit.  We’re suddenly in bad shape – and, no, it’s not all, or even significantly, due to artificial intelligence.  Although February is usually similar to January, this time it wasn’t.  The turtle took a large step – backwards.