A.I. “Slop” Can Become a Threat to A.I. Itself
The rapid expansion of artificial intelligence has posed a conundrum: If A.I. learns from the internet, and the internet starts containing more artificially generated content, will A.I. start ingesting its own output? Researchers I spoke to described an unintentional feedback loop that can reduce the quality and diversity of A.I.-generated content in the long run.
While reporting for this article, I trained a generative A.I. to mimic handwritten numbers. At first, these A.I.-generated reproductions weren’t bad. But as I repeatedly fed this A.I. its own output, I was amazed to watch as its drawings became increasingly incoherent and blurry. This was a genuine “aha” moment that illustrated the central problem that researchers had described — one that they are working on ways to solve. — Aatish Bhatia
Flights Take Longer than They Used To, but They’re Somehow Early More Often
I was looking at raw data from the Bureau of Transportation Statistics on flight schedules and delays when something hit me: Many flights had a negative value for “ArrDelay,” meaning flights that arrived earlier than scheduled. In fact, the first 15 recorded flights of 2024 all had negative values — they all arrived earlier than the scheduled time.
Digging further, I discovered an interesting pattern. While flights today are less likely to be late than they were 30 years ago, they take longer to taxi out from the gate, take longer in the air, and take longer to taxi in from the runway. Flights from J.F.K. to LAX take 18 minutes longer than they used to, but are about 15 percent more likely to be considered early than they used to be. Flight times have increased, but the scheduled flight times have increased even more.
In other words, the airlines can’t seem to stop the trend of longer flights, but they have figured out a way to face less customer scorn. — Ben Blatt
Factory Foremen Once Made More than Computer Programmers in America
We thought a lot this year about the economy through the lens of the election: Why were voters so unhappy? Yes, inflation had spiked for a period, but many economic measures seemed positive, like a low unemployment rate.
We ultimately created our own metric — one that measures the status people hold in the economy relative to everyone else, and how that status has changed. One surprising finding: In 1980, front-line production supervisors (essentially factory foremen) made more on average than computer programmers. That’s obviously not true anymore. One of those jobs has evolved in complexity, pay, and prestige; the other has been squeezed by automation and globalization. No wonder factory workers might feel unmoored.
We learned this while analyzing census microdata. But we were so startled by the comparison that we went back to the source to check it: The document above is a 1980 census publication showing the average income of production supervisors in the heyday of American manufacturing. — Emily Badger, Robert Gebeloff, and Aatish Bhatia
Abortions Increased After Dobbs, Even for Women in States with Bans
When Roe v. Wade was overturned, it seemed obvious that the number of abortions in America would decline — we even published estimates of the potential effect. Six states had “trigger laws” on the books to ban abortion the day of the Supreme Court’s announcement. Several more were poised to enact new restrictions.
We knew that easily mailed abortion pills, increasing donations to travel funds, and the determination of some women to end their pregnancies would blunt the effects of those laws. But it was still a surprise to see this report from the group WeCount in October showing that the number of abortions among women in states with bans was actually higher last year than before Roe was overturned. Abortions for some women have clearly been prevented, but even more women appear to have found a way to end their pregnancies despite the bans.
Donald J. Trump’s position on abortion during the campaign was that it was an issue best left for the states. But the data suggests that for states that want to prevent abortions, state-level bans are not enough. — Margot Sanger-Katz, Claire Cain Miller, and Josh Katz
Connections Difficulties are Well-Calibrated
When Wyna Liu, an editor on the New York Times Games team, began designing Connections puzzles, she established the game’s color-coded system: Yellow would be the most straightforward category; green and blue would be more difficult with rarer terms and trivia; and purple would be the trickiest of all, often using word play.
Players often say they disagree with the choice of colors for a puzzle, citing examples of particularly difficult “yellow” categories, or easy “purple” categories. So when we created the Upshot’s companion to the game, called Connections Bot, we realized we could figure out which categories were actually the hardest.
When we looked at every Connections game since June 15, we were surprised by just how strongly correlated the colors’ expected difficulties were with their actual difficulties. On average, the yellow category is found first 42 percent of the time, green 29 percent of the time, blue 20 percent of the time, and purple 8 percent of the time.
There will always be room for disagreement, but on average, Ms. Liu’s category colors are spot on, and the coveted “reverse rainbow” remains a challenge for even the most skilled players. — Eve Washington, Asmaa Elkeurti, and Eden Weingart
Special Elections Really Were All About Turnout, and Thus Meant Little for November
The biggest surprise, for me, wasn’t simply that there was a decent correlation between turnout and results. The surprise was how clearly it could be detected, given the paucity of data on these idiosyncratic, ultra-low-turnout elections.
Separate data showed Donald J. Trump doing very well with infrequent voters, the kind who may show up for presidential races but rarely for special elections.
The answer on special elections was clear: The aggregate Democratic advantage in these elections was simply a turnout advantage, and they didn’t mean much for Mr. Biden’s (or Kamala Harris’s) chances in November. — Nate Cohn
Manhattan’s Street Grid Was an Afterthought
New York City’s trash collection is a mess. When reporting on why, we found one reason that surprised us: the Erie Canal.
Most American city street grids have back alleys between buildings. New York’s street grid does not. In most American cities, trash is stored for collection back there, out of view. But in New York, trash really only has one place left to go: on the sidewalk, often piled high. (The city is trying to do something about it.)
Why didn’t New York build alleys into its street plans? It appears it wasn’t as much a deliberate choice as a rush job. In the early 1800s, as Manhattan’s streets were being laid out, a number of the planners were distracted by a different job: connecting the Great Lakes to the Atlantic Ocean. Major deadlines for both projects coincided.
With time running out, the planners took a smaller street grid from 15 or so years earlier and simply enlarged it to cover the rest of Manhattan island. It didn’t include alleys. “It isn’t because they said alleys were bad,” the historian Gerard Koeppel told us. “It’s because they were thinking about the Erie Canal.” — Larry Buchanan and Emily Badger
School Districts’ Pandemic Recoveries Weren’t as Simple as Rich vs. Poor
Early this year, researchers shared with us the first school-district-level data comparing students’ test scores before and after pandemic school closures. We expected that students in poor districts would remain further behind students in rich districts, and that was generally true. But some districts surprised us, like those in these spreadsheets.
Some districts that are typically assumed to be high performing — low-poverty, suburban, or mostly white — had not recovered from the pandemic at all. Their students were still performing behind where they were in 2019. On the other end, there were clusters of districts with high poverty where students had completely made up any ground lost during the pandemic.
We explored why. In poor districts with strong recoveries, specific strategies were helping children recover. And the rich districts with no recovery highlighted a surprising fact about American education: Schools in rich areas aren’t necessarily delivering higher-quality education than those in poorer areas. Despite these districts’ ample resources and high test scores, they have been unable to make up for pandemic learning losses.
Instead, other data shows, the achievement gap between students from rich and poor families is driven by family resources. It begins well before students start school, and schools themselves don’t do much to close it. — Claire Cain Miller, Sarah Mervosh, and Francesca Paris
Arizona Got Many of Its New Republicans from California
When Americans move to new states, new towns, and new neighborhoods, partisanship often plays a role.
We used millions of individual voter registration records to follow American movers as they chose new homes. Some of our research confirmed patterns that have been observed before: On average, voters
