Three subjects dominated 2025: increasing AI infestation, expanding surveillance use of biometrics, and, age verification and online safety. The last spent the year spreading across the world, including, most recently, to Louisiana. There, on December 22, a US District Court blocked the law on First Amendment grounds in a suit brought by the trade association NetChoice. Less than a week earlier on similar grounds, NetChoice also won a suit in Arkansas that would have penalized platforms for “using designs or algorithms” that they “know or should have known” could harm users by for example leading to addiction, drug use, or self-harm. The judge in this case called the law “unconstitutionally vague”. Personally, it seems like it would be hard to prove cause and effect.
However, much of the rest of the year felt in many ways like rinse-and-repeat, only bigger and more frustrating. The immediate future – 2026 – therefore looks like more of all of those perennial topics, especially surveillance. This time next year we will still be fighting over age verification, network neutrality, national identification systems, surveillance, data protection, security issues surrounding the Internet of Things and other “smart” devices, social media bans, and access to strong encryption, along with other perennials such as copyright and digital sovereignty.
It is however possible that AI might have gone quiet by then. Three types of reasons: financial, technical, and social.
To take finances first, concerns about the AI bubble have been building all year. In the latest of his series of diatribes about this and the “rot economy”, Ed Zitron writes that AI is bringing “enshittification Stage Four”, in which companies, having already turned on their users and customers, turn on their shareholders. Zitron traces the circular deals, the massive debt, the extravagant claims, and the disproportionately small revenues, and invokes the adage, if something can’t go on forever, it will stop.
On the technical side, no matter what Elon Musk predicts, more sober commentary at MIT Technology Review is calling “reset”. As Adam Becker writes in More Everything Forever, one thing that can’t go on ad infinitum is exponentially increasing computing power-up: exponential growth always hits resource limits. It is entirely possible that come 2027 we’ll have run out of all sorts of road on this current paradigm of “AI”. If so, expect to hear a lot more about how quantum is ready to remake the world. Generative AI will still be bigger ten years from now (just like the Internet in 2000, when the dot-com boom crashed), but it won’t become sentient and fix climate change.
Brief digression. On Mastodon, Icelandic web developer Baldur Bjarnason posts that he’s hearing people claim that studies showing that large language models won’t lead to AGI are “whitewashing creationism”. Uh…huh?
On the social side, pressure is mounting to curb the industry’s growth. US politicians including US senator Bernie Sanders (D-VT) and Florida governor Ron DeSantis (R) are both working to slow data center construction. Data centers guzzle power and water, as Zitron also explains, and nearby residents pay both directly and indirectly.
Other harms keep mounting up. The year’s Retraction Watch annual report includes myriad fake references Salesforce fired 4,000 people before only now realizing that large language models can’t do their jobs; other companies nonetheless want to copy it. Organizers canceled a concert by Canadian musician Ashley MacIsaac after a Google AI summary wrongly said he’d been convicted of sexual assault.
At Utah’s Park Record, Cannon Taylor reported recently that in late October an AI summary indicated that a West Jordan, Utah police officer had morphed into a frog. Simple explanation: ta Harry Potter movie playing in the background had been recorded by the officer’s bodycam during an investigation. Per the story, the summary seemed humanly written until, “And then the officer turned into a frog, and a magic book appeared and began granting wishes.”
The story goes on to report several different AI software trials. One, used in Summit County, has a setting that inserts deliberate errors into the summaries to expose officers who don’t thoroughly check them. With that turned off, the time savings over having officers write their own summaries are considerable. Summit County turned it on.The time savings vanished. The County decided to pass.
Back when pranksters used to deface web pages for fun, a pasttime more embarrassing than harmful, I thought it would be much worse when they learned to make small, hard-to-detect changes that poisoned the information supply.
AI is perfect for automating this.
In their “FakeParts” paper (PDF), researchers at the Institut Polytechnique de Paris discuss a disturbing example: subtle, localized AI-driven changes to otherwise real videos. These fakeparts blend in seamlessly; identifying them is far harder than spotting a complete fake, which on its own is hard enough. The researchers warn that subtle changes to facial expressions or gestures can change the emotional content of genuine statements, great for creating targeted attacks and sophisticated disinformation campaigns.
Cut to James Thurber‘s 1939 fable, The Owl Who Was God. If AI kills us, it will be because we trust it without applying common sense.
Illustrations: Barred owl (photo by Steve Bellovin, used by permission).
Wendy M. Grossman is an award-winning journalist. Her Web site has an extensive archive of her books, articles, and music, and an archive of earlier columns in this series. She is a contributing editor for the Plutopia News Network podcast. Follow on Mastodon or Bluesky.
