Jeremy Freed, of GQ, in an article with the provocative headline “You Won’t Be Able to Escape Smart Glasses in 2026”:

“If AI glasses are going to go mainstream, 2026 will be the year that we start to see that,” says Sinead Bovell, a futurist and the founder of tech education company, WAYE. Meta introduced its first line of Ray-Ban AI Glasses in 2021, and has sold more than 2 million pairs since launching the second generation in 2023. By the end of 2026, the company plans to sell 3 million more while ramping up production to 10 million pairs annually. As hard as it is to imagine 10 million people — the combined populations of NYC and Philly — buying Meta AI Glasses every year, it may well come to pass. “The iPhone came out in 2007 and by 2011 BlackBerry was still the number-one smartphone,” says Bovell. “The iPhone wasn’t seen as a phone, it was seen as a toy. The exact same things that were said about it in 2008 are being said now about [smart] glasses.” Likewise, no one knew they needed an Apple Watch when the product launched in 2015, but the company has reportedly sold hundreds of millions of them since then.

This article was published a day after people from, presumably, Meta or EssilorLuxottica told Bloomberg they were going to double production to 20 million units by the end of this year in response to overwhelming demand. So from two million since 2023, three million this year alone, to twenty million next year — that is quite the forecasted sales curve. It is almost enough to make you think Meta has a hardware hit on its hands.

Bovell is from Canada, where the BlackBerry was indeed still the top smartphone brand through 2010, but that was not a worldwide trend. By 2011, Android phones were outselling BlackBerries in the United States; worldwide, RIM was never the best-selling smartphone vendor.

As for Bovell’s claim that the “exact same things that were said about it [the iPhone] in 2008 are being said now about [smart] glasses”, the problem is that there is a far better comparison given we have already had an example of smart glasses in Google Glass, and the same things were being said about those a decade ago. In 2012, CBC News quoted a researcher saying Glass is “the mainstreaming of this kind of device”. In 2013, Jessica Guynn, of the Los Angeles Times, wrote that they “may still be on the fringes of mainstream consciousness. But they are not going to stay there very long”. The following year, Paul Saffo lamented for CNN that while “[i]nfo-glasses today are like PCs in 1984 – they look cool but perform a few functions that aren’t all that useful, such as taking pictures or surfing the Web while sitting in a bar with friends” — yes, “they look cool” is presented as a factual statement — in the very near future from 2014 “we are certain to be astonished by the capabilities of the device sitting on the bridge of our nose”.

Well, it has been nearly twelve years since Saffo wrote that, and the killer capabilities of smart glasses remain based entirely around the camera. And you know what? It is a pretty good feature — but it alone is not as compelling today as was a smartphone in 2008. I do not think the question is are smart glasses today akin to smartphones in 2008?; the question is more like what is different about today’s smart glasses compared to Google Glass?. To his credit, Freed attempts to answer this in noting that Meta’s Wayfarer shape instead of a sci-fi is an obvious upgrade, but I think he underplays the advancements in image, language, and speech detection since Google Glass by calling it “a Siri-like voice assistant”. GQ is not a technology publication, true, but that is among the biggest changes for a device so dependent on real-time interaction with the surrounding environment, like for translation features.

But there are problems with today’s smart glasses that remain unchanged from those that affected Google Glass. Most obviously, they are still a privacy nightmare for yourself and for others. Meta says the externally-visible recording LED must not be obstructed to record video, but people are modifying the glasses to remove that restriction. They must effectively be treated like spy glasses because they could be recording anywhere — in a public area running facial recognition software, to the apparent privacy of a massage room.

Meta is far from the only company producing glasses like these. Snap has its Spectacles and Xiaomi’s A.I. Glasses are available in China. All of these companies are responsible for developing a selfish future that prioritizes selling buyers on the advantages of an unobtrusive camera while barely acknowledging the societal impact of the same. Google is taking another kick at the can, and rumours consistently indicate Apple and Samsung are each working on their own, too. They may all say the right things about privacy, but the fundamental fact is that a barely-visible camera is a tool for abuse as much as it is entertainment.

Freed:

Whether the possibilities presented by smart glasses sound fun and appealing or like the tipping point into a dystopian nightmare is a matter of perspective. There are the obvious doubts about what happens if someone hacks your glasses and what companies like Meta are planning to do with your data (spoiler alert: it’s being used to train AI), but these aren’t so different from existing concerns around other internet-enabled devices. “Every piece of technology ever created has been used for good and bad things,” says Edward R. McNicholas, a Partner at Ropes & Gray in Washington DC who leads the firm’s global data, privacy and cybersecurity practice. “Just think of the Internet itself — it helps bad actors, but it brings the globe together, creates enormous economic opportunity, and inspires millions.” What will ultimately decide the fate of smart glasses, he says, is regulatory friction — and cultural embrace. “That is, what’s the rizz? Do the 20-somethings deem it based or cringe?”

McNicholas was admitted to the Maryland bar in 1996. His career as a lawyer is at least Millennial-aged and, as a Millennial myself, I feel pretty confident in saying he cannot use “rizz”, “based”, or “cringe” like this. It is not, in fact, lit.

I find it difficult to believe it is a coincidence there are two stories promoting Meta’s A.I. glasses appearing in the news the same week Meta laid off ten percent of its Reality Labs employees, and reallocating funds to the team developing those glasses. I am sure these things have their defenders, and may be more popular than Meta expected given the company’s long run of hardware flops. The relative success of the glasses means Meta can jettison its original messy concept of the metaverse and redefine it to suit its needs today.

But this does not feel like the nascent days of the iPhone, nor like we will not “be able to escape smart glasses” this year. I knew lots of people with smartphones in the mid-to-late-2000s, including some with original iPhones despite them not being available in Canada. Anecdotally, I do not personally know anyone who owns or is even thinking about buying smart glasses. Mind you, I know plenty of people with an Apple Watch today who did not consider it compelling even years after it launched. Maybe it is like the early days of smartwatch ownership, after all, and I simply do not notice because Meta’s glasses just look like Ray-Bans. That is, I guess, the whole point.

Despite the concept of smart glasses being the product of so much hype and excitement, it never seems to have materialized in something you can buy. Maybe that will change; maybe that has changed without me noticing it. But one of the other biggest shifts of the past ten years is how much people say they want more distance from technology. One of the predictions for 2026 in a list from the New York Times is the rise of the “dumb phone” as a status symbol. Some people who have tried smartwatches have found them more demanding over time than helpful; I stopped wearing one after four years because I need less technology in my life, not more. There is a vast gulf between what people say they want and their actual behaviour, of course, but I cannot shake the feeling this technology is still too much of an imposition. We will not need to “escape smart glasses” if people still choose not to buy them.

Riana Pfefferkorn, in an op-ed for the New York Times:

A.I. companies like xAI can and should do more not just to respond quickly and decisively when their models behave badly, but also to prevent them from generating such material in the first place. This means rigorously testing the models to learn how and why they can be manipulated into generating illegal sexual content — then closing those loopholes. But current laws don’t adequately protect good-intentioned testers from prosecution or correctly distinguish them from malicious users, which frightens companies from taking this kind of action.

To the extent A.I. companies are truly “red teaming” their models — this term has been misused by the industry, which often outsources the work to contractors in developing nations — current laws restrict the limits to which they can be taken. On this I agree with Pfefferkorn, and I think she is right to call for a change in policy.

But I am not convinced xAI is much interested in ensuring its model is that much safer. CSAM is obviously over the line and I would be surprised if anyone there were to defend Grok on that. Most anything else, however, is something I think xAI would find permissible albeit perhaps unseemly for Grok to generate. Remember: Grok is supposed to be “unfiltered”. Does it offend you? Because it should, buddy. That is the freedom you get when you look at the world through the lens of a mall-grade edgelord who will be turning 55 years old in June.

I made a mistake on Friday: instead of waiting to polish a more comprehensive article, I effectively live-blogged my shifting understanding of how StatCounter was collecting its iOS version number data by way of updates and edits to existing posts. In my own defence, I did not know the rate of users updating to iOS 26 would become as much of a story unto itself as it has. So allow me to straighten this out.

Here is the background: StatCounter publishes market share data by country and user technology based on statistics it collects from its web analytics package which, it says, is used by over a million websites totalling around five billion page views monthly. I have not heard of many of the sites using its analytics, but it seems to be a large enough and generic enough sample that it should be indicative — more so than, say, visitors to my audience-specific website. Ed Hardy, over at Cult of Mac, used StatCounter’s figures to report, on January 8, that “only about 15% of iPhone users have some version of the new operating system installed”. Hardy compared this to historical StatCounter figures showing a 63% adoption rate of iOS 18 by the same time last year, 54% on iOS 17 the year prior, and 62% on iOS 16 the year before that. If true, this would represent a catastrophic reluctance for iPhone users to update.

If true.

I do not think the iOS 26 uptake rate is about 15%. I think it is lower than the 54–63% range in previous years, but not by nearly that much. I think StatCounter has been misinterpreting iOS 26’s user base since last year because of a change Apple made to Safari.

If the phrase “user agent” does not make you respond by tipping your head to the side like my dog did when I asked him if he knew what I meant by that, you can skip this paragraph. A user agent string is a way for software to identify itself when it makes an HTTP request on behalf of a user. A user agent might describe the type and version of a web browser, the operating system, and have other information so that, in the old days, websites could check for compatibility. This leads to user agent strings that look a little silly:

Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/134.0.0.0 Safari/537.36 Edg/134.0.0.

This does not represent a Firefox user, despite starting with “Mozilla”, nor does it represent a Safari or Chrome user, despite the mentions of “Safari”, “Chrome”, and “AppleWebKit”. It is a user agent string for Microsoft Edge, which is begging to be treated like its competitors.

This is a simplified explanation, but it is important for how StatCounter works. When someone browses a website containing its analytics code, it reads the user agent string, and that is how StatCounter determines market share. The above user would be counted for Edge market share (“Edg/134.0.0”) and Windows. Which version of Windows? Well, while “NT 10.0” suggests it is Windows 10, it is also used by Edge running on Windows 11 — that part of the user string has been frozen. The Chromium team did the same thing and reduced the amount of specific information in the user agent string. This removes a method of fingerprinting and is generally fine.

This movement was spearheaded by Apple in 2017, when Ricky Mondello announced Safari Technology Preview 46 “freezes Safari’s user agent string. It will not change in the future”. But this remained a desktop-only change until September 2025, when Jen Simmons and others who work on WebKit, announced that the version of Safari shipping in iOS 26 would have its user agent stuck on the previous version of iOS:

Also, now in Safari on iOS, iPadOS, and visionOS 26 the user agent string no longer lists the current version of the operating system. Safari 18.6 on iOS has a UA string of:

Mozilla/5.0 (iPhone; CPU iPhone OS 18_6 like Mac OS X) AppleWebKit/605.1.15 (KHTML, like Gecko) Version/18.6 Mobile/15E148 Safari/604.1

And Safari 26.0 on iOS has a UA string of:

Mozilla/5.0 (iPhone; CPU iPhone OS 18_6 like Mac OS X) AppleWebKit/605.1.15 (KHTML, like Gecko) Version/26.0 Mobile/15E148 Safari/604.1

Apple justified this change only by implication, writing “we highly recommend using feature detection instead of UA string detection when writing conditional code”. But, as Jeff Johnson points out, this change does not eliminate version detection entirely:

[…] because Safari is always inseparable from the OS, so it’s possible to derive the iOS version from the Safari version, which continues to be incremented in the User-Agent. On macOS, in contrast, the latest version of Safari typically supports the three latest major OS versions, so Safari 26 can be installed on macOS 15 Sequoia and macOS 14 Sonoma in addition to macOS 26 Tahoe, and therefore the User-Agent — which actually says “OS X 10_15_7”! — is a little more effective at obscuring the OS version.

I noticed this, too, and it led to a mistake I made in my first guess at understanding why StatCounter was reporting some iOS 26 traffic, but not a lot. I thought StatCounter could have made a change to its analytics package to interpret this part of the user agent string instead, but that it may not have rolled out to all of its users. I was wrong.

What actually appears to account for iOS 26’s seemingly pitiful adoption rate is that third-party browsers like Chrome and Brave produce a user agent string that looks like this, on my iPhone running iOS 26.3:

Mozilla/5.0 (iPhone; CPU iPhone OS 26_3_0 like Mac OS X) AppleWebKit/605.1.15 (KHTML, like Gecko) CriOS/144.0.7559.53 Mobile/15E148 Safari/604.1

Safari, meanwhile, produces this user agent:

Mozilla/5.0 (iPhone; CPU iPhone OS 18_7 like Mac OS X) AppleWebKit/605.1.15 (KHTML, like Gecko) Version/26.3 Mobile/15E148 Safari/604.1

“iPhone OS 26_3_0” on Chrome, but “iPhone OS 18_7” in Safari. And iOS 18.7 also exists, with a similar user agent string as Safari, albeit with “Version/18.7” in place of “Version/26.3”. The operating system version is the same in both, however: “18_7”. StatCounter’s iOS 26 data is not reflective of all iOS users — just those using third-party browsers that still have the current iOS version in their user agent string.

Even though third-party browsers are available on iOS, most users browse the web through Safari. And that means StatCounter is almost certainly counting the vast majority of people on iOS 26 as iOS 18.7 users. I retrieved those user agent strings using StatCounter’s detection utility, which is how it says you can validate the accuracy of its statistics. And it seems they are not. (I asked StatCounter to confirm this but have not heard back.)

The actual rate of iOS 26 adoption is difficult to know right now. Web traffic to generalist websites, like the type collected by StatCounter, seems to me like it would be a good proxy had its measurement capabilities kept up with changes to iOS. Other sources, like TelemetryDeck, indicate a far higher market share — 55% as I am writing this — but its own stats reported nearly 78% adoption of iOS 18 at this time last year, far higher than StatCounter’s 63%. TelemetryDeck’s numbers are based on aggregate data from its in-app analytics product, so they should be more accurate, but that also depends on which apps integrate TelemetryDeck and who uses them. What we can see, though, is the difference between last year and this year at the same time, around 23 percentage points. For comparison, in January 2024, TelemetryDeck reported around 74% had updated to iOS 17 — iOS 26 is 19 points less.

If its reporting for this year is similarly representative, it likely indicates a 20-point slide in iOS 26 adoption. Not nearly as terrible as the misleading StatCounter dashboard suggests, but still a huge slide compared to prior years. Apple will likely update its own figures in the coming weeks for a further point of comparison. However, even though there are early indications iOS 26 is not as well-received as its predecessors, what we do not know is why that is. Fear not, however, for there are obvious conclusions to be drawn.

Hardy, later in the same Cult of Mac article:

It’s not that millions of iPhone users around the world have somehow overlooked the launch of iOS 26 followed by iOS 26.1 and iOS 26.2. They are holding off installing the upgrades because this is Apple’s most controversial new version in many years. The reason: Liquid Glass — a translucent and fluid new interface. Many elements of the UI go semi-transparent, while clever effects make it seem like users are looking through glass at objects shown on the screen behind the Control Center and pop-up windows.

David Price, of Macworld, made the same assumption based on Hardy’s story — twice:

It’s debatable whether the egregious design of last year’s OS updates falls under the category of arrogance or incompetence; perhaps it’s both. But the takeaway for Apple should be that customer loyalty is finite, and there are consequences when you consistently lower your quality-control standards. When your entire business is built on people liking you, it’s best not to take them for granted.

I have no particular affinity for Liquid Glass. I am not sure its goals are well-conceived, and I do not think it achieves those objectives.

Even so, I think the aversion to Liquid Glass is so strong among some commentators that erroneous stats are fine so long as they are confirmation of their biases. Put it this way: if just 15% of users had, indeed, upgraded to iOS 26 and the reason for so many people remaining on previous versions is Liquid Glass, surely that should mean a corrected percentage — perhaps 55%, perhaps lower — is indicative that most people are not actually bothered by Liquid Glass, right?

Yes, there is a likely 20-point gap and, if that is due to Liquid Glass, it should be cause for worry at the highest levels of Apple. iOS is a mass-market operating system. The audience is not necessarily obsessed with information density or an adequate contrast ratio. If a redesign of iOS were exciting, people would race to update, just as they did when iOS 7 was launched. They appear hesitant. Maybe the reason is Liquid Glass, or maybe something else. Or maybe there are further measurement errors.

Whatever the case, I would avoid believing articles making sweeping conclusions based on a single data point. After all, if that number is shown to be incorrect, it destabilizes the whole argument.

Jason Snell, Six Colors:

I’ve been thinking about Apple’s relationship with computer displays lately. Maybe it was the report that the iMac Pro might somehow return, combined with John Voorhees of MacStories detailing how he gave up the Studio Display for an ASUS monitor? And, of course, there’s the prospect that we may be seeing new Apple-made standalone displays in 2026.

I don’t want to go back to a world where Apple no longer makes standalone displays. But that said, I think the company’s approach to display technology needs a serious upgrade.

Co-signed. When I bought my iMac 5K in early 2019,1 it was the only big desktop display Apple sold. Now, as I think about an update, I am dismayed I will not officially be able to use it as an external display, though it seems like Luna Display might be worth a shot over Thunderbolt 3. If that does not work, it is gutting to me that I will have to effectively re-purchase this same 5K panel for my desk. And then there is the question of what I should do with my otherwise-good-just-outdated iMac.

I have been keeping an eye on Michael Tsai’s posts about displays, particularly those regarding non-Apple 27-inch 5K models. I know I have to adjust my expectations. Even so, these all look pretty tacky. Imagine being from Asus or ViewSonic and thinking I want to look at an ugly logo all day long. Still, a near-thousand-dollar difference is hard to dismiss, especially since Apple’s expensive display continues to have software problems.

And, of course, the only other display Apple sells is a 32-inch model that is over-engineered for most people, costs nearly eight thousand dollars in Canada with a stand, and has not been updated in over six years. Asus’ comparable 6K display is under $2,000, though it is not as bright and it looks pretty tragic.


  1. Arguably the worst time in modern Apple history to be shopping for a new Mac. Butterfly keyboards in all the laptops, stale desktops, and on the verge of the last batch of Intel-based Macs. ↥︎

Elizabeth Lopatto, the Verge:

Less than five years ago, I sat through the interminable Epic v. Apple antitrust trial. Real heads will remember that Apple’s lawyers heavily implied that a naked bananaman called Mr. Peely was somehow inappropriate for court. This came after a week where Apple argued that an indie storefront that users could install via Epic was a problem because it hosted porny games, calling games on Itch.io “offensive and sexualized.”

You know what’s “offensive and sexualized,” you worthless fucking cowards? Nonconsensual AI-generated images of women in bikinis spreading their legs, and of children with so-called “donut glaze” on their faces — which, by the way, were being generated at a rate of one per minute. I’d also call that “offensive, insensitive, upsetting, intended to disgust, in exceptionally poor taste” and *especially *“just plain creepy”! Do you need a back brace to stand up straight, buddy? Because at this point, I am certain you haven’t got a single vertebra.

Correctly righteous anger. That there has barely been any reckoning with this from Apple, Google, or even xAI despite at least a week of mainstream media coverage shows a callous indifference to the ongoing effects of the victims of this abuse. The best argument I can imagine — and it is a terrible argument — is that Apple’s lawyers advised the company against doing anything that looks anticompetitive since it is currently being sued by xAI. This is why it is a bad idea to rely on private corporations to do the job of regulators and law enforcement — but, still, Apple and Google should not be carrying apps from this company for as long as it continues to be a mass-scale abuse generator.

Hartley Charlton, MacRumors:

In the first week of January last year, 89.3% of MacRumors visitors used a version of iOS 18. This year, during the same time period, only 25.7% of MacRumors readers are running a version of iOS 26. In the absence of official numbers from Apple, the true adoption rate remains unknown, but the data suggests a level of hesitation toward iOS 26 that has not been seen in recent years.

The numbers being reported — 15% from StatCounter, 26% from MacRumors, and 55% from TelemetryDeck — are all over the place, but there is a clear-ish direction: people are not updating to iOS 26 like they have previous versions of iOS. There are lots of possible reasons why. Liquid Glass is the most visible explanation, but it is also possible the growing size of iOS plays a role. Apple Intelligence alone consumes nearly 7 GB of disk space on my iPhone.

Even so, the differences in these numbers are wild, especially compared to previous years where there was tighter agreement between different reports. For example, in January last year, TelemetryDeck reported about 78% were running iOS 18, StatCounter said it was 63%, and Apple itself said it was 68% of all iPhone users interacting with the App Store,1 rising to 76% adoption among users of devices four years old or newer. That is a fifteen point spread between. This year, with only third-party data so far available, it is a forty point spread between StatCounter’s 15% and TelemetryDeck’s 55%. Something is not adding up.

Update: My iPhone running iOS 26.3 is detected by StatCounter’s user agent detection tool as an iOS 18.7 device. This reflects how StatCounter says is how it collects its figures. Two other devices running iOS 26 were also detected by StatCounter as iOS 18.7 devices; however, on one of them in the Chrome browser, StatCounter correctly detected it as iOS 26.1. I also see this effect in my own limited analytics, where the only reports of iOS 26 versions are non-Safari browsers. If an analytics package relies on the OS version string in the user agent, it will also misreport iOS 26 Safari users.


  1. I corrected the description of iPhone users here thanks to a comment on Michael Tsai’s site. I am not sure what Apple means by “as measured by devices that transacted on the App Store”, though — does that include device users making in-app purchases, or recurring billing, or even automatic app updates? Or is it strictly devices purchasing or downloading apps from the App Store? ↥︎

Yesterday, I linked to a report from Ed Hardy, Cult of Mac, pointing to a shockingly low iOS 26 adoption rate compared to previous years. Hardy relied on date from StatCounter, which uses web traffic at massive scale to measure all kinds of stuff, including operating system versions.

Given that StatCounter’s data has been similar to Apple’s own reporting of version adoption in previous years, I wrote:

[…] StatCounter’s figures might be off, but it would be shocking if they were out by 40-plus percent. That would point to a serious measurement error that, somehow, did not impact previous reporting.

Well, it turns out there is likely a measurement difference that would not have impacted iOS 18 or before.

Jen Simmons and others who work on WebKit, in September:

Also, now in Safari on iOS, iPadOS, and visionOS 26 the user agent string no longer lists the current version of the operating system. Safari 18.6 on iOS has a UA string of:

Mozilla/5.0 (iPhone; CPU iPhone OS 18_6 like Mac OS X) AppleWebKit/605.1.15 (KHTML, like Gecko) Version/18.6 Mobile/15E148 Safari/604.1

And Safari 26.0 on iOS has a UA string of:

Mozilla/5.0 (iPhone; CPU iPhone OS 18_6 like Mac OS X) AppleWebKit/605.1.15 (KHTML, like Gecko) Version/26.0 Mobile/15E148 Safari/604.1

In both, you will notice iPhone OS is set to “18_6” despite only one of them actually running iOS 18.6. If StatCounter was relying on this part of the user agent string for calculating operating system version number, it could be inaccurate. There is still a Safari version number that could be a proxy for the operating system version in the latter part of the user agent string, however. On my iPhone, running iOS 26.3, the relevant section reads Version/26.3 Mobile/15E148 Safari. The iPhone OS string also reads “18_7”, which is also true for users running iOS 26.2.

It is not like StatCounter has no data for iOS 26. It shows traffic from iOS 26.1 and 26.2, indicating it likely updated its tracking metrics. It is possible some of the 18.6 and 18.7 traffic is also iOS 26 — we just do not know how much.

Data from TelemetryDeck seems more robust, and suggests about 55% of iOS users have updated to iOS 26, compared to about 78% of users one year ago running iOS 18. Not as bad as StatCounter’s figures, but still a twenty-point gap between latest version uptake last year and this year.

Thanks to Sam Gross for pointing me in this direction.

Emanuel Maiberg, of 404 Media, has been following a Telegram group in which members find workarounds to guardrails in generative A.I. products. Instead of finding interesting exploits to do clever things, though, the loopholes are being harnessed mostly to harass, abuse, and bully women. It is pretty revolting.

Maiberg:

It’s good and correct for people to be shocked and upset when they wake up one morning and see that their X feed is flooded with AI-generated images of minors in bikinis, but what is clear to me from following this Telegram community for a couple of years now is that nonconsensual sexual images of real people, including minors, is the cost of doing business with AI image generators. Some companies do a better job of preventing this abuse than others, but judging by the exploits I see on Telegram, when it comes to Grok, this problem will get a lot worse before it gets better.

It is clear Maiberg is just as disgusted with this as any person should be, so I am not trying to um, actually this, but I am not sure treating it as a “cost of doing business” is correct. The design and capabilities of these products matters immensely and, by permitting a relatively open-ended environment, xAI allows for experimentation to find its weak points. This is true of any generative A.I. product with a text box as its input. (As opposed to, say, a generative object removal tool in an image editor.) The degree of impact may also vary depending on the quality or style of the image — though, personally, I would still be alarmed if someone were harassing me with images even if they were cartoons.

Matt Burgess and Maddy Varner, Wired:

Unlike on X, where Grok’s output is public by default, images and videos created on the Grok app or website using its Imagine model are not shared openly. If a user has shared an Imagine URL, though, it may be visible to anyone. A cache of around 1,200 Imagine links, plus a WIRED review of those either indexed by Google or shared on a deepfake porn forum, shows disturbing sexual videos that are vastly more explicit than images created by Grok on X.

Caroline Haskins, Wired:

Over the past two years, Apple and Google removed a number of “nudify” and AI image-generation apps after investigations by the BBC and 404 Media found they were being advertised or used to effectively turn ordinary photos into explicit images of women without their consent.

But at the time of publication, both the X app and the stand-alone Grok app remain available in both app stores. Apple, Google, and X did not respond to requests for comment. Grok is operated by Musk’s multibillion-dollar artificial intelligence startup xAI, which also did not respond to questions from WIRED. In a public statement published on January 3, X said that it takes action against illegal content on its platform, including CSAM. “Anyone using or prompting Grok to make illegal content will suffer the same consequences as if they upload illegal content,” the company warned.

X’s threat of “consequences” would be more convincing if it had functional site moderation generally.

Apps have been kicked off the App Store for far less than what X is today. Removing it — and the rest of xAI’s apps — would be a good start, but we should not expect private companies to do the job of law enforcement and regulators. There is a good case for banning X as long as it continues to permit this poorly-moderated image generator. People should be criminally charged, too.

Ed Hardy, Cult of Mac:

Nevertheless, iOS 26 adoption is extremely low. Roughly four months after launching in mid-September, only about 15% of iPhone users have some version of the new operating system installed. That’s according to data for January 2026 from StatCounter. Instead, most users hold onto previous versions.

For comparison, in January 2025, about 63% of iPhone users had some iOS 18 version installed. So after roughly the same amount of time, the adoption rate of Apple newest OS was about four times higher.

As far as I can tell, Apple updates its own iOS version stats twice annually — once with summer numbers, and again in the winter. Last year, it updated its stats on January 24, and it indicated 68% of iPhone users who transacted with the App Store on January 21 were using iOS 18. This is pretty close to StatCounter’s 63%. In February 2024, 66% were using the then-newest iOS 17; as Hardy writes, StatCounter reported it was around 54% at the time. A greater gap, to be sure, but it was clear well over half of iPhone users had updated. StatCounter’s figures might be off, but it would be shocking if they were out by 40-plus percent. That would point to a serious measurement error that, somehow, did not impact previous reporting.

Update: Turns out there might actually be large measurement differences that did not impact previous years’ reporting.

Update: Relying on StatCounter’s data could be flawed because Safari and third-party browsers report different iOS version numbers due to a change in iOS 26.

Juli Clover, MacRumors:

Language learning app Duolingo has apparently been using the iPhone’s Live Activity feature to display ads on the Lock Screen and the Dynamic Island, which violates Apple’s design guidelines.

According to multiple reports on Reddit, the Duolingo app has been displaying an ad for a “Super offer,” which is Duolingo’s paid subscription option.

I saw this, too.

Clover points to Apple’s Human Interface Guidelines, which once advised “avoid using a Live Activity to display ads or promotions”, but now explicitly say “don’t use a Live Activity” to show ads. But the HIG is not the App Store Guidelines, and there is nothing in there expressly prohibiting this behaviour, as far as I can see. Tacky.

Kate Conger, New York Times:

Elon Musk’s artificial intelligence company, xAI, said on Tuesday that it had raised $20 billion from investors to fund its expansion in the race to train the most intelligent chatbot.

[…]

Investors in xAI’s latest funding included Fidelity, the Qatar Investment Authority and Valor Equity Partners, a firm led by Mr. Musk’s friend and former Tesla board member Antonio Gracias. Nvidia, the maker of A.I. chips, also participated. In total, xAI has raised more than $42 billion, according to PitchBook.

Twenty billion dollars is a ghastly sum of money to give anyone, let alone someone who, in his official capacity, cut funding and killed over half a million people. This feels less like an investment for direct financial gain, and more like a way for a select group of people to find influence through a company connected to a transparently corrupt government.

If you are wondering about the timing — immediately after a wave of coverage about xAI’s Grok generating a flurry of abusive and heinous imagery — it turns out this is not the first time Grok’s output has not affected major investment news. In July, shortly after Grok began parroting praise of Adolf Hitler, the U.S. government’s Chief Digital and Artificial Intelligence Office announced xAI would be one of several A.I. companies receiving up to $200 million. As of today, Grok’s stream of replies on X are still full of it generating sexualized images of real women, mostly without their consent. People with lots of money seem to think all of this is completely fine.

Jason Koebler, 404 Media:

Over the last week, users of X realized that they could use Grok to “put a bikini on her,” “take her clothes off,” and otherwise sexualize images that people uploaded to the site. This went roughly how you would expect: Users have been derobing celebrities, politicians, and random people—mostly women—for the last week. This has included underage girls, on a platform that has notoriously gutted its content moderation team and gotten rid of nearly all rules.

Ananya Bhattacharya, Rest of World:

On January 2, India’s IT ministry issued a 72-hour ultimatum to X over its artificial intelligence chatbot generating “obscene” content — specifically, sexualized or manipulated images of women and, in some cases, minors.

Maggie Harrison Dupré, Futurism:

In addition to the sexual imagery of underage girls, the women depicted in Grok-generated nonconsensual porn range from some who appear to be private citizens to a slew of celebrities, from famous actresses to the First Lady of the United States. And somehow, that was only the tip of the iceberg.

When we dug through this content, we noticed another stomach-churning variation of the trend: Grok, at the request of users, altering images to depict real women being sexually abused, humiliated, hurt, and even killed.

It is extraordinary yet, sadly, predictable to me that xAI is not treating this as a problem. Setting aside the fundamentally offensive capability that any user can tell Grok to generate photorealistic images based on someone else’s likeness — something it should be prohibited from doing — there appears to be no rush to fix this high-res misogyny on demand. A basic corporate response would be to turn off image generation capabilities until better safeguards are in place. Yet I just opened the Grok account on X, switched to the Replies tab, and it took almost no scrolling at all to find it generating images like these mere seconds ago.

The kinds of images Grok is generating should be criminal, and the people who oversee it should be held liable as should those who are prompting Grok. In my eyes, though I am not a lawyer, it is no less a case of harassment when a robot is directed by a user than when the user does it themself. I know A.I. is a nascent industry and there are things people are worried about over-regulating, but this should not be one of them. xAI is clearly overseen by adolescent-brained people who think it is sufficient to make it against company policy for users to make heinous requests, rather than sufficiently restricting its own software.

Elissa Welle, the Verge:

A viral Reddit confessional about a “major food delivery app” posted January 2nd is most likely AI-generated. The original post by user Trowaway_whistleblow alleged that an unnamed food delivery company regularly delays customer orders, calls couriers “human assets,” and exploits their “desperation” for cash, among other indefensible actions. Nearly 90,000 upvotes and four days later, it’s become increasingly clear that the post’s text is probably AI-generated.

The link from “viral Reddit confessional” is not to the thread itself, but to a link post from the Verge. That post, from Andrew J. Hawkins, still asks “[w]hich company do you think this is about?” and has not been updated to reflect the fictional status of this story.

Alex Schultz, Hard Reset:

But the allegations in the Reddit thread are not true. They are fabricated. The original post was generated by AI and was intended to defame Uber Eats. I know as much because I traded Signal messages with Trowaway_whistleblow, who additionally used AI to prepare a fake-but-sophisticated-looking “internal document” of other supposedly damning allegations; Trowaway_whistleblow circulated the document to journalists, hoping someone would take the bait and publish a libelous “scoop.”

I was shocked to see so many people posting a link to this thread credulously. I write that not to scold anyone, only out of genuine surprise that an evidence-free anonymous Reddit post with clear inconsistencies was being spread so widely. There is so much existing evidence that so-called “gig economy” workers are exploited, underpaid, and endangered, and that food delivery platforms specifically are bad for restaurants, that we simply do not need to add questionable Reddit posts to the mix. This scam played into that reputation perfectly. If anything, that is why it makes sense to be even more skeptical. Being too credulous undermines the documented failures of these platforms through the confusion it creates.

You may remember Jim Nielsen’s exploration last month — previously linked — of the various icons in MacOS Tahoe menus, and comparing them to Apple’s long-time recommendations for icon use.

Nikita Prokopov went even further:

In my opinion, Apple took on an impossible task: to add an icon to every menu item. There are just not enough good metaphors to do something like that.

But even if there were, the premise itself is questionable: if everything has an icon, it doesn’t mean users will find what they are looking for faster.

And even if the premise was solid, I still wish I could say: they did the best they could, given the goal. But that’s not true either: they did a poor job consistently applying the metaphors and designing the icons themselves.

This is a gallery of elementary problems. None of this should have shipped if someone with power internally had a critical eye for consistency and detail. If Apple deems it necessary to retain the icons, though I am not sure why it would, it should be treating this post as one giant bug report.

Update: Speaking of illegible, I think Prokopov should reconsider both the automatically-enabled snow animation and the searing yellow background.

Shalini GovilPai, Google:

At CES 2026, we’re previewing how Gemini is making Google TV even more helpful, and bringing more ways to interact with your TV across more brands and surfaces like projectors. Here’s what we announced:

[…]

  • Use Nano Banana and Veo to reimagine your personal photos or create original media directly on your TV.

I guess I am the dumb one because I cannot see any reason why I would want Google to generate new scenes from my photos — let alone “original media”, whatever that means. This seems like something that would be fun for about ten minutes, once, and then would never be touched again.

I am not sure which was the first of Google’s A.I. ads I saw that made me wonder if I was suffering from a fever. Maybe it was the one where “pandas in high fashion outfits walk the runway” — shown over thirty million times since early November, as of writing — but it just as easily could have been the one with “some crazy fruit creatures”. Maybe it was the ad with the strutting dogs. All are equally bizarre in the way a giant corporation’s ad company thinks it is offering something quirky and weird.

In any case, each ends with the same encouragement: “make more videos with a Google A.I. Pro subscription” for $27 per month. And I have to ask: who is this for?

These are not the only ads Google is running for its A.I. features. It shows them off in context for general users in Pixel phone ads, for students in ads for NotebookLM, for corporate clients with Cloud A.I., and — of course — for advertisers.

It also is not the only ad for its A.I. Pro subscription bundle. In a different one, different tasks are strung together for someone to do absolutely no work to put together a ’90s trivia night, leading to an invitation that twice tells recipients to wear platform shoes and “get ready”. The fine print might remind viewers to “check responses for accuracy”, but in the ad, the creator clicks send on a terribly written invitation immediately.

But it is those bizarre animal-themed ones which led to me looking up the advertising strategies of all the big A.I. players. I think this is a fascinating window. The ads they pay to show give us a glimpse of how they idealize the use of these products and services. While ads can sometimes be abstracted from the true function of a product or service, A.I. is already a new and confusing thing, so the narrative each company spins about its narrative seems telling of its own vision. I limited my search on this to ads of a typical video spot — no more than 30 seconds long. I did not include case studies or company-produced tutorials.

OpenAI runs a couple of types of ads. Those for ChatGPT mostly show its use as a personal assistant. There are several 30-second spots shot as oners with a pleasingly retro warmth. One has a training regiment for learning to do pull-ups; another is about siblings on a road trip. The same kind of messaging is used in a series of shorter spots. It also runs ads for its Codex agent, obviously targeted at developers, that are more clinical.

All of these seem practical to me. I could not find any current ads from OpenAI as disconnected from reality as Google’s. Just as notable is the focus of OpenAI’s spots — Google’s Ads Transparency Centre says the company is running about two hundred ads in Canada right now, most of which are variations in size, targeting, and placement of the shorter practical examples above, plus ads for Codex. For comparison, Google’s ads are all over the place. It is running around twenty thousand ads right now in Canada and, though not all of them are for A.I. features, many are, and you can tell from the examples above how much Google is just throwing stuff at the wall.

Anthropic’s ads are far more limited. All are for Claude and feature a video ad with no indication of how it is being used. It simply says “Claude is A.I. for […] all of us, anywhere” with overhead shots of different scenes representing different professions. This is basically the same sentiment as OpenAI’s ads, but executed without any specificity or examples. The company’s YouTube channel has plenty of case studies and demos, but no similar video spots.

If Anthropic is trying to mimic OpenAI’s quiet confidence, Perplexity has chosen overt aggression. Quasi-influencer types follow a similar script saying ChatGPT makes things up, and that is why you should trust Perplexity as it “searches the entire internet in less than one second and gives you one verified answer”. This explanation avoids acknowledging how much Perplexity depends on external A.I. models, including OpenAI’s GPT. In two of the ads, the narrator asks health-related questions, which is the boldest and maybe most reckless use case I have seen in any A.I. ad. There is nothing wrong with the answers it has generated, so far as I can tell, but it seems like a higher-stakes use than, like, a cake recipe, for example. In a variation of this ad type, the “influencer” looks generated, and contains Google search result screenshots that are at least twelve years old. Whatever the case, Perplexity must be spending huge amounts on its advertising. Each of these has been shown tens of millions of times, and two have hundreds of millions of views.

Meta’s ads are bizarre. I could not find any video ads for A.I. stuff — or metaverse-related things — through Google’s Ads Transparency Centre, but the company has a handful of 30-second spots on its YouTube channel. Almost like Google, Meta does not seem to know why people would use its A.I. assistant, so it comes up with offbeat premises that fall flat. In one, described as a way to “help sharpen your skills and transform the way you learn and understand complex subjects”, a guy is meeting his partner’s dad, a thermodynamicist. What a real person who is good at conversations would do is ask the dad a bunch of questions. What Meta thinks you should do is ask its assistant those questions instead and develop a fake level of expertise.

Also, not for nothing, but when I asked the same question in the web version of Meta A.I., it responded not with the tight and efficient explanation in the ad, but with:

😂 Thermodynamics-ist? Sounds like a real party crusher, eh? 🤣 Okay, serious face… A thermodynamicist is like, someone who’s all about the study of heat, energy, and how they transform into each other. They’re like the ultimate energy nerds, but in a good way! 🤓 They work with stuff like engines, refrigeration, and even black holes (mind blown, right? 🤯). Basically, they help us understand how to harness and use energy efficiently. Pretty cool, eh?

Am I meant to be charmed by “okay, serious face…” and four emoji in this barely coherent response? Because I am not.

In another ad, a guy somehow turns a large pan of what appears to be milk and two other pans of mystery liquid into pancakes for his kid. The kid pushes the plate off the counter, the dad asks “how can I get my two-year-old to eat his breakfast?”, and Meta A.I. knows that the kid likes when he makes silly faces. It is unclear to me how it knows this — whether it is automatically personalized to this user, or something the dad needed to teach the system, or if it is just an assumption that all two-year-olds like silly faces. The impression I get, though, is that it is distinctly unbelievable, and I do not want to give Meta knowledge of my entire life to find out if I am right.

To round this out, I looked for similar ads from Microsoft and Apple. I scrolled through hundreds of Microsoft ads on Google’s Ads Transparency tool, virtually all of which were Xbox related. I also checked its YouTube channel. I did not see any ads of the same type as above. I also found only a couple of ads for Apple Intelligence on Apple’s YouTube channel, both of which are for Clean Up in Photos. Apple seems to have cleaned up its YouTube channel overall, removing a whole bunch of older ads including some for Apple Intelligence.

I do not want to overstate how much these ads tell us — they are ads, you know? — but I think I learned something from the way each of these businesses thinks of its own products. In OpenAI, I see confidence; in Anthropic and Perplexity, I see an attempt to catch up. And in Google and Meta, I see established companies that are desperate to prove themselves — particularly in Google’s case, as I still cannot understand why generating arbitrary video is supposed to be compelling to a broad audience.

In the most practical and grounded ads, what I do not see are significant leaps beyond what a search engine today could do. OpenAI’s ads show ChatGPT summarizing a workout plan, but there are loads of those on external websites. Guides to road tripping through the Blue Ridge Parkway are plentiful. The same is true of the responses in Perplexity’s ads. What I see most in these ads are the big “pure” A.I. players normalizing their raison d’être, and established mega corporations entirely out of touch with what someone might want to do. Both are embarrassing in their own way for what is often pitched as the most revolutionary technology since the internet.

2025 was a tough year for pretty much everyone I know, and I imagine the coming year will bring more of the same dreary news. I thought I would start off 2026 with something much dumber. I have plumbed the depths and found the bottom of the deepest of barrels, and that is how you are reading an “I asked ChatGPT” post.

There is a minor point here. I needed to use up a cup of milk and a quarter-cup of sour cream lurking in my fridge, and I wanted to make some kind of quick bread or loaf with it. The problem is that I am an inexperienced baker and I do not have a great sense of the proportion of other ingredients I would need.

I think this is the kind of thing a more advanced search engine could help with. I have previously used recipe finder tools like Supercook and RecipeRadar but, while they often ask what ingredients are available, they usually do not do so in specific quantities, necessitating looking through a bunch of recipes to find one that fits. Something that can process longer natural language queries seems tailor made for this kind of task.

I told it “I would like loaf/quick bread recipes that will use up 1 cup of milk and ¼ cup of sour cream”, and ChatGPT gave me four options. The first, a “classic vanilla quick bread”, looked adequate, but boring. (Also, a careful reader will quibble with its “optional add-ins (1–1½ cups total)” after which is a list of, among other things, cinnamon, and I do not think you should add a whole cup of cinnamon to anything.) I do like cinnamon, though, and I settled on the fourth suggestion, a “cinnamon swirl coffee bread”. Here is the entirety of its recipe:

Batter

2 cups flour
¾ cup sugar
1½ tsp baking powder
½ tsp baking soda
½ tsp salt
1 cup milk
¼ cup sour cream
⅓ cup oil
2 eggs
1 tsp vanilla

Swirl

⅓ cup brown sugar
1½ tsp cinnamon

Bake

Layer batter + swirl in pan, gently knife once

350°F for 50–60 min

The ingredient list, to my untrained eye, seems fine. The instructions are obviously incoherent. More comprehensive directions were offered for the “classic vanilla quick bread” above it, which I missed because I only followed this recipe. Since I made a different loaf recently, however, I had a rough idea of what I should do. Also, I made a couple of minor changes:

  • I substituted vanilla extract for a few dashes of Fee Brothers’ cardamom bitters.

  • I had no ground cinnamon on hand and grating a cinnamon stick is tedious, so I stopped after about a teaspoon.

While these ingredient substitutions might affect the flavour, they would not materially affect the chemistry.

A cross-section of a golden loaf with a brown sugar swirl in the middle.

The resulting loaf is fine. I was hoping for either catastrophic failure or incredible success to more wholly justify this low-effort post, but it was just fine. Better than I expected, given where it came from, though the brown sugar swirl is achingly sweet and settled in the middle despite my best attempts. I still do not know what “gently knife once” is supposed to mean. I would not confuse this with a professional baker’s work, of course, but that is more like operator error. I wish the directions were, overall, clearer; if I had little to no previous experience baking a quick bread, I might have been lost.

I have experimented with ChatGPT and food before, particularly for weekly meal planning, and I have never been satisfied with its results. This, though, worked pretty well for me. I got to use up a couple of things in my fridge and made an okay dessert from it. Happy New Year.

Rindala Alajaji, the Electronic Frontier Foundation:

Age-verification mandates create barriers along lines of race, disability, gender identity, sexual orientation, immigration status, and socioeconomic class. While these requirements threaten everyone’s privacy and free-speech rights, they fall heaviest on communities already facing systemic obstacles.

This is a compelling list of reasons why age verification laws are bad for all of us. I have mixed feelings about their need and implementation so far. In theory, I think checking user ages can be justified in a number of circumstances. Many apps and websites have some boilerplate text claiming they have no interest in serving children, and rely on kids self-authorizing. This is obviously insufficient. Also, Apple wants your iPhone to replace your wallet, but is somehow uninterested in using the information you provide, which is bizarre.

But giving tech companies even more information and control seems similarly fraught. Existing operating system-level options for parents are frequently broken, so why would we entrust them with some form of valid identification? And the current patchwork of laws and proposals could mean a worst-of-both-worlds situation: depending on where the company is governed, you might have to provide documentation to centralized app distributors, and to individual websites and apps. And, as Alajaji writes, the people who will struggle most to satisfy these requirements are already discriminated against.

I have written posts in which I am swayed by proposals for age verification. I am also convinced by the ten arguments Alajaji lists, many of which feel U.S. specific but are probably even more burdensome in developing countries. These problems seem to be varying degrees of insurmountable. But the great freedom created by the web has not yet been met with commensurate responsibility, either.

Howard Oakley:

In real life, whiteouts are dangerous because they’re so disorienting. There’s no horizon, no features in the landscape, and no clues to navigation. We see and work best in visual environments that are rich in colour and tonal contrasts. Tahoe has continued a trend for Light Mode to be bleached-out white, and Dark Mode to be a moonless night. Seeing where controls, views and contents start and end is difficult, and leaves them suspended in the whiteout.

Oakley reviews several lingering problems with Liquid Glass in MacOS, but the above remains the most — and I use this word intentionally — glaring issue I have with it. It is a problem that becomes entirely clear as you scroll to the bottom of Oakley’s post and find a screenshot from — I think — Mac OS X Mavericks with evident precision and contrast. I do not think Apple should have frozen this interface in time, nor that there are no changes made since which have been an improvement. However, though the exact same elements remain in today’s MacOS, they lack a similarly rigorous structure and care. It is notable how many translucency controls have been added in the intervening years.

Rudy Fraser, founder and CEO of Blacksky, made two big announcements on the occasion of it being the first anniversary of its launch. The first is online cash payments in USD, which I am a little unclear about, but relies in part on the trust implicit in AT Protocol connections.

The second, though, is easier for me to understand:

As Mozilla has been saying, we need to decentralize the internet. Decentralization necessarily requires running your own servers and as mentioned above by Moxie, founder of the beloved Signal protocol, no one wants to run their own servers, and never will. If you take seriously that both are true, how do you square the two?

[…]

I’m excited to say: promises made, promises kept. Clinton has built a mobile-friendly web app that will allow users to create a new PDS [Personal Data Server] that is truly one-click and will be hosted on Blacksky’s infrastructure. […]

I am intrigued by the possibility of announcements like this one. It offers a meaningful layer of self-governance beyond monolithic social platforms, made possible only by open standards like AT Protocol and the rival ActivityPub.