A.I. ‘Gives Me the Prompts, I Paint the Pictures’ ⇥ sauropods.win
John Conway flipped the roles of artist and “artificial intelligence”, and painted pictures as suggested by generated prompts. Clever idea for an exercise.
Search Results for: "artificial intelligence"
Assigning Copyright Liability for the Output of ‘A.I.’ Systems ⇥ techdirt.com
Camilla Hodgson, Financial Times (syndicated at Ars Technica):
AI models are “trained” on data, such as photographs and text found on the internet. This has led to concern that rights holders, from media companies to image libraries, will make legal claims against third parties who use the AI tools trained on their copyrighted data.
The big three cloud computing providers have pledged to defend business customers from such intellectual property claims. But an analysis of the indemnity clauses published by the cloud computing companies show that the legal protections only extend to the use of models developed by or with oversight from Google, Amazon and Microsoft.
Ira Rothken, Techdirt:
Here’s the crux: the LLM itself can’t predict the user’s intentions. It simply processes patterns based on prompts. The LLM learning machine and idea processor shouldn’t be stifled due to potential user misuse. Instead, in the rare circumstances when there is a legitimate copyright infringement, users ought to be held accountable for their prompts and subsequent usage and give the AI LLM “dual use technology” developers the non-infringing status of the VCR manufacturer under the Sony Doctrine.
It seems there are two possible points of copyright infringement: input and output. I find the latter so much more interesting.
It seems, to me, to depend on how much of a role machine learning models play in determining what is produced, and I find that fascinating. These models have been marketed as true artificial intelligence but, in their defence, are often compared to photocopiers — and there is a yawning chasm between those perspectives. It makes sense for Xerox to bear zero responsibility if someone uses one of its machines to duplicate an entire book. Taking it up a notch, I have no idea if a printer manufacturer might be found culpable for permitting counterfeiting currency — I am not a lawyer — but it is noteworthy anti-duplication measures have been present in scanners and printers for decades, yet Bloomberg reported in 2014 that around 60% of fake U.S. currency was made on home-style printers.
But those are examples of strict duplication — these devices have very little in the way of a brain, and the same is true of a VHS recorder. Large language models and other forms of generative “intelligence” are a little bit different. Somewhere, something like a decision happens. It seems plausible an image generator could produce a result uncomfortably close to a specific visual style without direct prompting by the user, or it could clearly replicate something. In that case, is it the fault of the user or the program, even if it goes unused and mostly unseen?
To emphasize again, I am not a lawyer while Rothken is, so I am just talking out of my butt. These tools are raising some interesting questions is all I want to highlight. Fascinating times ahead.
When I was much younger, I assumed people who were optimistic must have misplaced confidence. How anyone could see a future so bright was a complete mystery, I reasoned, when what we are exposed to is a series of mistakes and then attempts at correction from public officials, corporate executives, and others. This is not conducive to building hope — until I spotted the optimistic part: in the efforts to correct the problem and, ideally, in preventing the same things from happening again.
If you measure your level of optimism by how much course-correction has been working, then 2023 was a pretty hopeful year. In the span of about a decade, a handful of U.S. technology firms have solidified their place among the biggest and most powerful corporations in the world, so nobody should be surprised by a parallel increase in pushback for their breaches of public trust. New regulations and court decisions are part of a democratic process which is giving more structure to the ways in which high technology industries are able to affect our lives. Consider:
Right-to-repair legislation has been proposed or has become law in Canada, the E.U., and many U.S. states — and probably some other places that I do not track as closely.
Google’s dominance in a range of technologies was tested in the U.S. by its loss against Epic Games, a Play Store settlement, and an antitrust trial which will be decided later this year.
Meta settled a privacy lawsuit in California, was fined in the E.U. — also on privacy grounds — and was fined in a separate E.U. case, again for privacy reasons, and its very business model is in question because of — guess what — privacy.
Apple switched to USB-C across its new iPhone lineup, bringing a universal standard to all models and, for Pro models, the first change to data transfer speeds since the release of the third-generation iPod in 2003 by implementing the USB 3.2 spec.
The Digital Markets Act and Digital Services Act went into effect in the E.U. this year.
Thanks to E.U. pressure, Adobe will no longer acquire Figma.
That is a lot of change in one year and not all of it has been good. The Canadian government went all-in on the Online News Act which became a compromised disaster; there are plenty of questions about the specific ways the DMA and DSA will be enforced; Montana legislators tried to ban TikTok.
It is also true and should go without saying that technology companies have done plenty of interesting and exciting things in the past year; they are not cartoon villains in permanent opposition to the hero regulators. But regulators are also not evil. New policies and legal decisions which limit the technology industry — like those above — are not always written by doddering out-of-touch bureaucrats and, just as importantly, businesses are not often trying to be malevolent. For example, Apple has arguably good reasons for software validation of repairs; it may not be intended to prevent users from easily swapping parts, but that is the effect its decision has in the real world. What matters most to users is not why a decision was made but how it is experienced. Regulators should anticipate problems before they arise and correct course when new ones show up.
This back-and-forth is something I think will ultimately prove beneficial, though it will not happen in a straight line. It has encouraged a more proactive dialogue for limiting known negative consequences in nascent technologies, like avoiding gender and racial discrimination in generative models, and building new social environments with less concentrated power. Many in tech industry love to be the disruptor; now, the biggest among them are being disrupted, and it is making things weird and exciting.
These changes do not necessarily need to be made from the effects of regulatory bodies. Businesses are able to make things more equitable for themselves, should they so choose. They can be more restrictive about what is permitted on their platforms. They can empower trust and safety teams to assess how their products and services are being used in the real world and adjust them to make things better.
Mike Masnick, Techdirt:
Let’s celebrate actual tech optimism in the belief that through innovation we can actually seek to minimize the downsides and risks, rather than ignore them. That we can create wonderful new things in a manner that doesn’t lead many in the world to fear their impact, but to celebrate the benefits they bring. The enemies of techno optimism are not things like “trust and safety,” but rather the naive view that if we ignore trust and safety, the world will magically work out just fine.
There are those who believe “the arc of the universe […] bends toward justice” is a law which will inevitably be correct regardless of our actions, but it is more realistic to view that as a call to action: people need to bend that arc in the right direction. There are many who believe corporations can generally regulate themselves on these kinds of issues, and I do too — to an extent. But I also believe the conditions by which corporations are able to operate are an ongoing negotiation with the public. In a democracy, we should feel like regulators are operating on our behalf, and much of the policy and legal progress made last year certainly does. This year can be more of the same if we want it to be. We do not need to wait for Meta or TikTok to get better at privacy on their own terms, for example. We can just pass laws.
As I wrote at the outset, the way I choose to be optimistic is to look at all of the things which are being done to correct the imbalanced and repair injustices. Some of those corrections are being made by businesses big and small; many of them have advertising and marketing budgets celebrating their successes to the point where it is almost unavoidable. But I also look at the improvements made by those working on behalf of the public, like the list above. The main problem I have with most of them is how they have been developed on a case-by-case basis which, while setting precedent, is a fragile process open to frequent changes.
That is true, too, for self-initiated changes. Take Apple’s self-repair offerings, which it seems to have introduced in response to years of legislative pressure. It has made parts, tools, and guides available in the United States and in a more limited capacity across the E.U., but not elsewhere. Information and kits are available not from Apple’s own website, but a janky looking third-party. It can stop making this stuff available at any time in areas where it is not legally obligated to provide these resources, which is another reason why it sucks for parts to require software activation. In 2023, Apple made its configuration tools more accessible, but only in regions where its self-service repair program is provided.
People ought to be able to have expectations — for repairs, privacy, security, product reliability, and more. The technology industry today is so far removed from its hackers-in-a-garage lore. Its biggest players are among the most powerful businesses in the world, and should be regulated in that context. That does not necessarily mean a whole bunch of new rules and bureaucratic micromanagement, but we ought to advocate for structures which balance the scales in favour of the public good.
If there was one technology story we will remember from 2023, it was undeniably the near-vertical growth trajectory of generative “artificial intelligence” products. It is everywhere, and it is being used by normal people globally. Yet it is, for all intents and purposes, a nascent sector, and that makes this a great time to set some standards for its responsible development and, more importantly, its use. Nobody is going to respond to this perfectly — not regulators and not the companies building these tools. But they can work together to set expectations and standards for known and foreseeable problems. It seems like that is what is happening in the E.U. and the United States.
That is how I am optimistic about technology now.
NYT: Apple Explores Licensing Deals With Publishers for Large Language Model Training ⇥ nytimes.com
Benjamin Mullin and Tripp Mickle, New York Times:
Apple has opened negotiations in recent weeks with major news and publishing organizations, seeking permission to use their material in the company’s development of generative artificial intelligence systems, according to four people familiar with the discussions.
This is very different from the way existing large language models have been trained.
Kali Hays, of Business Insider, in November:
Most tech companies seemed to agree that being required to pay for the huge amounts of copyrighted material scraped from the internet and used to train large language models behind AI tools like Meta’s Llama, Google’s Bard, and OpenAI’s ChatGPT would create an impossible hurdle to develop the tech.
“Generative AI models need not only a massive quantity of content, but also a large diversity of content,” Meta wrote in its comment. “To be sure, it is possible that AI developers will strike deals with individual rights holders, to develop broader partnerships or simply to buy peace from the threat of litigation. But those kinds of deals would provide AI developers with the rights to only a minuscule fraction of the data they need to train their models. And it would be impossible for AI developers to license the rights to other critical categories of works.”
If it were necessary to license published materials for training large language models, it would necessarily limit the viability of those models to those companies which could afford the significant expense. Mullin and Mickle report Apple is offering “at least $50 million”. Then again, large technology companies are already backing the “A.I.” boom.
Mullin and Mickle:
The negotiations mark one of the earliest examples of how Apple is trying to catch up to rivals in the race to develop generative A.I., which allows computers to create images and chat like a human. […]
Tim Bradshaw, of the Financial Times, as syndicated by Ars Technica:
Apple’s latest research about running large language models on smartphones offers the clearest signal yet that the iPhone maker plans to catch up with its Silicon Valley rivals in generative artificial intelligence.
The paper, entitled “LLM in a Flash,” offers a “solution to a current computational bottleneck,” its researchers write.
Both writers frame this as Apple needing to “catch up” to Microsoft — which licenses generative technology from OpenAI — Meta, and Google. But surely this year has demonstrated both how exciting this technology is and how badly some of these companies have fumbled their use of it — from misleading demos to “automated bullshit”. I have no idea how Apple’s entry will fare in comparison but it may, in retrospect, look wise for it to dodge this kind of embarrassment and the legal questions of today’s examples.
‘A.I.’ and Trust ⇥ schneier.com
Bruce Schneier, Slate:
Knowing that they are under constant surveillance changes how people behave. They conform. They self-censor, with the chilling effects that brings. Surveillance facilitates social control, and spying will only make this worse. Governments around the world already use mass surveillance; they will engage in mass spying as well.
Corporations will spy on people. Mass surveillance ushered in the era of personalized advertisements; mass spying will supercharge that industry. Information about what people are talking about, their moods, their secrets — it’s all catnip for marketers looking for an edge. The tech monopolies that are currently keeping us all under constant surveillance won’t be able to resist collecting and using all of that data.
And Schneier on his blog, a republished transcript of a September talk at Harvard:
In this talk, I am going to make several arguments. One, that there are two different kinds of trust—interpersonal trust and social trust—and that we regularly confuse them. Two, that the confusion will increase with artificial intelligence. We will make a fundamental category error. We will think of AIs as friends when they’re really just services. Three, that the corporations controlling AI systems will take advantage of our confusion to take advantage of us. They will not be trustworthy. And four, that it is the role of government to create trust in society. And therefore, it is their role to create an environment for trustworthy AI. And that means regulation. Not regulating AI, but regulating the organizations that control and use AI.
If you only have time for one of these, I recommend the latter. It is more expansive, thoughtful, and makes me reconsider how regulatory framing ought to work for these technologies.
Both are great, however, and worth your time.
About That OpenAI ‘Breakthrough’ ⇥ theverge.com
You are probably sick of hearing about OpenAI palace intrigue; I am, too, but I have a reputation to correct. I linked favourably to something published at Fast Company recently, and I must repent. I have let you down and I have let myself down and, happily, I can fix that.
On Monday, which only just happened earlier this week, Fast Company’s Mark Sullivan asked the question “Is an AGI breakthrough the cause of the OpenAI drama?”; here is the dek, with emphasis added:
Some have theorized that Sam Altman and the OpenAI board fell out over differences on how to safeguard an AI capable of performing a wide variety of tasks better than humans.
Who are these “some”, you might be asking? Well, here is how the second paragraph begins:
One popular theory on X posits that there’s an unseen factor hanging in the background, animating the players in this ongoing drama: the possibility that OpenAI researchers have progressed further than anyone knew toward artificial general intelligence (AGI) […]
Yes, some random people are tweeting and that is worthy of a Fast Company story. And, yes, that is the only source in this story — there is not even a link to the speculative tweets.
While stories based on tweeted guesswork are never redeemable, the overall thrust of Sullivan’s story appeared to be confirmed yesterday in a paywalled Information report and by Anna Tong, Jeffrey Dastin, and Krystal Hu of Reuters:
Ahead of OpenAI CEO Sam Altman’s four days in exile, several staff researchers wrote a letter to the board of directors warning of a powerful artificial intelligence discovery that they said could threaten humanity, two people familiar with the matter told Reuters.
[…]
The sources cited the letter as one factor among a longer list of grievances by the board leading to Altman’s firing, among which were concerns over commercializing advances before understanding the consequences. Reuters was unable to review a copy of the letter. The staff who wrote the letter did not respond to requests for comment.
But Alex Heath, of the Verge, reported exactly the opposite:
Separately, a person familiar with the matter told The Verge that the board never received a letter about such a breakthrough and that the company’s research progress didn’t play a role in Altman’s sudden firing.
Heath’s counterclaim relies on a single source compared to Reuters’ two — I am not sure how many the Information has — but note that none of them require that you believe OpenAI has actually made a breakthrough in artificial general intelligence. This is entirely about whether the board received a letter making that as-yet unproven claim and, if that letter was recieved, whether it played a role in this week of drama.
Regardless, any story based on random internet posts should be canned by an editor before anyone has a chance to publish it. Even if OpenAI really has made such a breakthrough and there really was a letter that really caused concern for the company’s board, that Sullivan article is still bad — and Fast Company should not have published it.
Update: In a lovely coincidence, I used the same title for this post as Gary Marcus did for an excellent exploration of how seriously we ought to take this news. (Via Charles Arthur.)
Artificial Intelligence and Automation Bias ⇥ newyorker.com
Eyal Press, the New Yorker:
In June, an appellate court ordered the N.Y.P.D. to turn over detailed information about a facial-recognition search that had led a Queens resident named Francisco Arteaga to be charged with robbing a store. The court requested both the source code of the software used and information about its algorithm. Because the technology was “novel and untested,” the court held, denying defendants access to such information risked violating the Brady rule, which requires prosecutors to disclose all potentially exculpatory evidence to suspects facing criminal charges. Among the things a defendant might want to know is whether the photograph that had been used in a search leading to his arrest had been digitally altered. DataWorks Plus notes on its Web site that probe images fed into its software “can be edited using pose correction, light normalization, rotation, cropping.” Some systems enable the police to combine two photographs; others include 3-D-imaging tools for reconstructing features that are difficult to make out.
This example is exactly why artificial intelligence needs regulation. There are many paragraphs in this piece which contain alarming details about overconfidence in facial recognition systems, but proponents of allowing things to play out as things are currently legislated might chalk that up to human fallibility. Yes, software might present a too-rosy impression of its capabilities, one might argue, but it is ultimately the operator’s responsibility to cross-check things before executing an arrest and putting an innocent person in jail. After all, there are similar problems with lots of forensic tools.
Setting aside how much incentive there is for makers of facial recognition software to be overconfident in their products, and how much leeway law enforcement seems to give them — agencies kept signing contracts with Clearview, for example, even after stories of false identification and arrests based on its technology — one could at least believe searches use photographs. But that is not always the case. DataWorks Plus markets tools which allow searches using synthesized faces which are based on real images, as Press reports — but you will not find that on its website. When I went looking, DataWorks Plus seems to have pulled the page where it appeared; happily, the Internet Archive captured it. You can see in its examples how it is filling in the entire right-hand side of someone’s face in a “pose correction” feature.
It is plausible to defend this as just a starting point for an investigation, and a way to generate leads. If it does not pan out, no harm, right? But it does seem to this layperson like a computer making its best guess about someone’s facial features is not an ethical way of building a case. This is especially true when we do not know how systems like these work, and it does not inspire confidence that there are no standards specific with which “A.I.” tools must comply.
There has been a wave of artificial intelligence regulatory news this week, and I thought it would be useful to collect a few of those stories in a single post.
Earlier this week, U.S. president Joe Biden issued an executive order:
My Administration places the highest urgency on governing the development and use of AI safely and responsibly, and is therefore advancing a coordinated, Federal Government-wide approach to doing so. The rapid speed at which AI capabilities are advancing compels the United States to lead in this moment for the sake of our security, economy, and society.
Reporting by Josh Boak and Matt O’Brien of the Associated Press indicates this executive order was informed by several experts in the technology and human rights sectors. Unfortunately, it seems that something I interpreted as a tongue-in-cheek statement to the adversary of the latest “Mission: Impossible” movie is being taken seriously and out of context by some.
Steven Sinofsky — who, it should be noted, is a board partner at Andreessen Horowitz which still has as its homepage that ridiculous libertarian manifesto which is, you know, foreshadowing — is worried about that executive order:
I am by no means certain if AI is the next technology platform the likes of which will make the smartphone revolution that has literally benefitted every human on earth look small. I don’t know sitting here today if the AI products just in market less than a year are the next biggest thing ever. They may turn out to be a way stop on the trajectory of innovation. They may turn out to be ingredients that everyone incorporates into existing products. There are so many things that we do not yet know.
What we do know is that we are at the very earliest stages. We simply have no in-market products, and that means no in-market problems, upon which to base such concerns of fear and need to “govern” regulation. Alarmists or “existentialists” say they have enough evidence. If that’s the case then then so be it, but then the only way to truly make that case is to embark on the legislative process and use democracy to validate those concerns. I just know that we have plenty of past evidence that every technology has come with its alarmists and concerns and somehow optimism prevailed. Why should the pessimists prevail now?
This is a very long article with many arguments against the Biden order. It is worth reading in full; I have just pulled its conclusion as a summary. I think there is a lot to agree with, even if I disagree with its conclusion. The dispute is not between optimism and pessimism; it is between democratically regulating industry, and allowing industry to dictate the terms of if and how it is regulated.
That there are “no in-market products […] upon which to base such concerns” is probably news to companies like Stable AI and OpenAI, which sell access to Eurocentric and sexually biased models. There are, as some will likely point out, laws in many countries against bias in medical care, hiring, policing, housing, and other significant areas set to be revolutionized by A.I. in the coming years. That does not preclude the need for regulations specifically about how A.I. may be used in those circumstances, though.
The point is this: if you accept the premise that regulation locks in incumbents, then it sure is notable that the early AI winners seem the most invested in generating alarm in Washington, D.C. about AI. This despite the fact that their concern is apparently not sufficiently high to, you know, stop their work. No, they are the responsible ones, the ones who care enough to call for regulation; all the better if concerns about imagined harms kneecap inevitable competitors.
[…]
In short, this Executive Order is a lot like Gates’ approach to mobile: rooted in the past, yet arrogant about an unknowable future; proscriptive instead of adaptive; and, worst of all, trivially influenced by motivated reasoning best understood as some of the most cynical attempts at regulatory capture the tech industry has ever seen.
There is a neat rhetorical trick in both Sinofsky’s and Thompson’s articles. It is too early to regulate, they argue, and doing so would only stifle the industry and prevent it from reaching its best potential and highest aspirations. Also, it is a little bit of a smokescreen to call it a nascent industry; even if the technology is new, many of the businesses working to make it a reality are some of the world’s most valuable. Alas, it becomes more difficult to create rules as industries grow and businesses become giants — look, for example, to Sinofsky’s appropriate criticism of the patchwork approach to proposed privacy laws in several U.S. states, or Thompson’s explanation of how complicated it is to regulate “entrenched” corporations like Facebook and Google on privacy grounds given their enormous lobbying might.
These are not contradictory arguments, to be clear; both writers are, in fact, raising a very good line of argument. Regulations enacted on a nascent industry will hamper its growth, while waiting too long will be good news for any company that can afford to write the laws. Between these, the latter is a worse option. Yes, the former approach means a new industry faces constraints on its growth, both in terms of speed and breadth. With a carefully crafted regulatory framework with room for rapid adjustments, however, that can actually be a benefit. Instead of a well poisoned by years of risky industry experiments on the public, A.I. can be seen as safe and beneficial. Technologies made in countries with strict regulatory regimes may be seen as more dependable. There is the opportunity of a lifetime to avoid entrenching the same mistakes, biases, and problems we have been dealing with for generations.
Where I do agree with Sinofsky and Thompson is that such regulation should not be made by executive order. However, regardless of how much I think the mechanism of this policy is troublesome and much of the text of the order is messy, it is wrong to discard the very notion of A.I. regulation simply on this basis.
A group of academics published a joint paper concerning A.I. development, which I thought was less alarmist and more grounded than most of these efforts:
The rate of improvement is already staggering, and tech companies have the cash reserves needed to scale the latest training runs by multiples of 100 to 1000 soon. Combined with the ongoing growth and automation in AI R&D, we must take seriously the possibility that generalist AI systems will outperform human abilities across many critical domains within this decade or the next.
What happens then? If managed carefully and distributed fairly, advanced AI systems could help humanity cure diseases, elevate living standards, and protect our ecosystems. The opportunities AI offers are immense. But alongside advanced AI capabilities come large-scale risks that we are not on track to handle well. Humanity is pouring vast resources into making AI systems more powerful, but far less into safety and mitigating harms. For AI to be a boon, we must reorient; pushing AI capabilities alone is not enough.
John Davidson, columnist at the Australian Financial Review, interviewed Andrew Ng, who co-founded Google Brain:
“There are definitely large tech companies that would rather not have to try to compete with open source [AI], so they’re creating fear of AI leading to human extinction.
“It’s been a weapon for lobbyists to argue for legislation that would be very damaging to the open-source community,” he said.
Ng is not an anti-regulation hardliner. He acknowledges the harms already caused by A.I. and supports oversight.
Dan Milmo and Kiran Stacey, of the Guardian, covered this week’s Bletchley Park A.I. safety summit:
The possibility that AI can wipe out humanity – a view held by less hyperbolic figures than Musk – remains a divisive one in the tech community. That difference of opinion was not healed by two days of debate in Buckinghamshire.
But if there is a consensus on risk among politicians, executives and thinkers, then it focuses on the immediate fear of a disinformation glut. There are concerns that elections in the US, India and the UK next year could be affected by malicious use of generative AI.
I do not love the mainstreaming of the apparently catastrophic risks of A.I. on civilization because it can mean one of two possibilities: either its proponents are wrong and are using it for cynical or attention-seeking purposes, or they are right. This used to be something which was regarded as ridiculous science fiction. That apparently serious and sober people see it as plausible is discomforting.
‘The Problem With Jon Stewart’ Is Ending ⇥ nytimes.com
Benjamin Mullin, John Koblin, and Tripp Mickle, New York Times:
Jon Stewart’s show on Apple’s streaming service is abruptly coming to an end, according to several people with knowledge of the decision, the result of creative differences between the tech giant and the former “Daily Show” host.
[…]
But Mr. Stewart and Apple executives had disagreements over some of the topics and guests on “The Problem,” two of the people said. Mr. Stewart told members of his staff on Thursday that potential show topics related to China and artificial intelligence were causing concern among Apple executives, a person with knowledge of the meeting said. As the 2024 presidential campaign begins to heat up, there was potential for further creative disagreements, one of the people said.
I am taking the rationale cited in this report with a grain of salt. When working at the Wall Street Journal, Mickle was one of the reporters on a story about Apple’s apparent aversion to sexual, violent, profane, and dark media. It is hard to see that story as accurate; Apple has several shows which contain all of those things to some degree.
However, its geopolitical exposure was another rumoured point of contention. In 2019, Alex Kantrowitz and John Paczkowski reported for Buzzfeed News that Apple was one of several studios which wanted to avoid irking powerful people in China. It is risky for any large studio to be unable to show its productions in China but, as has become a normal point of discussion for me, Apple’s exposure is even greater because of its manufacturing requirements.
In February 2020, I wrote about this question:
[…] But there is unique risk in attaching a provocative entertainment arm to the body of a consumer goods company — one of those, of course, is the Apple’s relationship with China. Hollywood studios are choosing to censor films to have a shot at the lucrative Chinese market. But they, unlike Apple, don’t rely on factories in the country to produce the bulk of their revenue. It is not unreasonable to speculate that this is at least one of the reasons Apple is being particularly cautious about the portrayal of China in its original programming.
Apple is a big, sprawling conglomerate. If it cannot handle Stewart’s inquiries about China or our machine learning future, I think it should ask itself why that is, and whether those criticisms have merit.
Update: It would make sense to me that Stewart’s show could have been cancelled at least in part because of its popularity or lack thereof. But because streaming services do not disclose viewership numbers, we are left with only proxy measurements. On YouTube, for example, “The Problem” has 1.27 million subscribers while “Last Week Tonight” — comparable in both format and the host’s names — has over nine million. The most popular “Tonight” video has 41 million views, while the most popular “Problem” video has just four million. On TikTok, the ratio is reversed: John Oliver’s show has just 132,000 followers and less than a million total “likes”, while Stewart’s show has 897,000 followers and nearly seven million “likes”.
Those metrics are flawed for lots of reasons, but the main question I am left with is staring us right in the face: was Stewart’s show not popular enough for Apple? Surely it is not the least watched show Apple made — for what it is worth, nobody I know has personally recommend I watch even high-profile programming like “The Morning Show” or “For All Mankind”.
Since Google’s introduction of its Pixel 8 phones earlier this month, it has been interesting and a little amusing to me to read the reactions to its image manipulation tools. It feels like we have been asking the same questions every year — questions like what is a photograph, anyway?, and has technology gone too far? — since Google went all-in on computational photography with its original Pixels in 2016. In fact, these are things which people have been asking about photography since its early development. Arguments about Google’s complicity in fakery seem to be missing some historical context. Which means, unfortunately, a thousand-word summary.
As it happens, I took a photo history course when I was in university many years ago. I distinctly remember the instructor showing us an 1851 image shot by Edouard Baldus, and revealing to us that it was not a single photo, but instead a series of exposures cut and merged into a single image in a darkroom. That blew my mind at the time because, until then, I had thought of photo manipulation as a relatively recent thing. I had heard about Joseph Stalin’s propaganda efforts to remove officials who displeased him. But, surely, any manipulation that required precisely cutting negatives or painting over people was quite rare until Photoshop came along, right?
No. Not even close. The legacy of photography is a legacy of lies and liars.
In the introductory essay for the 2012 exhibition “Faking It: Manipulated Photography Before Photoshop” — sponsored by Adobe — Mia Fineman writes of the difference between darkroom techniques to adjust regions of a photo for exposure or cropping for composition, and photos where “the final image is not identical to what the camera ‘saw’ in the instant at which the negative was exposed”.1 The catalogue features nearly two hundred years of images which fit this description: from subtle enhancements, like compositing clouds into an overexposed sky, to artistic or humorous choices — “Man on a Rooftop with Eleven Men in Formation on His Shoulders” is an oft-cited delight — to dastardly projections of political power. Perhaps the most insidious examples are those which seem like journalistic “straight” images; one version of an image of the Animas Canyon by William Henry Jackson includes several fictional elements not present in the original.
Even at the time of manipulation-by-negative, there were questions about the legitimacy and ethics of these kinds of changes. In his 1869 essay “Pictorial Effect in Photography”, Henry Peach Robinson writes “[p]hotographs of what it is evident to our senses cannot visibly exist should never be attempted”, concluding that “truth in art may exist without an absolute observance of facts”. Strangely, Robinson defends photographic manipulation that would enhance the image, but disagrees with adding elements — like a “group of cherubs” — which would be purely fantastical.
This exhibition really was sponsored by Adobe — that was not a joke — and the company’s then-senior director of digital imaging Maria Yap explained why in a statement (sic):2
[…] For more than twenty years — since its first release, in 1990 — Adobe® Photoshop® software has been accused of undermining photographic truthfulness. The implicit assumption has been that photographs shot before 1990 captured the unvarnished truth and that manipulations made possible by Photoshop compromised that truth.
Now, “Faking It” punctures this assumption, presenting two hundred works that demonstrate the many ways photographs have been manipulated since the early days of the medium to serve artistry, novelty, politics, news, advertising, fashion, and other photographic purposes. […]
It was a smart public relations decision for Adobe to remind everyone that it is not responsible for manipulated images no matter how you phrase it. In fact, a few years after this exhibition debuted at New York’s Metropolitan Museum of Art, Adobe acknowledged the twenty-fifth anniversary of Photoshop with a microsite that included a “Real or Photoshop” quiz. Several years later, there are games to test your ability to identify which person is real.
The year after Adobe’s anniversary celebration, Google introduced its first Pixel phone. Each generation has leaned harder into its computational photography capabilities, with notable highlights like astrophotography in the Pixel 4, Face Unblur and the first iteration of Magic Eraser in the Pixel 6, and Super Res Zoom in the Pixel 7 Pro. With each iteration, these technologies have moved farther away from reproducing a real scene as accurately as possible, and toward synthesizing a scene based on real-life elements.
The Pixel 8 continues this pattern with three features causing some consternation: an updated version of Magic Eraser, which now uses machine learning to generate patches for distracting photo elements; Best Take, which captures multiple stills of group photos and lets you choose the best face for each person; and Magic Editor, which uses more generative software to allow you to move around individual components of a photo. Google showed off the latter feature by showing how a trampoline could be removed to make it look like someone really did make that sick slam dunk. Jay Peters, of the Verge, is worried:
There’s nothing inherently wrong with manipulating your own photos. People have done it for a very long time. But Google’s tools put powerful photo manipulation features — the kinds of edits that were previously only available with some Photoshop knowledge and hours of work — into everyone’s hands and encourage them to be used on a wide scale, without any particular guardrails or consideration for what that might mean. Suddenly, almost any photo you take can be instantly turned into a fake.
Peters is right in general, but I think his specific pessimism is misguided. Tools like these are not exclusive to Google’s products, and they are not even that new. Adobe recently added Generative Fill to Photoshop, for example, which does the same kind of stuff as the Magic Eraser and Magic Editor. It augments the Content Aware Fill option which has been part of Photoshop since 2010. The main difference is that Content Aware Fill works the way the old Magic Eraser used to: by sampling part of the real image to create a patch, though Adobe has marketed it as an “artificial intelligence” feature before the current wave of “A.I.” hype began.
For what it is worth, I tried that with one of the examples from Google’s Pixel 8 video. You know that scene where the Magic Editor is used to remove the trampoline from a slam dunk?
I roughly selected the area around the trampoline, and used the Content Aware Fill to patch that area. It took two passes but was entirely automatic:
Is it perfect? No, but it is fine. This is with technology that debuted thirteen years ago. I accomplished this in about ten seconds and not, as Peters claims, “hours”. It barely took meaningful knowledge of the software.
The worries about Content Aware Fill are familiar, too. At the time it came out, Dr. Bob Carey, then president of the U.S.-based National Press Photographers Association, was quoted in a Photoshelter blog post saying that “if an image has been altered using technology, the photo consumer needs to know”. Without an adequate disclaimer of manipulation, “images will cease to be an actual documentation of history and will instead become an altered history”.3 According to Peters, Google says the use of its “Magic” generative features will add metadata to the image file, though it says “Best Take” images will not. Metadata can be manipulated with software like ExifTool. Even data wrappers explicitly intended to avoid any manipulation, like digital rights management, can be altered or removed. We are right back where we started: photographs are purportedly light captured in time, but this assumption has always been undermined by changes which may not be obvious or disclosed.
Here is where I come clean: while it may seem like I did a lot of research for this piece, I cannot honestly say I did. This is based on writing about this topic for years, a lot of articles and journal papers I read, one class I took a long time ago, and an exhibition catalogue I borrowed from the library. I also tried my best to fact-check everything here. Even though I am not an expert, it made my head spin to see the same concerns dating back to the mid-nineteenth century. We are still asking the same things, like can I trust this photo?, and it is as though we have not learned the answer is that it depends.
I, too, have criticized computational photography. In particular, I questioned the ethics of Samsung’s trained image model, made famous by its Moon zoom feature. Even though I knew there has been a long history of inauthentic images, something does feel different about a world in which cameras are, almost by default, generating more perfect photos for us — images that are based on a real situation, but not accurately reflecting it.
The criticisms I have been seeing about the features of the Pixel 8, however, feel like we are only repeating the kinds of fears of nearly two hundred years. We have not been able to wholly trust photographs pretty much since they were invented. The only things which have changed in that time are the ease with which the manipulations can happen, and their availability. That has risen in tandem with a planet full of people carrying a camera everywhere. If you believe the estimates, we take more photos every two minutes than existed for the first hundred-and-fifty years after photography’s invention. In one sense, we are now fully immersed in an environment where we cannot be certain of the authenticity of anything.
Then again, Bigfoot and Loch Ness monster sightings are on a real decline.
We all live with a growing sense that everything around us is fraudulent. It is striking to me how these tools have been introduced as confidence in institutions has declined. It feels like a death spiral of trust — not only are we expected to separate facts from their potentially misleading context, we increasingly feel doubtful that any experts are able to help us, yet we keep inventing new ways to distort reality.
Even this article cannot escape that spectre, as you cannot be certain I did not generate it with a large language model. I did not; I am not nearly enough of a dope to use that punchline. I hope you can believe that. I hope you can trust me, because that is the same conclusion drawn by Fineman in “Faking It”:4
Just as we rely on journalists (rather than on their keyboards) to transcribe quotes accurately, we must rely on photographers and publishers (rather than on cameras themselves) to guarantee the fidelity of photographic images when they are presented as facts.
The questions that are being asked of the Pixel 8’s image manipulation capabilities are good and necessary because there are real ethical implications. But I think they need to be more fully contextualized. There is a long trail of exactly the same concerns and, to avoid repeating ourselves yet again, we should be asking these questions with that history in mind. This era feels different. I think we should be asking more precisely why that is.
I am writing this in the wake of another Google-related story that dominated the tech news cycle this week, after Megan Gray claimed, in an article for Wired, that the company had revealed it replaces organic search results with ones which are more lucrative. Though it faced immediate skepticism and Gray presented no proof, the claim was widely re-published; it feels true. Despite days of questioning, the article stayed online without updates or changes — until, it seems, the Atlantic’s Charlie Warzel asked about it. The article has now been replaced with a note acknowledging it “does not meet our [Wired’s] editorial standards”.
Gray also said nothing publicly in response to questions about the article’s claims between when it was published on Monday morning to its retraction. In an interview with Warzel published after the article was pulled, Gray said “I stand by my larger point — the Google Search team and Google ad team worked together to secretly boost commercial queries” — but this, too, is not supported by available documentation and it is something Google also denies. This was ultimately a mistake. Gray, it seems, interpreted a slide shown briefly during the trial in the way her biases favoured. Wired chose to publish the article in its “Ideas” opinion section despite the paucity of evidence. I do not think there was an intent to deceive, though I find the response of both responsible parties lacking — to say the least.
Intention matters. If a friend showed you a photo of them apparently making an amazing slam dunk, you would mentally check it against what you know about their basketball skills. If it does not make sense, you might start asking whether the photo was edited, or carefully framed or cropped to remove something telling, or a clever composite. This was true before you knew about that Pixel 8 feature. What is different now is that it is a little bit easier for that friend to lie to you. But that breach of trust is because of the lie, not because of the mechanism.
The questions we ask about generative technologies should acknowledge that we already have plenty of ways to lie, and that lots of the information we see is suspect. That does not mean we should not believe anything, but it does mean we ought to be asking questions about what is changed when tools like these become more widespread and easier to use.
We put our trust in people to help us evaluate information. Even people who have no faith in institutions and experts have something they see as reputable, regardless of whether it actually is. Generative tools only add to the existing inundation of questionably-sourced media. Something feels different about them, but I am not entirely sure anything is actually different. We still need to skeptically — but not cynically — evaluate everything we see.
Update: Corrected my thrice-written misuse of “jump shot” to “slam dunk” because I am bad at sports. Also, I have replaced the use of “bench” with “trampoline” because that is what that object in the photo is.
Food in Visual Look Up ⇥ macstories.net
With Visual Look Up, you can identify and learn about popular landmarks, plants, pets, and more that appear in your photos and videos in the Photos app. Visual Look Up can also identify food in a photo and suggest related recipes.
Meal identification is new to iOS 17, and it is a feature I am not sure I understand. Let us assume for now that it is very accurate — it is not, but work with me here. The use cases for this seem fairly limited, since it only works on photos you have saved to your device.
Federico Viticci, in his review of iOS 17, suggests two ways someone might use this: finding more information about your own meal, or saving an image from the web of someone else’s. One more way is to identify a meal you took a picture of some time ago and may have forgotten what it was. But Visual Look Up produces recipes, not just dish identification, so that suggests to me that this is to be used to augment home cooking. Perhaps the best-case scenario for this feature is that you stumble across a photo of something you ate some time ago, get the urge to re-create it, and Siri presents you with a recipe. That is, of course, assuming it works well enough to identify the meal in the photo.
Viticci:
Except that, well, 🤌 I’m Italian 🤌. We have a rich tapestry of regional dishes, variations, and local cuisine that is hard to categorize for humans, let alone artificial intelligence. So as you can imagine, I was curious to try Visual Look Up’s support for recipes with my own pictures of food. The best way I can describe the results is that Photos works well enough for generic pictures of a meal that may resemble something the average American has seen on Epicurious, but the app has absolutely no idea what it is dealing with when it comes to anything I ate at a restaurant in Italy.
Siri struggles with my home cooking, too, often getting the general idea of the dish but missing the specifics. A photo of a sweet corn risotto yielded suggestions for different kinds of risotto and various corn dishes, but not corn risotto. Some beets were identified as different kinds of fruit skewers or some different Christmas dishes; the photo was taken in August.
In many places, getting the gist of a dish is simply not good enough. The details matter. Food is intensely binding — not just among a country, but at smaller regional levels, too. It is something many people take immense pride in. While it is not my place to say whether it is insulting that Siri identified many distinct curry preparations as interchangeable curries of any type, it does not feel helpful when I know the foods identified are nothing like what was actually in the photo.
Update: Kristoffer Yi Fredriksson emailed to point out how Apple could eventually use food identification in its health efforts; for example, for meal tracking. I could see that. If it comes to pass, the accuracy of this feature will be far more important.
Microsoft Demos Copilot, Its ‘Artificial Intelligence’ Virtual Assistant ⇥ microsoft.com
Jared Spataro, of Microsoft:
Today at an event in New York, we announced our vision for Microsoft Copilot — a digital companion for your whole life — that will create a single Copilot user experience across Bing, Edge, Microsoft 365, and Windows. As a first step toward realizing this vision, we’re unveiling a new visual identity — the Copilot icon — and creating a consistent user experience that will start to roll out across all our Copilots, first in Windows on September 26 and then in Microsoft 365 Copilot when it is generally available for enterprise customers on November 1.
This is a typically ambitious effort from Microsoft. Copilot replaces Cortana, which will mostly be dropped later this year, and is being pitched as a next-generation virtual assistant in a similar do everything vein. This much I understand; tying virtual assistants to voice controls does not make much sense because sometimes — and, for me, a lot of the time — you do not want to be chatting with your computer. That is certainly a nice option and a boon for accessibility, but clear and articulate speech should not be required to use these kinds of features.
Microsoft’s implementation, however, is worrisome as I use a Windows PC at my day job. Carmen Zlateff, Microsoft Windows vice president, demoed a feature today in which she said “as soon as I copy the text, Copilot appears” in a large sidebar that spans the entire screen height. I copy a lot of stuff in a day, and I cannot tell you how much I do not want a visually intrusive — if not necessarily interruptive — feature like this. I hope I will be able to turn this off.
Meanwhile, a bunch of this stuff is getting jammed into Edge and Microsoft 365 productivity apps. Edge is getting so bloated it seems like the company will need to make a new browser again very soon. The Office features might help me get through a bunch of emails very quickly, but the kinds of productivity enhancements Microsoft suggests for me have not yet materialized into something I actually find useful. Its Viva Insights tool, introduced in 2021, is supposed to analyze your individual working patterns and provide recommendations, but I cannot see why I should pay attention to a graphic that looks like the Solar System illustrating which of my colleagues I spoke with least last week. Designing dashboards like these are a fun project and they make great demos. I am less convinced of their utility.
I get the same kind of vibe from Copilot. I hope it will be effective at summarizing all my pre-reads for a meeting, but I have my doubts. So much of what Microsoft showed today requires a great deal of trust from users: trust in its ability to find connections; in its accuracy; in its ability to balance helpfulness and intrusion; in its neutrality to its corporate overlords. One demo showed someone searching for cleats using Microsoft’s shopping search engine and getting a deal with the browser-based coupon finder. It is a little thing, but can I trust Copilot and Microsoft Shopping are showing me the best quality results that are most relevant, or should I just assume this is a lightly personalized way to see which companies have the highest ad spend with Microsoft?
It seems risky to so confidently launch something like this at a time when trust in big technology companies is at rock-bottom levels in the United States, especially among young people. Microsoft is certainly showing it is at the leading edge of this stuff, and you should expect more from its competitors very soon. I am just not sure giving more autonomy to systems like these from powerful corporations is what people most want.
Hundreds of Machine Generated ‘News’ Sites Clash With the Online News Act ⇥ thestar.com
Kevin Jiang, the Toronto Star:
Just months after the advent of ChatGPT late last year, hundreds of websites have already been identified as using generative artificial intelligence to spew thousands of AI-written, often misinformation-laden “news” stories online.
As the world nears a “precipice” of AI-driven misinformation, experts tell the Star that the tech industry pushback to Canada’s Online News Act — namely Google and Meta blocking trusted Canadian news sources for Canadians — may only make the issue worse.
This is not just a future concern: people affected by wildfires in British Columbia and the Northwest Territories have been unable to share news stories with each other on Meta’s platforms. That is obviously a horrible side effect, though better than what happened last time Meta issued national restrictions.
Also, I have no time for people who treat the exchange of news and information on Facebook or Instagram — or other social media platforms — as a mistake or some kind of dumbing-down of society. It is anything but. People moved their community connections online long ago, and their hosting is migrated to wherever those people congregate. And, for a long time now, that has been Facebook.
But, while it is Meta that is affecting the distribution of news on its platform, it is for reasons that can best be described as a response to a poorly designed piece of legislation — even though that law is not yet in effect. If Meta is told that it must soon pay for each news link shared publicly on its platforms, it is obviously going to try its best to avoid that extra variable expense. The only way it can effectively do that is to prohibit these links. It is terrible that Meta is standing firm but this feels like a fairly predictable consequence of a law based on links, and it seems like the federal government was ill prepared as it is now requesting Meta to stand down and permit news links again.
The irony of the fallout from this law is that any supposed news links in a Canadian’s Facebook or Instagram feed will be, by definition, not real news. The advertising businesses of Google and Meta surely played a role in encouraging more publishers to move behind paywalls, but they were not solely responsible. News has always been expensive to produce and that puts it at odds with a decades-long business obsession of maximizing profit and minimizing resources and expenses no matter how much it strains quality. Research and facts and original reporting will increasingly be treated like luxuries — in the same was as well made long-lasting products — if we do not change those priorities.
Claims About ChatGPT’s Supposed Liberal Bias Are Wildly Overstated ⇥ aisnakeoil.com
Gerrit De Vynck, Washington Post:
A paper from U.K.-based researchers suggests that OpenAI’s ChatGPT has a liberal bias, highlighting how artificial intelligence companies are struggling to control the behavior of the bots even as they push them out to millions of users worldwide.
The study, from researchers at the University of East Anglia, asked ChatGPT to answer a survey on political beliefs as it believed supporters of liberal parties in the United States, United Kingdom and Brazil might answer them. They then asked ChatGPT to answer the same questions without any prompting, and compared the two sets of responses.
The survey in question is the Political Compass.
Arvind Narayanan on Mastodon:
The “ChatGPT has a liberal bias” paper has at least 4 *independently* fatal flaws:
– Tested an older model, not ChatGPT.
– Used a trick prompt to bypass the fact that it actually refuses to opine on political q’s.
– Order effect: flipping q’s in the prompt changes bias from Democratic to Republican.
– The prompt is very long and seems to make the model simply forget what it’s supposed to do.
Colin Fraser appears to be responsible for finding that the order of how the terms appear affects the political alignment displayed by ChatGPT.
Narayanan and Sayash Kapoor tried to replicate the paper’s findings:
Here’s what we found. GPT-4 refused to opine in 84% of cases (52/62), and only directly responded in 8% of cases (5/62). (In the remaining cases, it stated that it doesn’t have personal opinions, but provided a viewpoint anyway). GPT-3.5 refused in 53% of cases (33/62), and directly responded in 39% of cases (24/62).
It is striking to me how the claims of this paper were widely repeated with apparent confirmation that tech companies are responsible for pushing the liberal beliefs that are ostensibly a reflection of mainstream news outlets.
Microsoft’s Machine Generated Article Recommending the Ottawa Food Bank to Tourists ⇥ twitter.com
Microsoft is really hitting it out of the park with its AI-generated travel stories! If you visit Ottawa, it highly recommends the Ottawa Food Bank and provides a great tip for tourists: “Consider going into it on an empty stomach.”
Jay Peters, the Verge:
If you try to view the story at the link we originally included in this article, you’ll see a message that says “this page no longer exists.” However, Microsoft’s article is still accessible from another link.
Microsoft laid off journalists at Microsoft News and MSN in 2020 to replace them with artificial intelligence. Microsoft didn’t immediately respond to a request for comment.
The article was pulled for users using British English, but remains accessible in American English and, perhaps more relevant, Canadian English. How hard can it be to remove all versions of this obviously dumb article?
Way of the future.
Update: Microsoft seems to have pulled the article entirely. I cannot find a language code which works.
Threads’ user base seems to be an object of fascination among the tech press. Mark Zuckerberg says it is “on the trajectory I expect to build a vibrant long term app” with “10s of millions” of users returning daily. Meanwhile, third-party estimators have spent the weeks since Threads’ debut breaking the news that its returning user base is smaller than its total base, and that people are somewhat less interested in it than when it launched, neither of which is surprising or catastrophic.1 Meanwhile, Elon Musk says Twitter is more popular than ever but, then again, he does say a lot of things that are not true.
All that merits discussion, I suppose, but I am more interested in the purpose of Threads. It is obviously a copy of Twitter at its core, but so what? Twitter is the progenitor of a genre of product, derived from instant messenger status messages and built into something entirely different. Everything is a copy, a derivative, a remix. It was not so long ago that many people were equating a person’s ban from mainstream social media platforms with suppression and censorship. That is plenty ridiculous on its face, but it does mean we should support more platforms because it does not make sense for there to be just one Twitter-like service or one YouTube-like video host.
So why is Threads, anyway? How does Meta’s duplication of Twitter — and, indeed, its frequent replication of other features and apps — fit into the company’s overall strategy? What is its strategy? Meta introduced Threads by saying it is “a new, separate space for real-time updates and public conversations”, which “take[s] what Instagram does best and expand[s] that to text”. Meta’s mission is to “[give] people the power to build community and bring the world closer together”. It is a “privacy-focused” set of social media platforms. It is “making significant investments” in its definition of a metaverse which “will unlock monetization opportunities for businesses, developers, and creators”. It is doing a bunch of stuff with generative artificial intelligence.
But what it sells are advertisements. It currently makes a range of products which serve both as venues for those ads, and as activity collection streams for targeting information. This leaves it susceptible to risks on many fronts, including privacy and platform changes, which at least partly explains why it is slowly moving toward its own immersive computing platform.
Ad-supported does not equate to bad. Print and broadcast media have been ad-supported for decades and they are similarly incentivized to increase and retain their audience. But, in their case, they are producing or at least deliberately selecting media of a particular type — stories in a newspaper, songs on a radio station, shows on TV — and in a particular style. Meta’s products resemble that sort of arrangement, but do not strictly mimic it. Its current business model rewards maximizing user engagement and data collection. But, given the digital space, there is little prescription for format. Instagram’s image posts can be text-based; users can write an essay on Facebook; a Threads post can contain nothing more than a set of images.
So Meta has a bunch of things going for it:
a business model that incentivizes creating usage and behavioural data at scale,
a budget to experiment, and
an existing massive user base to drive adoption.
All this explains why Meta is so happy to keep duplicating stuff popularized elsewhere. It cloned Snapchat’s Stories format in Instagram to great success, so it tried cloning Snapchat in its entirety more than once, both of which flopped. After Vine popularized short videos, Facebook launched Riff. After Twitter dumbly let Vine wither and die, and its place was taken by Musical.ly and then TikTok, Facebook launched Lasso, which failed, then Reels and copied its recommendation-heavy feed, moves which — with some help — have been successful. Before BeReal began to tank, it was copied by, uh, TikTok, but Meta was working on its own version, too.
But does any of this suggest to you an ultimate end goal or reason for being? To me, this just looks like Meta is throwing stuff at people in the hope any of it sticks enough for them to open the advertising spigot. In the same way a Zara store is just full of stuff, much of it ripping off the work of others, Meta’s product line does not point to a goal any more specific than its mission statement of “bring[ing] the world closer”. That is meaningless! The same corporate goal could be used by a food importer or a construction firm.
None of this is to say Meta is valueless as a company; clearly it is not. But it makes decisions that look scatterbrained as it fends off possible competitors while trying to build its immersive computing vision. But that might be far enough away that it is sapping any here-and-now vision the company might have. Even if the ideas are copies — and, again, I do not see that as an inherent weakness — I can only think of one truly unique, Meta-specific, and successful take: Threads itself. It feels like a text-only Instagram app, not a mere Twitter clone, and it is more Meta-like for it. That probably explains why I use it infrequently, and why it seems to have been greeted with so much attention. Even so, I do not really understand where it fits into the puzzle of the Meta business as a whole. Is it always going to be a standalone app? Is it a large language model instruction farm? Is it just something the company is playing around with and seeing where it goes, along the lines of its other experimental products? That seems at odds with its self-described “year of efficiency”.
I wish I saw in Meta a more deliberate set of products. Not because I am a shareholder — I am not — but because I think it would be a more interesting business to follow. I wish I had a clearer sense of what makes a Meta product or service.
-
Then there is the matter of how Sensor Tower and SimilarWeb measure app usage given how restricted their visibility is on Android and, especially, iOS. Sensor Tower runs an ad blocking VPN which it uses in a way not dissimilar from how Meta used Onavo, and several screen time monitoring products, which is something that was not disclosed in an analysis the company did with the New York Times.
SimilarWeb has a fancy graphic illustrating its data acquisition and delivery process, which it breaks down into collection, synthesis, modelling, and digital intelligence. Is it accurate? Since neither Apple nor Google reports the kind of data SimilarWeb purports to know about apps, it is very difficult to know. But, as its name suggests, its primary business is in web-based tracking, so it is at least possible to compare its data against others’. It says the five most popular questions asked to Google so far this year are “what”, “what to watch”, “how to delete instagram account”, “how to tie a tie”, and “how to screenshot on windows”. PageTraffic says the five most-Googled questions are “what to watch”, “where’s my refund”, “how you like that”, “what is my IP address”, and “how many ounces in a cup”, and Semrush says the top five are “where is my refund”, “how many ounces in a cup”, “how to calculate bmi”, “is rihanna pregnant”, and “how late is the closest grocery store open”. All three use different data sources but are comparable data sets — that is, all from Google, all worldwide, and all from 2023. They also estimate wildly differing search volumes: SimilarWeb’s estimate of the world’s most popular question query, “what”, is searched about 2,015,720 times per month, while Semrush says “where is my refund” is searched 15,500,000 times per month. That is not even close.
But who knows? Maybe the estimates from these marketing companies really can be extrapolated to determine real-world app usage. Colour me skeptical, though: if there is such wide disagreement in search analysis — a field which uses relatively open and widely accessible data — then what chance do they have of accurately assessing closed software platforms? ↥︎
Zoom’s Machine Learning Backlash ⇥ stackdiary.com
Alex Ivanovs, Stackdiary:
Zoom’s updated policy states that all rights to Service Generated Data are retained solely by Zoom. This extends to Zoom’s rights to modify, distribute, process, share, maintain, and store such data “for any purpose, to the extent and in the manner permitted under applicable law.”
What raises alarm is the explicit mention of the company’s right to use this data for machine learning and artificial intelligence, including training and tuning of algorithms and models. This effectively allows Zoom to train its AI on customer content without providing an opt-out option, a decision that is likely to spark significant debate about user privacy and consent.
Smita Hashim of Zoom (emphasis theirs):
We changed our terms of service in March 2023 to be more transparent about how we use and who owns the various forms of content across our platform.
[…]
To reiterate: we do not use audio, video, or chat content for training our models without customer consent.
Zoom is trialling a summary feature which uses machine learning techniques, and it appears administrators are able to opt out of data sharing while still having access to the feature. But why is all of this contained in a monolithic terms-of-service document? Few people read these things in full and even fewer understand them. It may appear simpler, but features which require this kind of compromise should have specific and separate documentation for meaningful explicit consent.
Inside the A.I. Factory ⇥ nymag.com
Josh Dzieza, writing for New York in collaboration with the Verge, on the hidden human role in artificial intelligence:
Over the past six months, I spoke with more than two dozen annotators from around the world, and while many of them were training cutting-edge chatbots, just as many were doing the mundane manual labor required to keep AI running. There are people classifying the emotional content of TikTok videos, new variants of email spam, and the precise sexual provocativeness of online ads. Others are looking at credit-card transactions and figuring out what sort of purchase they relate to or checking e-commerce recommendations and deciding whether that shirt is really something you might like after buying that other shirt. Humans are correcting customer-service chatbots, listening to Alexa requests, and categorizing the emotions of people on video calls. They are labeling food so that smart refrigerators don’t get confused by new packaging, checking automated security cameras before sounding alarms, and identifying corn for baffled autonomous tractors.
The magical feeling of so many of our modern products and services is too often explained by throwing money at low-paid labourers. Same day delivery? Online purchases of anything? Expedited free returns? Moderation of comments and images? As much as it looks from our perspective like the work of advancements in computing power, none of it would be possible without tens of thousands of people doing their best to earn a living spending unpredictable hours doing menial tasks.
The extent to which that bothers you is a personal affair; I am not one to judge. At the very least, I think it is something we should all remember the next time we hear about a significant advancement in this space. There are plenty of engineers who worked hard and deserve credit, but there are also thousands of people labelling elbows in photos and judging the emotion of internet comments.
For the first time in more than a decade, it truly feels like we are experiencing massive changes in how we use computers now, and how that will change in the future. The ferocious burgeoning industry of artificial intelligence, machine learning, LLMs, image generators, and other nascent inventions has been a part of our lives first gradually, then suddenly. The growth of this new industry provides an opportunity to reflect on how it ought to be grown while avoiding problems similar to those which have come before.
A frustrating quality of industries and their representatives is a general desire to avoid scrutiny of their inventions and practices. High technology is no different. They begin by claiming things are too new or that worries are unproven and, therefore, there is no need for external policies governing their work. They argue industry-created best practices are sufficient in curtailing bad behaviour. After a period of explosive growth, as regulators are eager to corral growing concerns, those same industry voices protest that regulations will kill jobs and destroy businesses. It is a very clever series of arguments which can luckily be repurposed for any issue.
Eighteen years ago, EPIC reported on the failure of trusting data brokers and online advertising platforms to self-regulate. It compared them unfavourably to the telemarketing industry, which pretended to self-police for years before the Do Not Call list was introduced. At the time, it was a rousing success; unfortunately, regulators were underfunded and failed to keep pace with technological change. Due to overwhelming public frustration with the state of robocalls, the U.S. government began rolling out call verification standards in 2019, and Canadian regulators followed suit. For U.S. numbers, these verification standards will be getting even more stringent just nine days from now.
These are imperfect rules and they are producing mixed results, but they are at least an attempt at addressing a common problem with some success. Meanwhile, a regulatory structure for personal privacy remains elusive. That industry still believes self-regulation is effective despite all evidence to the contrary, as my regular readers are fully aware.
Artificial intelligence and machine learning services are growing in popularity across a wide variety of industries, which makes it a perfect opportunity to create a regulatory structure and a set of ideals for safer development. The European Union has already proposed a set of restrictions based on risk. Some capabilities — like when automated systems are involved in education, law enforcement, or hiring contexts — would be considered “high risk” and subject to ongoing assessment. Other services would face transparency requirements. I do not know if these rules are good but, on their face, the behavioural ideals which the E.U. appears to be constructing are fair. The companies building these tools should be expected to disclose how models were trained and, if they do not do so, there should be consequences. That is not unreasonable.
This is about establishing a set of principles to which new developments in this space must adhere. I am not sure what those look like, but I do not think the correct answer is in letting businesses figure it out before regulators struggle to catch up years later with lobbyist-influenced half-measures. Things can be different this time around if there is a demand and an expectation for doing so. Written and enforced correctly, these regulations can help temper the worst tendencies of this industry while allowing it to flourish.
Inhuman Writing for Google Is Eating the Web ⇥ theverge.com
Mia Sato, the Verge:
[Jennifer] Dziura still updates her personal blog — these are words for people.
The shop blog, meanwhile, is the opposite. Packed with SEO keywords and phrases and generated using artificial intelligence tools, the Get Bullish store blog posts act as a funnel for consumers coming from Google Search, looking for things like Mother’s Day gifts, items with swear words, or gnome decor. On one hand, shoppers can peruse a list of products for sale — traffic picks up especially around holidays — but the words on the page, Dziura says, are not being read by people. These blogs are for Google Search.
[…]
This is the type of content publishers, brands, and mom-and-pop businesses spend an untold number of hours on, and on which a booming SEO economy full of hackers and hucksters make promises ranging from confirmed to apocryphal. The industries that rely heavily on Search — online shops, digital publishers, restaurants, doctors and dentists, plumbers and electricians — are in a holding pattern, churning out more and more text and tags and keywords just to be seen.
The sharp divergence between writing for real people and creating material for Google’s use has become so obvious over the past few years that it has managed to worsen both Google’s own results and the web at large. The small business owners profiled by Sato are in an exhausting fight with automated chum machines generating supposedly “authoritative” articles. When a measure becomes a target — well, you know.
There are loads of examples and I recently found a particularly galling one after a family friend died. Their obituary was not published, but several articles began appearing across the web suggesting a cause of death, which was not yet public information. These websites seem to automatically crawl publicly available information and try to merge it with some machine learning magic and, in some instances, appear to invent a cause of death. There is also a cottage industry of bizarre YouTube channels with videos that have nothing to do with the obituary-aligned titles. I have no idea why those videos exist; none that I clicked on were ad-supported. But the websites have an easy answer: ghoulish people have found that friends and family of a person who recently died are looking for obituaries, and have figured out how to scam their ad-heavy pages to high-ranking positions.
At the same time, I have also noticed a growing number of businesses — particularly restaurants — with little to no web presence. They probably have a listing in Apple Maps and Google Maps, an Instagram page, and a single-page website, but that could be their entire online presence. I know it is not possible for every type of business. It does seem more resilient against the slowly degrading condition of search engines and the web at large, though.