- nexa - server-nexa.polito.it

Five things privacy experts know about AI
by Daniela Tafani Jan. 15, 2025

Jan. 15, 2025

Five things privacy experts know about AI: 1. AI models memorize their training data 2. AI models then leak their training data 3. Ad hoc protections don't work 4. Robust protections exist, though their mileage may vary 5. The larger the model, the worse it gets Bonus thing: AI companies are overwhelmingly dishonest In November, I participated in a technologist roundtable about privacy and AI, for an audience of policy folks and regulators. The discussion was great! It also led me to realize that there a lot of things that privacy experts know and agree on about AI… but might not be common knowledge outside our bubble. That seems the kind of thing I should write a blog post about! 1. AI models memorize their training data When you train a model with some input data, the model will retain a high-fidelity copy of some data points. If you "open up" the model and analyze it in the right way, you can reconstruct some of its input data nearly exactly. This phenomenon is called memorization. Memorization happens by default, to all but the most basic AI models. It's often hard to quantify: you can't say in advance which data points will be memorized, or how many. Even after the fact, it can be hard to measure precisely. Memorization is also hard to avoid: most naive attempts at preventing it fail miserably — more on this later. Memorization can be lossy, especially with images, which aren't memorized pixel-to-pixel. But if your training data contains things like phone numbers, email addresses, recognizable faces… Some of it will inevitably be stored by your AI model. This has obvious consequences for privacy considerations. 2. AI models then leak their training data Once a model has memorized some training data, an adversary can typically extract it, even without direct access to the internals of the model. So the privacy risks of memorization are not theoretical: AI models don't just memorize data, they regurgitate it as well. In general, we don't know how to robustly prevent AI models from doing things they're not supposed to do. That includes giving away the data they dutifully memorized. There's a lot of research on this topic, called "adversarial machine learning"… and it's fair to say that the attackers are winning against the defenders by a comfortable margin. Will this change in the future? Maybe, but I'm not holding my breath. To really secure a thing against clever adversaries, we first have to understand how the thing works. We do not understand how AI models work. Nothing seems to indicate that we will figure it out in the near future. 3. Ad hoc protections don't work There are a bunch of naive things you can do to try and avoid problems 1 and 2. You can remove obvious identifiers in your training data. You can deduplicate the input data. You can use regularization during training. You can apply alignment techniques after the fact to try and teach your model to not do bad things. You can tweak your prompt and tell your chatbot to pretty please don't reidentify people like a creep1. You can add a filter to your language model to catch things that look bad before they reach users. You can list all those in a nice-looking document, give it a fancy title like "Best practices in AI privacy", and feel really good about yourself. But at best, these will limit the chances that something goes wrong during normal operation, and make it marginally more difficult for attackers. The model will still have memorized a bunch of data. It will still leak some of this data if someone finds a clever way to extract it. Fundamental problems don't get solved by adding layers of ad hoc mitigations. 4. Robust protections exist, though their mileage may vary To prevent AI models from memorizing their input, we know exactly one robust method: differential privacy (DP). But crucially, DP requires you to precisely define what you want to protect. For example, to protect individual people, you must know which piece of data comes from which person in your dataset. If you have a dataset with identifiers, that's easy. If you want to use a humongous pile of data crawled from the open Web, that's not just hard: that's fundamentally impossible. In practice, this means that for massive AI models, you can't really protect the massive pile of training data. This probably doesn't matter to you: chances are, you can't afford to train one from scratch anyway. But you may want to use sensitive data to fine-tune them, so they can perform better on some task. There, you may be able to use DP to mitigate the memorization risks on your sensitive data. This still requires you to be OK with the inherent risk of the off-the-shelf LLMs, whose privacy and compliance story boils down to "everyone else is doing it, so it's probably fine?". To avoid this last problem, and get robust protection, and probably get better results… Why not train a reasonably-sized model entirely on data that you fully understand instead? It will likely require additional work. But it will get you higher-quality models, with a much cleaner privacy and compliance story. Understanding your training data better will also lead to safer models, that you can debug and improve more easily. 5. The larger the model, the worse it gets Every privacy problem gets worse for larger models. They memorize more training data. They do so in ways that more difficult to predict and measure. Their attack surface is larger. Ad hoc protections get less effective. Larger, more complex models also make it harder to use robust privacy notions for the entire training data. The privacy-accuracy trade-offs are steeper, the performance costs are higher, and it typically gets more difficult to really understand the privacy properties of the original data. Bonus thing: AI companies are overwhelmingly dishonest I think most privacy experts would agree with this post so far. There are divergences of opinion when you start asking "do the benefits of AI outweigh the risks". If you ask me, the benefits are extremely over-hyped, while the harms (including, but not limited to, privacy risks) are very tangible and costly. But other privacy experts I respect are more bullish on the potentials of this technology, so I don't think there's a consensus there. AI companies, however, do not want to carefully weigh benefits against risks. They want to sell you more AI, so they have a strong incentive to downplay the risks, and no ethical qualms doing so. So all these facts about privacy and AI… they're pretty inconvenient. AI salespeople would like it a lot if everyone — especially regulators — stayed blissfully unaware of these. Conveniently for AI companies, things that are obvious truths to privacy experts are not widely understood. In fact, they can be pretty counter-intuitive! >From a distance, memorization is surprising. When you train an LLM, sentences are tokenized, words are transformed into numbers, then a whole bunch of math happens. It certainly doesn't look like you copy-pasted the input anywhere. LLMs do an impressive job at pretending to be human. It's super easy for us to antropomorphize them, and think that if we give them good enough instructions, they'll "understand", and behave well. It can seem strange that they're so vulnerable to adversarial inputs. The attacks that work on them would never work on real people! People really want to believe that every problem can be fixed with just a little more work, a few more patches. We're very resistant to the idea that some problem might be fundamental, and not have a solution at all. Companies building large AI models use this to their advantage, and do not hesitate making statements that they clearly know to be false. Here's OpenAI publishing statements like « memorization is a rare failure of the training process ». This isn't an unintentional blunder, they know how this stuff works! They're lying through their teeth, hoping that you won't notice. Like every other point outlined in this post, this isn't actually AI-specific. But that's a story for another day… Additional remarks and further reading On memorization: I recommend Katharine Jarmul's blog post series on the topic. It goes into much more detail about this phenomenon and its causes, and comes with a bunch of references. One thing I find pretty interesting is that memorization may be unavoidable: some theoretical results suggest that some learning tasks cannot be solved without memorizing some of the input! On privacy attacks on AI models: this paper is a famous example of how to extract training data from language models. It also gives figures on how much training data gets memorized. This paper is another great example of how bad these attacks can be. Both come with lots of great examples in the appendix. On the impossibility of robustly preventing attacks on AI models: I recommend two blog posts by Arvind Narayanan and Sayash Kapoor: one about what alignment can and cannot do, the other about safety not being a property of the model. The entire blog post series is worth a read. On robust mitigations against memorization: this survey paper provides a great overview of how to train AI models with DP. Depending on the use case, achieving a meaningful privacy notion can be very tricky: this paper discusses the specific complexities of natural language data, while this paper outlines the subtleties of using a combination of public and private data during AI training. Acknowledgments Thanks a ton to Alexander Knop, Amartya Sanyal, Gavin Brown, Joe Near, Liudas Panavas, Marika Swanberg, and Thomas Steinke for their excellent feedback on earlier versions of this post. <https://desfontain.es/blog/privacy-in-ai.html>

1 0

You don’t need a smartphone: a practical, personal guide to downgrading your device | Well actually | The Guardian
by Alberto Cammozzo Jan. 15, 2025

Jan. 15, 2025

<https://www.theguardian.com/lifeandstyle/2025/jan/14/switching-smartphone-f…> I have spent years of my life online. I have nearly crashed my car because I was composing a message. I have had panic attacks over losing my data connection. I have scrolled for hours, and then failed to recall anything I saw during those hours. I have chosen a screen over a conversation, a view, a book, a good night’s sleep, a moment of quiet reflection. In other words, I have owned a smartphone. For the first half of my 20s, I was an online art influencer. That was my job. I documented all aspects of my life for an audience of almost 200,000 on Instagram. Nothing was off limits: creative process, mood swings, love life, hospital stays. My life was a one-woman tabloid, generating traffic for my art business. I needed the attention almost as much as I needed the money. I had few friends. I had a huge following. The balance was not sustainable. I became isolated and addicted to social media, and the two states perpetuated each other. One day, I broke down and deactivated my account. I ordered an old Nokia and everything changed. The effects of ditching my smartphone have been innumerable and nothing short of life-changing. It took going cold turkey to see just how distracted I’d become, how anxious, how easily bored, how resistant I’d become to silence, presence, pain and other natural human states. The benefits of downgrading accumulated gradually, over many months of withdrawal and adjustment. Even years after my downgrade, I am still noticing new benefits, aspects of myself that are only now emerging from beneath the shadow of tech addiction. I am never bored now. Everything is interesting. I read long books. I go on long walks without headphones. When I have to wait around – when a bus is late or a friend heads to the bathroom – I just sit there. I check my emails and such when I’m at my computer. I’ve found that when the internet is out of reach, it rarely crosses my mind. The most easily quantifiable benefit of downgrading is time. You will get hours of each day back when you stop scrolling. You can use that time to be healthy or social or productive, but you can also use it to be idle. Idleness is a lost art, and is in my opinion essential to mental health and creativity. I spend a good part of the day just flipping through periodicals, drinking cups of tea, poking clumsily at the piano. I don’t believe that it is possible to waste time when you’re fully present. If you’re interested in switching to a dumbphone, you might have questions about how practical or even possible it is. How do we go about our lives without Google Maps, calling over wifi, and dual-factor authentication? These questions can make it hard to fathom downgrading. But just remember that people lived for millennia without smartphones. With time, patience and clever workarounds, we can all find our way back to that. Here are my answers to some of the most common questions I’ve received. How do you use two-factor authentication? For security purposes, certain online platforms might require you to confirm your logins on a smartphone app. There are a few ways to deal with this, depending on the platform you’re using: Purchase a physical security device to plug into your computer while authenticating. One popular option is the YubiKey. Ask if you can receive SMS verification, if you work or study somewhere with an IT department. This will mean you can get an authentication code sent to your dumbphone, and sometimes this can be manually enabled – but be aware that this is a less secure method. Some dumbphones have browser functions that support dual-factor authentication. Use a backup phone (this might seem at first to defeat the purpose of a downgrade – we’ll discuss it later). If you find yourself stuck on this particular problem, you are not alone! This is one of the trickiest, and most essential, smartphone features to replicate. You can also check out online forums (such as the dumbphones subreddit) for more ideas. How can you stay in touch with friends, family and co-workers? After you downgrade, your new phone will still allow you to send text messages and place calls. The only potential complication is that many of us use online messaging to stay in touch, especially internationally. Fortunately, almost all of these – iMessage, Telegram, Facebook, Instagram – are accessible from your desktop computer. And while laptops are internet-connected devices, they will never prey on your attention the way a smaller, hand-held device will. In my experience, the only people you really need to contact on the go are those you’re coordinating with. What time will you be there? or I’m running late – that’s urgent information. Less urgent are group chats, messages to friends and family abroad, and DMs to social media acquaintances. These things might feel urgent, but once you step away from them, you may realize that the sense of urgency was a product of software engineering, your own anxiety, or a combination. How are you supposed to navigate without a maps app? Here’s something you might not have realized: most dumbphones feature a maps app. In fact, unlike the bricks of the early 2000s, some modern dumbphones feature Bluetooth, MP3 players, even a tiny, awkward-to-use search engine. If those amenities feel necessary to you, then by all means find a dumbphone that provides them. (I used to have a flip phone with Google Maps, but I recently downgraded to an even dumber phone because I was using the maps function obsessively, checking my ETA every few blocks to see whether I’d beat the estimated walking time. It’s incredible what suffices for entertainment once you’ve broken your smartphone addiction.) But it’s also possible to live without digital navigation tools. I personally prefer to navigate using a combination of: looking up directions before leaving the house, consulting posted transit maps, asking strangers for help, and generally acquainting myself with the area so I can get around intuitively. Interestingly, I find that I rarely need to consult my hand-drawn maps: the process of writing down directions serves to embed them in my brain. Over time, my accumulated knowledge of local geography and transportation has enabled me, for the most part, to navigate map-free. What about listening to music and podcasts? Even before the smartphone, people listened to audio on the go. You can find old MP3 players for cheap online and certain dumbphones feature audio players. You will, however, have to figure out a way to download what you want to listen to. Libraries are a great resource for CDs and audiobooks. Podcasts are accessible for free download and offline listening via Apple. If you’re looking for more recent releases, I strongly suggest purchasing them on Bandcamp, putting the money you’ve saved on your phone plan (not to mention the phone itself) toward supporting musicians. All that said, your need for audio stimulation will certainly decrease once you downgrade. Soon, you’ll find that you can happily move through the world with only your thoughts for entertainment. What about taking photos? I carry a film camera with me everywhere I go. Receiving my film scans is the highlight of my month – the photos are higher quality and strike me as far more special than any of the 60,000 photos I took on my smartphone before downgrading. The key difference is not just the film. You may well prefer a digital camera, and I know many dumbphone users who do. Regardless, using a real camera is a vastly different experience for two reasons: the friction of retrieving a separate device will make photo-taking a more deliberate act, and the mono-purpose nature of the camera will mean no distractions or intrusions as you use it. Is there anything you can’t use without a smartphone? As insistent as I am about the viability of smartphone-free living, even I have to admit that there are certain things that are impossible without one – impossible, that is, when you’re on the go, away from your computer and internet connection. These include WhatsApp, Spotify, tickets for certain concerts and sports matches, electric car charging, and location sharing. It’s easy to get hung up on these losses, but I encourage you to think instead about all you are gaining in the process: time, presence, peace of mind. Aren’t those things worth any inconvenience? … do I need a backup phone? A backup phone is a smartphone that ideally remains off and put away outside of special circumstances. Many dumbphone users still keep a smartphone in a drawer, or even at a friend’s house, just in case it’s required for verification. There are certain services and applications that just assume everyone has access to a smartphone, and the complication involved in circumventing these obstacles may not feel worth it. So, if you feel like you need a backup phone, then there’s no shame in keeping one for those moments of need (if you have an electric car and need to charge it, for example). However, for other everyday tasks, you will find they feel less urgent the less accessible they are. Do you really need to check your email on the train? Do you really need to buy concert tickets at the doctor’s office? Or can these things wait until you’re at your computer? It is easy to tally up the inconveniences involved in a downgrade. I encourage you to think instead about what you will gain, which is less quantifiable. Isn’t your newfound presence, attention and free time worth occasionally missing an email or showing up late to an event? Yes, there will be downsides, and you might be tempted to view them as justification for switching back. In those moments, try to reconnect with your original motivation for downgrading. Who did you want to become? Isn’t it worth it? It’s been years since I downgraded, and I’ve so fully adapted to the dumbphone that I forget all about it until someone else brings it to my awareness. “Do you miss having a smartphone?” they ask, and I think back to the height of my addiction. How could I miss that time? I was hardly even there.

1 0

Dan McQuillan - Labour's AI Action Plan
by Daniela Tafani Jan. 14, 2025

Jan. 14, 2025

Labour's AI Action Plan - a gift to the far right Critical computing expert Dan McQuillan argues that, on top of the clear social and environmental harms associated with the technology, Labour's vapid fixation on AI-led growth in lieu of real change will further enable the far right. Instead, he proposes an alternative strategy of 'decomputing'. By Dan McQuillan - 14 Jan 2025 Labour has published its 'AI Opportunities Action Plan' (The Plan). The Prime Minister is very bullish about The Plan, and peppers his foreword with muscular terms like growth, revolution, ambition, strength and innovation. In itself, The Plan is full of claims that AI is essential and inevitable, and urges the government to pour public money into the industry so as not to miss out. In the style of tech entrepreneurs, The Plan likes to put 'x' after things, so investment must go up by 20x (meaning twenty times), the amount of compute AI requires has already gone up by 10,000x and so on. The Plan claims that Britain is already leading the world through the AI Safety Institute (of which more later) and infuses the usual AI hype with nationalist vibes via terms like world leader, world-class, national champions and 'Sovereign AI'. Above all, The Plan emphasises the need to scale. The significance of scale for AI and its technopolitical impacts will be explored below. This article addresses the poverty of The Plan and the emptiness of its claims about AI but, rather than a point-by-point rebuttal, it's about the underlying reasons why this Labour government supports measures that will harm both people and the environment. In between invocations of speed, pace and scale, there's some recognition in The Plan that the UK is not a wholly happy place right now. While recommending a high-tech form of land enclosures via 'AI Growth Zones' (AIGZs), which are about handing data centre developers "access to land and power", it gestures towards the idea that these could drive local innovation in post-industrial towns. While The Plan's claims about AI's inevitable progress and the oncoming wave of agentic systems that will reason, plan and act for themselves already seem dated and discredited, what hasn't changed is that the very regions targeted for growth via AIGZs have already seen violent anti-immigrant pogroms accompanied by fascist rhetoric, and those sentiments have not gone away. Ultimately, it's argued here, the misstep represented by The Plan and its total commitment to AI will reinforce and amplify the threat of the far right, as well as connecting it to the extremely reactionary ideas that are in the ascendency in Silicon Valley. This article proposes instead 'decomputing'; the refusal of hyperscale computing, the replacement of reductive solutionism with structures of care, and the construction of alternative infrastructures based on convivial technologies. Labour and AI In some ways, it's fairly obvious why this Labour government would want to prioritise AI. Kier Starmer's single identifiable political belief is the idea of 'growth', so demonstrating economic growth supersedes all other government concerns. Growth will demonstrate that Kier and colleagues are serious politicians who are no threat to the establishment, and at the same time win over Mr. & Mrs Red Wall voter who are disillusioned with orthodox politics. Boosting GDP is Labour's answer to all the wicked problems that beset the UK's public services and infrastructure, and avoids having to actually challenge the underlying logic of Thatcherite neoliberalism which has dominated for decades. And if there's one area of economic activity which is growing, it's certainly AI. All the graphs are going up; the venture capitalist investment, the stock market valuations, the size of the AI models and the number and scale of the data centres. Latching on to this growth is already working for the government, as a full 10% of the £63b promised by their 'record-breaking' international investment summit was earmarked for data centres. Having apparently decided to pin their hopes on AI, the Labour government have been aligning with demands from Big Tech. Not long after the election that brought Labour to power, Google issued a report titled 'Unlocking the UK's AI Potential' laying out their conditions for AI growth in the UK, including investment in data centres and a loosening of copyright restrictions. Of course, these aren't Google-specific requirements; the foundation of all contemporary AI is scale, meaning more and larger data centres to house all the servers and bigger pools of data to train them. The collision course with copyright comes from the fact that these data sets have always been too large to pay for, so the AI industry just grabs them from the internet without asking. Google's report was accompanied by a media round from their UK & Ireland vice president warning that the UK risked "losing leadership" and "falling behind" if their advice wasn't followed. It seems the message was received loud and clear; since the election, Labour have designated data centres as 'critical national infrastructure' which means that ministers can override any local planning objections, and the government is also floating a relaxation of copyright protections. It's not just Google that the Labour government is prepared to doff its cap to; Peter Kyle, the current secretary of state for Science, Innovation and Technology, has repeatedly stated that the UK should deal with Big Tech via 'statecraft'; in other words, rather than treating AI companies like any other business that needs taxing and regulating, the government should treat the relationships as a matter of diplomatic liaison, as if these entities were on a par with the UK state. This awe of Big Tech reflects deeper currents of commitment within the Labour government. Certainly, any ministers from the Blairite faction are going to be influenced by the absolute belief in AI expressed by influential think tank, the Tony Blair Institute (TBI). It's hard to exaggerate the centrality of AI to the TBI world view, but the title of their 2023 report is pretty representative: 'A New National Purpose: AI Promises a World-Leading Future of Britain'. According to this report, becoming a world leader in AI development is "a matter becoming so urgent and important that how we respond is more likely than anything else to determine Britain's future", and its 2024 sequel, 'Governing in the Age of AI: A New Model to Transform the State', opens with "There is little doubt that AI will change the course of human progress." The breathless rhetoric is accompanied by policy demands; a factor of ten increase in the UK's compute capacity, the diversion of major spending commitments to AI infrastructure, reducing regulation to US levels and, of course, enacting all this in close relationship with the private sector. A core promise is that turning the public sector over to AI will deliver huge savings and improved delivery, although one might question the reliability of their research, given that it was based on asking ChatGPT itself how many government jobs it could do. While this sketchy approach has echoes of the Iraq ('dodgy') Dossier, it's reflecting a realpolitik that sees both AI companies and rhetoric about AI as incredibly powerful at the current moment. This is perhaps the hole that AI fills for the Labour government; having long abandoned any substantive belief in the transformative power of socialism, it is lacking a mobilising belief system. At the same time, it's obvious to all and sundry that the status quo is in deep trouble and that being the party of continuity isn't going to convince anyone. Ergo, the claim that AI has the power to change the world becomes a good stand-in for a transformative ideology. The bonus for the Labour government is that relying on AI to fix things avoids the need for any structural changes that might upset powerful business and media interests, and rhetoric about global AI leadership has a suitably 'Empire' vibe to appeal to nationalistic sentiments at the grassroots. Harms at scale Beneath all the policy gloss and think tank reports, though, lurk the real harms of AI in the here-and-now, starting with environmental harms. The Labour government's vision for AI takes concrete form in the shape of more data centres. However, as some previously tranquil localities are starting to discover, this comes with significant impacts. Generative AI, in particular, is driving the computational scale of AI models through the roof. The rate at which these models are increasing in size outpaces any other recent tech revolution, from smartphone adoption to genome sequencing. In turn, this is driving massive increases in energy demand. To service AI and the internet cloud, the fastest growing type of data centre in the UK is the so-called hyperscale data centres run by the likes of Google, Microsoft and AWS. These are typically at least 10,000sq ft and contain upwards of 5,000 servers, but the industry wants them to be much larger, and filled with the energy-guzzling GPU chips that train and run AI. Sam Altman, CEO of ChatGPT's parent company OpenAI, has pitched plans for 5GW data centres in the US, which is the equivalent of about five nuclear reactors' worth and enough energy to power a large city. These voracious demands for electricity come with immediate consequences for national grids and for climate emissions. The Greater London Authority (GLA) has already had to impose a temporary ban on new housing developments in West London because a cluster of existing data centres was using the available grid supply. Because AI's energy demands are outpacing the development of bigger electricity grids, there's currently a push for bringing back fossil fuel sources, especially gas-powered turbines. Connecting directly to the natural gas network to overcome local power constraints is less efficient than grid-scale generation and increases unmonitored carbon emissions. Of course, Big Tech is already aware that driving climate emissions is a bad look and has previously tried to look 'green' via the use of 'renewables' and the cover story of carbon offsets. However, the scale of generative AI has blown this away to the point where both Google and Microsoft have admitted an inability to meet their own climate targets. Locally, it's not just potential power cuts that an AI data centre brings to an area, but a huge demand for cooling water to stop all the servers overheating and the pervasive presence of a background hum from all the cooling systems. The question is whether a pursuit of 'AI greatness' will make the UK more like Ireland, which has already been recolonised as a dumping ground for Big Tech's data centre infrastructure. The here-and-now harms of AI are also social. Never mind the sci-fi fantasies about AI taking over the world, the mundane reality of AI in any social context are forms of ugly solutionism that perpetuate harms rather than reducing them. The claim that more computation will improve public services is hardly new, and algorithmic fixes for everything from welfare to education have already left a trail of damage in their wake. In Australia, the 'Robodebt' algorithm wrongly accused tens of thousands of people of welfare fraud, and was only halted by a grassroots campaign and an eventual public inquiry, while in the Netherlands an algorithm falsely labelled tens of thousand of people as defrauding the child benefits system, causing crippling debts and family break-ups. What the UK's notorious Horizon IT system and contemporary AI have in common is the tendency to generate falsehoods while appearing to be working properly. What AI adds is the capacity to scale harms in way that makes the Horizon scandal look like small beer. The insistence that AI will reverse the rot in education and healthcare systems also has a tired history. Back in 2018, Facebook's non-profit arm inserted an online learning platform into a California public school system on the basis that it offered 'personalised learning', the central mantra of all AI-driven educational technology. It took mass resistance by 17-year old students to get rid of it. In the open letter they sent to Zuckerberg they said "Most importantly, the entire program eliminates much of the human interaction, teacher support, and discussion and debate with our peers that we need in order to improve our critical thinking. Unlike the claims made in your promotional materials, we students find that we are learning very little to nothing. It's severely damaged our education, and that's why we walked out in protest". Meanwhile the Nobel Prize-winning godfather of AI, Geoffrey Hinton, made the claim back in 2016 that thanks to the superior accuracy of AI's image classification there was no need to train any more radiologists. As it turned out, of course, that claim was just as specious as the more recent hype about ChatGPT passing medical exams. Labour's minister for Science, Innovation and Technology is continuing to use the trope of an AI-powered solution to cancer detection to push for more AI in public services while ignoring calls by leading cancer specialists to "concentrate on the basics of cancer treatment rather than the 'magic bullets' of novel technologies and artificial intelligence". Forcing AI into services in lieu of fixing underlying issues like decaying buildings and without funding more actual teachers and doctors is a form of structural violence - a form of violence by which institutions or social structures harm people through preventing them from meeting their fundamental needs. The political continuity here is that a commitment to AI solutions also enacts a kind of Thatcherite 'shock doctrine' where the sense of urgency generated by an allegedly world-transforming technology is used as an opportunity to transfer power to the private sector. The amount of data and computing power required to create on of today's foundation models, the big generative and supposedly general purpose systems, is beyond the reach of all but the biggest tech companies. Whether it's in welfare, education or health, a shift to AI is a form of privatisation by the backdoor, shifting a significant locus of control to Silicon Valley. Like the original Thatcherism, this is also going to be accompanied by job losses and a change to more precarious forms of algorithm-driven outsourcing. It's Deliveroo all round, not because AI can actually replace jobs but because its shoddy emulations provide managers and employers with a means to pare away employment protections. What marks the Labour government out from earlier forms of neoliberalism is its emphasis on total mobilisation, that all material and human resources should be aligned behind the national mission for growth. This translates into a rhetorical and legislative intolerance for the idea that people should be 'idle', no matter their states of mental distress or other disability. Unfortunately, while AI systems are themselves unproductive in a practical sense, they excel at exactly the functions of ranking, classification and exclusion that are required for forms of social filtering at scale. Turning more services over to AI-assisted decision-making will indeed facilitate the differentiation of the deserving from the undeserving in line with this productivist ideology. This alignment of social and technical ordering will show that Labour is indeed still the 'party of the workers', but only in the sense of squeezing the last drop of work out of people while using algorithmic optimisation to decide who is relatively disposable. Far right Ultimately, the Labour government's capitulation to AI in the vain hope of meaningful growth is a gift to the far right. Most governing parties across Europe are making the mistake of incubating the far right while complaining about their seemingly inexorable rise. Governments seem oblivious to the political fact that trying to distract people with the spectre of the far right while completely failing to address the structural failings of neoliberalism that leave people feeling angry and abandoned, only serves to empower polarising and post-truth politics. Moreover, the fact that populist rhetoric gains support leads the same governing parties to mainstream their reactionary narratives, as the Labour government has done around immigration and the so-called 'small boats crisis'. Using AI to distract from structural problems while failing to deliver actual solutions follows a similar pattern. The only thing these algorithms will do is filter and classify groups of people to blame for the way AI itself degrades the already shoddy state of public services. The double whammy that comes with AI is the way that the industry itself is a catalyst for more extreme right ideologies. The seeds of this can be seen in the apparently innocuous turn to 'AI safety', which was initiated by Rishi Sunak but has been endorsed and continued by the current government. The rationale of AI safety is not to protect people from the everyday harms of dysfunctional AI in their everyday lives, but to head off the imagined potential for AI to trigger human extinction by developing bioweapons or by becoming completely autonomous and simply taking over. This, in turn, derives from the underlying belief in AGI or Artificial General Intelligence; the belief that the humungous but stumbling AI of today is a step to superintelligent systems that will be superior to humans. As Hollywood as this might seem, it's the position of many in AI, from godfather Geoffrey Hinton to the founders of most of the main companies like DeepMind and OpenAI. Such powerfully warping beliefs spawn real world consequences because, ultimately, it's built on a eugenicist mindset. The very idea of 'general intelligence' comes from Victorian eugenics, when scientists like Francis Galton and Karl Pearson were rationalising the racial supremacy that legitimised the British Empire. The idea of superior intelligence always comes with its corollary of inferior intelligence, whether that's defined racially or in terms of disabilities, and always pans out as assessing some lives as more worthy than others. Part of the motive for a belief in AGI is self-serving. If superintelligent AI is our only hope to solve climate change then we shouldn't be limiting the development of AI through small minded measures like carbon emissions targets, but should be mobilising all available resources, fossil fuel or not, behind its accelerated development. This is also good news for fossil fuel oligarchs who, not coincidentally, are some of the biggest funders of far right think tanks. Similarly, if the future of humanity is to join such superintelligence inside computers themselves, and this leads to vastly multiplied numbers of virtual humans, then facilitating the emergence of AGI and the 1054 virtual future humans becomes morally more important than any collateral harms to actual humans in the present moment. Again, while this might seem deranged, it's the stance of a set of beliefs known variously as 'effective altruism' (EA) or 'long termism' which are very influential in Silicon Valley. As a world view, it elevates the self-styled mission of people in AI above the meagre concerns of ordinary folk. Disturbingly, it seems that the infiltration of US and UK policy circles by people with EA beliefs was responsible for the shift to an AI Safety agenda and, in the UK, the creation of an AI Safety Institute. Lurking behind long termism, but equally influential in Silicon Valley, are the darker and more explicitly fascist beliefs of neoreaction . These kinds of ideas, as espoused by the likes of Peter Thiel, argue that democracy is a failed system, that corporations should take over social governance with monarchical models, and that society should optimise itself by weeding out inferior and unproductive elements. It's this strand of far right tech accelerationism that merged with the MAGA movement in the run up to Trump's 2024 re-election. While Elon Musk's support for anti-immigrant pogroms and his direct attacks on Starmer et al have been the most visible consequences for the UK so far, the real threat is the underlying convergence of Big Tech and reactionary politics. AI is not simply a technology but a form of technopolitics, where technology and politics produce and reinforce each other, and in the case of contemporary AI this technopolitics tends towards far right solutionism. Alternatives Neither Labour nor any other political party is going to defend us against this technopolitics. We can, however, oppose it directly through 'decomputing'. Decomputing starts with the refusal of more data centres for AI, on the basis that they are environmentally damaging and because they run software that's socially harmful. Hyperscale data centres are the platform for AI's assaults on workers' rights, through precaritisation and fake automation, but also for the wider social degradations of everything from preemptive welfare cuts to non-consensual deepfake porn. Decomputing opposes AI because it's built on layers of exploitative labour, much of which is outsourced to the Global South, and because its reductive predictions are foreclosing life chances wherever they're applied to assess our future 'value'. What's needed to salve pain and suffering isn't the enclosure of resources to power the judgements of GPUs but acts of care, the prioritisation of relationships that acknowledge our vulnerabilities and interdependencies. Decomputing is a direct opposition to the material, financial, institutional and conceptual infrastructures of AI not only because they promote an already-failed solutionism but because they massively scale alienation. By injecting even more distancing, abstraction and opacity into our lives, AI is helping to fuel our contemporary crisis, furthering the bitter resentments that feed the far right and the disenchantments that separates us from the more-than-human lifeworld. What we urgently need, instead of a political leadership in thrall to AI's aura of total power, is a reassertion of context and agency by returning control to more local and directly democratic structures. Decomputing argues that, wherever AI is proposed as 'the answer', there is a gap for the self-organisation of people who already know better what needs to be done, whether it's teachers and students resisting generative AI in the classroom or healthcare workers and patients challenging the algorithmic optimisation of workloads that eliminates even minimal chances to relate as human beings. Decomputing claims that the act of opposing AI's intensification of social and environmental harms is at the same time the assertion that other worlds are possible. It parallels contemporary calls for degrowth, which also opposes runaway extractivism by focusing on the alternative structures that could replace it. As much as contemporary AI is a convergence of off-the-scale technology and emergent totalitarianism, decomputing offers a counter-convergence of social movements that brings together issues of workers' rights, feminism, ecology, anti-fascism and international solidarity. Where AI is another iteration in the infrastructuring of Empire, decomputing recognises the urgency of starting to develop alternative infrastructures in the here-and-now, from renewable energy coops to structures of social care based on mutual aid. The murmurings in the financial pages of mainstream media that AI's infinite growth is actually a bubble prone to collapse misses the point that a wider collapse is already upon us in one form or another. Climate change is happening in front of our eyes, while it's pretty clear that liberal democracy is allowing itself to be eaten from the inside by the far right. The more that AI is allowed scale its reactionary technopolitics, the more it will have the effect of narrowing options for the rest of us. Decomputing is the bottom-up recovery of alternatives that have been long buried under techno-fantasies and decades of neoliberalism; people-powered visions of convivial technologies, cooperative socialities and a reassertion of the commons. <https://www.computerweekly.com/opinion/Labours-AI-Action-Plan-a-gift-to-the…>

1 0

Esistono "tecnologie" "politicamente neutrali"?
by Giacomo Tesio Jan. 14, 2025

Jan. 14, 2025

Ciao Giuseppe, Alfredo, Juan Carlos e tutti. Colgo l'assis di Giuseppe per riportare in lista una discussione che è proseguita in privato e che potrebbe essere di interesse perché ricorrente in Nexa. Ad esempio, il Professor Cerroni, nell'intervista condivida da Juan Carlos [1], dice esplicitamente: > la tecnologia non è neutra, contiene dentro di sé un’anima politica > anche al di là dell’uso che se ne può fare. Una affermazione che io trovo al limite dell'ovvietà (forse perché creo tecnologie per mestiere e partecipo alla definizione degli effetti politici che queste avranno, fino al limite di dichiararmi incapace di realizzare certi software, qualora li ritenga dannosi per la società in cui vivranno i miei figli). Per contro, Giuseppe ed Alfredo sostengono che quanto meno la maggioranza delle tecnologie sono neutrali e gli effetti sociali e politici che hanno dipendono esclusivamente dalle scelte di coloro che le usano. Si tratta di posizioni antitetiche, che credo possa essere interessante discutere alla ricerca di una sintesi che le comprenda al proprio interno e ci fornisca un modello migliore della realtà. On Mon, 13 Jan 2025 09:27:51 +0100 Giuseppe Attardi wrote: > Temo che non siamo d’accordo sul significato di “neutrale” per una > tecnologia. Allora perché non provare a partire da definizioni rigorose? Definiamo "tecnologia" un determinato insieme di artefatti concreti costruiti dall'uomo. In questo modo possiamo ragionare su qualcosa di "tangibile", non di un'idea astratta, sia essa platonica ("coltellini svizzeri", "LLM"), o pura fantasia ("morte nera", "intelligenze artificiali" etc..). Naturalmente molte idee (più precisamente, informazioni), e come vengono descritte / diffuse, hanno profonde implicazioni politiche e sociali, ma per quanto interessante, l'analisi degli effetti politici di determinate idee ci porterebbe lontano dai temi della lista. Non parliamo però dell'idea astratta di una tecnologia (i "LLM") ma delle sue applicazioni concrete. Definiamo poi "politicamente neutrale" un artefatto non intrinsecamente orientato politicamente, ovvero che 1. non presuppone alcuna specifica organizzazione sociale 2. non favorisce né sfavorisce alcun uso specifico, non favorendo o sfavorendo, di conseguenza, alcuna evoluzione della società nel tempo Molte tecnologie (insiemi di artefatti costruiti dall'uomo) che non sono neutrali per il secondo punto, lo sono anche per il primo, ma non viceversa. Una tecnologia che presuppone, ad esempio, un forte accentramento di ricchezza e potere per essere creata, non è politicamente neutrale per il primo punto. Infatti, come minimo, l'esistenza di artefatti tecnologici con tale caratteristica giustifica e rafforza quella specifica organizzazione sociale, rendendo la società nel suo complesso dipendente dal perpetuarsi di tale organizzazione. Una tecnologia che favorisce certi usi socialmente rilevanti e ne sfavorisce altri, non è neutrale per il secondo punto. Ad esempio un coltellino svizzero non è facile da utilizzare come arma, ma incentiva una cultura del fai da te per piccole riparazioni. La scheda ethernet o l'automobile sono altri esempi di tecnologie politicamente orientate da entrambi i punti di vista: - presuppongono una società industriale per essere prodotte e enormi investimenti infrastrutturali (strade / cablaggi) per essere utili - velocizzando le comunicazioni e i trasferimenti, riducono le distanze, facilitando determinate evoluzioni della società e ostacolandone altre. Vi sembrano definizioni ragionevoli? Se no, perché? Una volta individuate definizioni esplicite, attinenti e condivise, potremo ragionare se esistano davvero tecnologie politicamente neutrali, ovvero che rispondano a tali definizioni, nonché valutare in modo rigoroso la neutralità di nuove tecnologie. Giacomo [1] https://server-nexa.polito.it/pipermail/nexa/2025-January/053820.html PS x Giuseppe: per quanto possibile (sono un po' a corto di tempo libero) ti rispondero sul resto della mail nell'altro thread, ma credo che trovare definizioni condivise e cercare una sintesi sulla questione della "neutralità della tecnologia" può meritare un thread separato.

7 8

Dagmar Monet su IA e istruzione (da LinkedIn)
by J.C. DE MARTIN Jan. 14, 2025

Jan. 14, 2025

*Dagmar Monet* /Director of Computer Science Dept., Prof. Dr. Computer Science (Artificial Intelligence, Software Engineering)/ https://www.linkedin.com/posts/monettdiaz_ai-activity-7284433996137046016-T… Sometimes the evidence is there, before our eyes, in the form of *papers*. The only thing we need is to go out of our bubbles and *read them*! There is mounting evidence against the use of AI in education. This is why I agree with Chollet and celebrate his being open about it. It could mean he is at least aware of part of the issues! These are three examples *among many*: (1) "Exploring the Impact of ChatGPT on Business School Education: Prospects, Boundaries, and Paradoxes" (Valcea et al., 2024) https://lnkd.in/dPR5UXdj <https://lnkd.in/dPR5UXdj> "AI is likely to have a net-negative effect on learning. This stance is based on our trials with ChatGPT on various cognitive tasks organized around the revised Bloom’s Taxonomy of learning. We identify areas ... where [students] exhibit significant limitations, such as logical reasoning and critical thinking. We caution against the potential deskilling in critical thinking due to students’ overreliance on AI for basic tasks." (2) "Cognitive ease at a cost: LLMs reduce mental effort but compromise depth in student scientific inquiry" (Stadler et al., 2024) https://lnkd.in/dJ4kETS7 <https://lnkd.in/dJ4kETS7> "The study aimed to investigate potential differences in cognitive load, as well as the quality and homogeneity of the students' recommendations and justifications. Results indicated that students using LLMs experienced significantly lower cognitive load. However, despite this reduction, these students demonstrated lower-quality reasoning and argumentation in their final recommendations compared to those who used traditional search engines. Further, the homogeneity of the recommendations and justifications did not differ significantly between the two groups, suggesting that LLMs did not restrict the diversity of students’ perspectives. These findings highlight the nuanced implications of digital tools on learning, suggesting that while LLMs can decrease the cognitive burden associated with information gathering during a learning task, they may not promote deeper engagement with content necessary for high-quality learning per se." (3) "AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking" (Gerlich, 2025) https://lnkd.in/dthFBNwT <https://lnkd.in/dthFBNwT> "The findings revealed a significant negative correlation between frequent AI tool usage and critical thinking abilities, mediated by increased cognitive offloading. Younger participants exhibited higher dependence on AI tools and lower critical thinking scores compared to older participants." Among others! [[Added: We cite other examples in our preprint "The commodification of education and the (generative) AI-induced scam-like culture" Cc<https://www.linkedin.com/in/ACoAAAPiUt8BhbVRSOXbKYFTfZnIEOrHQZVysMI>Gilbert Paquet, M.A. <https://www.linkedin.com/in/gilbertpaquet/> https://lnkd.in/dV--9MXB <https://lnkd.in/dV--9MXB> and there are many other reasons, e.g. related to unethical practices: https://lnkd.in/eF9K-sSK <https://lnkd.in/eF9K-sSK> ]]

3 2

Re: [nexa] Esistono "tecnologie" "politicamente neutrali"?
by Giuseppe Attardi Jan. 13, 2025

Jan. 13, 2025

> On 13 Jan 2025, at 12:16, nexa-request(a)server-nexa.polito.it wrote: > > Non parliamo però dell'idea astratta di una tecnologia (i "LLM") ma > delle sue applicazioni concrete. Proprio questo è il punto cruciale. Quando sostengo che bisogna distinguere tra tecnologia e le sue applicazioni, intendo proprio questo. Per tecnologia intendo i principi di base, la conoscenza tecnico scientifica utilizzabile per scopi pratici. Ogni applicazione concreta ha le sue finalità e di conseguenza è orientata a scopi e interessi di chi la realizza. Chi le produce deve esserne cosciente e si assume la responsabilità delle conseguenze. Chi le utilizza accetta e condivide la responsabilità degli effetti di tale utilizzo. Questa condivisione di responsabilità è coerente con le affermazioni che strumenti e loro utilizzatori costituiscono a volte un’accoppiata inseparabile di responsabilità. Ho fatto come esempio di tecnologia separato dalle applicazioni le tabelle dei logaritmi, ma anche per le altre applicazioni citate da qualcuno come controesempi di tecnologie non neutre si può fare la stessa distinzione: - smartphones: la tecnologia è semplicemente la combinazione di personal computer con connettività internet (GSM o WiFi). BTW, Tablet e PC usano la stessa tecnologia, per cui è artificiosa la critica agli uni e non agli altri. - StarLink, la tecnologia è quella dei satelliti LEO e della connettività Internet via radio e tramite link dinamici. Anche qui la critica al monopolio di Musk riguarda questa particolare istanza di servizio. - whatsapp, la tecnologia è un semplice la combinazione di tre protocolli: Signal, WebRTC e XMPP. Tutte queste tecnologie sono disponibili e altre implementazioni sono possibili. Il fatto che alcune di queste si siano imposte rispetto ad altre è una questione ben poco tecnologica (come è stato fatto notare non sempre si afferma la tecnologia migliore), ma principalmente antropologica (winner takes all o rich gets richer, che dipende dai limiti cognitivi della nostra mente: possiamo ricordare al più 3-4 cose per settore) che viene sfruttato dai sistemi economici e finanziari per promuovere un vincente e stroncare i concorrenti. La tecnologia ha sempre dei limiti ed è perfettamente legittimo metterli in evidenza. Per le attuali tecnologia degli LLM ci sono limiti riguardo la stima a priori dell’accuratezza, la scalabilità dei modelli, la quantità di dati necessaria all’apprendimento, il catastrophic forgetting, ecc. Ma direi che nessuno di questi costituisca una violazione di valori morali tale da metterle al bano o di sottoporle alla generale esecrazione. Esistono tecnologie che incorporano violazioni di principi morali. Ad esempio le camere a gas, anche se usate per eseguire pene capitali irrogate da un tribunale, sono un esempio: molti paesi hanno abolito la pena capitale e questo rende tali tecnologie contrarie ai principi morali a cui le leggi di quei paesi fanno riferimento. Lo stesso vale per le armi nucleari. Legittimo quindi da parte degli scienziati di rifiutarsi di sviluppare tecnologie destinate a tali usi o di sviluppare applicazioni che violino principi morali. —

3 2

Assistant/Associate Professor Position in Computational Sociology of Work - Institut Polytechnique de Paris
by Antonio Casilli Jan. 13, 2025

Jan. 13, 2025

Dear all, The Institut Polytechnique de Paris (more precisely its telecommunication "Grande École", Télécom Paris) is recruiting a Maître de conférences (a position equivalent to Assistant/Associate Professor) working at the intersection of computational sociology and sociology of work. The successful candidate will join the Department of Economic and Social Sciences and become a member of the Interdisciplinary Institute on Innovation (i3), a joint research unit of the CNRS (National Centre for Scientific Research). This is also where my own research team, DiPLab (Digital Platform Labor), is based. The Institut Polytechnique de Paris stands among France's most prestigious public research universities. Its remarkable legacy includes five Nobel Prize winners and two Fields medalists. Within it, Télécom Paris ranks among the top 5 engineering schools in France. Its place in history was secured when the term 'telecommunications' was first coined at the school in 1904. Télécom Paris's mission is to design, innovate, and develop digital technologies and solutions that benefit society and the environment. Located on the Palaiseau campus alongside other Grandes Écoles, its human-scale and inclusive environment maintains a strong international presence while fostering close ties with industry through an incubator in Paris at the heart of the French ecosystem. The Department of Economic and Social Sciences has been at the heart of Télécom Paris for forty years, championing the integration of social sciences into engineering education. Within the department, sociology has contributed to building the school's national and international reputation in social sciences. We are now looking to align the department's sociological work with emerging themes in the educational and scientific community. The position we are offering combines expertise in sociology of work with computational social sciences, with a particular focus on artificial intelligence. The successful candidate will contribute to both research and teaching missions. Teaching responsibilities will include courses in digital sociology, sociology of work, AI sociology, and computational social sciences within both Télécom Paris and Institut Polytechnique de Paris engineering programs. We expect candidates to hold a doctorate in sociology or computer science. We also welcome applications from candidates with doctorates in related disciplines (economics, philosophy, mathematics, information & communication, media studies, law, or anthropology) or relevant specializations (human-computer interaction, computational social sciences, statistical learning, or data science). Please read the presentation attached and before applying, be sure to contact me (Antonio Casilli) and my colleague Dr. Tiphaine Viard (<firstname.lastname>@telecom-paris.fr). Final applications must be submitted before 02/17/2025 at: https://institutminestelecom.recruitee.com/l/en/o/assistantassociate-profes… Best regards, Antonio -- Antonio A. Casilli Professor, Institut Polytechnique de Paris My new book: Waiting for Robots. The Hired Hands of Automation, University of Chicago Press, 2025. My new documentary: In the Belly of AI, Federation Studios, 2025. *We respect your right to disconnect. My workflow efficiency accounts for the timing of this email. Taking action or responding outside of office hours is not required.*

1 0

Papa Francesco al corpo diplomatico
by don Luca Peyron Jan. 13, 2025

Jan. 13, 2025

Alcuni passaggi mi paiono parecchio interessanti, ad esempio in tema di educazione. https://www.vatican.va/content/francesco/it/speeches/2025/january/documents… Buona lettura dl _________________________ don Luca Peyron Pastorale Universitaria - Apostolato Digitale Arcidiocesi di Torino www.universitari.to.it via XX settembre 83, Torino tel. 011 5156239

1 0

Re: [nexa] Intervista ad Andrea Cerroni
by Giuseppe Attardi Jan. 13, 2025

Jan. 13, 2025

Non ho capito cosa c’entri l’AI col progetto Cybersyn del governo Allende. È vero che ormai se non si cita l’AI non si è à la page, ma forse si esagera. Oltretutto mi pare che le ricostruzioni sociologiche a posteriori lascino il tempo che trovano e cadono in inesattezze. Per esempio la Operation Room di cui si mostra una foto, non è mai stata realizzata, era solo un progetto futuro. Ho conosciuto di persona Fernando Flores e l’ho invitato a Pisa negli anni ‘90 a raccontare il suo modello di organizzazione aziendale, basato sulla teoria degli Speech Act. Flores aveva collaborato con Terry Winograd, uno dei pionieri dell’AI linguistica. Dopo la caduta di Allende si era trasferito in California dove aveva fondato una società che forniva consulenze aziendali sfruttando il suo modello. Ma non esisteva al tempo una tecnologia di AI che potesse essere usata in pratica. Cybersyn era infatti sostanzialmente un sistema distribuito (o meglio centralmente interconnesso) di controllo e gestione della produzione, basato su tecniche di Ricerca Operativa. Straordinario ed entusiasmante a quel tempo, specialmente perché fatto tutto in casa con l’obiettivo di mantenere un controllo pubblico della produzione. — > On 13 Jan 2025, at 09:28, nexa-request(a)server-nexa.polito.it wrote: > > Subject: [nexa] Intervista ad Andrea Cerroni > Message-ID: <9c60f2d5-f175-4227-be9e-3f95fcc5610b(a)polito.it> > Content-Type: text/plain; charset=UTF-8; format=flowed > > Segnalo questa intervista al Prof. Andrea Cerroni, recentemente ospite > del Centro Nexa: > > "CILE 1973: IL GOLPE DELL’ INTELLIGENZA ARTIFICIALE. INTERVISTA AL PROF. > ANDREA CERRONI" > https://giubberossenews.it/2025/01/07/cile-1973-il-golpe-dell-intelligenza-…. > > jc >

1 0

Re: [nexa] In a mirror
by Giuseppe Attardi Jan. 13, 2025

Jan. 13, 2025

4 4