May 2025 - nexa - server-nexa.polito.it

[Responsible AI] When Abstraction Fails: Navigating Uncertainty in High-Stakes Systems

by Daniele Quercia

Title: When Abstraction Fails: Navigating Uncertainty in High-Stakes Systems Join the meeting<https://teams.microsoft.com/l/meetup-join/19%3ameeting_YzczNmM0YWUtNWU4ZC00…> Abstract: With high-stakes systems, safety and security engineering depends heavily on abstractions — models, simulations, bench tests — that stand in for the system itself. But these abstractions have limits. They inevitably rely on imperfect assumptions: simplifications that hold until they don’t, especially under conditions of novelty, extremity, or scale. In systems that push the boundaries of performance, therefore, abstraction becomes brittle, and the pretense of total foresight becomes a liability. This talk examines what happens when engineers reach the limits of their abstractions. Focusing on testing and modeling under conditions of extremity, I explore how ‘rational’ accidents can arise from the limits of what can be known, tested, or anticipated in advance. From there I examine how one domain — civil aviation — manages these accidents by leveraging decades of accumulated operational experience. The talk is an invitation to reconsider what counts as credible knowledge in high-consequence engineering, and to value operational experience not as a fallback when models fail, but as a foundation for safety and security when they inevitably do. <https://arxiv.org/abs/2407.02191>Bio: John Downer is Associate Professor of Science and Technology Studies at the University of Bristol. He has written extensively about the idiosyncrasies of technoscientific knowledge, the limits of proof, and their implications for technology governance. His 2024 monograph ‘Rational Accidents: Reckoning with Catastrophic Technologies,’ examines the logical dilemmas of establishing ‘ultra-high’ reliabilities in highly complex systems such as jetliners and reactors, and unpacks the practical ramifications of those dilemmas. His broader work looks at epistemological and regulatory issues in a range of technological domains, from civil aviation and nuclear, to autonomous systems and artificial intelligence. He has a PhD in Science and Technology Studies (STS) from Cornell University, and prior to joining Bristol he held positions at Kings College London, the London School of Economics (LSE), and Stanford University. He is currently a visiting researcher at the Centre for the Governance of Artificial Intelligence (GovAI) in Oxford.

May 12, 2025

[Responsible AI] When Abstraction Fails: Navigating Uncertainty in High-Stakes Systems

by Daniele Quercia

Title: When Abstraction Fails: Navigating Uncertainty in High-Stakes Systems Join the meeting<https://teams.microsoft.com/l/meetup-join/19%3ameeting_YzczNmM0YWUtNWU4ZC00…> Abstract: With high-stakes systems, safety and security engineering depends heavily on abstractions — models, simulations, bench tests — that stand in for the system itself. But these abstractions have limits. They inevitably rely on imperfect assumptions: simplifications that hold until they don’t, especially under conditions of novelty, extremity, or scale. In systems that push the boundaries of performance, therefore, abstraction becomes brittle, and the pretense of total foresight becomes a liability. This talk examines what happens when engineers reach the limits of their abstractions. Focusing on testing and modeling under conditions of extremity, I explore how ‘rational’ accidents can arise from the limits of what can be known, tested, or anticipated in advance. From there I examine how one domain — civil aviation — manages these accidents by leveraging decades of accumulated operational experience. The talk is an invitation to reconsider what counts as credible knowledge in high-consequence engineering, and to value operational experience not as a fallback when models fail, but as a foundation for safety and security when they inevitably do. <https://arxiv.org/abs/2407.02191>Bio: John Downer is Associate Professor of Science and Technology Studies at the University of Bristol. He has written extensively about the idiosyncrasies of technoscientific knowledge, the limits of proof, and their implications for technology governance. His 2024 monograph ‘Rational Accidents: Reckoning with Catastrophic Technologies,’ examines the logical dilemmas of establishing ‘ultra-high’ reliabilities in highly complex systems such as jetliners and reactors, and unpacks the practical ramifications of those dilemmas. His broader work looks at epistemological and regulatory issues in a range of technological domains, from civil aviation and nuclear, to autonomous systems and artificial intelligence. He has a PhD in Science and Technology Studies (STS) from Cornell University, and prior to joining Bristol he held positions at Kings College London, the London School of Economics (LSE), and Stanford University. He is currently a visiting researcher at the Centre for the Governance of Artificial Intelligence (GovAI) in Oxford.

May 12, 2025

Re: [nexa] I 23 “no” per riprendersi Internet

by Marco A. Calamari

On gio, 2025-05-08 at 21:06 +0200, Andrea Trentini wrote: > Ho appena ascoltato questo podcast di Attivissimo: > > https://attivissimo.me/2025/05/07/podcast-rsi-perche-devo-dare-una-password… > > (a questo link anche il testo trascritto, per chi non ama gli audio) > > è carino, non sapevo della lista dei 23 "no", invero se ne > potrebbero aggiungere un bel po' :) Confermo. Mi ha colpito molto come Attivissimo ha trattato e riassunto i contenuti. Didatticamente molto ben fatto. > Andrea Trentini ⠠⠵ - http://atrent.it - public key ID: 0xA7A91E3B - > Dip.to di Informatica - > Università degli Studi di Milano > Cittadinanza Digitale e Tecnocivismo - > https://www.ledizioni.it/prodotto/cittadinanza-digitale-tecnocivismo > - > https://edizionithemis.it/catalogo/digitale-e-societa/dal-tecnocivismo-alla… > Sistemi Embedded - > https://www.ledizioni.it/prodotto/a-carraturo-a-trentini-sistemi-embedded-t… >

May 8, 2025

I 23 “no” per riprendersi Internet

by Andrea Trentini

Ho appena ascoltato questo podcast di Attivissimo: https://attivissimo.me/2025/05/07/podcast-rsi-perche-devo-dare-una-password… (a questo link anche il testo trascritto, per chi non ama gli audio) è carino, non sapevo della lista dei 23 "no", invero se ne potrebbero aggiungere un bel po' :) -- Andrea Trentini ⠠⠵ - http://atrent.it - public key ID: 0xA7A91E3B - Dip.to di Informatica - Università degli Studi di Milano Cittadinanza Digitale e Tecnocivismo - https://www.ledizioni.it/prodotto/cittadinanza-digitale-tecnocivismo - https://edizionithemis.it/catalogo/digitale-e-societa/dal-tecnocivismo-alla… Sistemi Embedded - https://www.ledizioni.it/prodotto/a-carraturo-a-trentini-sistemi-embedded-t…

May 8, 2025

Dopamina: Quando un'app diventa droga

by Giacomo Tesio

Ciao a tutti, volevo segnalarvi questa breve serie di corti divulgativi sui social (credo a produzione francese): https://www.arte.tv/it/videos/RC-017841/dopamina/ Ne ho visti un paio e li ho trovati chiari e divertenti (sebbene, come spesso accade, superficiali dal punto di vista tecnico [1]) Non mi pare fossero già passati in lista, ma credo che possano aiutare molti a comprendere un po' meglio le caratteristiche progettuali di questi software e le intenzioni di chi li realizza e distribuisce. Giacomo [1] ad esempio, quando vengono elencati i dati condivisi dalle varie app, l'elenco non tiene conto delle correlazioni fra coordinate satellitari (e dunque luoghi), orari etc.. di tutte le altre persone ivi presenti nello stesso posto etc...

May 8, 2025

AI Is Reshaping American Cities - But Not Equally

by Daniele Quercia

May be of interest… AI Is Reshaping American Cities - But Not Equally Daniele Quercia [cid:image001.png@01DBBFF9.91B8F060] Darker colours represent higher AI impact within MSAs. Grey areas indicate regions outside the boundaries of MSAs, which generally have lower population densities and are not included in the analysis due to limited data availability. Artificial intelligence promises to transform how we work, automating tasks, boosting productivity, and opening up new economic frontiers. But beneath the optimism lies a stark reality: AI's impact is far from evenly spread. In a new study<https://epjdatascience.springeropen.com/articles/10.1140/epjds/s13688-025-0…> published in Springer EPJ Data Science, we examined how AI is likely to affect labour markets across U.S. metropolitan areas. The results reveal a troubling pattern: one that risks deepening America’s existing economic divides. By analysing 24,758 AI-related patents filed between 2015 and 2022 and linking them to 17,879 occupational tasks, we developed a fine-grained map of where AI is likely to hit hardest. The findings defy the assumption that only major tech hubs are vulnerable. In fact, many of the most at-risk places are mid-sized cities whose economies hinge on a single sector. Dalton, Georgia, for example, is tightly bound to carpet manufacturing. Columbus, Indiana depends on the automotive industry. Rochester, Minnesota, renowned for its healthcare sector, also faces elevated risk. In each case, a heavy reliance on routine-intensive industries leaves these places exposed to automation, with few alternative sectors to soften the blow. In contrast, some of the least-impacted cities in our analysis such as Daphne-Fairhope-Foley in Alabama, Coeur d’Alene in Idaho, and Grants Pass in Oregon share a different trait: economic diversity. These cities spread employment across a broader range of industries, making them better equipped to absorb shocks and adjust as AI transforms specific sectors. We also found that the East Coast, on average, is more exposed to AI disruption than the West. Though highly innovative hubs like the Bay Area and Seattle remain high-impact outliers. Cities experiencing rapid growth in “creative class” jobs such as engineers, scientists, designers are also more likely to face disruption. Knowledge work, it turns out, is not immune to automation; in some cases, it is on the front lines. AI is not just a story of technological progress. It is a story of geography, resilience, and inequality. Left unchecked, it could entrench existing divides, fuelling growth in already-advantaged regions while sidelining those with fewer resources to adapt. The policy response must match the scale of the challenge: support for economic diversification, investment in re-skilling, and a renewed focus on regions at risk of being left behind. To help individuals and policymakers better understand what AI might mean for them, we’ve also made our data public. On our interactive platform<http://social-dynamics.net/aii>, users can explore how AI is projected to affect different occupations, and even drill down to specific tasks within those jobs to see which are most susceptible to automation. The AI revolution is underway. With the right tools and the right foresight, we can ensure it lifts more communities than it leaves behind.

May 8, 2025

European Data Sovereignty, iniziative italiane?

by Marco Fioretti

Buongiorno a tutti, devo scrivere (entro domani sera...) un breve commento, molto generale, su questo tema. Qualsiasi segnalazione di iniziative e proposte italiane in proposito, nonche' ad esempio di posizioni e commenti dei politici, e' benvenuta. Idem per vostre opinioni sul tema. NON so se riusciro' a usare quanto ricevero' stavolta, ma potrei proporlo come seguito del primo post. Grazie in anticipo, Marco

May 8, 2025

AI and Internet Hygiene

by Daniela Tafani

Non mi pare sia stato segnalato a suo tempo AI and Internet Hygiene we are going to have to re-learn how to use the computer Kate Wagner Oct 01, 2024 I’m from a very specific generation, the first one in which learning how to use the open Internet was a skill taught in childhood. There’s not much that’s good about being from this generation (I will never own a house, be out of student loan debt, and the idea of having children takes on an air of fantasy) but at least I learned how to be proficient at the computer. This proficiency was out of necessity. You see, back in the days before total market saturation, the big tech companies had an incentive to educate. They couldn’t sell a technology people were so intimidated by. Classes were taught in schools, libraries, and community centers, and a cottage industry of computer instruction sprung up everywhere. A wealth of educational materials flooded the market targeted towards new computer and Internet users, which, in the early 1990s, was pretty much anyone not working in a previously computerized industry. Some of these were even for children: <https://www.youtube.com/watch?v=SzQLI7BxfYI> The Internet of the 1990s and 2000s and the Internet of the present couldn’t be more different. Both the early and middle Internet shared a commonality of being mostly decentralized and much less closed-source than it presently is under platformization. While so many of us are nostalgic for this now, in an era when five corporations own the entire web, at the time it was a double-edged sword. In some ways, the basic HTML and blinking GIFs Internet of the past was utopian in scope — millions of people learning how to express themselves and communicate with one another in real time, each around a certain skill level. But in other ways, it was an experimental void in which a mix of anonymity and technological prowess could be used not to contribute knowledge or become better-connected but to exploit those just figuring out how to use this brand new technology. Later, when the Internet reached more people, when it evolved into more expansive communities, blogs, content aggregators and early social media sites such as YouTube or MySpace, and when storage space became more conducive to hosting big files, threats and dangers adapted to this new environment in ever-more savvy ways. In other words, the old Internet was not exactly safe. This was especially true during the Windows XP era when that operating system achieved an unprecedented and long-lasting market share, making it a common target. Pop-up ads, scams, drive-by downloads, and wretched computer viruses were extremely common. If you ended up somewhere you weren’t supposed to, the consequences could be dire: your computer would be crippled, hijacked and made slow; sometimes you could even lose everything, all from clicking a link you weren’t supposed to. People my age (30) remember this well. It was a rite of passage in those days to give your family computer a particularly terrible virus (such as one of those fake anti-virus programs that is actually malware) while trying to download custom content for The Sims 2. Hence, because of this fraught webscape, by the age of ten, I knew how to clean up after these kinds of messes — going into the registry, system restoring, booting the computer into Safe Mode so as to run an anti-virus or uninstall a harmful program — all of which required understanding on a basic, structural level, how the computer worked. But more importantly, I learned how to avoid getting into all that trouble in the first place by being able to spot suspicious search engine results, inspect links, and generally discern through visual cues and precedence what parts of online seemed legitimate versus illegitimate. I practiced what is sometimes called Internet hygiene. While computer viruses never stopped being a threat (everyone should still employ the services of an antivirus software) the nature of that threat has changed. Organized hackers are less interested in giving your PC a fake antivirus and more interested in targeting large scale legacy systems such as those used in governments, hospitals, and corporations. Rather than THEMATRIX_(1999).mov.exe, social engineering, ransomware, click fraud, cryptojacking, and data leaks are the orders of the day. Meanwhile, the platformized, app-driven Internet, through moderation and centralization, have made websites into closed systems more impervious to old-school threats such as pop-up ads or drive-by downloads. New Internet-based licensing protocols (increasingly based on monthly subscriptions instead of one-time software licenses) such as Adobe Creative Cloud and Steam, and streaming media services such as Spotify and YouTube that provide swaths of content for little to no cost have not only decimated physical media (and price gouged creators and users alike) — they’ve also chipped away at piracy and the once-overwhelmingly common practice of peer-to-peer sharing, another virus hotbed. Ad-blockers, while a wicked problem for the old-web economy of click-based revenues, also do their part in keeping users safe. I can remember very clearly when I stopped being able to use the computer. It was around 2012, when I switched to MacOS, which (I feel) is a much less transparent operating system than Windows, one that held less of a market share and was thus less commonly targeted by widespread Internet booby traps. (This is not to say Macs don’t get viruses — they absolutely do.) Meanwhile, Internet consumption shifted to smartphones, and the app-based ecosystem only reinforced the closedness and opaqueness of platforms as other websites gradually disappeared. I myself have no idea how a smartphone works, nor the full extent to which my data is being tracked. In fact, we are so inured to sharing data that the very concept of Internet (and personal) privacy itself is eroding more and more. As a result of these massive transformations, computer and Internet usage have been remarkably deskilled among ordinary users. The knowledge of how the Internet or an operating system works, things that used to be necessary simply for using both, has become obfuscated and, as a result, relegated to those in the privatized technology sector. Now, however, the Internet is changing massively yet again. The premature unleashing of AI and large language models (LLMs) in particular onto the open Internet is already having dire consequences for what were, for years, considered stabilized, centralized (if flawed) systems, such as the search engine. AI, thanks to the scope of its spread, its rogue unreliability (it lies — often), the way it poisons search results, hijacks SEO, and is increasingly being used for disinformation and fraud, is reintroducing a fundamental and destabilizing distrust back into the Internet. Once more, I can no longer trust the results Google provides me. On a daily basis I have to ask myself increasingly familiar questions: Is this first result a legitimate news source? Is this image of a protest real? Is that picture of a Kandinsky painting really his or is an AI forgery of his work? Across the board, it’s becoming increasingly hard to tell. For me and countless others, what used to be rote Internet usage has now turned into a nightmarish amount of wasted time spent discerning what is and isn’t real. As far as I, the ordinary user, am concerned, AI is evolving not into a life-changing and labor-saving technology as was promised by its capitalist overlords, but rather into a form of malware that targets, whether unwittingly or not, critical Internet infrastructure. The problem is, the deskilling of Internet usage has made many users helpless against this new threat. While it’s (barely) possible for the savvy among us to use prior knowledge of existing resources and databases, or search engine filters (e.g. by date) to avoid AI-generated content, the more AI becomes enmeshed in critical infrastructure, the more we will once more have to develop a new type of Internet hygiene out of necessity. As I had to navigate this new web more and more, I began noticing that being able to tell the signs of whether a publication is legitimate or AI slop was not so different on its face than being able to tell whether that link to a repository of videogame content is giving virus. AI generated content has a visual and rhetorical trace that is becoming more and more consolidated as the software feeds on its own entrails. Just like landing pages with five DOWNLOAD links and janky websites gave pause in the aughts, fake journalism sites that didn’t exist before 2021 and are just ads and links should give pause now. The good news is, we’ve developed Internet hygiene before and we can do it again. Faced with this threat, new guides to healthy and safe Internet use will have to be written for the common user — especially since it looks like the big tech companies aren’t going to do anything to stem the rot. This is a project that should, in a just society, be taken on by libraries and other public outreach programs. Universities and schools should develop curricula not about how to use or embrace AI but to teach a new form of Internet literacy and usage protocol in order to protect students from scams and disinformation. Hopefully, new tools — a kind of AI anti-virus — will be developed to block AI’s web crawlers or even scan for ersatz content, though this might be wishful thinking considering how the technology works. This task will become especially critical in fields such as medicine, engineering, and law, where AI poisoning has the potential to cause incredible damage that puts human life at risk. But first, all of this requires actually recognizing the existential, infrastructural threat AI poses to the tools modern society now runs on. Those systems faced threats at the dawn of their creation that were no less existential. Call me nostalgic, but the old-school virus prevention framework, I think, is a helpful and time-tested one. Finally, as terrible as the AI infection of the Internet at large is, perhaps a re-skilling of the computer-using population is a silver lining. After two generations of Internet illiteracy, an embrace of new tools for empowerment and understanding can help salvage not only the Internet (if such a thing is still possible) but maybe even some of that old utopianism it once promised. At its core, the computer and the Internet are tools for creation, education, commerce, play, and communication. We as a society shouldn’t have to throw the baby out with the bathwater. <https://www.late-review.com/p/ai-and-internet-hygiene>

May 7, 2025

Curl sotto DDoS da AI

by Giacomo Tesio

Gli sviluppatori di Curl banneranno chi segnala bug individuati da AI. <https://archive.is/IZOug> That's it. I've had it. I'm putting my foot down on this craziness. 1. Every reporter submitting security reports on #Hackerone for #curl now needs to answer this question: "Did you use an AI to find the problem or generate this submission?" (and if they do select it, they can expect a stream of proof of actual intelligence follow-up questions) 2. We now ban every reporter INSTANTLY who submits reports we deem AI slop. A threshold has been reached. We are effectively being DDoSed. If we could, we would charge them for this waste of our time. We still have not seen a single valid security report done with AI help.

May 7, 2025

Re: [nexa] AI as Normal Technology

by Giacomo Tesio

Ma diamogli ragione, Daniela! Il 6 Maggio 2025 15:34:05 UTC, Daniela Tafani ha scritto: > Just the other day I was at a lunch in Palo Alto and there were some young > AI scientists there who were saying that they would never have a “bio baby” > because as soon as you have a “bio baby,” you get the “mind virus” of the > [biological] world. > And when you have the mind virus, you become committed to your human baby. > But it’s much more important to be committed to the AI of the future. > And so to have human babies is fundamentally unethical. Per chi crede nella favola della AI, riprodursi è immorale. Peccato che la selezione naturale sia così lenta... :-) Giacomo

May 6, 2025