iiinova — where your digital self keeps evolving

And meanwhile copilot... :-)

110

Good. The dataset of counterfactuals we used has been out there for a while and I'm not surprised things have been "fixed" . The question is "how this was done" and if it can be generalized to any similar cases in the future. Unfortunately many of these things can't be verified with closed models.

010

You can tests it yourself. With some differences these models are largely similar and I'd expect to operate in similar way. Nevertheless, there are many strategies in which this can be fixed in commercial models and they don't offer the same level of transparency needed for what we wanted to test.

110

really important to keep in mind: VLLMs don't see.

220

This holds across model types and sizes (for what we could test). Even more interesting is that the model is "sure and confident" about the wrong answer. Considering how VLLMs are being positioned more and more as consumer products in a number of devices (e.g. wearables) these results are

130

at #ECCV. By using using sequence-level margins, PMI correction, token-step margins, and entropy diagnostics they show how models consistently fall back on their textual prior when confronted with counterfactual visuals. To say it clearly, the model will count 4 legs even if your your camel has 5.

161

Luca Rossi @lrossi.bsky.social · 08/09/2026

VLLMs don't "see", they remember (and they are pretty stubborn about their memories too!). I had the pleasure to supervise two master students at @itu.dk that decided to investigate this fascinating problem of VLLMs. Ayat Khudoir and Rabia Sufian just presented their work at the FAILED workshop

1285

Very lovely weather in Kraków for #ecpr26 where we'll present some initial work on how European climate movements engage with MAPA themes. Spoiler: not a lot.

000

Luca Rossi @lrossi.bsky.social · 07/09/2026

When in the future someone will ask how was it that we got to a point when the European project was in danger as well as peace on the continent... The answer is "we couldn't be bother to take care of the county side and to spread a little bit the was wealth we had accumulated".

011

Reposted by Luca Rossi

Daniel Moser @dmoser.bsky.social · 03/09/2026

Pollution levels in London after 10 years of advancing sustainable transport.

164115623176

Luca Rossi @lrossi.bsky.social · 31/08/2026

ChatGPT, Reddit and Roblox are now VLOSE. These platforms will have several obligations and will be hold to a higher standard of responsibility. Extremely interesting, from a research point of view, is the data access part. ROBLOX! here we come! Give us the data! ec.europa.eu/commission/p...

ec.europa.eu

Commission designates ChatGPT, Reddit, Roblox under Digital Services Act*

Today, the Commission has designated ChatGPT as a Very Large Online Search Engine (VLOSE), as well as Reddit and Roblox as Very Large Online Platforms (VLOPs), under the Digital Services Act (DSA).

073

Claes de Vreese @claesdevreese.bsky.social · 27/08/2026

"willing" is at least debatable. This agreement doesn't introduce anything that wasn't there in some form before. In the meantime it takes no responsibility or accountability for the content or the business model it supports while delegating full responsibility (and conflicts) to families.

010

Reposted by Luca Rossi

Update: the party now labels their AI generated image as AI on their party website. One may fundamentally disagree with such AI usage by political parties or politicians, at least it is now compliant with rules for labelling AI deepfakes.

062

If you're Meta this is a great move, you still don't take any responsibility, you delegate that to the families while limiting your interventions to things that were largely already there. If you're everyone else, you'll have to keep waiting.

000

Cracking Open the European Newsfeed | Journal of Quantitative Description: Digital Media

[here some work that we did ourselves journalqd.org/article/view... ]. Framing it in this way misses the point and fails at addressing the actual risks that algorithmic platforms designed to maximize views and engagement regardless of any notion of "quality" poses to society.

journalqd.org

100

Cracking Open the European Newsfeed | Journal of Quantitative Description: Digital Media

Older generations show no better online behavior or wisdom than your average 14 years old with the added problem that grown ups vote and can influence their kids. That these were not teens problems has been known for years in different domains

journalqd.org

100

we (as a society) missed a huge opportunity but we managed to placate the actual anxious generation — not teens, but the parents whose anxiety this was designed to soothe. With few exceptions, everything that is dangerous and problematic for a teenager is dangerous and problematic for a grown up.

110

Luca Rossi @lrossi.bsky.social · 18/08/2026

Most of what Meta committed to in this settlement was already available, and what's new still just delegates the conflict — and the responsibility — to families. But that's not even the worst part. The worst part is that by framing this as a measure for "teens"

111

University Rankings are bad. Even the ranking that puts your workplace in the Top10 European Universities for #ComputationalSocialScience (and in the top 50 worldwide). cssrankings.org#metric=weigh... Jokes aside, rankings are seriously bad and problematic.

cssrankings.org

Computational Social Science (CSS) Rankings

041

Reposted by Luca Rossi

NERDS @nerdsitu.bsky.social · 10/08/2026

Carl Bergstrom @carlbergstrom.com visiting NERDS and SODAS Save the dates: Wed 19 Aug, 14:00 & Fri 21 Aug, 11:00 nerds.itu.dk/2026/08/07/c...

Photograph of an attentive scientist wearing glasses and a grey pullover

0143

Reposted by Luca Rossi

Axel Bruns @snurb.info · 07/07/2026

Meanwhile, I also have a new chapter in the Handbook of Social and Communication Networks, by @toledobastos.bsky.social and @lrossi.bsky.social. Here I explore the way gatewatching and newssharing practices are affected by platform policies, using the example of the Facebook News Bans in 🇦🇺/🇨🇦.

132

Luca Rossi @lrossi.bsky.social · 26/06/2026

The problem with #AI for office workers is that it doesn't really work. The morning brief that @microsoft.com Copilot was so eager to offer me: missed a couple of important emails that needed follow-up and, instead, recommended me to "confirm my participation" to a meeting I had yesterday.

000

Luca Rossi @lrossi.bsky.social · 25/06/2026

I used to think that the view for my coffee break in Urbino (1) was hard to beat, but I must admit that the view from @itu.dk is also something.

160

Luca Rossi @lrossi.bsky.social · 15/06/2026

from an accessible, yet comprehensive, guide to a very open and hardly defined field of studies. I'm pretty happy with the result and how it came out. Let me know what you think. #networks #communication

000

Luca Rossi @lrossi.bsky.social · 15/06/2026

written by great scholars, that provides both a theoretical frame for how networks and communication can be combined (spoiler: it's not that simple) as well as many examples of practical research. When we decided to work on this project we had in mind students and young scholars who could benefit

100

Luca Rossi @lrossi.bsky.social · 15/06/2026

I'm happy to announce that the Handbook of Social and Communication Networks is finally out: www.e-elgar.com/shop/gbp/han... The work, co-edited with @toledobastos.bsky.social , attempts to define what happens when network science meets communication studies. It's a collection of works,

e-elgar.com

Handbook of Social and Communication Networks

This comprehensive Handbook explores the application of social network analysis and network science in the field of communication studies. Leading experts examine the interactions between social and c...

1178

Reposted by Luca Rossi

Claes de Vreese @claesdevreese.bsky.social · 28/05/2026

🔊 The magnificent @ddc-sdu.bsky.social x TrygFonden Fellowship Program is open, seeking the 2027 cohort 🚀 3 types of fellowships: 1️⃣ academic 2️⃣ journalist 3️⃣ policy fellows 🗓️ deadline September 1 Read the call for details and instructions 🚨🚨Spread the word 🚨🚨 www.sdu.dk/en/forskning...

sdu.dk

Digital Democracy Centre x TrygFonden Fellowship Program

11918

Reposted by Luca Rossi

Miguel Sicart @miguelsicart.bsky.social · 27/05/2026

So, is done, and I am happy and exhausted and ready to start planning Playing Futures 2027.

playingfutures.today

2026 :: Playing Futures Conference

1176

Luca Rossi @lrossi.bsky.social · 21/05/2026

Ok, #aoir26 does anyone know a reasonable solution for child care in Mexico City in October?

021

semantically connected with the themes of justice and exploitations. The problem is - again - how we frame this type of research feeding the hyped narratives of sentients model that experience and suffer traumas.

000

If we talk about "trauma" one would expect - for example - that the trauma would be transmitted in the skill file regardless of the efforts done in the prompt to mitigate that. If models were able to experience exploitation they would not need to be fed with a schema or rewards and stakes that is

100

or information stored in skill files is clearly interesting for scenarios where one imagines multiple agents deployed for long periods of time, but the main observed results is the specific product of the context window and the language doesn't really help a reasonable discussion.

100

(please read the actual instructions). Obviously at this point the skill file will incorporate all the data and semantic connections that were elicited by the totally unnecessary initial design and the "trauma" 🤦 will persist. Now this is clearly not entirely useful. The issue of unwanted data

100

rather than anything else. The authors smartly acknowledge this but then suggest that this could be inherited by future agents through a skill file that might be created. They test this idea by asking the models to create a skill file that saves the "lessons learned"

100

one model shows an effect (suggesting stronger guardrails in the other two) because the result obtained is straightforwardly produced in their growing context window and it's not the result of a "reflection" on the exploitative nature of work. This is the result of the design of the experiment

100