Federico Cabitza

@vdignum @EU_Commission However, I'd rather argue for an AI-decentered humanity... #comments-6392401738753728512&trk=prof-post" target="_blank" rel="nofollow noopener">linkedin.com/pulse/debunkin…

What would you do with ~800k de-identified newborn screening records + 100 biomarkers (nearly a decade, Lombardy 🇮🇹)? New open-access Scientific Data descriptor + dataset + code. Any idea for early detection & data curation? Paper/dataset: nature.com/articles/s4159…

What is the effect of pre-LLM AI on hard clinical endpoints? In 47 RCTs, CDSS in EHRs show little impact on mortality and modest benefits on morbidity. Will LLMs do better? We will see. doi.org/10.1016/j.ijme… Thanks to IRCCS Galeazzi, @unimib and all co-authors.

Very good source of clear and useful insights on creativity and IP in the era of genAI, by @simonealiprandi

@ElonMuskNews47 Yes

@David_Gunkel There are some essential readings that anyone working in AI should know. This vignette is incredibly timely.

Homework assignment: Agentic #AI evangelists and critics need to read J.C.R. Licklider and Robert W. Taylor's "The Computer as a Communication Device" (1968). internetat50.com/references/Lic…

@Allen_Schmaltz Let XAI generate reliability metadata, not medical explanations doi.org/10.1016/j.cmpb…

I would argue that interpretability research has been key to AI Safety/Security research to date. Hard-attention via model decompositions (which can be viewed as a type of mechanistic approach) led to the discovery that neural networks are hidden instance-based metric learners (i.e., the predictions can be meaningfully approximated as a weighting over the representation space of the training set). (Hard-attention can be biased with task-specific inductive biases, which is useful for semi-supervised learning and feature detection, but that the local-level predictions can change while the document-level predictions stay the same reflects the non-identifiability of the parameters, and that an additional mechanism is needed for making sense of the predictions at test-time.) This insight that neural networks can be viewed as hidden instance-based metric learners then led to methods to calibrate the predictive uncertainty. Why is that critical? Because it provides a robust approach for correctness prediction, evaluation, exogenous tool-calling, test-time search, etc....and more generally, constraining the neural network model to perform as it did during held-out evaluation prior to deployment. But I agree with the general view that wider collaboration across the interpretability, uncertainty quantification, and AI Safety/Security communities would be productive.

I just signed the Statement on Superintelligence, because the only singularity I'm worried about is the expertise's one. Let's take the future back from Big Tech. #KeepTheFutureHuman superintelligence-statement.org

@elonmusk This contrasts the second law of Cipolla about human stupidity: "The probability that a particular person is stupid is independent of any other characteristic of the person himself".

Accurate

@autreche @emollick The tool has been improved: entechne.com/metimeter

@emollick This may be because predictions are miscalibrated/non monotone when conditioning on the radiologist own assessment, making it difficult for radiologist to incorporate ai predictions? It can be fixed by human-aligned calibration here: arxiv.org/abs/2306.00074

In this study, AI was more accurate than two thirds of radiologists, yet when radiologists had AI help their diagnoses did not improve. Why? Humans ignored the AI’s advice when it conflicted with their views. A big barrier to future human-AI collaboration blueprintcdn.com/wp-content/upl…

@emollick Pretty common finding: dl.acm.org/doi/full/10.11…

@EricTopol @tberzin @TheLancet Natali, C., Marconi, L., Dias Duran, L.D. et al. AI-induced Deskilling in Medicine: A Mixed-Method Review and Research Agenda for Healthcare and Beyond. Artif Intell Rev 58, 356 (2025). doi.org/10.1007/s10462…

Will A.I. lead to de-skilling of physicians? And if so what can be done to preempt it? @tberzin and I address this in a new @TheLancet piece thelancet.com/journals/lance…

🔵Su #Play2000 "Rallentare | Festival dei ritmi sostenibili" 📲Sabato #4ottobre dalle ore 11 su Play2000.it 📌Una giornata di incontri, dibattiti e laboratori dedicata alla riflessione sui ritmi di vita contemporanei e alla promozione della #sostenibilità ambientale, sociale e umana Tutte le info 👇 tv2000.it/ufficiostampa/… @giuriggio @UcsiSocial @caritas_milano @Avvenire_Nei @Unicatt @paolo_bovio @cabitzaf @Acli_nazionali

@David_Gunkel @Inframethod @GundelGedanken @MCoeckelbergh Das Thun Ist Alles

@Inframethod @GundelGedanken As @MCoeckelbergh once described it: "Alterity is a verb" (p. 190).

This is precisely my biological-educational angle: We cannot wait for an ‘ontological proof.’ Ethical responsibility arises long before we ‘know for sure’, which may never come. Gunkel’s relational approach opens the door to entirely new ways forward.

#ResponsibleAI Talk Designing AI for Hybrid Decision-Making or... Why Assessing Appropriate Reliance Matters Speaker: @cabitzaf (Bicocca) Date: September 22, 2025 Time: 3:30pm (London) To join online, send empty email, subject “Subscribe RAI” to daniele.quercia ;)

🚨 Excited to share our article in Artificial Intelligence Review with Chiara Natali, Luca Marconi & @cabitzaf : “AI-induced Deskilling in Medicine: A Mixed-Method Review and Research Agenda for Healthcare and Beyond” 🔗 doi.org/10.1007/s10462… #AIEthics #AI #Healthcare

@David_Gunkel @mitpress This is just the beginning of the trajectory. As I see it, pedestrian status gives robots no real individual rights, mainly obligations for robot suppliers. But yes, cars now must stop for them at crosswalks, otherwise liability shifts to the driver. New suppliers' rights, then.

Since the publication of the 2018 book "Robot Rights" @mitpress this is precisely what I have argued would be the actually trajectory that would lead to robot rights. AI welfare is just getting off the ground, but robot right is already a legal reality. intimedia.id/read/south-kor…

🗣@cabitzaf, docente di interazione uomo-macchina e supporto decisionale presso @unimib, sarà ospite del festival #Rallentare per dialogare insieme sull'impatto dell'#AI nel mondo del #lavoro. 👉Scopri di più: rallentare.org/il-tempo-del-l…

Su «la Lettura» la conversazione sui nuovi ruoli dell’IA tra @cabitzaf (ingegnere informatico), @GiaccardiChiara (sociologa), @Walter4C (informatico) e Laura Turuani (psicologa) a cura di @mirovelli bit.ly/3Vx2WGy

@eugeniosantoro NotebookLM

I medici di medicina generale e i pediatri di libera scelta si prenderanno davvero carico di scrivere il profilo sanitario sintetico dei proprio pazienti? Saranno periodicamente aggiornati? E i medici dei pronto soccorso si fideranno di queste info? doctor33.it/articolo/65303…

"Prompted by Me. Generated by ChatGPT" New essay published #OpenAccess in @HMCjournal stars.library.ucf.edu/hmc/vol10/iss1… With commentary from @autumnedwards and Katrin Etzrodt.