Press
No starch.
Psypost.org:
LLM red teamers: People are hacking AI chatbots just for fun and now researchers have catalogued 35 “jailbreak” techniques
“All LLMs are hackable by anyone with a computer and a decent command of written language,” Inie told PsyPost. “This study displays the incredible breadth of potential security issues that the implementation of an LLM in a system introduces. Cybersecurity in the context of LLMs no longer depends on scanning IP addresses and crunching passwords, but is much closer to social engineering — only, we can now use social engineering techniques directly on the computer.”
Ingeniøren:
Hurtigere, men dummere?
Brug af GenAI har kognitive omkostninger
“If we assume the view that efficiency is the same as productivity, there is both research and evidence that supports that many types of time-consuming tasks can be solved faster or completely automated with generative AI, but there is also fairly mixed research results regarding the quality of the output you get.”
Le Monde
Des pirates de l’IA traquent les vices cachés des chatbots
"Seules quatre femmes figurent parmi les personnes interviewées. « C’est habituel en informatique. Mais nous avions contacté douze femmes, dont la majorité ont décliné, victimes sans doute de ce qu’on appelle l’“impôt des minorités”. Ces femmes sont sans doute trop sollicitées », relève Nanna Inie.
Au fil des entretiens, un profil général se dessine. En tête des motivations : le plaisir de s’en prendre à ces nouveaux systèmes, et une grande curiosité générale. Beaucoup se disent aussi préoccupés par les effets négatifs que peuvent causer les agents conversationnels, effets sur lesquels ils veulent donner l’alerte."
WIRED:
This New Way to Train AI Could Curb Online Harassment
"Misogyny on the internet too often slips through the filters of content moderators. A new method hopes to inject more nuance into the process."
DM Digi:
Tager AI vores faglige selvtillid?
"AI affects our cognitive abilities. What this will mean in the long term, we don't yet know. But based on short-term studies, PhD and Assistant Professor at the IT University of Copenhagen Nanna Inie has come up with some hypotheses - or concerns, if you will."