AI generates covertly racist decisions about people based on their dialect | Nature
Posted by jpluimers on 2024/10/17
LLM are just statistic text generators which depend on the texts they have been trained which and alleviating this usually makes things worse: [Wayback/Archive] AI generates covertly racist decisions about people based on their dialect | Nature
Related:
- [Wayback/Archive] LLMs produce racist output when prompted in African American English
- [Wayback/Archive] AI worse than humans at summarising information, trial finds “A test of AI for Australia’s corporate regulator found that the technology might actually make more work for people, not less.”
Of course these issues are not limited to natural language LLM: artificial computer language LLM are also full of training issues that are likely very hard to resolve. What if covert organisations succeed poisoning LLM platforms with malicious code?
Via
- [Wayback/Archive] Mary Elizabeth on X: “Are language models racist? Though their overt answers may not be, a @nature paper shows that multiple LLMs are prejudiced against African American English speakers, and that this bias can’t easily be removed. Congrats @vjhofmann @ria_kalluri @jurafsky”.
- [Wayback/Archive] Mary Elizabeth on X: “It’s accompanied by an excellent @NatureNV by @sulin_blodgett and @ZeerakTalat”
- [WaybackSave/Archive] Jeroen Wiert Pluimers @wiert@mastodon.social on X: “@meharpist @Nature @vjhofmann @ria_kalluri @jurafsky I wonder solving issues with language models of artificial languages (especially programming ones) is just as hard (or even make it worse) than for natural languages. There is already so much trust in LLM results. I am afraid that’s only because of the money loads thrown at them.”
- [Wayback/Archive] Fefes Blog: Aus der beliebten Kategorie “mit KI wird alles schlechter”, heute: Eine australische Regulierungsbehörde hat mal “KI” zum Zusammenfassen von Textinput ausprobiert.
Aus der beliebten Kategorie “mit KI wird alles schlechter”, heute: Eine australische Regulierungsbehörde hat mal “KI” zum Zusammenfassen von Textinput ausprobiert. Ergebnis:
Artificial intelligence is worse than humans in every way at summarising documents and might actually create additional work for people, a government trial of the technology has found.
Ja komm, höre ich den umgeschulten Blockchain-Bro im Hintergrund meckern, dann haben die halt nicht genug Blockc… äh… nicht genug “KI” genommen! Mehr hilft mehr!!
Amazon conducted the test
Siehste! Die Amazonier haben bestimmt bloß ihre eigenen Schrott-“KI”s genommen!1!!
The test involved testing generative AI models before selecting one to ingest five submissions from a parliamentary inquiry into audit and consultancy firms. The most promising model, Meta’s open source model Llama2-70B, was prompted to summarise the submissions with a focus on ASIC mentions, recommendations, references to more regulation, and to include the page references and context.
Oh. Nee. Sie haben nicht ihre eigene minderwertige “KI” genommen, sie haben fünf Modelle gegeneinander getestet. Achtung: ASIC ist der Name der Behörde, das hat nichts mit Schaltkreisen zu tun.
Ten ASIC staff, of varying levels of seniority, were also given the same task with similar prompts. Then, a group of reviewers blindly assessed the summaries produced by both humans and AI for coherency, length, ASIC references, regulation references and for identifying recommendations. They were unaware that this exercise involved AI at all.
Das ist schonmal ein guter Versuchsaufbau. Ergebnis:
These reviewers overwhelmingly found that the human summaries beat out their AI competitors on every criteria and on every submission, scoring an 81% on an internal rubric compared with the machine’s 47%.
Lacher am Rande: Die “KI”-Zusammenfassungen waren so scheiße, dass drei der fünf Prüfer Vermutungen äußerten, dass sie hier gerade “KI”-Müll lesen.
This was just after AI generated text where a German journalist was accused of committing the very bad things he had been reporting about. For LLM, text is just statistics without meaning. LLM generated text is just imagination without meaning.
- [Wayback/Archive] Microsoft Bing Copilot blames reporter for crimes he covered • The Register
- [Wayback/Archive] KI-Chat macht Tübinger Journalisten zum Kinderschänder – SWR Aktuell
- [Wayback/Archive] „Der Totmacher“ – Warte, warte nur ein Weilchen… – Cul Tu Re
- [Wayback/Archive] Home – Cul Tu Re
--jeroen
The Wiert Corner - irregular stream of stuff 
Leave a comment