Study: AI edits can change meaning of users' drafts
Academics “examined the behaviour of mainstream large language models provided by Elon Musk’s xAI, Meta, Google, China’s Alibaba and France’s Mistral and found the introduction of bias happens even when the AI tool is instructed to preserve the original sense.”
Hypervisible 65“One of the co-authors, Prof Sandra Wachter, compared the effect of AIs introducing bias into social media posts to ‘polluting the forest’”.
Kevin Gannon (now a Moo Deng fan account)33"instructed to preserve the original sense." Ah yes; good thing LLMs were built to comprehend tone, nuance, and intent.
Lil Mr Big Boy Man22How are people not understanding that the machine cannot conceptualize? Words are just a series of squiggles, they have no ability to conceive of what those words represent as symbols.
C.V. Danes7I wouldn't want to accuse the model builders of intentionally adding bias to the code, but sometimes [most times] things are exactly as they appear.
Kai TechnoLust (he/him)4I worked at a company last year that went “AI Native” and all the meetings were recorded and summarized by LLMs and the summaries went into a “Codex” that was the “source of truth” for the company. Except it would frequently leave out “not” in summaries so the opposite information was recorded.
Djoran Keil2Wait... so... you're saying... the 'AI' is making me hate them MORE?! I mean, that kinda makes me like... oh wait, there it is!
Oisín McGann106You can't tell an LLM not to 'hallucinate', to be honest, or avoid mistakes, or stay faithful to the source material. The wrong answers are created by exactly the same process as the right ones, it can't tell them apart. It's like giving it the instruction: 'Do this, but don't do it by doing it.'
Dr. Nic, Namer of the Universe93I cannot believe how many times we need to remind people that there is no such thing as neutral technology, but especially not the current crop of LLMs which have some incredible underlying biases and agendas built into how they are constructed
eleniel88It can’t “preserve the original sense”! It doesn’t comprehend meaning!!!! AaaaaaaaAAAAAAAAA
Not (A.) I.83When the AI boosters say “AI is the future” this is what they mean. A form of homogeneous “knowledge" production that is not knowledge at all, but words in a casing of biases that privileges the powerful, regurgitated so frequently that it flattens the intricacies of embodied experience.
Joshua Foust 🪖🎮81I keep hearing from people that these models reduce bias and just about every study I’ve seen shows the opposite 🤷
Elizabeth (H.) Bonesteel (she/her)40Please inderstand LLMs don’t know what you mean when you prompt “preserve the original sense.” They are probability generators. They literally do not understand what *any* words mean.
Thomas Cheney 6shocked to discover that the assembly of words by virtue of probability rather than you know intellect, has an impact on meaning
Stacy King4Ah yes. This is really going to help reverse our pancakes/waffles reading comprehension problems. /s, just in case
Jen Wunder3Because AI cannot think. It cannot “make sense” of words, so telling it to preserve “the original sense” doesn’t work. The bias is in the base text patterns of the LLMs. It is intrinsic.
Bijan Parsia3I would very very much love to have a good tool to correct typos and things like tense wandering. I want LLMs to be good at proofreading but I have trust issues.
