The biggest thing I’ve found is limiting the inputs with a filter and vetting outputs results in higher quality results. One project I’m working on takes highly complex language and simplifies it for users. There’s no user input and it’s not being used to create anything that isn’t already there. It takes the highly technical language with lots of acronyms and breaks it down into more understandable units for normal people. Of course my company is heavily regulated so we’re extremely focused on QA and ensuring it will never output something that doesn’t align correctly.
Comment on NYC’s AI Chatbot Tells Businesses to Break the Law
pdxfed@lemmy.world 2 months agoWork in HR. Have a very smart boss. Asked me about AI for recruiting, screening and other purposes. Told my boss, wait 5 years, we’ll see the catastrophic lawsuits and early adopters, then after 5 more there will be some plug and play usable solutions.
Anyone eating up Big4 and startups own horseshit deserve what they get. They’ve fully demonstrated they don’t QC, and especially on critical, difficult to parse, contextual or changing info.
Daxtron2@startrek.website 2 months ago
p03locke@lemmy.dbzer0.com 2 months ago
LLMs are still good for the kind of flowery language you need in HR, but not for any sort of fact-based generation.
Think of it as being creative, not logical.
pdxfed@lemmy.world 2 months ago
Yeah not talking about flowery language.
HR needs automated text and audio(pretty much the same with transcription) screening/interviewing. It needs to be able to ask industry or role-specific questions, generate follow ups based on responses, and rank the answers accurately or at least better than humans do(which is miserable coin flip). 75% of interviewing could be done away with.
Right now however, the quality would be poor and the risk astronomical. I’m sure we’ll get there in 5-10 years. The risk will still be there to a degree, but just like autonomous cars, at some point it will be statistically proven that the AI is less biased and of course more efficient.
The crap part will be subscribing to updates for your AI. "Oh, you want it to ask questions about that new software? $$$