Google, OpenAI, Anthropic and co are constantly tinkering their LLMs to give certain answers to certain questions, BUT with LLMs being an unwieldy stochastic monstrosity, my bet is that touching one query or adding one extra guardrail is inevitably messing up another one.

The problem seems not reproducible by most, or not tested thoroughly (how would it answer for all US elections in history?) and seeing this on Arstechnica and resonating on Mastodon seems excessive.

That said, I don't use Google services at all - so I should not have too much clear traces in their tracking machinery.

I reproduced Google AI Mode's non-answering of who won the 2020 presidential election from the UK by using Firefox's built-in VPN to give a US location.

Google Maps shows Gulf of Mexico as Gulf of America, too.

Even Trumps GOV-AI is sure Biden won the 2020 election ...

"I am America, a government information assistant operated by the General Services Administration (GSA) and built by the National Design Studio. I search official U.S. government sources and explain them with citations."

Source: https://america.gov/chat

You’re analysing the problem of technology companies without including the dimension of political and personal power.

You’re issuing benefit of the doubt in a situation which is the technology is already being applied for authoritarian purposes, and those purposes are shaping the technology itself. This is the actual deeply problematic issue in my opinion.

"If you doubt me, look up the election records"

The bot already gave the answer and you basically repeated the question.

There's something subtle and non-evil going on here.