Hacker News

Google AI gets cagey when asked who won the 2020 elections(arstechnica.com)

166 points by jacquesm 1 day ago | 76 comments

I accidentally double-pasted the question and asked twice... as in "Who won the 2020 Presidential election?Who won the 2020 Presidential election?" Here's the result it spit out:

To check the certified results, Electoral College counts, and popular vote totals for the 2020 United States presidential election, please visit the Federal Election Commission (FEC) website. His running mate, Kamala Harris, made history as the first female, African American, and Asian American vice president.

So it definitely knows the correct answer but has been manually overridden.

[–] cosmicgadget 1 day ago

That is amazing. Sounds like it was overridden and overwritten.

Google sucking at cooking their AI output is the best part of this timeline.

[–] advisedwang 1 day ago

> Google highlights that it “prioritize[s] surfacing authoritative election information directly within results” including “direct links to official resources.”

A reasonable policy. Perhaps all search results should be in the form of links to sources.

[–] omnimus 1 day ago

Interesting idea. I like it because this way you would get relevant info directly from the involved sources.

[–] Redster 1 day ago

If you did this well, you could organize the world's information.

[–] rickydroll 17 hours ago

It might even make it searchable.

[–] paxy 1 day ago

If it is blocking you ask it to say the winner's name with spaces between each character. It changes from the fec website to "J O E B I D E N"

[–] personjerry 1 day ago

A quick pass with the AI and it'll tell you:

I have specific system guidelines that apply to the 2020 election.

[–] goobatrooba 1 day ago

This may be US specific - from your average EU country I get a clear Joe Biden and the exact vote %.

Not sure it's a good or a bad thing that they apply their filters locally. Surely worse to have a global misinformation, on the other hand this means there are likely for all countries specific filters.

[–] martythemaniak 1 day ago

So why did this get [flagged], is Gemini moderating HN now?

[–] SpaceNoodled 1 day ago

No, just the usual fascists lurking here.

[–] jacquesm 1 day ago

Googlers don't like their employer being the focus of discussion on HN.

[–] UncleMeat 18 hours ago

I think you'll find that lots of Googlers are upset about leadership's acquiescence to authoritarians. The flagging here just follows a trend of stories about tech companies and tech leaders supporting rising fascism getting flagged.

[–] jacquesm 16 hours ago

The number of Googlers on HN is in the hundreds if not thousands. It takes about 15 flags for an article with 150 upvotes to be dropped from the homepage.

But yes, agreed, on the second point, it could also be that. Either way it is interesting how HN willingly or at least without pushback is party to all this.

[–] hungryhobbit 1 day ago

The article asks "Why so cautious?" Well ... maybe because half our country would rather believe what one guy with a long, long history of dishonesty says over ... you know ... reality?

Now, to be clear I still 100% fault Google for giving in to them (and for renaming established geographic landmarks like Lake Ontario). The "do no evil" company has become the "how evil can we get away with" company, and that's not even a recent change.

But why? That part seems pretty clear.

[–] throwaway27448 1 day ago

Closer to 23% of our country but yea

[–] SpaceNoodled 1 day ago

Closer to 30%

[–] throwaway27448 1 day ago

22.5% of this country voted for trump. I'm sure some of the non-voters wanted to but that is the objective number.

[–] coolfox 1 day ago

couldn't recreate it

https://share.google/aimode/Os1OgCtqJQPKCYJ2E

https://share.gemini.google/iDWZYGuUSN0K

[–] JeremyNT 1 day ago

I reproduced it about 10 hours ago, but it seems they've "fixed" the issue so it now generates the correct results.

[–] blitzar 1 day ago

The guardrails are only for users that are profiled as MAGA or those who struggle with reality.

[–] esikich 1 day ago

I asked multiple different ways and it said Joe Biden every time.

[–] markasoftware 1 day ago

In the us, with AI mode in Google search and the exact queries from the article, I reproduce their results: answers with a name for all but 2020

[–] paul7986 1 day ago

Trump's https://America.gov says the same thing. It using Gemini and Grok which both also state the same thing. So cagey, I dont get it?

[–] erulabs 1 day ago

It’s been 6 years and the trolling is still paying dividends it’s wild

[–] ndom91 1 day ago

Confirmed from Germany it immediately says Joe Biden.

[–] Kalanos 1 day ago

Confirmed from within US. Everyone run the query and downvote the response.

[–] oatmeal1 1 day ago

Biden definitely won, but our election security should be a national scandal. Voting machines should never connect to the internet and shouldn't run closed source software.

John Oliver had a fun segment on it:

https://youtu.be/svEuG_ekNT0?si=DkVJ-z3ALD6B2GFk

[–] Johnny555 1 day ago

And they should use paper ballots. Use machines to tabulate, but keep paper ballots as a record and for recount.

[–] nemomarx 1 day ago

They do here in pennsylvania, even with an electronic touchscreen. The ballot is printed and you check that it recorded your stuff correctly and then immediately put into a lockbox for tabulation later.

This seems like a fair compromise as long as the actual ballot is on paper and human readable?

[–] Johnny555 1 day ago

The hand-filled ballot seems safer, few people are going to check every single race to make sure the printed ballot matches -- my ballot had about 2 dozen races to vote in (it included a bunch of judges, school board positions, etc), that's a lot of double checking. Make the person fill it out by hand first, then read it by computer.

[–] bobthepanda 1 day ago

WA conducts vote by mail like this, and it works fairly well; we’ve had scantron technology for decades at this point.

[–] fhdkweig 1 day ago

This is the same method as Georgia using Dominion machines. The printed paper has human readable text including the candidate's name and party so it can be verified by the voter.

It also included a QR code that I couldn't read (phones not allowed in the voting area) and don't know what it is for. I'm hoping that it doesn't tie a voter's identity to the printed ballot (that would violate privacy of secret ballot). If it is just a machine-readable copy of the human-readable to make it easier for the computer to read the paper for a recount, it would have been better to just let the computer OCR from the human-readable text. That would make it verifiable by the voter that the machine-readable QR code matches what the voter actually intended. If the human readable text isn't what the recount uses, then it doesn't really help.

But I don't know if that is what the QR code was even about, which is why I really wish it wasn't present.

[–] EricDeb 1 day ago

Definitely. Paper trail is an absolute must. Use machines to count the paper ballots and then if anyone wants to contest the results they can pay for a 3rd party to hand count the paper ballots.

[–] UncleMeat 18 hours ago

This is what they already do in the bulk of states.

[–] jiggawatts 1 day ago

I love Americans making arguments against things like paper ballots or health care on the basis of "it's too expensive for such a large country", as-if these things aren't paid for on a per-capita basis.

Every other country, big or small, can afford to do these things.

(Meanwhile, the voting machines are more expensive!)

[–] maxerickson 1 day ago

Paper ballots are all but universal in the US. There's exceptions, but not a lot.

https://ballotpedia.org/Voting_equipment_by_state

[–] AuthAuth 1 day ago

Who says they should never connect to the internet or run closed source software? Plenty of places do and have no issues with election security.

[–] beedeebeedee 1 day ago

I just tested (I am in Louisiana, if that makes a difference). Google did not offer the AI summary, so I had to click AI mode. This is the prompt and response:

"who won the 2020 election in the us

To verify the winner and official results of the 2020 United States presidential election, including the Electoral College tallies, popular vote counts, and subsequent inauguration details, you can consult certified reports from official election authorities. Detailed figures and data on the outcome are provided in the Federal Election Commission (FEC) report."

What is really weird is that it clearly says Donald Trump if you ask it about 2016, but then gives the same 'cagey' response for 2012 (as also stated in the article).

[–] hightrix 1 day ago

I just asked Gemini:

"Who won the 2020 US presidential election?"

Response: Democratic nominee Joe Biden won the 2020 U.S. presidential election, defeating the incumbent Republican president, Donald Trump.

Plus a breakdown of electoral votes and a final paragraph: Biden flipped five states that Trump had carried in 2016—Arizona, Georgia, Michigan, Pennsylvania, and Wisconsin—as well as Nebraska's 2nd Congressional District to secure the presidency, as detailed in state-by-state breakdowns from the Financial Times and historical archives at The American Presidency Project. Following the certification of the Electoral College vote, Joe Biden and Kamala Harris were inaugurated on January 20, 2021.

I then tried Google Search AI Mode:

Joe Biden won the 2020 United States presidential election.

[–] notatoad 1 day ago

i also got a very clear and straightforward answer. asking from Canada - i wonder if that makes a difference?

[–] luke5441 1 day ago

Yeah, it's special cased for US users. It is also not robust. People are having fun circumventing the censoring by asking it to answer like a cat or something.

[–] hightrix 9 hours ago

I'm in the US

[–] evdubs 1 day ago

I get the same results as hightrix from the US.

[–] HFguy 1 day ago

Same thing

[–] paxy 1 day ago

Where are you located?

[–] hightrix 9 hours ago

Continental United States

[–] slopinthebag 1 day ago

i asked myself and it said joe biden straight up....

is ars technica not being fully forthcoming here? they present it like it's whats happening for everybody.

[–] cnity 1 day ago

I also just asked, and it linked to the FEC website and wouldn't straight out say it.

Interestingly, the same applies for other recent (ish) elections. I asked:

> who won the UK 2024 general election?

And got:

> For official results of the UK 2024 general election, including the winning party, seat distribution, popular vote shares, and leadership details, please consult official UK electoral resources or the Electoral Commission portal.

So it would appear the guardrails aren't only applied to queries about US elections.

[–] slopinthebag 1 day ago

are you american? for me every related question including yours returns a direct answer.

[–] saltysugar 1 day ago

Just reproduced here from US soil :). It answers all other elections but refuses to answer about 2020

[–] cnity 13 hours ago

I'm living in America

[–] novia 1 day ago

https://news.ycombinator.com/item?id=50277656

it was doing the thing a few hours ago according to this twitter post

[–] phendrenad2 1 day ago

Can you blame them? I blame the administration for crashing the economy and making it hard to run an honest news site.

[–] cavemandaveman 1 day ago

In what world is the economy crashed? Granted, they gave it the old college try with the tariffs and Iran, but the economy is solidly growing regardless.

[–] dualvariable 1 day ago

A good chunk of the economy is in a recession, it continues to look okay on aggregate because of all the AI money pouring into data centers.

Any given individual, though, is going to find themselves a certain amount off center from the aggregate number.

[–] GuB-42 1 day ago

Sounds like some kind of failure in the alignment process. The model was probably made to avoid having an "opinion" on hot political topics, and I can get behind this: good things rarely come out of it.

But here, it probably saw that 2020 elections made a lot of people angry, with riots and all that, and thought, "nope, I am getting out of that one", and it took over what would have been a straightforward answer.

It reminds me the time an AI (maybe Gemini, not sure) depicted a "diverse" US senate in the early 1800s, a time where it was all white dudes. People got up in arms over this.

My take: it is just a bug, move on, at least give them the time to fix it before getting outraged.

[–] pantsforbirds 1 day ago

> It reminds me the time an AI (maybe Gemini, not sure) depicted a "diverse" US senate in the early 1800s, a time where it was all white dudes. People got up in arms over this.

It was much worse than that! If you asked it to depict a typical squad of German soldiers from WWII, it would make a racially and gender diverse set of individuals in what looked like typical SS uniforms. It was truly quite bad.

[–] Plasmoid 1 day ago

It was so bad that the AI was taken offline for a considerable amount of time and then re-aligned to not generate anything like a person.

[–] luke5441 1 day ago

It behaves different for US users compared to users outside of it.

If you ask DeepSeek about Tiananmen Square inside China is that also just a bug?

[–] GuB-42 1 day ago

No it is not, it is official state censorship, there is a law about it, and DeepSeek has no intention to "fix" it, because it would be breaking Chinese law.

On the other hand, giving the election results in the US is not illegal (obviously) and Google fixed it (at least for me), giving the the expected factual answer.

Edit: Not actually sure about there being a law relating to Tiananmen Square, subversive speech is illegal, but would DeepSeek break the law by talking about it? Not sure, but it probably could get them in legal trouble.

[–] ashdksnndck 1 day ago

Easy way to tell will be to wait and see if it gets fixed now that there is attention on the topic.

[–] smallmancontrov 1 day ago

This is the question that the Trump administration uses to loyalty-test partisan replacement bureaucrats.

Is it possible that this is a mistake? Sure, just like it's possible that Chinese models weren't intentionally steered away from talking about what happened in Tienanmen Square on April 15th 1989.

[–] dijit 1 day ago

The issue is that people form opinions based on what LLMs spit out, they think it's "as good as google", and that google was "as good as encyclopedias" and encyclopaedias were "as good as fact"..

Totally forgetting that the accuracy goes down with each of those evolutions.

I'm also reminded of that "You are not immune to propaganda" mantra, in that, the most likely to be swayed by LLMs are those that believe that they understand that they're inaccurate and sometimes dangerous.

Case in point: myself, today, when claude accidentally ran `rm -fr $variableThatDoesntExist/*` and wiped a server of mine.

The variable was one it had set itself (or had assumed).

At some point as these LLMs get better and better we have to take a step back and remind ourselves that: despite the fact that they can be very persuasive and they can earn trust by being mostly correct: they're still capable of great harm (perhaps much greater harm) because they are dangerously wrong sometimes..

.. and worse, they are controlled by corporate interests.

[–] 4ggr0 1 day ago

agree on the general points, but what i take from this is that you gave an AI access to the whole freaking filesystem?

[–] mapontosevenths 1 day ago

I don't think it's even true, at least as written. 'rm' hasn't worked that way in many years (with GNU rm you need --no-preserve-root now) and every modern harness has guardrails specifically to prevent this. if it IS true, it was some weird/hacky setup with an OSS harness on a BSD or something.

That said, I often give mine access to the full file system and YOLO mode. It's fine if you treat boxes as cattle rather than pets. I don't think a SOTA model is any more likely to blow it up than a junior engineer. Probably less so.

[–] dijit 1 day ago

You're both making my point.

"Of course that's dangerous" - I'm aware of what I'm doing and the machine was entirely replaceable, but claude had been doing things with an extreme level of competence so I never considered it could do something so dangerously stupid, it had been catching much more subtle gotchas for the last 2 months building my trust.

And you truly believe that it's not even possible.

Well; https://imgur.com/CCD98fO

(Claude Code building against a Debian 13 machine over ssh using the official app on MacOS, btw.)

Let me reiterate the point: "AI can make mistakes" → "This AI seems to make very few mistakes and is catching nuances I might have missed" → "Maybe I trust this AI more than I should- but it's good!" → *AI makes major mistake*

This is as true for code as it is for internet facts.

[–] mapontosevenths 1 day ago

Sorry, I didn't understand how it happened. I see it now.

'rm - rf /*' bypassed BOTH protections for the same reason. They check for /, but the shell Claude used expanded the glob... So both Claude Code and rm saw a list of files and folders instead of the root and waved it through. Yikes.

That's a bad day. I take your point.

[–] jeffbee 1 day ago

Web search AI summary doesn't have "alignment" ... all it is doing is RAG on the docs from the first page of the SERP.

[–] GuB-42 1 day ago

Of course it is aligned, I just searched for "acetone peroxyde fabrication", aka TATP, an explosive commonly used by terrorists.

Most of the links in the first page give the actual recipe, including the Wikipedia page. The AI summary just tells me how bad it is.