The Other Election Night Winner: Perplexity

AI startup Perplexity demonstrated strong performance in real-time during Tuesday election coverage, while rivals failed by predicting wrong outcomes before polls closed, marking the first major test of AI systems in U.S. election reporting, TechCrunch reports.

Perplexity launched an election hub featuring live maps powered by Associated Press and Democracy Works data, contrasting with major competitors like OpenAI's ChatGPT and Google's Gemini, which declined to provide election information. Despite some minor data display issues and occasional inaccuracies in state-level analysis, Perplexity's coverage largely matched traditional media outlets, potentially intensifying its ongoing legal battle with Dow Jones over audience competition.

Finally a good use for AI

[Ed Tice]

This seems like a *great* use for AI as it takes out some emotional bias. By about 7:30pm Eastern Time, I already knew (as did most viewers) that Trump was going to win the election. But the pundits kept saying "We don't know the outcome yet." That's true because anything can happen. But if county after county shows Trump running 3% better than 2020 and there's no reason to believe they are outliers, well, Trump was clearly going to win.

Part of the coverage was legitimate (Nobody wants to predict too early and look silly). Part of it was, of course, wanting to keep viewers watching for ratings. But part if it also seemed to be that the broadcasters were holding out false hope since most of the media liked Harris better. (No surprise as she did exceptionally well among educated voters) Having an AI run models keeps the emotion out of it.

As always the NYT needle was pretty accurate as well.

Re:

[timeOday]

I too like the NYT needle. Assuming the hard part (polling) has been done, I think, really, only a fairly simple statistical model is needed to factor in all the relevant information, and update it when a state is called. Then sample it to get the outcome probabilities.

Re:

[timeOday]

What I meant was, combining the available polling into a statistical model (without AI) seems a doable task. But yes, if the underlying polling isn't very good, as seems to be the case, the resulting model isn't going to be very accurate until a lot of the results are in anyways. If the polling were perfect, the needle wouldn't move all evening.

Re:

[Mitreya]

Part of it was, of course, wanting to keep viewers watching for ratings.

Bah, they have also spent weeks claiming "It's too close to call", "It's a coin toss".
I actually did not know who is going to win, but I was pretty sure it won't be as close as the polls claim.
Either they really have no idea how to poll or they spent weeks messing with the polls to make it look so 50-50.

Re:

[Ed Tice]

I don't think we can blame the polls here. The election was, as many are, largely decided on turnout. It's hard to get final numbers. But from what I can tell, there were more votes cast in 2020 than 2024. There were also many first-time voters who came out for Trump. We know this because they would come into the polling locations and then leave like a second later because the only bubble they filled out was the top of the ticket.

JD Vance was right about one thing. Much or rural America had disengag

Re:

[timeOday]

I do feel like for all the scientific polling, and talking heads for days with all the interesting theories, they sure don't seem to actually pass the test of predictive power.

It would be interesting to know if the campaigns' internally-commissioned polling is more accurate.

Intermediate results should be banned

[Okian Warrior]

Posting intermediate results (vote totals) should be banned, it opens the door to election cheating.

Note: Not saying that cheating happens or did happen, only that in a security sense posting intermediate results makes cheating easier.

Also Note: This is relevant to the article, an AI system that predicts the outcome in real time before the polls close.

If you know your candidate is losing, you can calculate by how much and arrange to have the minimum number of fake ballots delivered to push your candidate over the winning line. A smaller number of fake ballots makes it less likely that the ballots will be discovered as fake, and the "very slight margin of winning" reduces suspicion.

We already have accusations of fake ballots entered, ballot box dumps in the middle of the night, extra boxes found late in the evening... all of these give the viewer low confidence in the election outcome. We also have cases where registrations exceed the number of people living in the county, zillions of examples of registrations from non-residences and so on.

If you don't have real-time intermediate results it's much harder to gauge how much effort you need to swing the ballot count, or even if you need to cheat or whether cheating will do any good.

Don't start counting until *all* boxes from remote polling have arrived, don't accept extra boxes once you start counting, make the "percentage of votes counted" public, but don't publish the actual counts until you're done.

That rule alone would go a long way towards installing confidence in our elections.

Re:

[bobthesungeek76036]

This argument has been going on since television started broadcasting election results.

Re:

[Tora]

Ranked Choice would resolve this. You can't start counting things early, you have to do it all in one pass.

And it'd resolve a TON of other problems with our election system. Like it'd all but remove the whole "vote against a candidate you dislike" behavior we see, rather than voting FOR one you want.

But it's scary and confusing and isn't what people know, so of course we can't get it. Despite the fact it really would resolve a TON of problems.

I've never understood

[SuiteSisterMary]

I've never understood the whole 'we, a random news outlet with incomplete information, are now calling X for Y' thing.

Let the votes be counted.

Re: I've never understood

[LindleyF]

Statistics. The math says the outcome is sufficiently likely given the data.

Re:

[_xeno_]

It's not that hard to understand. In an election that isn't particularly close, eventually the vote gap between first and second becomes insurmountable, even if every remaining vote is for the second place candidate. At that point, you can call it without counting the remaining votes.

Of course, there are points before that where in order for an election result to flip, the remaining votes would have to skew so far for the second place candidate to be outside the realm of statistical possibility. It's that p

Re:

[SethJohnson]

If you look closely at the numbers behind their decisions, they're examining the number of votes that have already been counted, and the percentage of the split per that total. Let's say state X has a historical tendency for rural voters to vote by 60% for party R and metro voters to support party B by 60%. The rural polls will have generally completed their counting earlier than the population centers, so that number of votes is already split approximately 60/40 between the two candidates there. Maybe it s

Re:

[LordAba]

I've never understood the whole 'we, a random news outlet with incomplete information, are now calling X for Y' thing.

Let the votes be counted.

It's the electorial college, so it can be pretty easy to call it once the candidate has enough states going for them that there is no way the other candidate can catch up. California is being... super slow to count (currently still 55% according to google) and might take weeks, so why wait for them? It'll be interesting to see how the popular vote shakes out (California certain has enough remaining votes to flip Trumps current 4.5 million lead for what its worth) but that's about it

You can make the argument