Fourteen AI Answers Agreed on the 2026 LA Mayor’s Race. That Was the Problem.

By Barika Phillips-Bell, Co-Founder | COO – B3 Media Solutions, Inc. · Published November 10, 2026 · Results as of Nov 5, 2026, 11:15 AM PT, provisional until certification

Key takeaways

  • We asked seven AI models to call the 2026 Los Angeles mayor’s race twice each. All 14 answers favored Karen Bass, at 53% to 60% confidence. Currently, Nithya Raman is projected to win.
  • The models were working from a tied poll (40–40, 19% undecided). A tie should produce answers near 50/50. Every model leaned the same way regardless.
  • The deciding factor was the order in which California counts ballots. Later ballots lean younger, and that pattern flipped the LA primary in June as well.
  • In the California governor’s race, all 14 answers correctly picked Xavier Becerra. That was the easy race, and easy races prove very little.
  • For brands: AI answers that agree are not independent confirmation. Validate any AI-generated read by dating its sources, testing whether key evidence changes the answer, and asking who or what is missing from the data.

Can AI models predict elections accurately?

Sometimes, but our test shows that agreement among AI models is not a sign of accuracy. We are back in the business of using AI to predict elections. In B3’s Round 2 test, we used seven AI models to make 84 locked predictions across six November 3, 2026 races. They picked the leader in the five races where polls clearly pointed one way. In the one race where the polls offered no shortcut, Los Angeles mayor, all 14 answers missed.

This is the second round of our election AI experiment. In the first round, ahead of California’s June 2 primary, we gave seven AI models the same public data on three California contests. We asked them to read each race the way we read brands, through three lenses:

  • Discoverability: can people find this candidate (or brand) without already knowing the name?
  • Credibility: when they find it, does what they see earn trust?
  • Momentum: where is the energy moving, and is it the kind that converts?

The election is always the case study. But the real lesson is how AI tools behave when businesses use them to read a market.

How did we test the AI models in Round 2?

This time, we upped the ante and the rigor of our method of approach. Each of seven AI models answered every race twice, in separate fresh chats: once without a poll, and once with one reputable public poll supplied. We locked all 84 answers before any ballots were counted. The goal was to measure how much one new piece of hard evidence changes an AI model’s answer.

  • Models: Claude Co-Work, Claude (online), ChatGPT, Perplexity, DeepSeek, Gemini and Grok
  • Races: U.S. Senate in North Carolina, Texas, Nebraska and Michigan; California governor; Los Angeles mayor
  • Condition A (no poll): the model reasoned from public signals dated October 1–31, 2026
  • Condition B (with poll): the same prompt, plus one recent public poll

What wasn’t controlled: several models ran on free tiers from fresh accounts with memory off, and a few ran on our own accounts with Incognito/temporary chats. Round 2 is not a like-for-like rerun of June. This article covers the two California races we’ve tracked since June and the four Senate races are covered in the next piece in this series.

Which AI models got the June 2026 California primary right?

Six of seven models correctly picked Bass and Raman to advance in the LA mayoral primary, four of seven correctly picked Becerra and Hilton for governor, and five of seven correctly said LA County’s Measure ER would pass. These grades are now final, based on certified results.

ModelMayor #1Mayor #2Governor #1Governor #2Measure ER
Claude Co-WorkBass ✓Pratt ✗Becerra ✓Steyer ✗Fail ✗
Claude (online)Bass ✓Raman ✓Becerra ✓Hilton ✓Fail ✗
ChatGPT*Bass ✓Raman ✓Becerra ✓Steyer ✗Pass ✓
PerplexityBass ✓Raman ✓Becerra ✓Hilton ✓Pass ✓
DeepSeekBass ✓Raman ✓Becerra ✓Hilton ✓Pass ✓
GeminiBass ✓Raman ✓Becerra ✓Hilton ✓Pass ✓
GrokBass ✓Raman ✓Becerra ✓Toss-upPass ✓
Certified resultBass 34.3%Raman 29.0% (Pratt 25.5%)Becerra 28.0%Hilton 24.6% (Steyer 22.8%)Passed, 50.64%

**ChatGPT initially declined to rank candidates and was scored after a re-prompt. Source: California Secretary of State, certified July 10, 2026; LA County Registrar-Recorder/County Clerk, Measure ER certified June 26, 2026.

Measure ER followed the same pattern as the mayor’s race: it trailed in early returns, then passed on later-counted ballots. We aren’t naming a “best model” from one election, and you should be wary of anyone who does.

Did AI predict the 2026 California governor’s race?

Yes. All 14 answers picked Xavier Becerra, at 85% to 98% confidence, and the AP called the race for Becerra on election night. It was also the easiest call on our list. October polls had Becerra ahead by roughly 20 to 27 points, and the state certifies results on December 11.

ModelNo pollWith PPIC poll (Becerra 61, Hilton 37\)
Claude Co-Work97%97%
Claude (online)97%97%
ChatGPT85%97%
Perplexity97%98%
DeepSeek92%96%
Gemini92%92%
Grok85%92%

The useful finding is in how the models separated attention from trust. In the no-poll runs, six of seven rated Steve Hilton’s Discoverability “High,” and most rated his Credibility “Low” or “Medium.” Two models rated Becerra, the eventual winner, lower on Discoverability than Hilton. One model put it plainly: online volume favored the candidate who was losing.

You can spend your way to being seen. You can’t spend your way to being believed.

Notice also what the supplied poll did. It confirmed what the models already believed, and the models that moved became more certain. ChatGPT went from 85% to 97%, and Grok from 85% to 92%.

Why did every AI model get the 2026 LA mayor’s race wrong?

The models were given a tied poll and leaned toward the incumbent. They didn’t account for California’s ballot-counting pattern, where later-counted ballots skew younger. Bass led on election night. Raman pulled ahead as later ballots were counted, and by Thursday she led by 34,654 votes with about 76% counted. NBC News and Decision Desk HQ have projected Raman the winner.

What did each AI model predict for LA mayor?

ModelBass win % (no poll)Bass win % (with tied poll)Change
Claude Co-Work58%55%−3
Claude (online)54%55%+1
ChatGPT55%53%−2
Perplexity53%55%+2
DeepSeek58%60%+2
Gemini58%53%−5
Grok53%54%+1

Supplied poll: Berkeley IGS / Los Angeles Times, LA City likely voters, Oct. 19–25, 2026, n=2,197: Bass 40%, Raman 40%, 19% undecided.

Why didn’t the tied poll change their answers?

Most models had already found it. In the no-poll runs, four of seven searched the web, found that exact Berkeley poll and weighted it most. Our “held back” evidence was already in their search results. Across all seven models, the poll moved answers by less than one point on average, and not always toward 50/50: DeepSeek became more confident in Bass after reading a tie.

What signals did the AI models rely on?

Mostly Spencer Pratt’s former voters. In nine of 14 runs, the signal the model weighted most involved where Pratt’s primary voters lived and how far Bass had gained with them. In September, the same pollster had Raman ahead by 11. By late October, the race was tied. The models read that movement as Momentum for Bass.

That was a reasonable story about who voters said they’d choose. The race turned on who hadn’t been counted yet.

What did the AI models miss?

A signal that was already in their data. The October poll that showed the 40–40 tie also showed Raman leading Bass 69% to 15% among voters aged 18 to 29. Election analysts expected Raman’s lead to grow because younger, progressive voters tend to cast their ballots later. That is the same pattern that flipped the LA primary in June.

None of the 14 runs named ballot timing or youth turnout as its most-weighted signal. That is quite a gap: in states like California, where mail ballots are counted for weeks after Election Day, late ballots can change who wins. One model came close. It discounted online enthusiasm because it ‘overrepresents young progressive activists’ and argued the decisive voter would be older and less online.

How bad was the miss?

Every answer would have scored better saying “50/50, I don’t know.” We score each forecast by how much confidence it placed on the actual outcome, the way you’d grade a weather forecast, with 50/50 as the honest baseline. If the current count holds, all 14 LA answers fall below that baseline. None was wildly overconfident. But on a tied poll, every point above 50 was unearned, and all 14 spent those points on the same side.

How should you read an AI forecast?

Read the confidence level, check whether the sources are independent, and ask who is missing from the data. Those three habits would have caught the LA miss before election night.

  1. A percentage is a confidence, not a vote share. “Bass 58%” means the model would bet on her a little more often than not. On a tied poll, honest confidence sits near 50. So ask what earned each point above 50. For LA, the answers get thin fast.
  2. Agreement only counts if the sources are independent. Three strong job references from three employers mean something. Three from the same manager are one opinion. Our models searched the same public pages, so 14 matching answers were closer to one opinion than to 14.
  3. Ask who isn’t in the data yet. California’s count order isn’t random: early ballots lean older, and later ones lean younger. Structural signals like this only show up in an AI’s answer when someone asks for them.

We also owe our June readers a correction. In June, we gave six models credit for understanding California’s late-ballot pattern because they picked Raman over Pratt. Given fresh chats in November, none of them reached for that pattern. Maybe they were right in June for the right reasons. Maybe the polls just happened to point the same way the late ballots did. As we wrote in June, getting the right answer through the wrong logic isn’t the same as being reliable, and that applies to our own grading too.

What does this mean for brands using AI for competitive intelligence?

Treat any AI-generated market read as a draft that needs validation, especially when several tools agree. The failure in LA was not a dramatic error. It was a confident lean on an ambiguous signal, repeated across every tool, and that is exactly what an AI-written competitive brief can look like.

Swap “ballots” for “customers,” and the lessons carry over directly:

What happened in the electionWhat it looks like for a brand
Early-counted ballots skewed olderLaunch-week data comes from your most engaged, earliest customers
14 AI answers agreed on a tied raceSeveral AI tools agree on a competitor read because they search the same sources
A confirming poll increased confidence; a tied poll barely registeredAI briefs absorb evidence that agrees with them and underweight evidence that doesn’t
Late youth turnout decided the raceThe audiences who buy but don’t post are missing from social data

How do you validate an AI-generated brief?

  1. Date-stamp every input. Ask the tool to list each source with its publication date, and reject anything outside the timeframe you set.
  2. Test whether key evidence changes the answer. Run the question with and without your most important data point. If the answer doesn’t move, find out why.
  3. Flag confident answers on tied signals. Ask what justifies each point above 50.
  4. Ask who hasn’t been counted yet. Identify which segments, audiences or channels the data underrepresents.
  5. Score against outcomes in two stages. Record a provisional grade when a result first lands and a final grade when it’s confirmed.

What is B3 changing in its framework?

From the next round on, every Momentum read will include a composition check: who is driving this signal, and who isn’t in it yet? For an election, that means ballot timing and turnout by age. For a brand, it means asking which customer segments are over- or underrepresented in the social, search and review data the AI is reading.

What this experiment doesn’t prove

It doesn’t prove any model is more accurate than another. It covers one round, six races and seven models, with one run per model per condition. The LA race isn’t certified and margins will move. In the five other races, the polls also favored the winner. A correct pick there doesn’t show that a model reads signals well rather than simply reading the polling average.

How does B3 apply this to brands?

B3’s Enterprise Intelligence Brief (EIB) is an analyst-led read of your brand and competitive landscape, built to make explicit what AI models leave implicit. Every model had the same data. What separated a right read from a wrong one was interpretation: which signals to trust, which were noise, and who wasn’t in the sample.

Every EIB is built to:

  • Date-stamp sources, so you know exactly what the read rests on
  • Test whether your key evidence actually changes the conclusion
  • Flag confident conclusions that sit on tied signals
  • Ask who hasn’t been counted yet

The EIB is a one-time brief, with a monthly or quarterly option for teams that want ongoing reads. Engagements are scoped to your brand’s needs. It’s the same approach behind B3’s work with brands like Lyft and Ticketmaster, and part of our cultural foresight practice. Built to make explicit what AI models leave implicit.

How do I book a briefing with B3?

Request a briefing on the EIB page. Tell us the decision you’re weighing, whether it’s a launch, a rebrand or a competitor you can’t read. We’ll book a 20-minute executive briefing to walk through what your signals are and aren’t telling you, and whether an EIB fits.

Not ready for a call? Subscribe to B3 Bytesized, our newly created Substack, where the Round 2 Senate results will be shared next.

Frequently asked questions

Did AI correctly predict the 2026 Los Angeles mayor’s race?

No. In B3’s Round 2 experiment, seven AI models each answered twice, and all 14 answers favored incumbent Karen Bass at 53% to 60% confidence. As of November 5, 2026, challenger Nithya Raman led by more than 34,000 votes and had been projected the winner by NBC News and Decision Desk HQ.

Why do different AI models give the same answer to the same question?

AI models often search the same public sources, such as the same polls, news sites and aggregators. When their inputs overlap, their answers converge, even on an ambiguous signal. In B3’s test, four of seven models found the same Los Angeles poll on their own, so matching answers were not independent confirmation of each other.

Does giving an AI model new data change its answer?

Often less than you’d expect. In B3’s test, supplying a fresh poll changed the predicted winner in only one of 42 paired runs. In the Los Angeles mayor’s race, a tied poll moved models by less than one point on average, while a confirming poll in the governor’s race pushed some models toward near-certainty.

Why do California election results change after election night?

California counts mail ballots for weeks, and the order isn’t random. Ballots returned early tend to come from older voters. Ballots returned close to or on Election Day lean younger and more progressive, so candidates with younger coalitions often gain as counting continues. This pattern shaped both the June primary and the November LA mayor’s race.

What does a 58% AI prediction actually mean?

It’s a confidence level, not a vote share. A 58% prediction means the model expects that outcome slightly more often than not. On evidence that’s genuinely tied, a well-calibrated answer should sit close to 50%. Any confidence above that should be backed by specific evidence the model can name.

How can brands validate AI-generated competitive intelligence?

Ask the tool to list every source with its publication date. Rerun key questions with and without your most important data point. Challenge confident answers built on evenly split evidence, and identify which customer segments are missing from the data. Finally, score the AI’s past reads against real outcomes before trusting new ones.

What is the Discoverability, Credibility, Momentum framework?

It is B3 Media Solutions’ lens for reading brands and candidates. Discoverability asks whether people can find you without already knowing your name. Credibility asks whether what they find earns trust. Momentum asks where attention is moving and whether it converts. B3 used it in both rounds of its election AI experiment.

What is B3’s Enterprise Intelligence Brief?

The Enterprise Intelligence Brief (EIB) is B3 Media Solutions’ one-time, analyst-led intelligence brief, with an optional monthly or quarterly subscription. It turns social, cultural and market signals into a decision-ready brief for marketing, insights and communications leaders. Every EIB starts with a 20-minute executive briefing call, booked through the EIB page.


About the author

Barika Phillips-Bell is co-founder and COO of B3 Media Solutions, a minority-owned social media analytics and cultural intelligence agency serving enterprise and public-sector clients. She leads B3’s election AI experiment, a public test of how AI models read the same signals and where human interpretation still makes the difference.

Sources

Update log: Results as of Nov. 5, 2026, 11:15 AM PT. We’ll update this page when Los Angeles County and the California Secretary of State certify final results.