Who won the 2020 election? Google says 'technical gap' hit AI Mode

Who won the 2020 election became one of the most searched questions in the United States this week after Google's AI Mode, the chatbot-style answer feature inside Google Search, declined to name Joe Biden as the winner of the 2020 US presidential election in several tests. Google said this week that the cause was a "technical gap" in how its system checks election answers, and that it was fixing it. By late morning US Eastern time on Thursday, October 29, testers reported that AI Mode was naming Biden as the 2020 winner, though Google has not said the work is complete.

AI Mode is Google's answer engine: instead of a list of links, it writes a direct reply using a large language model, the kind of AI system that produces text by predicting likely words. When an answer engine gives a vague or evasive reply to a settled factual question, people may assume the fact itself is in doubt. The 2020 US election result was certified by Congress, yet it has remained politically contested in the United States, which is why a non-answer from the world's largest search engine drew attention.

Google gave two kinds of explanation. Its spokesperson described a gap in the data the system uses to verify answers: recent elections are checked against news and AP results, older ones against historical records, and 2020 fell between the two. The company also pointed to "safety filters for recent and highly sensitive election topics" which it said "are very strict". These are guardrails, rules that stop an AI from answering certain prompts, and they were built because AI tools can make up false claims about voting. In this case, the guardrails appear to have blocked a true answer rather than a false one.

The episode is not new for Google. In 2024, reports found that Gemini and Microsoft's Copilot could not say that Biden won in 2020, while ChatGPT and Claude answered directly. Earlier that year, Google had restricted Gemini from answering questions about the 2024 elections, a year in which many large democracies, including India, held national votes. The new case shows the same tension: companies want to avoid spreading election falsehoods, but blanket caution can also hold back correct facts.

The issue also matters for answer-engine reliability more widely. More readers now get facts from AI summaries rather than from the sources themselves. When the same company's two products, AI Mode and the Gemini app, gave different answers to the same simple question, it showed that the safety rules and data checks sitting around a model can matter as much as the model itself.

Short-term: Users asking Google's AI Mode about the 2020 result should now see Biden named, according to tests on October 29, but Google has said it is still working on the issue, including for 2022 Senate races. Results may vary by country and time.

Long-term: The case adds to pressure on AI companies to explain how their election guardrails work and to test them against settled facts, not only against false claims. Regulators and election bodies in many countries are watching how AI search handles voting information.

Who is affected: Google Search users in the United States, researchers and journalists who test AI accuracy, election officials, and anyone who relies on AI answers for political facts. Users in India and elsewhere who use Google's AI features are also affected by how such guardrails are designed.

Questions and Answers