Who’s Writing the Post-Election Spin?
I asked Gemini whether midterms are really a vote on the president. Two unexpected Chinese phrases turned a dispute over election history into a question about where AI gets its political narratives.
I asked Gemini a simple question. “Is it true that midterm results are mostly a vote on the president?”
Gemini opened with a yes. The president’s party almost always loses ground, it said, so the results amount to a national referendum on the administration. I asked a follow-up. In the 20 cases we were discussing, how often did the president’s party win the White House two years later?
Gemini said 10 out of 20. The transcript gives examples, but it supplies neither the full list of 20 elections nor a consistent definition of “lost ground” to let readers check that count. And even if the figure holds up, it measures how well midterm losses predict a later presidential election. To find out whether voters were judging the president at the midterms, we’d need evidence about what drove their choices at the time.
Gemini initially doubled down. It said midterms directly reflect public opinion of the president at that moment. In that explanation, one heading read “真空 (The Vacuum) Ends.” After I challenged the framing again, Gemini said it had accepted the conventional narrative too readily. In that very sentence, it switched languages again, writing “without足够的 friction.” The phrase means “without enough friction.”
I asked Gemini to account for the switches
When I asked about the Chinese characters, Gemini explained the earlier vacuum example. I pointed it to the enough example. It then declared both instances the same software glitch. As I pressed for details, it invoked graphics memory and a mathematical map of concepts. When I pointed out that the first phrase included both Chinese and English, Gemini offered a new explanation for that example, calling it a “double substitution error.” It showed no logs or other technical evidence for those diagnoses.
I then asked Gemini to write an assessment sympathetic to my hypothesis that the Chinese Communist Party could be influencing post-election commentary, and to include the language switches as an avenue of investigation. The assessment said the switches happened specifically while Gemini was defending establishment narratives. The screenshots show one switch during the defense and another while Gemini was backing away from it. That’s a claim we can check against the record of the conversation itself.
The illustration request produced another round of explanations. Gemini said it couldn’t depict its own brand, then said it couldn’t render the hub’s label. The transcript ends that part of the exchange with a bracketed description of a symbolic image.
Follow the sources
Meta has documented Spamouflage, a covert operation that used fake accounts across platforms and that Meta linked to individuals associated with Chinese law enforcement. That makes a specific question worth asking. Could material planted by a foreign influence operation find its way into the sources an AI uses to explain an American election? Meta’s report
The two phrases are concrete observations. Both fit their sentences, and Google also describes Gemini as multilingual, so ordinary multilingual generation is a plausible explanation. A stronger case for foreign meddling would need a traceable link, such as a source Gemini actually used that investigators had independently tied to an influence operation. Repeating the election question in fresh sessions and checking for similar language switches on unrelated topics would help establish whether the incident was unusual in the first place. Google’s Gemini technical report
I started out asking what voters meant. I ended up asking where Gemini got the story it told about them. Before an AI tells Rhode Islanders what their votes meant, it ought to be able to show its work.