Gemini 4 has proved to be so good at annotating our television news coverage timelines that we now have a standard workflow of uploading every timeline PNG directly to Gemini with the prompt below. However, like all AI models, Gemini 4 still hallucinates and with our news coverage timelines, those hallucinations typically take the form of Gemini altering the timeline to add bursts of coverage where events might have warranted those surges in attention, but where the selected television news channels did not actually increase their coverage. Thus, for every timeline Gemini generates for us, we go back and verify it against the source image to vet that Gemini didn't alter the timeline in any way and rerun it a second time if we see that it has been altered.
Take this timeline of coverage of the White House Ballroom:
We give that PNG image to Gemini 4 along with this prompt:
Annotate this timeline with brief explanations of the major trends. Make the annotations look like professional journalistic annotations. Annotated ALL of the sustained surges in coverage with the events that most likely drove them. Do NOT change the timeline itself at all, just add your annotations to the image.
Gemini returns the timeline below, which at first glance seems to be exactly the timeline we uploaded, just with descriptive annotations explaining the drivers of each surge. However, look closely at the end, in October 2026: Gemini added a massive surge of MSNOW coverage making it the second-highest day of MSNOW coverage of the ballroom. The problem is that this surge is entirely fabricated. In essence, Gemini expected a huge surge in MSNOW coverage due to the leaked ballroom funding documents, but this surge never happened: Gemini expected an event that never actually materialized.
To fix this, we simply ran our prompt a second time in a fresh session and this time Gemini left the timeline as-is and didn't try to "correct it" to what it expected:
