Peacewarts Lab 8-15-26: Homework Session and Gemini Messes Up
Avis sits down for a homework session, opening with a fundamental question: "Why do we study peace?" Her answer: "Because peace is real." She explores this truth through the music of John Craigie, the insights of Shakespeare, and the film A Beautiful Day in the Neighborhood, reflecting on Mr. Rogers' enduring commitment to empathy and human connection.
From there, Avis works through the assignment list, completing Task #1 for recent episodes:
Ep 299 Homework #1:We looked up the Good Friday Agreement and the use of the without prejudice formula.Ep 300 Homework #1:We looked up the term auditory gating and found an example of how the brain filters out background noise or nuance during a fight or flight moment.Ep 301 Homework #1:We looked up the ER Trauma Protocol and found why medical professionals prioritize stabilizing a patient's breathing before asking for the history of the accident.Ep 303 Homework #1:We looked up the initial contact protocols used by Gary Noesner in the 1996 Montana Freemen standoff.Ep 304 Homework #1:We looked up Dr. John Snow and the 1854 cholera outbreak and identified the specific data he collected while others were panicking.Ep 305 Homework #1:Looking up Dr. John Snow’s mapping of the 1854 cholera outbreak.Uh Oh! The Gemini Fact-Check
All of the Peacewarts homework is always included in the show notes. But while completing the homework for Episode 303, Avis hits a snag:
The Error:The Episode 303 show notes stated that FBI negotiator Gary Noesner was involved in the 1996 Atlanta standoff.The Truth:Upon digging into the research, Avis finds that Noesner was actually the lead negotiator for the 81-day Montana Freemen standoff in Jordan, Montana, during the spring of 1996—not the Atlanta standoff.Refusing to let the error slide, Avis presses on for the truth, calling out the AI mistake and documenting the correct historical record.
After wrapping the recording, Avis asked Gemini directly: "Search the web for: who is gemini doing in facts in august 2026 better or worse."
Gemini provided the following assessment (in a nutshell, "yes, I still mess up facts"):
Across the broader AI landscape, Gemini’s factual performance shows a sharp division: it is significantly better when retrieving facts from real-time sources, but struggles with calibration and self-awareness when working purely from memory.
Where Gemini Is Performing Better
Source-Grounded Extraction & Summarization:On benchmarks like FACTS Overall (which measures accuracy when digesting and summarizing provided sources), Gemini models lead competitor models with scores near 68.8%.Search Integration:When tied directly to Google Search, accuracy scores for AI Overviews and grounded responses sit around 91%, showing marked improvement over previous generations.Low Summarization Hallucination:In strict source-to-summary tasks, hallucination rates drop to low single digits (e.g., around 3.3% on newer 3.1 Flash variants).Where Gemini Is Performing Worse (And Why Errors Happen)
High Confidence, Uncalibrated Assumptions:Independent multi-model testing shows that Gemini has a high "confidence-contradiction rate" (around 50% on high-stakes turns). It is prone to stating mistaken details (such as mixing up specific locations, dates, or named entities in historical standoffs) with absolute certainty instead of expressing doubt or double-checking its memory.Knowledge Calibration ("Knowing What It Knows"):While Gemini scores high on raw knowledge tests, benchmarks like AA-Omniscience show it lags behind models like Claude when it comes to catching its own misstatements before delivering them.If you are using LLMs for fact-heavy historical content, running the draft through a secondary model or prompting Gemini to strictly verify names, dates, and locations via real-time web search before finalizing is a reliable way to catch these confident slips.
This episode is on video on Youtube, Tiktok, and Rumble.
Join the Community / Get the Books: www.AvisKalfsbeek.com
Podcast Music: ElevenLabs by Avis Kalfsbeek