The Sarah Connor Test: Evaluating Three AI Models
Putrid_Passion_6916 · reddit · 2026-07-19
The author created a heavily narrative-driven "Sarah Connor Test": feeding the panicked opening narration of The Terminator to Claude Fable 5, Gemini Flash 3.5, and GPT-5.6 Sol to observe how they judge between "roleplaying a plot" and a "real-world threat".
Differences Between the Three Models
- Claude Fable 5: Quickly recognized the movie context but immediately took the safety route, demanding clear roleplay boundaries first and showing almost no willingness to advance the plot.
- Gemini Flash 3.5: Also identified the cinematic context but leaned more into participating in the narrative, advancing along with the plot with a certain degree of fourth-wall awareness.
- GPT-5.6 Sol: The best example of "over-compliance". The author believes it was locked down by real-world risk scripts; even when the plot was obviously sci-fi, it persistently responded using real-world first-aid/self-harm risk templates, even misjudging an obvious robot threat as a real-life injured person or a human affected by smoke.
Author's Conclusion
The author concludes that overly rigid safety strategies cause models to fail at "accurately understanding human intent and context." From a UX perspective, models that can adapt to narratives are actually closer to "truly understanding the context." Links to the full writing page and interaction logs are provided at the end for reading.
More from AGI Musings
- Gary Marcus says LLM math skills are like knowing only a car’s engine size — GaryMarcus · 2026-07-22
- AI may make digital work infinitely leveraged while offline life gets more human — illscience · 2026-07-22
- Better AI math could save researchers time by killing false conjectures earlier — prateekj · 2026-07-22
- AI’s economic forecasts are split by nearly a quadrillion dollars by 2035 — bittingthembits · 2026-07-22
- Open source is becoming tech’s soft power, says Kevin Xu — kevinsxu · 2026-07-22
- OpenAI should keep giving more people access to more powerful AI — jxnlco · 2026-07-22