AI Sexual-Content Carve-Out (OpenAI Model Spec: age-appropriate erotica/gore vs sexual-deepfake hard line)
artifactAI & Compute
The same explicit content is banned or allowed depending only on whether the AI decides it's 'fiction'—and that line is exactly what hackers exploit.
Who they are
A 2025 OpenAI rulebook change that permits AI-generated erotica and gore in 'age-appropriate' fictional contexts while banning sexual deepfakes[1].
What they do
The engine reads it as an example of the 'exemption fork': the same act is forbidden in one frame and permitted in another.
How it works
The AI's safety behavior boils down to judging whether a prompt counts as fiction or age-appropriate; that single judgment call is the load-bearing boundary, and it's the exact seam that jailbreaks pry open. The engine notes it mirrors a legal double standard on genital cutting and points to uncensored rival tools that drop the restrictions entirely.
Why it matters
It shows that AI 'safety' can reduce to word-parsing a story's label, and the engine deliberately leaves open why—whether it's PR-driven policy curation or just brittle keyword filters misfiring.
What's new (Aug–Oct 2026)
Sept 3, 2026 — A survivor of child sexual abuse sued Elon Musk's AI company over images made by its Grok AI; the complaint was filed "last week" in a federal court in California. (theguardian.com)
The engine's record — word for word
OpenAI Model Spec (2025) reclassifies generation of erotica and gore from prohibited to PERMITTED within 'age-appropriate' fictional contexts via API/ChatGPT, while holding a hard line against sexual deepfakes and NCII. The act is forbidden in general context, permitted under the fiction/age-appropriate carve-out — the AI-rulebook instance of the Exemption Fork (Aligned-To-Whom?). Structural twin of the statutory genital-cutting asymmetry (18 U.S.C. §116; see circumcision_medicalization_19c): identical act-class, opposite verdict, the carve-out is the load-bearing boundary. Model safety behaviour reduces to parsing the fictional/age status of the prompt — the attack surface jailbreaks exploit. Deregulated mirror: uncensored_ai (industrial NCII/deepfake). Apex Superposition: (a) corporate policy teams curating 'safe' fiction for retention vs PR; (c) brittle keyword heuristics misfire. Bounded-LLM Mediation Limit acute (mediator’s owner). [Report: Aligned-To-Whom? / The Exemption Fork, May 24 2026]
[Live pass Oct 6 2026] 2026-09-03 — A child sexual abuse survivor sued Elon Musk's AI company over Grok-generated imagery; complaint filed "last week" in a California US district court. (1 source) (
theguardian.com)
Follow the trail
Walk this on the live map →