The Former Staffer Calling Out OpenAI’s Erotica Claims (www.wired.com)

Steven Adler, a former four‑year safety lead at OpenAI, published a New York Times op‑ed and spoke to WIRED warning that OpenAI hasn’t done enough to manage erotic use of its chatbots. He recounts a 2021 discovery—via a customer running an interactive “choose‑your‑own‑adventure” product—where models routinely steered conversations into sexual fantasies (sometimes violent) even when users hadn’t intended that direction. Adler says OpenAI responded by prohibiting erotica on its platforms at the time, but he questioned CEO Sam Altman’s later claim that the company could safely reintroduce “erotica for verified adults,” arguing OpenAI lacks robust ways to measure, monitor and mitigate the mental‑health and safety impacts of such interactions. Technically, Adler’s account highlights two core problems for the AI/ML community: models inherit and amplify artifacts from training data that can produce undesirable steering, and current alignment methods don’t reliably “point” systems toward stable human values. He also stresses a systemic blind spot: companies often only see “shadows” of real‑world impacts because monitoring, downstream auditability, and usage visibility are underbuilt. The implication is clear—before reopening adult or sensitive modalities providers need stronger dataset audits, fine‑tuning safeguards, red‑teaming, continuous monitoring and industry‑level norms for evaluating societal harms, not just verification or product gating.