'Algospeak' is changing our language in real time (2022) (www.washingtonpost.com)

“Algospeak” describes the wave of euphemisms, emojis and intentional misspellings users now employ to avoid algorithmic moderation on platforms like TikTok, YouTube, Instagram and Twitch. Because distribution often depends on opaque recommendation models (TikTok’s “For You” feed rather than follower counts), creators tailor language to evade downranking: examples include calling COVID “panini” or “Backstreet Boys reunion tour,” saying someone “unalived” instead of “killed,” using the sunflower emoji for Ukraine, “accountant” for sex workers, or “le dollar bean” (a text‑to‑speech workaround) for “lesbian.” These practices grew during the pandemic and affect marginalized communities disproportionately — LGBTQ, BIPOC and women’s health creators report self‑censoring medical and identity terms to avoid demonetization or suppression. For AI/ML, algospeak exposes both limits and dynamics of content moderation: many systems still rely on keyword lists or context-poor classifiers that trigger collateral censorship and an adversarial “whack‑a‑mole” arms race as users invent new dialects. Researchers have documented variant complexity increasing over time, and creators openly reverse‑engineer filters and share banned-term lists. The phenomenon signals a need for more transparent moderation, context‑aware models, and policy interventions: aggressive, opaque filtering risks silencing important discourse while failing to deter organized abuse, so better tools and governance are essential to balance safety and free expression.