Generative AI Art Walls: Moderation and Safety at Scale

A single content filter is not a moderation system. We run four gates in sequence and treat each as independently fallible. Gate one is input normalisation and blocklisting, and in this market that means English, Hindi, Arabic and their transliterations. Hinglish and Arabizi defeat naive wordlists because the offensive term never appears in the script the list was written in. The blocklist has to cover Roman-script Hindi, common leetspeak substitutions and spaced-out characters.

Gate two is a classifier on the intent of the input, which catches what the blocklist misses: political content, religious provocation, requests for real people, and attempts to invoke third-party brands or protected characters. Requests naming a real person or a copyrighted character are rejected outright, and the wall says so plainly rather than silently ignoring the input. Gate three runs on the generated output before it is composited, because a clean prompt can still produce an unusable image.

Gate four is a human. Every asset passes through a hold queue on a moderator screen with an approve, reject and blur control, and nothing reaches the wall unreviewed. We run a rolling display buffer of roughly 30 seconds so the moderator is always working ahead of the wall rather than chasing it. There is a physical kill switch at the operator position that drops the wall to a safe pre-rendered loop in one press. Every decision is logged with a timestamp, the input, the output hash and the moderator ID.

The most common practical failure is that moderation and throughput are in direct conflict, and nobody priced that in. Each gate adds latency. A human hold queue adds more, and it adds a person per shift. At a high-footfall expo in Mumbai or a mall activation in Dubai you may see 400 submissions an hour, and one moderator reviewing at that rate will start approving on autopilot within 40 minutes. Either staff two moderators in rotation, or throttle ingest to a rate a human can genuinely review. Throttling feels like a compromise until you consider the alternative, which is an unreviewed image on a nine-metre wall in front of a client's customers.

The second failure is the false positive. Aggressive gates reject harmless submissions, and a guest who watched their word get silently dropped will assume the installation is broken. Tell them. A clear on-screen message that the input could not be used, with an invitation to try another word, preserves the experience and reduces the number of people who conclude the wall ignored them. The third is the assumption that the crowd is the only risk. Staff, crew and the client's own team will test the system for amusement, usually during rehearsal when the gates are still being tuned. Treat every input path as public from the moment it exists. Fourth and least discussed: the wall must degrade gracefully. Generation queues stall, GPUs throttle, networks drop. If nothing new arrives for 90 seconds, the wall should keep evolving on already-approved content rather than sitting frozen with a spinner.