After Being Named by CCTV, “Libu Libu AI” Apologizes — Can It Really Restore the Bottom Line This Time?

The nature of this incident is very clear. This is not an ordinary mishap; it is a breach of content safety.

On April 14, after being exposed by CCTV for loopholes in AI-generated content involving pornographic and sexually suggestive material, LiblibAI issued a statement publicly apologizing and announcing the launch of a special rectification campaign. The platform admitted that under certain complex prompt combinations and evasive expression scenarios, it did indeed generate non-compliant content. Fixing it only after being called out does not count as a plus; it can only be considered making up missed homework.

The measures the platform has announced this time are not insignificant. They include internal special investigations, technical fixes, blocking risky pathways, strengthening red team-blue team exercises, and upgrading the review mechanism. It has also initiated an internal accountability review and plans to continue improving management and review processes. It has listed everything that should be done, but the key is not in the statement itself — it is whether it can consistently block similar content going forward.

The problem is also very straightforward. What AI products fear most is not the occasional error, but users already knowing how to bypass the rules. The platform itself mentioned that “in certain boundary scenarios involving complex prompt combinations and evasive expressions, the platform generated content that did not comply with regulations.” This sentence actually makes it clear that the loophole did not appear out of nowhere; rather, the safety strategy failed to keep pace with the intensity of adversarial attempts.

Viewed across the industry, this kind of incident is nothing new. CCTV had previously exposed more than one AI application, which shows that the issue is not a mistake by a single platform, but that many products prioritized generation capabilities first and only later tried to make up for safety capabilities. When the order is reversed, the cost will be high. Especially for products that accept public user input, review, identification, and interception should have gone live alongside the core functions from the start.

For ordinary users, when looking at platforms like this at the current stage, the first thing they should not focus on is how flashy the generated output can be, but how stable and reliable the platform’s boundary controls are. Content safety is not an optional add-on; it is the threshold that determines whether a product can continue to exist in the long term. The platform also stated that “content safety is the bottom line of the platform.” There is nothing wrong with that sentence, but a bottom line is never something proven by words — it only counts when it is enforced time and again.

My judgment is very simple. This rectification by LiblibAI is necessary, but it is still far from passing the test. Whether it can regain trust will not depend on how quickly it apologized, but on whether it will be breached by similar problems again later. If it still has to rely on post-exposure patchwork repairs, then this safety system still has not truly found its footing.