100 and 0. That's the score one developer hit when using Tencent Cloud's AI recognition capabilities to build an "AI anti-fake detective": one config field filled in correctly meant 100% accuracy on AI-generated image detection; leave it blank and every image gets labeled "real." This detail tells us: the real barrier for AI detection tools isn't the algorithm — it's those config fields the docs don't make clear.
What this is
The author stitched together four capabilities from Tencent Cloud's AI Skills matrix — image recognition (ims.ImageModeration), text recognition (tms.TextModeration), video recognition (vm), and face-swap protection (faceid) — into an Agent (an AI assistant that autonomously invokes tools to complete tasks): feed it images, text, or video, and it automatically decides "human-made" vs. "AI-made." The architecture is straightforward — the Skill handles perception, WorkBuddy (Tencent Cloud's dev assistant) handles orchestration. What actually blocked him was leaving the "scenario" field blank (news, fiction, blog, commercial copy, etc.) when creating an audit policy called Biztype. Without that scenario, the system has no basis for judgment, so it just default-labels everything as real — rendering the entire detection capability useless.
Industry view
Supporters argue that AI-generated content has already filtered down to family group chats and Moments feeds, and that mandatory platform labeling plus accessible verification tools for ordinary users are must-haves; the developer's open-source project is itself an attempt at "tool democratization" worth cheering.
But we should also see three risks. First, detection will always lag generation — today it can spot SD/Midjourney output, but a Veo3 upgrade tomorrow could break it. Second, misjudgments happen — the author himself found that the same block of text returned inconsistent results across checks, meaning "AI-written" labels could be misused to wrongly flag real content. Third, the fact that "one config field" can zero out an entire capability shows vendors still have a developer-friendliness gap; the activation bar is far higher than a single line of code.
Impact on regular people
For enterprise IT: Content moderation and platform compliance can plug into off-the-shelf APIs, but understand that "integrated" doesn't mean "working" — you need someone in-house who knows how to watch the config and the false-positive rate.
For individual careers: Self-media operators, content marketers, and content moderation roles should realize AI detection itself produces false negatives and false positives; "AI-written" cannot become a label for scapegoating or blame-shifting.
For the consumer market: The "real-vs-fake gut feel" in Moments and family group chats will get harder and harder to maintain; lightweight, consumer-grade verification tools are going to be a business.