A student submits an essay. A newsroom receives a photo that looks almost too clean. A brand screens a batch of product images before they go live.
In each case, the same question now hangs over the work: was a machine involved? And if so, does that change whether we trust it?
China's platforms have started answering that question with software, and one of the most accessible tools comes from an unexpected place — a free detector built by Tencent.
The tool anyone can open
Tencent's Zhuque (朱雀) AI Detection Assistant lives on a public web page and checks three kinds of content: text, images, and video. You paste writing or upload a picture or clip, and it returns a probability that the material was machine-made, sometimes with the suspicious passages or regions flagged.
The interface is deliberately plain. There is no account wall for basic use, and Tencent also exposes the same text-detection model through its cloud platform, where developers get a monthly free allowance of about 500,000 tokens via an API. That makes it one of the easier detectors to fold into a workflow rather than just try once.
Zhuque is not the only such tool, but it is notable because it backs a consumer page with an enterprise API and sits inside one of the largest Chinese internet companies.
How it claims to tell human from machine
The detector does not look for a watermark or a signature stamp. According to Tencent's own explainer, it studies writing and image patterns — how predictable the wording is, how evenly sentences are built, and which model-specific word habits show up.
On the text side, the company says it covers mainstream models, including GPT, Claude, and DeepSeek, alongside several major Chinese systems. On images and video, it says it can flag output from common generators across photography, painting, and short-video styles. The public page states image detection on social-media content reaches an accuracy "above 98%," and overall text accuracy "exceeds 92%."
Those numbers are Tencent's own claims, not independent lab results, and the tool itself prints a warning that the output is only a reference and must not be the sole basis for penalties or takedowns.
Why a detector is also a business
The rise of generative AI (生成式人工智能) has created a strange mirror market: for every system that writes or draws, another arises to catch it. Zhuque is part of that second wave, aimed at uses where the stakes are practical rather than philosophical.
- Academic integrity: students and teachers screen essays for machine-written passages.
- Content moderation: platforms filter fake or low-trust AI imagery before it spreads.
- Data hygiene: model builders use detectors to strip machine-generated text out of training sets and avoid model collapse.
Tencent is not alone here, but a free, well-distributed detector from a major platform changes the default. When the check is one paste away, more people will run it.
The large model behind the cat-and-mouse
Detection is a moving target because the generators it chases keep improving. A large model (大模型) that produces more human-sounding text today can fool a detector trained on yesterday's patterns, so the detection model has to keep updating just to stay even.
Tencent says its detector is refreshed to track new generators and methods, and the public page explicitly warns that results can contain false positives and false negatives. That honesty is the useful part: the tool is framed as an aid to human judgment, not a verdict machine.
For creators, the implication is double-edged. A detector can protect your reputation by proving authenticity — or damage it with a wrong call. Either way, the burden of proof is shifting onto the content itself.
What this means for global creators
You do not have to be in China to feel this. Detection tools trained on both Western and Chinese models are now free and fast, and publishers everywhere are adopting them as a first screen.
If you publish drafts shaped with AI tools, expect them to be scanned. If you publish human work, expect occasional false flags and keep your process evidence. The safer move is to treat detection as a signal to verify, never as a final answer.
Honest limitations
The accuracy figures (above 98% for images, above 92% for text) are Tencent's own statements on its detector page, not results from an independent benchmark we re-ran, so they describe the company's claims rather than proven performance. Detection is probabilistic by nature; Tencent's own disclaimer notes false positives and false negatives are possible, and the tool should not be the sole basis for any penalty. Coverage of specific generators can lag behind the newest releases, and a clean "human" score is not proof of authorship. This article is current as of 22 October 2026 and is not legal or compliance advice.
What readers can do now
- Run your own published writing and images through Zhuque and a second detector before a high-stakes submission, so you know how they read your work.
- If you manage a platform or classroom, set a rule that a detector flag triggers human review, never automatic punishment.
- Keep lightweight provenance for important content — drafts, shoot notes, source files — because a wrong "AI-made" call is best answered with evidence, not argument.