Fanqie Novel AI Creation Tool Research

1. Introduction

1.1 Platform Background and Release History

Platform scale:

• Fanqie Novel MAU exceeds 240 million (QuestMobile, as of end of June 2025).
• Compared with competitors, Fanqie's AI features are "more systematic": covering rewriting, expansion, continuation, synopsis/outline/idea generation, and continuation at plot hold-points; by contrast, Yuewen's "Writer Assistant" embeds DeepSeek-R1/V3, which focuses on Q&A and reasoning, while Qimao's "AI Assistant" has only 3 features.

Fanqie is the only subject in this group that simultaneously holds the dual identity of "creation tool" and "distribution + review + contract platform". This dual identity makes its Harness design fundamentally different from all pure-tool platforms: the output of the AI tools must ultimately pass the platform's own review. The tool and the umpire belong to the same entity, which is both Fanqie's unique strength (rules come from a single source, with no need to guess the standards) and its unique risk (authors lack an independent third-party assessment).

1.2 Official Positioning

The official phrasing from Fanqie Novel's announcement "AI Writing Tool Feature Launch Notice":

"Fanqie's AI writing tool is an AI writing product built on large language models. During the author's process of creating a novel, it provides the author with assistance such as content reference and information lookup. Whether the author ultimately uses the content generated by the tool, and how they use it, is decided by the author themselves."

Three keywords are worth unpacking:

This positioning is consistent with its L6 governance actions (mandatory disclosure, disabling for guaranteed-income works): the platform gives you the tool, but the responsibility is yours.

1.3 Pricing

Editorial note: The official announcement did not disclose a standalone price. Both claims in the third-party articles come from self-media sources and contradict each other, so this group does not trust either and uniformly marks them [To be filled].

Given Fanqie's business model (free reading + ad monetization, 240 million MAU), the AI tool is more likely a free investment to boost author output rather than a paid item — but this is inference, not fact, and is not cited as a conclusion.

1.4 Open Form and Entry Points

General interaction: "If you're not satisfied, you can click 'Regenerate' to generate again".

Open form: usable only within the official Fanqie App / Workbench; no API, no external integration, and no third-party tool access. This is the most closed platform in this group (comparable to NovelAI), but for a different reason — NovelAI is closed for privacy, while Fanqie is closed for governance control.

2. Glossary

3. Feature Description

3.1 Official Feature List

Source: Fanqie Novel's official announcement "AI Writing Tool Feature Launch Notice" (high credibility, from the official announcement page)

Three characteristics of the feature design:

1. Controlled quantities: 3 ideas, 5 plot developments, 10 names — all give fixed candidate counts. This aligns with Sudowrite's Wormhole (5 directions): give finite options rather than unlimited generation, to lower the cost of choosing.
2. Split feature layout across the two ends: Opening-book inspiration and outline generation are desktop-only (heavier planning, suited to large screens), while writer's-block help and naming are on both ends (light operations, suited to mobile).
3. Permission tiering: Expansion and continuation are only for authors of non-guaranteed-income works — the only case in this group where capability is tiered by contract type.

3.2 Orchestration by Creation Stage

This is a clear pipeline orchestrated by creation stage (L3). Compared with Sudowrite's Story Engine 3.0 (premise → synopsis → outline → Beats → prose draft), Fanqie's pipeline is closer to the actual rhythm of Chinese web novels — with "writer's block" as the core scenario rather than "structure".

3.3 Capabilities Not Officially Provided

This section directly corresponds to this group's core assertion: Fanqie's L4 is a weak state layer that uses "current novel content" as its state source, with no explicit persistent worldview library. This means that within the Fanqie system, long-form consistency problems rest entirely on the author — the platform provides neither a Bible nor a foreshadowing ledger.

4. Platform Architecture

图 4-1｜番茄小说平台架构：生成回路与审核回路同平台闭环

数据来源：基于本文分析绘制的示意图。

4.1 Overall Architecture

[作者] --> [番茄作家助手 App / 小说编辑器] --> [AI 工具箱]
                                                    |
                                    +---------------+---------------+
                                    |                               |
                            [大语言模型]                    [当前小说正文]
                                    |                               |
                                    +---------------+---------------+
                                                    |
                                            [生成结果]
                                                    |
                                    +---------------+---------------+
                                    |                               |
                            [换一换 重生成]                 [作者采纳/修改]
                                                                    |
                                                                    v
                                                    [作品提交] --> [签约审核]
                                                                    |
                                                        [AI 低质内容识别]
                                                                    |
                                                        [AI 申报校验]

The key characteristic of the architecture: the generation loop and the review loop are closed inside the same platform. The author generates with the platform's tool → the platform reviews the content the tool generated → the platform judges based on the declaration. This closed loop is why Fanqie has the strongest L6 governance.

4.2 Context Composition

Context strategy: Fanqie adopts a "minimal sufficient context" strategy — each tool takes only the minimum information needed to complete its task. Opening-book inspiration uses "idea + category" as minimal context, and continuation uses the "current body text" as context.

The advantage of this strategy is extremely low token cost and fast responses; the drawback is complete absence of long-range information — the model cannot see foreshadowing planted 30 chapters earlier, nor the character cards.

4.3 Coupling with the Review System

5. Harness Design

5.1 L1 Context Engineering Layer

Fanqie's L1 is a typical "enough to get by" strategy: it does no long-range context engineering)Skip, only ensuring immediate relevance for the operation at hand. For a web-novel production rhythm of daily 6,000 characters with dense single-chapter scenes, this strategy is effective at the single-chapter level — writing a chapter only requires knowing what happened in the previous one.

But the cost emerges in long-form works: when a work reaches one million characters, the model still only "sees" the current paragraph.

5.2 L2 Tools and Execution Layer

The tool layer has 9 items in total (including Regenerate), second only to Sudowrite in quantity. Among them, writer's-block help is the design that best matches the real pain point of web novels — it corresponds directly to the industry term "writer's block".

Comparison: Sudowrite's Wormhole generates 5 directions, while Fanqie's writer's-block help generates 5 follow-up plot developments — the same number, the same idea, showing that "giving 5 options" is the empirically optimal solution for creative assistance.

5.3 L3 Orchestration and Control Layer

A pipeline orchestrated by creation stage:

[开书前] 开书灵感 --> 生成大纲
                          |
                          v
[写作中] +--> 卡文锦囊（卡住时）
         +--> AI 续写 / AI 扩写（需延伸）
         +--> AI 改写 / 自定义描写（需修改）
         +--> AI 查询 / AI 助手（需资料）
         +--> AI 起名（需命名）
                          |
                          v
[提交前] 是否使用 AI 申报 --> 签约审核

The most distinctive part of Fanqie's L3 is not the pipeline but permission tiering: the same tool opens different capabilities to authors of different contract types. This is an implementation that pushes the L6 governance strategy down to the L2 tool layer — governance relies not on post-hoc review but on the capability switch itself.

5.4 L4 Memory and State Layer

Fanqie's L4 is the weakest among the 8 platforms in this group. This is a judgment that must be stated honestly.

Reason analysis (inference, not official description): Fanqie's core business metrics are daily-update output and first-showcase quantity, and its AI tools are designed to "help the author write today's chapter" rather than "help the author complete the whole book". The absence of L4 is a reasonable result under this objective function, but it is a substantial shortcoming for long-form authors.

Echo of the core assertion: Fanqie happens to be the counter-example evidence for the core assertion — it achieves the strongest L6 in this group, yet the weakest L4. The result: the platform can ensure "compliant content, truthful declaration, and controlled low quality", but cannot ensure "the work does not collapse at one million characters". These two things are independent — strong governance ≠ strong long-form capability.

5.5 L5 Evaluation and Observation Layer

Signals for platform-side evaluation (source: Toutiao self-media, must not be cited as fact): it is claimed that Fanqie has internally integrated multi-dimensional text-fingerprint technology that captures signals such as abnormal word frequency, monotonous sentence structure, and flat emotional fluctuation.

These three signals deserve separate mention — they are typical statistical features of AI-generated text:

This is the same phenomenon as the AI Slope proposed by HackerNoon (output sliding toward statistical averages), seen from two angles: creators treat it as a quality problem to fight against (using the Perplexity Gate), while platforms treat it as a compliance signal to detect (using text fingerprints).

Must emphasize: the above description of text fingerprints comes from a self-media source, is marked [To be verified], and has not been publicly confirmed by Fanqie.

5.6 L6 Governance and Security Layer

Fanqie is the sample with the strongest L6 governance in this group, and the most important compliance reference for the whole group.

5.6.1 Governance Timeline

5.6.2 Three Governance Red Lines

Red line 1: Capability tiering — AI expansion and AI continuation disabled for guaranteed-income works

The AI continuation and AI expansion capabilities are only available to authors of non-guaranteed-income works (disabled for guaranteed-income works). This is the only design in this group that tiers capability by contract type, moving governance forward from "post-hoc review" to a "capability switch".

Red line 2: Mandatory AI declaration

On September 23, 2025, a "Use AI?" checkbox popped up in the Fanqie author dashboard, prompting: "If the work you upload was synthesized/generated with AI-related tools, you must check the box truthfully"; a failure to declare truthfully may prevent passing review.

Red line 3: Long-term governance of AI low-quality content

On April 7, 2025, the "Fanqie Contract Process Optimization Notice" was published, starting substantive review of contracted works to crack down on "low-quality content generated with large amounts of AI". Effect: after launch, the daily number of male-channel first-showcase works once fell below one thousand, and has long stayed stable at 1000–1500.

5.6.3 Compliance Basis

Fanqie's AI declaration action directly corresponds to Article 6, Paragraph 4 of the "Measures for Identifying AI-Generated Synthetic Content" — platforms must provide a labeling function and remind users to proactively declare.

Industry comparison:

5.6.4 Governance Conflict and Concession: The 2024 AI Training Supplementary Agreement Incident

This event was the first large-scale conflict over L6 governance in the Chinese web-novel industry; its outcome established an industry default rule: using authors' works to train AI requires separate, revocable authorization. Yuewen's subsequent independent deployment of DeepSeek-R1 is precisely an avoidance of this risk.

5.7 Six-Layer Capability Summary

Echo of the core assertion (important): Fanqie is the counter-example specimen for this group's core assertion. It proves that "strongest L6" and "strongest L4" are mutually independent — a platform can be an industry benchmark in governance while being nearly blank in long-range state management. For long-form authors, this means: when writing long-form on Fanqie, compliance is the platform's responsibility; consistency is your own.

6. Practical Case Studies

6.1 Case One: The 2024 AI Training Supplementary Agreement Incident

Background: In July 2024, Fanqie required authors to sign an "AI Training Supplementary Agreement", consenting to the platform training AI models on their novels.

Conflict: The agreement touched authors' most core asset — the data rights of their works. Its publication triggered collective condemnation and halted updates by authors.

Result:

1. Fanqie issued a statement saying it could "help authors rescind the clause";
2. Official clarification said that only "pen names" are used for AI training.

Engineering and governance interpretation:

This is the only case in this group that documents the complete "governance conflict — concession — rule establishment" process, and it has direct reference value for any AI product that collects user data for model training.

6.2 Case Two: 2025-04-07 Contract Process Optimization and First-Showcase Quantity Change

Background: Nearly a year after the AI toolbox launched, the platform faced an influx of "low-quality content generated with large amounts of AI".

Plan: On April 7, 2025, the "Fanqie Contract Process Optimization Notice" was published, starting substantive review of contracted works, with a machine judging whether content is AI-generated "low-quality content"; in May 2025 the contract mechanism was again adjusted, raising the bar for contracting.

Effect:

Interpretation: This is an observable governance-effect indicator — the decline in first-showcase quantity directly reflects the scale of low-quality content being blocked. Working back from "long-term stable at 1000–1500 works", this figure should have been significantly above 1500 before governance (the exact number was not disclosed [To be filled]).

Also note: this review is a machine judgment and carries a risk of misjudgment. For authors, this means that "writing like AI" has itself become a risk — this is the same problem as Sudowrite using Muse to counter the AI Slope, from two sides: the platform is penalizing the AI Slope, while the tool is fighting the AI Slope.

6.3 Case Three: 2025-09-23 Mandatory AI Declaration and Compliance Implementation

Background: The "Measures for Identifying AI-Generated Synthetic Content" came into effect on September 1, 2025; its Article 6, Paragraph 4 requires platforms to provide a labeling function and remind users to proactively declare.

Plan: On September 23, 2025, a "Use AI?" checkbox popped up in the Fanqie author dashboard, with the prompt text: "If the work you upload was synthesized/generated with AI-related tools, you must check the box truthfully".

Effect: A failure to declare truthfully may prevent passing review.

Compliance chain:

《人工智能生成合成内容标识办法》
        |  2025-09-01 施行
        |  第六条第四款：平台须提供标识功能并提醒用户主动声明
        v
番茄作者后台"是否使用 AI"勾选项
        |  2025-09-23 上线（法规施行后 22 天）
        v
未如实申报 --> 可能无法通过审核

This is the most important compliance specimen in this group, and a landmark event in which the Chinese web-novel industry moved from "platform self-discipline" to "legal obligation". Cross-checking: Zongheng and Qimao ahead of Fanqie had already required authors to mark their AI usage, showing that leading platforms' compliance actions preceded legal compulsion.

6.4 Case Four: Capability Tiering — Disabling Expansion and Continuation for Guaranteed-income Works

Background: Guaranteed-income and non-guaranteed works differ in contract nature — authors of guaranteed-income works have income security, and the platform bears greater commercial risk.

Plan: AI continuation and AI expansion are only available to authors of non-guaranteed-income works (disabled for guaranteed-income works).

Design interpretation:

This is one of the most refined governance designs in this group. It does not impose a one-size-fits-all "ban on AI", but instead tiers by "how much each capability contributes to the final text": inspiration and outline are aids (low risk, fully open), while continuation and expansion are ghostwriting (high risk, restricted by contract type).

This tiering approach has direct reference value for AI governance in other scenarios: governance granularity should be refined to "specific capability × specific risk level", rather than a binary "used AI / did not use AI" judgment.

7. Summary

7.1 Strengths

1. L6 governance is the strongest in this group and an industry benchmark: expansion/continuation disabled for guaranteed-income works + mandatory AI declaration + long-term governance of low-quality content + revocable training clause — four layers of governance in place.
2. Fastest compliance response: the "Measures for Identifying AI-Generated Synthetic Content" took effect on 2025-09-01, and Fanqie launched its declaration feature on 2025-09-23, just 22 days later.
3. Finest granularity of capability tiering: a two-dimensional tiering by contract type (guaranteed / non-guaranteed) × capability risk (producing body text vs. aiding).
4. High credibility of official feature descriptions: the feature list comes from the official announcement page, not third-party elaboration.
5. Highly systematic toolbox: 9 tools cover the entire creation process, compared with Qimao's "AI Assistant" with only 3 features.
6. Writer's-block help hits the core pain point of Chinese web novels (corresponding to the industry term "writer's block").
7. Tool and review belong to the same platform, rules come from a single source, and authors do not need to guess the review standards.

7.2 Limitations and Known Shortcomings

1. L4 Memory and state layer is the weakest in this group: no persistent worldview library, no foreshadowing ledger, no timeline, no versioned snapshots. Long-form consistency rests entirely on the author.
2. Pricing is completely opaque: the two third-party claims contradict each other, and neither can be trusted.
3. Review is a machine judgment, carrying a risk of misjudgment; "writing like AI" itself becomes a risk.
4. No external integration capability: limited to the official App / Workbench, with no API.
5. Major governance conflicts in history: the 2024 "AI Training Supplementary Agreement" incident led to collective halts in updates by authors.
6. Contracting bar keeps rising: tightened twice in 2025-04 and 2025-05, squeezing authors' room to survive.
7. No structure templates: compared with Sudowrite's 4 outline templates, Fanqie offers only the single "Generate outline" capability.

7.3 Applicable Boundaries

7.4 Selection Recommendations

• Must-choose scenario: you submit your work on the Fanqie platform. In this case the official AI toolbox is the only choice with the same source as the review standards, and you must use its declaration process.
• Capability combination recommendation: Fanqie's L4 is blank, so long-form authors should build their own external state layer — you can refer to the Bible/State/Timeline four-directory structure of 05-claude-fiction.md, or the project constitution of Lingxie Creation described in 07-chinese-webnovel-ai.md, maintaining worldview and foreshadowing ledger outside Fanqie.
• Compliance mandate reminder: since 2025-09-01, AI labeling is a legal obligation. When content generated by any third-party AI tool (NovelAI, Sudowrite, Claude, etc.) is published on Fanqie, you must truthfully check the "Use AI?" box. These tools themselves do not provide AIGC labeling capability; the obligation rests entirely on the author.
• Special note for authors of guaranteed-income works: AI expansion and AI continuation are disabled for guaranteed-income works. If your work is under a guaranteed-income contract, avoid using equivalent capabilities on any platform beyond the official tools, so as not to conflict with the contract terms.

Information Gap Declaration

1. The pricing of Fanqie's AI writing tool: the official announcement did not disclose a standalone price; the third-party claims of "19 CNY/month bundled with membership" and "completely free" contradict each other, neither can be trusted, and both are marked [To be filled].
2. Persistent worldview / character library: no official description found; treated as "none" and marked [To be filled].
3. The three signals of text-fingerprint technology (abnormal word frequency, monotonous sentence structure, flat emotional fluctuation): from a Toutiao self-media source, not confirmed by the official side, marked [To be verified], described only as a phenomenon and never cited as fact.
4. The "three contract opportunities" rule: found only in self-media sources, marked [To be verified].
5. The number of first-showcase works before governance: not disclosed, so the decrease cannot be computed, marked [To be filled].
6. "Four usages safe within the rule framework" (AI builds the framework and the author fills in the flesh / AI names, looks up materials, and expands inspiration / AI pre-reviews plot holes and optimizes pacing / checking "Use AI" to declare truthfully): from self-media sources, marked [To be verified], not cited as fact.
7. Third-party descriptions of "pacing tips" and "satisfaction-beat advice": the research report explicitly determined these to be elaboration, not to be cited as fact, and this article does not adopt them.
8. Whether the AI declaration is written into the AIGC metadata fields: no technical-implementation description found, marked [To be filled].
9. Fanqie's underlying AI model: the official side only states "based on large language models" without disclosing the specific model, marked [To be filled].

8. References

1. Fanqie Novel official announcement "AI Writing Tool Feature Launch Notice" — Fanqie Novel, 2024. https://fanqienovel.com/writer/zone/article/7327136545129906238
2. Haike Caijing "Fanqie Novel's AI Problem" (Sina Finance) — Haike Caijing, 2025. https://finance.sina.com.cn/search/2025-10-09/doc-infthsqh9655363.shtml
3. Sohu Culture "2024-2025 Biennial List, Preface One: The AI Era, the 'Greater Literary View', and the Gentleness of the World's Dark Side" — Sohu Culture, 2025. https://cul.sohu.com/a/1058758339_121124749
4. Zhongxin Jingwei "Creative Effort Becomes AI Nourishment? Web-novel authors 'rise up', Fanqie Novel rushes to clarify" — Zhongxin Jingwei, 2025. https://www.jwview.com/jingwei/html/07-26/601732.shtml
5. Toutiao "Fanqie's 2026 AI New Policy: Pure AI generation results in a permanent ban, but using AI this way actually earns extra points" — Toutiao, 2026. https://m.toutiao.com/article/7626281118299357747
6. Toutiao "2026 AI Tool Recommendations Suited to Fanqie Novel Authors" — Toutiao, 2026. https://www.toutiao.com/article/7654053329475043846
7. China Youth Daily "'Miaobi Tongjian' and 'Manju Assistant' Released, AI Empowers Web-novel Creation and IP Adaptation" (Tencent News) — China Youth Daily, 2025. https://new.qq.com/rain/a/20251017A08J6J00
8. Baidu Baike "Writer Assistant Miaobi Edition" — Baidu Baike. https://baike.baidu.com/item/%E4%BD%9C%E5%AE%B6%E5%8A%A9%E6%89%8B%E5%A6%99%E7%AC%94%E7%89%88/63227250
9. HackerNoon "Claude Book: A Multi-Agent Framework for Writing Novels with Claude Code" — HackerNoon. https://hackernoon.com/claude-book-a-multi-agent-framework-for-writing-novels-with-claude-code
10. Sudowrite official blog "How Sudowrite Works: Story Bible, Muse, and the Tools" — Sudowrite. https://sudowrite.com/blog/how-sudowrite-works
11. Meipian "How Beginners Start Writing Web Novels" — Meipian. https://www.meipian.cn/5gn73n9w
12. Toutiao "Notes on Novel-Writing Techniques: From Beginner to Advanced" — Toutiao. https://m.toutiao.com/article/7666250379276927531
13. Geek Park "Can AI Write 'Joy of Life'?" — Geek Park. https://so.html5.qq.com/page/real/search_news?docid=70000021_10168f9fb6122152