Xiaohongshu's cautious approach to AI has often been misconstrued by outsiders as being overly 'restrained'.

According to 36Kr, for a considerable period, Xiaohongshu neither shied away from technological exploration nor aggressively pursued AI productization like its contemporaries. While continuously investing in model capability development, it carefully managed AI's integration into the community ecosystem.

Many perceive this balanced strategy as 'the aesthetic of resolve unique to a company that moves at a deliberate pace'.

In my opinion, this resembles a pursuit propelled by external pressures. It only began to sprint when it realized the road ahead was nearly obstructed.

Two Years of Indecision

Xiaohongshu's stance on AI over the past two years has been marked by deep division and ambivalence.

In 2024, no groundbreaking AI products emerged, easing Mao Wenchao's anxieties related to AI and leading to a significant drop in internal reporting frequency. Insiders disclosed to the media that he believed Xiaohongshu should refocus on its core business—AI projects continued but received no special emphasis.

Compared to AI, Xiaohongshu placed greater emphasis on commercialization and e-commerce. Media speculation suggested a concern that relying on community atmosphere and a 'human touch' for operation, AI and its generated content might lack warmth and trigger natural user resistance.

Against this backdrop, the standalone AI app 'DianDian' was launched, aiming to compete with user entry points like Doubao, Qianwen, and Yuanbao. However, the strategy failed to connect, yielding lackluster outcomes.

As a standalone entry point, it couldn't compete with tech giants; on the main platform, it remained just a feature. DianDian's awkward positioning foreshadowed Xiaohongshu's AI dilemma.

Xiaohongshu's AI turning point arrived in autumn 2025. After Mao Wenchao's Silicon Valley inspection, he declared at a closed-door meeting: 'The AI technological path has converged; we can invest more.'

The so-called 'technological path convergence' refers to the industry gradually forming a universal model for AI implementation: a large model foundation supporting multimodal inputs, paired with a vertical data knowledge base and RAG, culminating in Agent-style interaction.

This shift was catalyzed by DeepSeek. Internal tests revealed that Xiaohongshu's user search volume was being siphoned off, necessitating urgent AI action.

Xiaohongshu's AI ambivalence over the years stems from three key factors.

Firstly, the leader's lukewarm commitment to AI. After ChatGPT's 2022 debut, Mao Wenchao once used an employee's phone to ask GPT: 'Will Xiaohongshu be disrupted?' At that time, he closely monitored AI developments, demanding biweekly team reports.

However, in an August 2023 internal letter, he reaffirmed Xiaohongshu's core strengths. In his view, people turned to ChatGPT for life experience questions due to a lack of such content overseas, whereas Xiaohongshu boasted massive authentic sharing—his perceived content moat, difficult for AI to undermine.

In other words, he viewed AI as a potential threat but not an immediate battle to be won.

Questions were raised, answers provided—but for two years after, Xiaohongshu continued its operations as usual.

This indecision was directly reflected in its organizational structure: AI lacked a unified home until structural adjustments placed it under Konan's technical system. However, outsiders interpreted this as 'prioritizing product innovation over technological breakthroughs'.

Secondly, path dependency—a historical burden. On one hand, its community DNA made it hesitant to act swiftly. DianDian's entry point remained buried on the search page's fourth column; 'Ask' query coverage gradually increased from 1%-2% to 3%-4%, fearing irreversible damage to the community atmosphere.

Xiaohongshu also wavered in foundation model selection—initially self-developing models, later adopting DeepSeek before abandoning it. This revealed that when model performance clashed with product experience, Xiaohongshu lacked the engineering optimization capabilities to bridge the gap, resorting to downgraded solutions. In this context, 'restraint' was actually a limitation—'unable to do better'.

On the other hand, its linguistic assets made it feel no urgency. The 250 billion searches on Xiaohongshu generated authentic content—some of the best training material in this era.

Large model giants lacked such vivid data, explaining why Doubao and Qianwen recently launched voice input methods, targeting the web's most authentic, unfiltered human expressions.

But linguistic assets are both a boon and a constraint. Could this treasure operate independently of the community? DianDian already provided an answer: as a standalone app, it ranked 186th on download charts with only 45 ratings, eventually returning to the main platform.

Xiaohongshu's AI value relies heavily on the community context—stripped away, it loses competitiveness. The term 'path dependency' was confirmed after hitting this wall.

Xiaohongshu also faces systemic contradictions. AI search rankings based on authentic discussions clash with major clients' ad placements, leading commercial teams to demand that AI not proactively generate shopping lists.

The conflict between 'authentic community' and 'commercialization' intensifies in the AI era. Without technical solutions, it relies on slow interest negotiations and rule rebuilds. This knot cannot be untied by models—only by people and time.

Thirdly, strategic wishful thinking led to insensitivity toward AI evolution. Before 2025, Xiaohongshu internally agreed that ChatGPT-like apps wouldn't replace community search. But DeepSeek shattered this 'illusion'.

By 2026, industry narratives shifted—dialogue-centric Chat products receded, while OpenClaw ('Longxia') pushed Agents to the forefront, becoming the new attention center. FOMO (fear of missing out) surged across the internet, even among conservative organizations, making AI's importance undeniable.

Xiaohongshu realized it was lagging again. Without catching up, it might lose its seat at the table.

Combined, these factors make speeding up difficult—unless top leadership fully commits to AI.

Xiaohongshu's AI: Forced to Accelerate

But now, it must run faster, propelled by three forces.

Firstly, empirical panic over search entry point diversion. Search is Xiaohongshu's lifeline—73% of monthly active users engage in searches. Despite AI tools like Da Vinci, DianDian App, and in-platform 'Ask,' efforts to defend the search entry point yielded minimal results while maintaining a cautious, small-scale rollout, fearing large-scale replacement of traditional note streams.

Secondly, Xiaohongshu's listing needs an AI narrative. In late 2025, rumors surfaced about a potential April 2026 IPO. Former employees noted, 'Without listing, Xiaohongshu needs an AI story even more.'

AI plays a dual role: a technological tool on the surface, a valuation prop underneath. If business growth falls short of expectations, AI must fill the gap. This urgency reeks of capital pressure.

Thirdly, suspicions of 'performative' organizational mobilization. In March 2026, internal calls for 'full staff transition to AI product managers' emerged, with non-technical roles learning 'vibe coding' and campus recruitment focusing solely on AI positions with top salaries.

Consider this: tech giants invest billions in foundation models, betting on the next era; Xiaohongshu's response is organizational mobilization. This contrast reveals its true fear: not spending money, but missing the table entirely.

Today, Xiaohongshu's reactive approach has split into three paths.

The overt path uses AI to strengthen community moats. 'Ask' reached 10 million DAUs, with retention up 2%-3%—but answers must come from authentic notes, following an 'embedding' logic, not 'replacement'.

The covert path seeks mere presence: the open-source model 'dots' received a muted industry response, while secret project teams concealed Xiaohongshu's identity. Self-developed models serve as entry tickets; stealth projects act as trial balloons, securing participation in the next tech shift.

The middle ground is a horse race: DianDian and 'Ask' overlap in functions, with integration plans pending—like a company without a commander, where each platoon digs its own trenches.

The overt path defends, the covert path explores, and the horse race fills gaps. A company without a clear strategy scatters in haste. Each path keeps separate books, lacking a unified ledger.

Moreover, Xiaohongshu remains constrained. Regarding computing power, internal sources say, 'Simple tasks have enough cards, but scaling up might fall short.'

This 'shortfall' directly bars Xiaohongshu from the foundation model arms race, confining it to application layers and vertical scenes for differentiation.

The talent hurdle is equally high. Xiaohongshu's CTO position has been vacant since 2020. Its engineers 'do the work of three,' offering efficiency but lacking the talent density and technical depth for AI competition.

Governance-wise, community safety always trumps AI expansion. For example, over 300,000 daily content tags and nearly 180,000 AI-faked content disposals occur, yet clear standards distinguishing 'AI-assisted creation' from 'AI-generated content' remain elusive.

Thus, Xiaohongshu's AI approach isn't about choosing to move slowly or exercising 'restraint.' Its ambivalence stems from advancing with the brakes on.

This reflects a sober AI survival strategy—a pragmatic choice for a resource-constrained, genetically unique company pushed by external pressures amid multiple constraints.

Whether Xiaohongshu avoids elimination in the AI race hinges on two variables: Will AI-generated content erode the scarcity of 'authenticity'? Can vertical scene barriers withstand tech giants' dimensionality-reducing attacks?

One thing is clear: Xiaohongshu, once a believer in 'slow,' has been pushed onto the AI fast track. Once aboard, there's no exiting.

It must now answer: When community DNA meets AI waves, is restraint a talisman or a ceiling?

References:

36Kr, 'Xiaohongshu's Four-Year AI Journey: FOMO, Hesitation, and Sudden Acceleration'

Tang Chen's Classmate, 'Will the Input Method Table Be Overturned by Doubao and Qianwen?'