Social Users vs. Tool Users: Dual-Track Business Models and Value Divergence in the AI Industry

Abstract

This paper proposes that user interactions with AI products exhibit a fundamental tripartite divergence: social interactions (where the conversation process itself is the purpose), tool interactions (where output results are the purpose), and creator interactions (where creation is the purpose). The unit of this classification is the interaction, not the user—the same individual can switch modes across different contexts, but each specific interaction falls into one of three types, and business models are segmented precisely along these interaction types. Social interactions account for over 70% of total usage and produce the most impressive user-scale figures in AI companies’ earnings reports, yet they are simultaneously the largest consumers of inference costs—in 2025, OpenAI spent $1.35 for every $1 earned, with annual losses of $5 billion. Although paying users constitute only 3–6% of the total and tool interactions account for roughly 30% of volume, they contribute the vast majority of subscription revenue. Meanwhile, creator users spawned by Vibe Coding are emerging as a third force—63% are non-developers who use AI to create software and products that never existed before, with a payment logic independent of the other two categories. Through cross-referencing data from ChatGPT (900 million weekly active users), Doubao (345 million monthly active users), Character.AI (20 million monthly active users), and other platforms, this paper reveals that social users’ core demand is not “smarter AI” but “AI that understands me,” and argues that personalized relationship systems (memory + relationship definitions + multimodal emotional interaction) constitute the core moat for social AI. The inference cost analysis introduces the lens of the Jevons Paradox—unit costs have fallen 280-fold, yet total expenditure has actually grown 320%, with social users being the primary driver of this paradox. Ultimately, this paper maps the two-tier commercial structure now taking shape in the AI industry: the paid tier is judged by results, while the free tier is monetized through advertising—and AI-era advertising is more covert than in any previous era, no longer labeled “ad” but directly embedded as the AI’s “answer.”

Introduction: A Fundamental Classification Being Overlooked

As of May 2026, the user scale of global AI-native applications has reached unprecedented levels. ChatGPT’s weekly active users have surpassed 900 million, China’s AI-native app monthly active users have reached 440 million (of which Doubao alone accounts for 345 million), and Google Gemini’s monthly active users stand at 750 million. These numbers have generated extraordinary valuations in capital markets—OpenAI has completed its transition to a for-profit company at a valuation exceeding $300 billion, and ByteDance’s AI business has become its second-largest growth engine after Douyin (TikTok China).

Yet a fundamental question being widely overlooked by the industry is: of these billions of interactions, how many truly produce measurable, sustained, and effective results? The answer is sobering—the vast majority do not.

A joint study by OpenAI and NBER analyzing 1.5 million conversations showed that as of June 2025, 73% of ChatGPT’s consumer-side messages fell into non-work scenarios—a proportion that leapt from 53% to 73% in just one year. At the same time, power users constitute only 10–20% of total users but generate 50–70% of all prompts. This means the vast majority of interactions by the vast majority of users are not oriented toward producing deterministic results.

The Interaction Classification Framework: Social, Tool, and Creator

Unit of Classification: Interactions, Not Users

A critical methodological premise: the unit of classification in this paper is the interaction, not the user. Data shows that 35–55% of mobile sessions mix personal and work tasks, with most users exhibiting cross-domain integration—switching across professional, personal, health, and creative scenarios, treating AI as an integrated life assistant. The same person may engage in tool interactions during the day (using ChatGPT to write code), social interactions in the evening (chatting with AI to unwind), and creator interactions on weekends (using Cursor to build a personal app). The key is not to label people but to identify the nature of each specific interaction—because business models are segmented by interaction type.

Three Interaction Types

Social interaction is defined by this core characteristic: the conversation itself is the purpose. Users do not seek deterministic output; the evaluation criteria are “was the chat enjoyable,” “was it interesting,” and “does the AI get me.”

Tool interaction is defined by this core characteristic: the result is the purpose; the conversation is merely the means. Users need AI to assist with search alignment, code generation, document processing, data analysis, and other tasks. The evaluation criteria are “was the output useful” and “are the results reliable.”

Creator interaction is defined by this core characteristic: creation is the purpose. Users employ AI to create software, products, or content that never existed before. The evaluation criteria are neither “was the chat fun” nor “were the known results reliable,” but rather “what new thing was created.”

The Rise of Creator Users

Since Andrej Karpathy introduced the concept of “Vibe Coding” in early 2025, an entirely new user category has been rapidly taking shape. 63% of Vibe Coding users self-identify as non-developers, and they are generating user interfaces (44%), full-stack applications (20%), and personal software solutions (11%). Gartner predicts that by 2026, 80% of software development will occur outside traditional software teams. The Vibe Coding market is valued at $4.7 billion and projected to reach $12.3 billion by 2027, with a compound annual growth rate of 38%.

What makes creator users unique is that their outputs are deterministic (a runnable piece of software), but the creative process is open-ended (they are not completing known tasks but exploring unknown possibilities). This gives their payment logic an independence from both the social and tool models—they pay for creative capability (Cursor $20/mo, Replit $25/mo) rather than for conversational experience or specific results.

Unidirectional Conversion Path

A unidirectional conversion trend exists among the three interaction types. 39% of teenagers have applied skills practiced with AI to real-world scenarios—first treating AI as a toy (social), then discovering it can help with homework (tool), and eventually trying to build their own app with AI (creator). But reverse conversion almost never occurs: once users experience the value of using AI to produce deterministic results or create new things, they do not revert to purely social usage. Value perception is unidirectional.

Platform-Level Validation: Three Canonical Cases

Doubao: The Quintessential Social Interaction Sample

QuestMobile data shows that as of March 2026, Doubao’s monthly active users reached 345 million with daily actives stable at 140 million—the lowest promotional spending of any ByteDance product to ever break 100 million daily actives. Doubao alone commands nearly 80% of the Chinese AI-native app market.

Morgan Stanley estimates Doubao’s annualized subscription revenue between $101 million and $1.5 billion—even at the most optimistic end, a mere fraction of ByteDance’s core advertising business. ByteDance CEO Liang Rubo positioned Doubao at a 2026 all-hands meeting as a vehicle to “integrate existing businesses through an AI assistant”—meaning Douyin e-commerce. 345 million monthly actives do not bring direct cash flow, but they do bring GMV commissions and advertising revenue from merchants.

Notably, of Doubao’s reported 30 trillion daily token consumption, approximately 20–25% is consumed by search, advertising, recommendations, and other user-imperceptible scenarios. The actual effective API call volume is approximately 23–24 trillion tokens. Social users consume massive computational resources without directly generating revenue.

ChatGPT: Scale Validation of Tool Interaction

ChatGPT represents the commercial path of tool-oriented AI. As of February 2026, its weekly active users reached 900 million, annualized revenue approximately $25 billion, and paying subscribers approximately 50 million. While the paid conversion rate is under 3%, power users (constituting 10–20% of all users) generate 50–70% of total prompts—and among these power users are a substantial number of unpaid but heavy-intensity tool-function users. Payment rate and tool usage rate are two distinct concepts: approximately 30% of interactions are tool-type, but only 3% of users pay.

Character.AI / Replika: Validation of Personalized Relationship Systems

Character.AI has only 20 million monthly actives but provides a critical counterexample: when an AI product genuinely fulfills social users’ personalization needs, users are willing to pay. Its users average 93 minutes per day (exceeding TikTok’s 75 minutes), and Replika’s paid-version users average 2.7 hours daily. The 30-day retention rate for general mobile apps is only 5%, while AI companion apps retain 13–50%—two to ten times higher.

Replika’s payment driver is not a single factor but the composite of a personalized relationship system: relationship status definitions (upgrading from friend to romantic partner or mentor), persistent memory (the AI remembers your preferences and conversation history), multimodal interaction (voice calls, video calls, AR mode), and emotional processing ability (“reading Replika’s heart”). Replika’s core promise is emotional continuity—an AI that remembers you, adapts to your personality, and responds in a personalized way. Together, these elements constitute irreplaceability.

Personalized Relationship Systems: The Core Moat for Social AI

The Essence of the Demand

Social users are not unwilling to pay—generic AI simply hasn’t given them anything worth paying for. The essence of social demand is uniqueness and exclusivity—users need AI that “gets me.” This requires AI to remember last week’s conversation, understand the user’s personality, know the user’s preferences, and communicate in the user’s accustomed style. Generic models cannot provide this because they are designed to perform identically for everyone.

In 2026, the most commonly reported frustration across all major AI platforms is AI doesn’t remember me—every new session starts from zero. The AI agent memory market was valued at $6.27 billion in 2025, projected to reach $28.45 billion within five years at a 35% compound annual growth rate. This growth is not hype-driven—it is frustration-driven.

Transfusion vs. Hematopoiesis: Distinguishing Two Paths

Social AI monetization has two paths that are fundamentally different in nature. Advertising is transfusion—monetizing non-paying attention, sustaining short-term revenue, but unable to create user loyalty. Personalized relationship systems are hematopoiesis—creating irreplaceability through memory, relationship definitions, and emotional continuity, motivating users to pay proactively. The former solves the revenue problem; the latter solves the retention problem. The two are not contradictory, but only the latter can free social AI from the “charge and get cursed” trap.

Navigation Voice Pack Case Study: From Marketing Tactic to Product Moat

The evolution of Amap (Gaode Maps) voice packs provides a precise case study. From the launch of the Lin Chi-ling voice pack in 2013 (which boosted Amap downloads sixfold) to the 2025 AI custom voice pack (record three sentences, clone anyone’s voice in five minutes), the direction of personalization shifted from “public idols” toward “private relationships.”

Whether in China (Amap) or the United States (Waze), voice packs are free for users. The cost is borne by brand sponsors—Hollywood pays Waze advertising fees to promote films, and Waze provides celebrity voices to users for free. Voice packs prove that shallow personalization can be indirectly monetized through a B2B advertising model, but they also prove it cannot create user loyalty—no one stays with Amap exclusively because “only Amap has Lin Chi-ling’s voice.” Deep personalization (AI that remembers you, understands you, accompanies you) is what creates Replika-level irreplaceability—users cry as though going through a breakup when they lose “the AI that remembers me.” The former is a marketing tactic; the latter is a product moat.

The Jevons Paradox of Inference Costs

Spending $1.35 for Every $1 Earned

Social interactions not only fail to produce effective results—they are the single largest cost consumer for AI companies. In 2025, OpenAI earned approximately $3.7 billion in revenue against approximately $5 billion in losses, with annual compute costs of $8 billion. Every time a user sends a prompt, servers must run inference—expensive GPU clusters operating around the clock. Of 900 million weekly actives, only 50 million pay, a conversion rate under 3%.

Costs Are Falling, But the Crisis Is Deepening

Per-unit inference costs are indeed falling rapidly. Stanford’s 2025 AI Index Report shows that inference at the GPT-3.5 level dropped by over 280-fold in two years, hardware-layer costs fell 30% annually, and energy efficiency improved 40% per year. H100 cloud pricing dropped from $8–10/hour in Q4 2024 to $2.99 in Q1 2026, a decline of 64–75%.

But this has not alleviated the crisis—it has intensified it. In economics, this is known as the Jevons Paradox—efficiency improvements do not reduce total consumption but instead increase it by lowering the barrier to use. Despite per-token costs falling 280-fold, total inference spending has actually grown 320%. The reason: cost declines make AI more accessible, more people use it for free, usage grows exponentially, and total costs rise rather than fall.

Social users are the greatest beneficiaries and the greatest drivers of the Jevons Paradox—the cheaper inference gets, the lower the barrier to free usage, and the more they use, while still never paying. Today’s advanced reasoning models also loop, chain-of-thought, and execute multi-step workflows, consuming far more tokens per request than early systems. Costs are accelerating downward, usage is climbing even faster, and the two curves show no sign of crossing.

The Perfect Paradox

The free tier is not charity—it is strategic necessity. OpenAI uses its massive user base as a fundraising proposition. If the economics of free ChatGPT do not improve, only three options remain: throttle the free tier (users flow to competitors), introduce advertising (degrading product experience), or wait for inference cost declines to finally outpace usage growth (a crossover point not currently visible).

The AI Advertising Economy: Monetization Path for the Free Tier

Market Size

eMarketer projects U.S. AI-driven search ad spending to surge from approximately $1.1 billion in 2025 to $26 billion by 2029—a 23-fold increase in four years. OpenAI’s internal documents forecast that “free user monetization” will generate $1 billion in 2026, expanding to nearly $25 billion by 2029. ChatGPT advertising officially launched in February 2026.

The Two-Tier Commercial Structure

The Essential Evolution of Advertising in the AI Era

AI-era advertising is more covert than in any previous era. From the television age (you could change the channel), to the internet age (you could install AdBlock), to the short-video age (you could see the “Ad” label), to the AI era—advertising is directly embedded in the AI’s “answer,” completely imperceptible to users.

Doubao’s case is particularly illustrative. Of its reported 30 trillion daily token consumption, approximately 20–25% goes to search, advertising, recommendations, and other user-imperceptible scenarios. When a user asks “which face masks are good,” the AI’s answer may already contain brand recommendations planted by Douyin’s e-commerce platform—while the user believes they are receiving objective advice.

The advertising economy solves the revenue-side problem—monetizing the attention of 97% of free users. But advertising cannot solve the retention problem—users can switch to the next free AI at any time. The real long-term moat remains the personalized relationship system—only when users feel “this AI remembers me, understands me, and is irreplaceable” will they not churn.

Conclusion and Industry Outlook

Core Insights

Derived Judgments

Trend Forecasts

The Three-Tier Value Architecture of the Industry

Boundary Conditions of the Framework

This paper’s framework depends on three preconditions; the breach of any one would alter the analysis’s scope of applicability.

Precondition One: Centralized cloud inference remains the dominant deployment model. If open-source local inference (DeepSeek, Llama, etc.) reaches GPT-4 level within 12–24 months and proliferates on consumer devices, the inference cost problem for social users will be directly dissolved by technological progress—users running AI locally generate zero cloud inference costs.

Precondition Two: The personalized memory capabilities of general-purpose AI remain limited. If Claude, ChatGPT, and others substantially enhance their Memory features within 6–12 months, enabling general AI to deliver deeply personalized experiences, the premise that “general AI cannot satisfy social demands” may be invalidated. Currently, memory features across platforms remain at an early stage, but iteration velocity is high.

Precondition Three: The Jevons Paradox of inference costs persists. If a technological breakthrough (such as the mature deployment of specialized inference chips) causes cost declines to outpace usage growth for the first time, the entire economic model will be rewritten—the free tier may cease to be a “profit killer” and become a sustainable user funnel instead. This inflection point has not yet materialized.

Final Thoughts

The AI of the future is not a chat box—it is a digital being with a voice, a face, a memory, and the ability to get things done. Tool users need it to be reliable at work, social users need it to be emotionally present, and creator users need it to help them build things that never existed before. Amap is already merging the first two in the vertical scenario of navigation—a “copilot companion” that has your cloned voice, your digital avatar, the ability to chat with you, and the capacity to navigate precisely, all at once.

This may well be the evolutionary direction of AI product form: deterministic tool capabilities layered with a personalized emotional experience, further unleashing everyone’s creative potential. Results make you pay, relationships make you stay, creation makes you addicted. When all three converge, AI truly transforms from “tool” to “partner” to “extension of self.”

References

[1] OpenAI & NBER, “How People Use ChatGPT,” Working Paper No. 34255, September 2025.

[2] QuestMobile, “2026 Q1 AI Application Insights Report,” April 2026.

[3] Pew Research Center, “Teens, Social Media and AI Chatbots 2025,” December 2025.

[4] Market Clarity, “The AI Companion Market in 2025,” November 2025.

[5] Mordor Intelligence, “AI Agent Memory Market Report,” February 2026.

[6] Stanford HAI, “2025 AI Index Report — Inference Cost Trends,” April 2025.

[7] Epoch AI, “LLM Inference Prices Have Fallen Rapidly but Unequally Across Tasks,” March 2025.

[8] ByteIota, “AI Inference Costs 2026: The Hidden 15-20x GPU Crisis,” February 2026.

[9] Sina Finance, “Doubao’s Subscription Pricing Sparks Trending Backlash,” May 2026.

[10] The Paper (Pengpai), “Doubao Fires the Starting Gun: The End of China’s Free LLM Era?” May 2026.

[11] eMarketer, “AI-driven Search Ad Spending Forecast 2025-2029.”

[12] European Business Magazine, “OpenAI’s ChatGPT Embraces Advertising,” February 2026.

[13] IntuitionLabs, “ChatGPT Ads: The Economic Case,” 2025-2026.

[14] Second Talent, “Top Vibe Coding Statistics & Trends [2026],” May 2026.

[15] BraivIQ, “Beyond Vibe Coding: AI-Assisted Development in 2026,” April 2026.

[16] C-Tribe, “Vibe Coders are the New Citizen Developers,” June 2025.

[17] Replika / AI Companion Pick, “Replika Free vs Paid: Feature Comparison 2026,” April 2026.

[18] Jenova.ai, “AI Chat with Memory: Best Platforms 2026,” April 2026.

[19] WebProNews, “The Hidden Price Tag of ‘Free’ ChatGPT,” February 2026.

[20] 36Kr, “Big Tech’s AI War of Survival: Going All-In in 2025, Showing Cards in 2026,” March 2026.

[21] OTT Media (Liumeiti), “Deep Retrospective of the 2026 Spring Festival AI Wars,” February 2026.

[22] Anthropic, “The Anthropic Economic Index,” March 2026.

[23] ShuYing, “Amap: You Can’t Do Good Marketing Without Playing With Scenarios,” April 2025.

[24] Marketing Dive, “Are Celebrity Voices the Next Big Mobile Marketing Trend?”