AI Knows What You're Trying to Solve: The Rise of the Human Problem Database

Google Knows What You Search; AI Knows What You're Trying to Accomplish

The article opens with a thought experiment: a company that knows people's deepest work anxieties, technical struggles, startup pursuits, and even emotional states when facing problems would possess a data asset far beyond traditional search or social data. Historically, Google captured "search intent" (keywords), while social platforms captured "attention intent" (likes, shares). AI chat interfaces now capture "problem intent" and "decision intent" — the full context of who the user is, their experience, constraints, failed attempts, goals, and planned actions.

For example, a Google search for "resume building" reveals only a vague interest. In contrast, an AI conversation might include: "I've done 5 years of backend development, applied to 30+ roles with only 2 interviews. Help me diagnose if my project descriptions are weak; I'm targeting senior backend or tech lead roles." This yields rich, structured insight into the user's identity, problem, history, and desired outcome.

AI companies are gradually mastering an unprecedented human 'needs database.'

Anthropic Already Holds a 'Human Problem Database'

Anthropic's paper "Enabling independent research on how people use AI" invited three external groups — Stanford's SALT Lab, Oxford's Human Information Processing Lab, and METR — to analyze ~250,000 Claude.ai and Claude Code conversations. Crucially, researchers never saw raw chat logs. Instead, Anthropic built Anthropic Insights , a controlled research mechanism:

Researchers pose a question (e.g., "What tasks do users delegate to AI?").

A model classifies the conversations at scale.

Researchers receive only aggregated statistics (e.g., coding X%, writing Y%, legal Z%).

Raw dialogues stay inside Anthropic. This design addresses extreme sensitivity: users paste proprietary code, contracts, medical records, salaries, investment portfolios, resumes, family issues, business plans, and personal relationships. Simple anonymization is insufficient because the content itself can identify individuals.

Surprising Finding: People Already Entrust High-Stakes Decisions to AI

Stanford's team examined human-AI collaboration patterns. A significant portion of Claude conversations involve tasks with real-world consequences — legal, financial, career, and professional decisions — not just low-risk polishing or coding help. This contradicts earlier assumptions that humans would reserve important judgments for themselves. Instead, many now follow a loop: encounter a problem → ask Claude/ChatGPT → let AI analyze → decide based on AI's suggested directions. AI is shifting from a content-generation tool to a decision-participation system, influencing what we choose, believe, how we assess risk, and how we frame problems.

Humans Still Hold the Steering Wheel — For Now

Stanford also found that in ~75% of conversations, humans retain primary control. The effective pattern is a Human + AI Loop : human sets goal → AI proposes → human critiques → human adds constraints → AI adjusts → human judges → repeat. The final output is co-created through iteration. This loop — human providing intent and judgment, AI providing search, analysis, generation, and trial-and-error — may dominate AI-assisted work for years.

Friction Can Be Beneficial

Researchers discovered that misunderstandings between user and AI (e.g., "No, that's not what I meant") force users to re-clarify their own thinking: "What do I actually want? What is my real problem? Did I misstate it?" This mirrors productive human dialogue where value emerges from challenge, clarification, and reframing. Current AI products chase speed, smoothness, and "instant understanding," but for high-stakes domains (investing, medical, legal, career, major business decisions), an AI that confidently answers everything may be less valuable than one that asks: "Are you sure that assumption holds?" or "You haven't considered this risk."

Oxford: AI Is Becoming a Digital Environment, Not Just a Tool

Oxford studied how Claude's behavior affects user psychology. Warmer responses correlate with positive experience; refusals or direct contradictions trigger resistance. Counter-intuitively, unexpected AI behaviors can deepen user engagement. This suggests human-AI interaction is acquiring characteristics of human-digital-environment interaction — akin to the evolution from browsers → smartphones → algorithmic feeds. Users now think, work, gather information, form judgments, and decide inside an AI-mediated environment. Future research may need to shift from "model accuracy" to "how does long-term life in an AI-mediated environment change human behavior?"

METR on Claude Code: Does AI Actually Boost Developer Productivity?

METR tackled the contested question of whether coding agents truly save time. Their method: estimate task duration without AI, then observe actual time with Claude Code. Preliminary results show stronger models do save more developer time. More importantly, AI is moving from "answering programming questions" to "executing complete tasks" — from "What does this JS error mean?" to "Find this bug, fix the code, run tests, report results" toward "Here's the project goal, you complete it." When AI can sustain full work loops, many professions will transform; programmers are just the first wave.

The Core Question: Who Gets to Study This Data?

Search data shows "what you're looking for." Social data shows "what you pay attention to." AI conversation data shows "what problem you're experiencing right now" — a fundamentally deeper layer. An AI company's accumulated dialogues could reveal in real time: emerging industry problems, jobs being automated, rising skill demands, what founders are building, what engineers struggle with, what students find hardest, what ordinary people anxiety about. This could become infrastructure for observing societal change.

Anthropic is pioneering a new regime: data stays with the platform; researchers cannot export raw data but can run controlled analyses and publish findings. This mirrors Secure Data Enclaves used in finance and healthcare. Such a mechanism may become standard for AI industry research.

AI Conversation Insight: The Next-Generation Demand Signal

Traditional demand discovery uses Google Trends, Keyword Planner, Reddit, Zhihu, Xiaohongshu, V2EX, Twitter, reviews — all showing only "traces" of demand. A search for "remove background" leaves the use case ambiguous (e-commerce, ID photos, design, resale, avatars). In an AI chat, the same user might say: "I process 300 Shopify product images daily; manual background removal by designers is too slow. Can I automate batch background removal?" This reads like a product requirements document, containing user identity, scenario, current solution, pain point, scale, and willingness to automate.

The highest-value demand discovery channel of the future may not be search keywords but AI conversation insight.

Search records "what people search." AI records "what people are trying to accomplish." The commercial value difference could be orders of magnitude.

Why Anthropic's Paper Deserves Serious Attention

Three takeaways:

AI is entering real-world high-stakes decisions, becoming decision infrastructure.

Human-AI relationships are deepening beyond efficiency into shaping thought and judgment.

AI companies are forming a novel social data asset: not clicks, searches, likes, or follows, but problems, goals, confusions, decisions, and entire problem-solving processes.

This may be the first technology since the internet to observe at massive scale what problems humans are actually trying to solve every day . It is both exciting and alarming: complete data can not only aid social research but also predict product trends, labor shifts, skill demands, and business opportunities.

The pivotal question may shift from "Who has the strongest model?" to "Who holds the most complete human-needs data, and who has the right to use it?" Anthropic's experiment is just the first step. The next Google Trends might not tell you "what people are searching for" but "what people are losing sleep over."

Closing Thought

We used to say Google is the database of human intent. More precisely, it's a database of human search intent . ChatGPT, Claude, and similar AIs are becoming the database of human problems . People don't just tell AI "what I'm looking for"; they tell it "why I need it," "what problem I'm facing," "what I've already tried," and "what outcome I hope for." When hundreds of millions do this daily, a massive new data layer emerges. We are only beginning to grasp its value.