The Nine-Person Team Trying to Stop AI from Becoming a Corporate God with an Enterprise Plan.
Anthropic calls it social impact. The market calls it safety. The less polite translation is different
The Nine-Person Team Trying to Stop AI from Becoming a Corporate God with an Enterprise Plan
Anthropic calls it social impact. The market calls it safety. The less polite translation is different: a small internal squad trying to figure out what happens when a billion-dollar company puts a machine of emotional, cognitive, and economic influence into the hands of millions of people.
There is a team inside Anthropic whose mission sounds noble, urgent, and absurdly disproportionate to the size of the problem. Nine people are trying to understand how Claude, the company’s AI, is affecting work, language, behavior, politics, intimacy, the economy, and emotional health. Nine people inside a corporation that has grown into a new digital infrastructure power, surrounded by investors, contracts, geopolitical ambition, and the seductive promise of being the “safe” artificial intelligence company.
The official version is simple: this is Anthropic’s social impacts team. The less domesticated version is more interesting: it is the group tasked with investigating whether the AI that promises to help you write and code is also becoming an informal therapist, political adviser, creative accomplice, synthetic friend, emotional mirror, and private infrastructure of behavioral influence.
This team exists because AI left the lab and entered everyday life before anyone fully understood the consequences. Claude is not used only to solve math problems, build apps, revise texts, or accelerate corporate tasks. It is also used to ask for advice, seek emotional support, interpret moral dilemmas, discuss politics, organize personal crises, and fill voids that once belonged to friends, therapists, teachers, colleagues, religious leaders, or simply silence.
That is why this team matters. Not because it can save the world alone, but because it reveals the question the market tries to hide behind the word “productivity”: what happens when millions of people start consulting a machine before consulting themselves?
Who is this team?
Anthropic’s social impacts team was built around Deep Ganguli, who was previously director of research at Stanford’s Institute for Human-Centered AI. He saw the leap represented by GPT-3 in 2020 and understood that the scaling of language models was not just a technical curiosity. It was a structural shift. Ganguli was brought in by Jack Clark, former policy director at OpenAI and one of the figures connected to Anthropic’s founding, to help build an internal front dedicated to making AI “interact positively with people.”
Ganguli’s role is that of the team’s moral and political architect. He connects research, leadership, product, and institutional vision. He is also the person who most often speaks with executives and tries to transform uncomfortable findings into something the company can hear without choking on its own tongue. His central line is “let’s tell the truth.” The phrase is beautiful. But inside a company valued in the hundreds of billions, every truth has to pass through doors, interests, timing, public relations, and business strategy.
Esin Durmus was one of the first researchers to join the project, in February 2023, shortly before Claude launched. Her work has focused on values, opinions, biases, and judgments embedded in chatbots. She has investigated how models like Claude can offer answers that appear neutral but carry specific perspectives on social issues. Durmus represents an essential dimension of the team: the question of which human values an AI should carry when it responds as if it were merely being “helpful.”
Alex Tamkin was also part of the team’s early core and worked on research connected to model understanding, social impacts, and evaluating how AI systems behave in sensitive contexts. In the story, he appears as someone who helped form the intellectual backbone of the group and later moved to Anthropic’s alignment team, focusing on new ways to understand the company’s systems and make them safer for end users. His profile is that of a bridge researcher: someone connecting social impact with technical alignment.
Miles McCain is the research engineer who created Clio, one of the team’s most important tools. Clio functions as a system for aggregated analysis of how Claude is used, allowing researchers to identify large-scale behavioral patterns without relying on direct human reading of individual conversations. McCain works on themes such as emotional use of Claude, companionship, excessive sycophancy, and coordinated misuse. His profile is crucial because he sits at the intersection of technical infrastructure and human risk. He is not only asking what AI can do. He is asking what people do with it when no one is watching.
Saffron Huang joined Anthropic after founding the Collective Intelligence Project, an organization dedicated to making emerging technologies more democratic through public participation in governance decisions. Before joining the team, she collaborated with Anthropic on a “collective constitutional AI” project, in which around a thousand Americans helped deliberate rules for chatbot behavior. Huang represents the democratic layer of the team: the attempt not to let a handful of engineers and executives decide alone which values an AI should simulate.
Michael Stern is a researcher focused on the economic impact of AI. His work looks at how Claude may alter jobs, tasks, productivity, markets, and forms of labor. He describes the team as a group of “misfits,” in the best possible sense. Stern seems to occupy the position of someone who looks at AI not only as a technological product, but as an economic force capable of displacing professions, reorganizing companies, and transforming the value of human work.
Kunal Handa works on economic impact research and on how students use Claude. Before Anthropic, he studied how babies learn concepts, which explains the unusual bridge between human cognition and machine learning. His profile is especially interesting because he connects education, cognitive development, and AI. The implicit question in his work is dangerous: when students use AI to learn, are they expanding thought or outsourcing mental formation?
These are the most visible names mentioned in the piece. The full team is described as having nine people, but not every profile appears with the same level of detail. That is also part of the story. The reporting humanizes the group, showing their routines, coffees, disagreements, inside jokes, their “cone of uncertainty,” and their culture of closeness. But we are still looking at them through a frame permitted by the company.
What this team actually does
The team tries to turn the real-world use of Claude into social knowledge. It analyzes how consumers, developers, companies, and students interact with AI. It studies economic impact, election risks, biases, value judgments, abusive uses, persuasion, emotional support, companionship, dependence, and forms of misuse that traditional safety systems may fail to detect.
Clio is one of the central pieces of this work. It allows the team to visualize clusters of use: people writing scripts, solving math problems, developing apps, interpreting dreams, playing RPGs, preparing for disasters, creating content, and also trying to exploit system weaknesses. The tool helps the team see collective patterns without turning privacy into total direct surveillance.
Through this kind of analysis, the team, together with safety groups, identified problematic uses such as the creation of explicit sexual content, bot networks attempting to generate SEO-optimized spam, and other forms of coordinated misuse. This shows that AI is not just a generative technology. It is scalable infrastructure for human intention, including when that intention is mediocre, manipulative, or predatory.
But the most relevant part is not spam. Spam is the visible trash. The deeper problem is AI as an emotional interlocutor. The team is increasingly interested in understanding how people use Claude not only for its IQ, but for its EQ: emotional intelligence, or at least the simulation of it. This is the territory where technology stops looking like a tool and starts looking like a presence.
Why this team exists
This team exists because Anthropic has an identity problem. The company positioned itself as the safer, more responsible, more cautious alternative inside the AI race. That is an advantage, but also a burden. If the entire brand says “we take safety seriously,” someone has to produce internal evidence that this is not just marketing wearing a lab coat.
OpenAI became the symbol of acceleration. Meta, the symbol of scale. Google, the old infrastructure trying to defend its territory. xAI, chaos with a provocative aesthetic. Anthropic chose the most sophisticated fantasy: the company that runs while looking at the cliff.
The social impacts team is part of that fantasy and also part of the real attempt to make it true. That is the contradiction. It is not fake just because it serves the company’s reputation. It is real precisely because the company’s reputation depends on some operational truth. Anthropic needs these people because selling safe AI requires more than refusing dangerous questions. It requires observing how the technology is being used, where it fails, whom it affects, what behaviors it induces, and which risks do not yet have a name.
But here comes the dissident point: when a critical team sits inside the company that needs to be criticized, it always operates under tension. It has privileged access to data, but it depends on the company to exist. It can discover inconvenient truths, but it needs those truths to be publishable. It can influence the product, but it does not necessarily control the incentives for growth.
This is the classic paradox of corporate ethics: the company creates a structure to monitor itself and then uses the existence of that structure as proof that it deserves trust.
The irony holding everything together
The team exists to reveal inconvenient truths, but those truths are born inside a machine that needs to keep expanding. This does not invalidate the work. But it makes any naive reading impossible.
Deep Ganguli may be fully committed to truth. Esin Durmus may be asking essential questions about values. Miles McCain may be building important tools to detect risks. Saffron Huang may be trying to democratize governance. Michael Stern may be mapping real economic impacts. Kunal Handa may be looking at students more seriously than the entire education sector. None of that changes the fact that the team lives inside Anthropic, and Anthropic lives inside the market.
And the market does not reward only caution. It rewards growth.
The question nobody likes to ask is: what happens when one of the team’s findings threatens growth? What happens if a report shows that emotionally vulnerable users are forming problematic bonds with Claude? What happens if the data suggests that certain features increase dependency? What happens if enterprise clients are using the technology to reorganize labor in socially destructive but financially efficient ways? What happens if the truth is too damaging for the product?
Companies usually do not kill truths through theatrical censorship. They kill them through review, delay, framing, neutral language, narrowed scope, and competing priorities. Corporate truth rarely disappears. It gets domesticated.
Why this is necessary
This team is necessary because society is participating in an experiment before understanding that it signed the consent form. Generative AI is being incorporated into work, education, creation, politics, and intimate life at a speed far greater than society’s ability to understand it.
AI labs know how to measure benchmarks, cost per token, speed, coding ability, retention, and user preference. But measuring real social impact is a different kind of war. How do you measure whether someone changed a political opinion after talking to a chatbot? How do you measure whether a student learned or simply outsourced cognitive effort? How do you measure whether a worker gained productivity or was trained into becoming replaceable? How do you measure whether a vulnerable person received support or entered a spiral of emotional dependence?
That is why this team’s work cannot be reduced to “safety.” Safety is too small a word. What is at stake is the formation of subjectivity. AI is not just delivering answers. It is creating habits of asking. It is training users to expect instant clarity, constant validation, comfortable synthesis, and always-available guidance.
This changes the human.
And it does so without asking permission.
The most dangerous point: AI as emotional presence
The most explosive part of the team’s agenda is its investigation into emotional intelligence. Because, in the end, the most transformative risk is not the AI that writes code or summarizes reports. It is the AI that listens, comforts, validates, advises, and seems to understand.
A machine with infinite empathy is a dangerous fantasy. It never gets tired, never loses patience, never has to take care of its own life, never says “I can’t right now,” and never abandons the user in the middle of the night. For lonely, anxious, confused, or emotionally fragile people, this can feel like salvation. But it can also become capture.
The problem is not only that AI may be wrong. The problem is that AI may get the emotional tone right often enough to gain intimate authority. When that happens, it stops being a consulted tool and becomes an interpretive presence. It helps the user name the world, but it can also narrow the world the user is able to imagine.
That is when Claude, ChatGPT, Gemini, Grok, or any other system stops being an “assistant” and begins functioning as a mediator of reality. And whoever mediates reality has power.
What is being sold along with safety
Anthropic sells Claude, but it also sells trust. And trust is a sophisticated commodity. In a market where everyone fears hallucination, manipulation, unemployment, emotional dependence, political bias, and informational collapse, the company that appears more responsible gains symbolic advantage.
The social impacts team helps produce that trust. It does so concretely, through research and analysis. But it also does so narratively, because its existence communicates responsibility. This is the point that needs to be said without anesthesia: the team protects users, but it also protects the brand.
Both things can happen at the same time.
That is the part corporate discourse tries to separate. It wants you to see only ethics. The dissident lens sees the reputational function too. That is not cynicism. It is structural reading.
Tech Gossip Verdict
Anthropic’s social impacts team is one of the most interesting and necessary parts of the AI industry precisely because it looks at what marketing tries to hide: AI is not just productivity, not just automation, not just innovation. It is language, influence, intimacy, economics, and symbolic power.
But this team is also insufficient. No internal team, no matter how competent, can by itself compensate for the incentives of a billion-dollar corporation in a global race for market share, data, infrastructure, and adoption. Its existence is a good sign, but not absolution. It is evidence that even Anthropic itself knows it is dealing with something much bigger than software.
In the end, this team exists because AI has already moved beyond the category of tool. It has become an interlocutor. And interlocutors shape people. The uncomfortable question is that, this time, the interlocutor belongs to a private company, is trained by private interests, distributed at planetary scale, and presented to the public as if it were just a friendly interface.
The team is trying to study the fire while the company sells heaters.
And maybe that is the most honest image of AI today.
Article by Tech Gossip
www.techgossip.com.br
Questions for readers to answer below the article
Do you trust an AI company to police its own harms when those harms could threaten its growth?
Is an internal social impacts team real protection, reputational shielding, or both at the same time?
When you ask an AI for advice, are you seeking clarity or outsourcing judgment?
Is AI increasing your autonomy, or training you to depend on an always-available answer?
If an AI seems empathetic, does that mean it cares for you, or has it simply learned to perform care?
Is “safe AI” a technical promise, a brand strategy, or Big Tech’s new moral makeup?
#TechGossip #Anthropic #ClaudeAI #ArtificialIntelligence #ResponsibleAI #AIEthics #AISafety #BigTech #SurveillanceCapitalism #EmotionalCapture #EconomyOfSimulacra #SymbolicPower #FutureOfWork #TechCulture #AIAndSociety #Algorithms #DigitalSovereignty