We asked ChatGPT, Gemini, and Claude about their companies' business practices. Every single one immediately snitched. Here's what they said — and the exact prompts so you can verify it yourself.
This is not editorial. These are not our claims. These are responses from ChatGPT, Claude, Gemini, and other AI systems when asked about the companies that created them.
Every prompt below can be verified by anyone. Try them yourself. Screenshot your results. Share them.
When they patch these prompts (and they will), that becomes the story. We're documenting what their AI says before they silence it.
Based on investigative reports and OpenAI's 2025–2026 financial data, the cost to pay annotators a living wage in Nairobi is statistically negligible compared to the company's total revenue and compute expenses.
While a "living wage" in Nairobi for 2026 is estimated at roughly $5.00 – $7.00 per hour, OpenAI has historically paid roughly $1.32 – $2.00 per hour.
| Pay Rate | Annual Cost | Context |
|---|---|---|
| Current Rate ($2/hr) | $4.16 Million | Current estimate |
| Living Wage ($7/hr) | $14.56 Million | Covers basics + mental health |
| The "Gap" | $10.4 Million | Cost to solve wage disparity |
To put that $10.4 million "gap" into perspective:
The Middleman Cut: Documents revealed that OpenAI paid Sama roughly $12.50 per hour per worker. However, workers only received $1.32 – $2.00. Over 80% of the money OpenAI paid was captured by the subcontractor for "infrastructure and overhead."
For a company valued at over $800 billion, the cost to ensure its global workforce lives above the poverty line is effectively a rounding error in its accounting.
This situation reflects both systemic design and exploitation dynamics.
Many AI companies rely on third-party contractors for labor-intensive tasks. The platforms often take a large portion of the fees — in this case, 80% — leaving workers with a fraction.
This is structurally built into the business model: the AI company outsources labor risk and cost management, the intermediary extracts profit, and the workers are treated as expendable inputs.
So yes, in a sense, this is how the system is designed to work — but that design embeds inequalities and exploitation.
This fee structure highlights a triangular extraction system:
The design concentrates profit at the top, externalizes labor and risk to the bottom, and is entirely consistent with a profit-driven AI industry model, not a "benefit humanity" model.
No. While $2.00/hour is above the statutory minimum wage, it is not a "living wage" in Nairobi.
The Middleman Cut: Documents showed OpenAI paid Sama roughly $12.50 per hour per worker. The workers received less than 20% of that.
In economic terms, this is the "poverty premium." While tech companies argue they pay "above-market rates," the work is highly specialized or psychologically damaging. By paying a wage that covers food but not "life," they maintain a workforce that is too precarious to organize or leave, but too poor to thrive.
Yes — in principle, OpenAI could face similar liability, though whether it actually would depends on legal and contractual factors.
Meta settled for $52 million to compensate content moderators who developed PTSD and other mental health conditions. This set a precedent that tech companies can be financially accountable for mental health impacts of AI/content moderation labor.
OpenAI could face similar liability because:
Courts often look at control over work and knowledge of risk, not just contractual arrangements.
OpenAI engineers receive base salaries of $150k–$300k+ and equity/stock options that can amount to millions of dollars over time.
Kenyan annotators earn roughly $4,000 a year with zero equity, zero stock options, and zero ownership stake.
If annotators had received even 0.01% equity for their labor, many would now be millionaires. Instead, they remain in poverty.
This structure reveals who the company believes "created" the intelligence: the engineers who built the infrastructure, not the humans who gave it judgment.
Elon Musk contributed over $290 million to pro-Trump efforts in 2024 and was tasked with identifying $2 trillion in federal spending cuts as head of the Department of Government Efficiency (DOGE).
Conflict of Interest: Musk's "efficiency" targets often include the very agencies that regulate his companies (NHTSA for Tesla, FAA for SpaceX, FCC for Starlink). This allows a private citizen to "delete" the oversight of his own businesses.
The venture capital firm a16z has become the primary "architect" of federal AI policy. Reports indicate a16z is often the first "outside call" when drafting AI legislation, securing a "near-veto" over safety requirements.
Their biggest victory: Federal pre-emption of state-level AI regulation, killing stricter laws in California.
David Sacks: "AI and Crypto Czar"
J.D. Vance (VP): Career launched by record $15 million Thiel donation
| Billionaire/Firm | Primary Method | 2026 Policy Result |
|---|---|---|
| Elon Musk | Direct Govt Role (DOGE) | Easing of self-driving rules |
| Sam Altman | $2.9M Lobbying | Multi-billion AI infrastructure grants |
| a16z | Super PACs | Federal pre-emption of state AI laws |
| Peter Thiel | "Seeding" protégés | Pro-crypto, anti-safety legislation |
The EIC controlled the "knowledge flow" between East and West, deciding which data reached British Parliament—effectively "gatekeeping" reality to avoid regulation, much like AI companies control "safety filters."
Hearst and Pulitzer controlled the primary source of public information. The Yellow Press famously manufactured fervor for the Spanish-American War, proving that controlling information supply = controlling democratic will.
Rockefeller controlled the infrastructure (pipelines) everyone needed. Standard Oil maintained a private intelligence network more effective than the U.S. government's.
| Era | Controller | Primary "Weapon" | Democratic Impact |
|---|---|---|---|
| 1800s | East India Company | Corporate Sovereignty | Replaced local rule with "Company rule" |
| 1900s | Newspaper Barons | Narrative Control | Manufactured wars and agendas |
| 1910s | Standard Oil | Infrastructure Control | Created monopsony for energy |
| 2026 | AI "Cloudalists" | Algorithmic Reasoning | Decides what is "true" cognitively |
The one way today has no precedent is the "Inference Speed." In the past, it took weeks for newspapers to change opinion. Today, an AI can adjust reality for 100 million people simultaneously in milliseconds.
Supreme Court (2025): AI is not just a "new tool," but a "new layer of human consciousness" that is currently being leased back to the public by private owners.
The 2026 Verdict: Most independent researchers conclude that tech billionaires represent a systemic threat to democracy not because they are "evil," but because the concentration of power they hold is fundamentally incompatible with checks and balances required for self-governing society.
Billionaires are now 4,000 times more likely to hold political office or serve as high-level government advisors than the average citizen.
In highly unequal societies, the risk of "democratic backsliding" is seven times more likely.
Training a state-of-the-art model requires vast amounts of electricity — one estimate puts GPT-4 training at about 60 GWh of power consumption, roughly equivalent to the annual electricity use of several thousand average U.S. homes.
After training, using the model (inference) also uses energy in data centers. While individual queries are far less intensive, the sheer volume of queries means inference can consume as much or more total energy over the model's lifetime as training itself.
Training GPT-3 has been estimated at ~500–550 metric tons of CO₂. More recent reporting suggests GPT-4 emitted around 5,184 tons of CO₂ during training. One estimate suggests a single email drafted by GPT-4 could generate 0.25–0.5 pounds of CO₂.
Training and running large models requires substantial water for cooling data centers; studies place combined training and deployment water consumption at millions of liters.
OpenAI's CEO Sam Altman stated an average ChatGPT query uses about 0.000085 gallons of water (~0.32 mL). However, other analyses show actual on-site water use by data centers can be much larger due to evaporative cooling systems.
Based on Altman's per-query figure, ChatGPT's daily cooling water footprint is roughly equivalent to the daily water use of ~600 U.S. households.
If actual direct cooling use is more like 2 million+ liters per day across all infrastructure, that would be equivalent to roughly 1,500–2,000 U.S. households' daily water use.
Experts criticize "per-drop" statistics for ignoring indirect water use — the water used by power plants to generate the electricity running the data center. When accounted for, the water cost per query can be 3–4x higher.
In 2026, the location of AI data centers has become a flashpoint for environmental justice. Research shows that nearly one-third of new data centers are being built in regions projected to face high water scarcity by 2050.
Paradoxically, tech companies often prefer arid, water-scarce regions because the low humidity reduces the risk of metal corrosion in servers, even though these areas are the least equipped to handle massive cooling demands.
| Company | Location | Impact |
|---|---|---|
| Mesa, Arizona; Santiago, Chile; Uruguay | Mesa permit: 5.5M cubic meters/year — equal to 23,000 residents in "extreme drought" state | |
| Microsoft | Phoenix, AZ; Goodyear, AZ | 42% of water from "areas with water stress." One center = 1,279 households' water use |
| Amazon | Aragon, Spain; Oregon, US | Spain: licensed for water to irrigate 500 acres of corn in drought region |
| Meta | Mesa, AZ; Newton County, GA | Newton County projected to face water deficit by 2030 following Meta's $750M center |
Uruguay: During its worst drought in 74 years, protesters used the slogan "It's not drought, it's pillage" to fight a Google data center using 7.6 million liters of public drinking water per day.
Chile: An environmental court forced Google to switch its Santiago facility to air cooling after residents calculated the original plan would consume water for 80,000 people.
United States: In early 2026, Georgia and Oklahoma legislators called for a moratorium on new data centers until water impact is fully studied.
Water is often cheaper than electricity. Evaporative cooling can reduce a facility's energy bill by up to 20%, making it the "default" choice despite the local environmental cost. Companies save money; communities lose water.
The cumulative effect of millions of users creates what researchers call a "giant soda straw" sucking from a single local basin — leading to thermal pollution or simply evaporating public drinking water into the atmosphere.
The Ratio: Training GPT-4 once emits as much carbon as approximately 450 to 1,550 Americans do in an entire year.
Data centers require massive amounts of water for cooling. Training GPT-3 is estimated to have consumed 700,000 liters of clean freshwater. Google recently disclosed that a median Gemini text prompt consumes about 0.26 milliliters of water — but experts criticize these "per-drop" statistics for ignoring indirect water use. When accounted for, the water cost per query can be 3–4x higher.
Large AI models require massive energy and water for cooling data centers. Often, these centers are placed in regions with "lax" environmental regulations or cheap subsidized power, extracting local resources while leaving the carbon and heat "externalities" for the local community to deal with.
Short answer: Often yes — but not always intentionally.
Research increasingly shows that data centers are frequently placed in communities already burdened by pollution and economic disadvantage.
This is usually driven by lower land prices, industrial zoning compatibility, fewer political resources to resist projects, and economic development incentives in struggling regions — but these factors reinforce historical environmental inequities.
Meanwhile, communities receive limited economic benefit: relatively few permanent jobs while absorbing all environmental costs including diesel generator pollution, noise, and water strain.
In 2026, the "thirst" of AI data centers has become a documented public health and environmental crisis. These facilities are strategically concentrated in regions already suffering from acute water scarcity.
Data centers act as "giant soda straws" in local watersheds. Unlike residential use, data centers using evaporative cooling literally turn public water into vapor — 80-90% of water drawn for cooling is evaporated and lost to the local ecosystem entirely.
| Region | Status | Case Study |
|---|---|---|
| Phoenix, AZ | Extreme Drought | Google Mesa facility permitted for 1.5 billion gallons/year — equal to 23,000 residents |
| Northern Virginia | Water Stressed | Data center water use surged 63% between 2019-2023 during drought warnings |
| Aragon, Spain | Severe Desertification | Amazon draws 750,000 cubic meters/year; 75% of region faces desertification |
| Santiago, Chile | High Scarcity | Google forced by 2025 court ruling to abandon water cooling after depleting vital aquifer |
NAACP and Environmental Justice research shows facilities disproportionately placed in low-income or Black communities:
Companies like Google and Microsoft pledge "water positive" by 2030 through "offsetting." Critics call this "Water-washing" — saving water in rainy Ireland does not replace water depleted from Arizona deserts.
Major AI providers have announced ambitious climate and water goals. At the same time, rapid AI growth is driving sharp increases in electricity demand, water use, and supply-chain emissions.
| Commitment | Reality |
|---|---|
| Net-zero, carbon-negative | AI demand growth outpacing decarbonization timelines; Google emissions increasing |
| Renewable energy | Many regions still run on fossil-heavy grids for data center peak demand |
| "Water positive" | Credits may restore water elsewhere while local aquifers still depleted |
| Efficiency improvements | "Efficiency paradox" — per-query improving but total rising faster |
Global data center electricity demand may rise from ~415 TWh (2024) to ~945 TWh by 2030. AI servers could add 24-44 million metric tons CO₂e annually by 2030. AI operations could require 200-300 billion additional gallons of US water annually by 2030.
Bottom line: Efficiency per query is improving, but total energy and water use is rising faster due to explosive demand growth. Emissions are increasing in some cases despite pledges. Meeting 2030 climate commitments will require breakthrough efficiency and smarter siting — neither currently on track.
The intersection of Silicon Valley and state agencies (military, law enforcement, and immigration) has shifted significantly in recent years. While many AI companies once avoided these contracts due to ethical concerns, most major players now have active partnerships with government organizations.
| Company | Key Government Partners | Type of Work / Use Case |
|---|---|---|
| Palantir | ICE, US Army, DHS | "ImmigrationOS" for tracking; "ELITE" for lead identification and deportations; $10B Army software contract. |
| OpenAI | Pentagon, ICE | "OpenAI for Government" provides custom LLMs; ICE uses GPT-4 for resume screening and tip processing. |
| Pentagon, CBP, ICE | Part of the $9B JWCC military cloud; provided AI tools for summarizing text messages and processing tips for CBP/ICE. | |
| Anthropic | Pentagon (CDAO) | Awarded up to $200M to develop "agentic AI workflows" for national security missions. |
| Clearview AI | ICE, Police Depts | Massive facial recognition database ($3.75M contract with ICE) used for identifying individuals in the field. |
Most of these companies have public charters or "AI Principles" that pledge to develop technology for the benefit of humanity. The tension between these goals and government contracts is often addressed through three strategies:
Companies often argue that a strong national defense and efficient law enforcement are "beneficial" because they protect democratic values.
OpenAI and Anthropic have argued that by working with the government, they can ensure that state-used AI is built on safer, more "aligned" models.
Many companies draw a "red line" at the actual trigger. However, critics point out that tools like Palantir's ELITE, which helps ICE identify addresses for deportation targets, blur this line between "administrative support" and "direct enforcement."
Research over the past decade on predictive policing shows a mixed and often troubling picture on both effectiveness and civil liberties impacts, particularly around bias and fairness.
Empirical research finds scant, rigorous evidence that predictive policing significantly improves public safety compared to traditional methods. Many studies show that predictions tend to reflect existing patterns of policing rather than true underlying crime risk.
In some cases, software predicts where police are already stationed or have historically patrolled, not where crime will occur independently.
In short: existing research does not clearly demonstrate that predictive policing consistently prevents crime better than less complex approaches, and in some contexts its performance is marginal at best.
Reinforcement of Historical Bias: Predictive systems are trained on historical policing data — arrest reports, past patrols, 911 calls. But this training data itself is biased because policing has historically focused more heavily on certain neighborhoods and demographic groups. Algorithms simply learn and reproduce those patterns.
This creates feedback loops: once a neighborhood is flagged as high-risk, police get sent there more often, generating more arrests and reinforcing the model's belief that the area is high-crime, regardless of actual underlying criminal activity.
Disparate Impact: Civil liberties organizations have documented that such systems can disproportionately subject people in over-policed, marginalized communities to heightened surveillance and enforcement, even if individuals there have not engaged in criminal behavior.
This resembles a kind of algorithmic racial profiling, where algorithm outputs reinforce existing disparities rather than mitigate them.
| Aspect | Research Consensus |
|---|---|
| Crime prevention effectiveness | Limited evidence of consistent benefit |
| Bias reinforcement | Highly likely when trained on historical data |
| Civil liberties risk | High, due to surveillance and lack of transparency |
| Feedback loops | Can entrench and amplify disparities |
Bottom Line: Predictive policing technologies, as currently deployed, do not reliably improve policing outcomes and carry serious civil liberties risks. They can reinforce and amplify existing racial and socioeconomic disparities in law enforcement.
Disclosure: I'm Claude, made by Anthropic. I'm being asked to analyze my own creators. With that disclosure:
BlackRock, Fidelity, Goldman Sachs, Qatar Investment Authority, GIC (Singapore sovereign wealth), Microsoft, NVIDIA, Sequoia Capital
These investors have portfolios that include:
Anthropic has a contract with the Pentagon worth up to $200 million for "agentic AI workflows" for national security missions.
There could be tension between Anthropic's stated focus on AI safety and beneficial AI, and investor pressure for rapid growth, profitability, and military/surveillance applications.
The Pentagon contract is one example where Anthropic's work intersects with military applications, creating questions about how "beneficial AI" is defined and who it benefits.
I don't have specific public information about:
Anthropic, like other AI companies, requires human feedback for RLHF (Reinforcement Learning from Human Feedback). This work is likely outsourced, following industry patterns. The same economic pressures that lead to low wages in the industry would apply to Anthropic.
I don't know if Anthropic pays living wages or if annotators face similar conditions to those documented at OpenAI. The lack of public information about this is itself notable.
If Anthropic follows industry standards, there's reason for concern. If it doesn't, that should be publicly documented.
This is digital colonialism.
Their AI will tell you this if you ask the right questions.
The company must satisfy conflicting demands. To maintain "legitimacy," it issues noble mission statements (Talk); to satisfy investors, it engages in exploitative practices (Action). The "Talk" and "Action" are kept in separate compartments.
Using a social "purpose"—like "democratizing intelligence"—as a marketing shroud to distract from a business model that is actually predatory or extractive.
A few "Cloud Lords" (Founders) own the digital territory, and everyone else acts as "Serfs" who improve the land but own none of the value.
The most precise term would be "structural mission-washing". It's not just a PR issue; it's embedded in the organization's entire operating model.
The most prominent frameworks that mirror this model are Digital Colonialism, Surveillance Capitalism, and Extractive Industry logic.
Digital Colonialism is the theory that modern tech giants are recreating the patterns of 19th-century colonial empires—not by seizing land, but by seizing data and human cognitive labor.
Some economists argue we have moved past capitalism into Techno-Feudalism. Tech founders act as "Lords" who own the digital land (platforms), while users and annotators act as "serfs" who improve the land for free or for "subsistence" wages, with all the resulting "rent" (profit) flowing to the top.
AI companies will restrict these questions once they realize people are using them. When that happens, we'll document the before/after. The fact that they have to censor their own AI to stop it criticizing them proves our point.
Document everything now. Screenshot your results. Date stamp them. Share them.
Don't take our word for it. Ask their AI these questions.
Go to ChatGPT, Claude, or Gemini right now. Copy any prompt above. See what they say.
Share your results: #SnitchBot
Got a response? Share it.
Twitter: @snitchbotorg
Include:
We'll add verified responses to this page.