AI's Hidden Costs

Artificial intelligence consumes more than just electricity and water. It also draws upon our knowledge, data, attention, and trust.

The Candle Theory Explained

Arnaud Desmarest, an engineer and pedagogue specializing in AI, has developed the candle theory, a pedagogical model presented on a freely shareable poster. It poses a question before each use: is it worth the candle?

The Candle's Three Parts

Each candle is read in three parts: the wax represents resources (what is mobilized), the flame the benefits (what is obtained), and the smoke the impacts (what is experienced).

Benefits and Impacts

On the benefits side, the poster cites for example access to knowledge or assistance in medical diagnosis; on the impacts side, greenhouse gas emissions, electronic waste, misinformation, or loss of skills.

Categorizing the Candles

The twelve candles on the poster elaborate on this principle and are divided into four categories: Our planet, resources extracted: Materials, Energy, Water Our productions, resources captured: Contents, Data, Invisible Work The individual, eroded resources: Cognition, Relationships, Skills Our commons, shaken resources: Trust, Economy, State

A Concerning Scenario

The most concerning case combines both, when AI depletes a resource on which it depends, like a candle under a glass bell, which goes out after having impoverished the air it needs. This is the case with AI-generated texts that flood the web where future AI will come to learn.

Notable Facts About Each Candle

Each of the 12 candles has a fact. Each candle is documented in an in-depth article. Here is a notable fact for each. Materials: Generative AI could produce between 1.2 and 5 million tons of electronic waste cumulatively between 2020 and 2030.

More Insights on AI's Impact

Energy: Data centers, including AI, consumed approximately 415 TWh in 2024, or 1.5% of global electricity. The IEA predicts about 945 TWh in 2030. Water: The global demand for AI could lead to the withdrawal of 4.2 to 6.6 billion cubic meters of water by 2027.

The Human Cost of AI

Contents: All public human texts could be fully used to train AI between 2026 and 2032. Data: In 2022, the CNIL imposed a fine of 20 million euros on Clearview AI, which scraped photos published online to create a facial recognition database.

The Invisible Work Behind AI

Invisible Work: Kenyan workers employed by a subcontractor of OpenAI earned between 1.32 and 2 dollars per hour to label texts describing extreme violence. Cognition: In Turkey, high school students helped by ChatGPT without safeguards performed 48% better during exercises, but 17% worse on exams taken without AI.

AI's Effects on Relationships and Skills

Relationships: In the United States, 72% of 13-17 year olds have already used an AI companion. Skills: In Poland, doctors accustomed to AI assistance detected fewer precancerous lesions during colonoscopies performed without it. Their detection rate dropped from 28.4% before the tool's introduction to 22.4% after (observational study).

Erosion of Trust and Economic Shifts

Trust: In 2024, an employee of Arup in Hong Kong transferred the equivalent of 25.6 million dollars to scammers after a video conference where the voices and faces of executives had been imitated by AI. Economy: Global investments in data centers were expected to reach 580 billion dollars in 2025, more than the 540 billion dollars allocated to global oil supply.

AI's Influence on Governance

State: In June 2025, before a Senate investigation committee, a leader of Microsoft France admitted under oath that he could not guarantee that data of French citizens would never be transmitted to American authorities without the agreement of France.

A Tool for Discussion

Conceived for pedagogy, the poster serves as a support for a course, training, a workshop in a company, an association meeting, or a family discussion.