維基百科與微軟、Meta、亞馬遜簽署AI訓練授權協議

Wikipedia signs AI training deals with Microsoft, Meta, and Amazon

Wikimedia Enterprise signs Microsoft, Meta, Amazon, Perplexity, and Mistral AI to paid deals.

On Thursday, the Wikimedia Foundation announced licensing deals with Microsoft, Meta, Amazon, Perplexity, and Mistral AI, expanding its effort to charge major tech companies for using Wikipedia content to train the AI models that power AI assistants like Microsoft Copilot and OpenAI’s ChatGPT.

While these same companies previously scraped Wikipedia without permission, the deals mean that most major AI developers have now signed on to the foundation’s Wikimedia Enterprise program, a commercial subsidiary that sells API access to Wikipedia’s 65 million articles at higher speeds and volumes than the free public APIs provide. The foundation did not disclose the financial terms of the deals.

The new partners join Google, which signed a deal with Wikimedia Enterprise in 2022, as well as smaller companies like Ecosia, Nomic, Pleias, ProRata, and Reef Media. The revenue helps offset infrastructure costs for the nonprofit, which otherwise relies on small public donations while watching its content become a staple of training data for AI models.

“Wikipedia is a critical component of these tech companies’ work that they need to figure out how to support financially,” Lane Becker, president of Wikimedia Enterprise, told Reuters. “It took us a little while to understand the right set of features and functionality to offer if we’re going to move these companies from our free platform to a commercial platform… but all our Big Tech partners really see the need for them to commit to sustaining Wikipedia’s work.”

The cost of “free” knowledge

The push for paid licensing follows years of rising infrastructure costs as AI companies scraped Wikipedia content at an industrial scale. In April 2025, the foundation reported that bandwidth used for downloading multimedia content had grown 50 percent since January 2024, with bots accounting for 65 percent of the most expensive requests to core infrastructure despite making up just 35 percent of total pageviews.

By October, the Wikimedia Foundation disclosed that human traffic to Wikipedia had fallen approximately 8 percent year over year after the organization updated its bot-detection systems and discovered that much of what appeared to be human visitors were actually automated scrapers built to evade detection.

The traffic decline threatens the feedback loop that has sustained Wikipedia for a quarter century: Readers visit, some become editors or donors, and the content ostensibly improves. But today, many AI chatbots and search engine summaries answer questions using Wikipedia content without sending users to the site itself.

Meanwhile, the foundation’s own experiments with generative AI have met resistance from the volunteer editors who maintain the site. In June, Wikipedia paused a pilot program for AI-generated article summaries after editors called it a “ghastly idea” and warned it could undermine trust in the platform.

Wikipedia founder Jimmy Wales told The Associated Press that he welcomes AI models training on Wikipedia data. “I’m very happy personally that AI models are training on Wikipedia data because it’s human curated,” Wales said. “I wouldn’t really want to use an AI that’s trained only on X, you know, like a very angry AI.” But he drew a line at free access: “You should probably chip in and pay for your fair share of the cost that you’re putting on us.”

Ars Technica has been separating the signal from
the noise for over 25 years. With our unique combination of
technical savvy and wide-ranging interest in the technological arts
and sciences, Ars is the trusted source in a sea of information. After
all, you don’t need to know everything, only what’s important.