Quietly, in the fine print of a privacy policy revised in early July 2023, Google made explicit something that had been half-visible for years: the open web, the shared commons of text people have posted and published, may be drawn upon to teach the company's AI systems. The change is framed by Google as clarification rather than novelty, and the phrase "AI models" replacing "language models" widens the field of what that teaching can build. Yet the question that lingers is not only what Google may do, but whether people who wrote the words ever meant them to become raw material, and whether t
Google Confirms Its AI Models Are Trained on Publicly Available Web Data
The policy is silent on how copyrighted work is kept out.
Mark
So what actually changed in Google's policy?
Mimi
The policy now says Google may use publicly available information to train AI models, and it names Bard and Cloud AI along with Translate. It took effect July 1st, 2023, and Gizmodo spotted it on Monday.
Mark
Google says this is just a clarification. Is that fair?
Luke
That's Google's characterization, relayed by spokesperson Christa Muldoon. The revision history does show one concrete change, though: "language models" became "AI models." Whether that widens what Google can do in practice, the reporting doesn't establish.
Mimi
The wider wording is the part worth noticing. It isn't limited to chatbots anymore.
Mark
And where does a reader find this disclosure?
Mimi
Under the "Your Local Information" tab, behind a click, and then an embedded link about publicly accessible sources. It's disclosed, but not prominently.
Mark
What about copyrighted work?
Luke
The policy is silent on it. It doesn't say how or whether copyrighted material is kept out, and we shouldn't fill that silence with assumptions about what Google does.
Mark
Why does GDPR come up?
Mimi
GDPR protects people from misuse of their data without express permission. The article says it'll be interesting to see how this approach fares, which is a question, not a finding.
Luke
Right. No regulator action is reported here. The legal status of fair use for AI training is described as a gray area, and the lawsuits mentioned, like Gannett's, concern ad market monopoly claims, not this policy directly.
Le Pouls
- A privacy policy edit surfaced on a Monday, and suddenly the public web looked less like a library and more like a training ground.
- Google insists this is clarification, while critics see a wider phrase, "AI models," quietly granting room to build far beyond chatbots.
- The disclosure hides behind a tab that must be clicked open and an embedded link, raising doubts about how openly the change was announced.
- Copyright holders, publishers, and regulators are circling, with lawsuits and GDPR questions still unresolved.
- The company's answer so far leaves the central matter open: how copyrighted material is kept out, and whether it is kept out at all.
Quietly, in the fine print of a privacy policy revised in early July 2023, Google made explicit something that had been half-visible for years: the open web, the shared commons of text people have posted and published, may be drawn upon to teach the company's AI systems. The change is framed by Google as clarification rather than novelty, and the phrase "AI models" replacing "language models" widens the field of what that teaching can build. Yet the question that lingers is not only what Google may do, but whether people who wrote the words ever meant them to become raw material, and whether the law, still unsettled on fair use, will decide that question before the habit becomes custom.
On a Monday in early July 2023, Gizmodo noticed that Google had revised its privacy policy, effective July 1st. The new text said the company may use publicly available information from the web to train its AI products, naming Bard, Cloud AI capabilities, and Google Translate among them. The language also spoke of developing new products and technologies that benefit users and the public.
Google spokesperson Christa Muldoon told The Verge that this was a clarification, not a change in direction. The company had long said it used open web material to train language models for services like Translate, she explained, and the update simply made clear that newer services such as Bard were included. Google, she added, builds privacy principles and safeguards into its AI work under its AI Principles.
The revision history tells a slightly different story. The old phrase "language models" became "AI models," a broader term that allows training well beyond large language models. And the disclosure itself is easy to miss, sitting under a "Your Local Information" tab that must be opened and behind a link about publicly accessible sources.
Many questions remain. The policy does not say how copyrighted material is excluded, or whether it is. Many websites ban scraping for training purposes, and it is unclear how Google's approach fits with GDPR's protections against misuse of personal data. The wider industry has grown secretive about its sources, and fair use for such training remains a legal gray area that has already brought lawsuits and proposed legislation.
Google faces other pressures too, including a suit from Gannett, which accuses the company of using AI advances to sustain an ad monopoly, and critics who call its AI search a plagiarism engine that starves websites of traffic. What comes next will depend on how GDPR regulators and the courts handling copyright and fair use respond.
On Monday, Gizmodo noticed a change in Google's privacy policy. The company had revised the document, effective July 1st, 2023, to state that it may use publicly available information from the web to train its AI products. The named services include Bard and Cloud AI capabilities, alongside Google Translate.
The new language says Google uses information to improve its services and to develop new products, features, and technologies that benefit users and the public. It adds that the company may use publicly available information to help train Google's AI models and build products and features like Google Translate, Bard, and Cloud AI capabilities.
Google spokesperson Christa Muldoon described the change to The Verge as a clarification, not a shift. The policy, she said, has long made clear that Google uses publicly available information from the open web to train language models for services like Google Translate. The update, in her words, simply confirms that newer services such as Bard are included. She added that Google builds privacy principles and safeguards into its AI development, in line with its AI Principles.
The policy's revision history shows what did change. Earlier wording referred to "language models"; the new text says "AI models." That broader phrase gives Google room to train and build systems beyond large language models on public data. The disclosure is also not prominent. It sits under the policy's "Your Local Information" tab, which has to be clicked open, and the relevant note is tucked behind an embedded link labeled for publicly accessible sources.
The policy leaves several questions open. It says publicly available information is used to train Google's AI products, but it does not say how, or whether, the company will keep copyrighted material out of that pool. Many publicly accessible websites have policies that ban data collection or scraping for the purpose of training large language models and similar tools. How Google's approach squares with global rules such as GDPR, which protect people from having their data misused without express permission, remains to be seen.
The backdrop is a industry-wide reluctance to talk about sources. Those laws, together with growing market competition, have made the makers of popular generative AI systems, including OpenAI's GPT-4, very guarded about where their training data came from. They have been equally guarded about whether it includes social media posts or copyrighted work by artists and authors.
Whether fair use extends to this kind of training currently sits in a legal gray area. That uncertainty has prompted lawsuits and pushed lawmakers in some countries to propose stricter laws on how AI companies collect and use training data. It also raises questions about how that data is processed so it does not feed dangerous failures in AI systems. The people who sort through these large pools of data are often subjected to long hours and extreme working conditions.
Google faces pressure on other fronts as well. Gannett, the largest newspaper publisher in the United States, is suing Google and its parent, Alphabet. It claims advances in AI have helped the company hold a monopoly over the digital ad market. Google's AI search beta has also been called a "plagiarism engines" by critics, who say such products starve websites of traffic.
What to watch is how regulators enforcing GDPR and the courts handling copyright and fair use cases respond. Their answers will help determine whether training AI on scraped public data remains permissible.
Citations marquantes
The update simply clarifies that newer services like Bard are also included in training on publicly available web information.— Christa Muldoon, Google spokesperson, to The Verge
Envie de l'histoire complète ?
Lire l'original sur The Verge ↗
Technology