Quietly, in the way that so much of the digital age's business is conducted, Google revised the words that describe what it does with the open web. Its privacy policy now states that publicly available material is mined to improve AI services such as Bard, and the disclosure sits where few readers would stumble upon it. The episode belongs to a long human argument about who owns the commons of shared knowledge, and whether a text once written for people can be quietly repurposed to teach machines. Underneath it lies a deeper worry: that the wells of information being drawn upon are themselves
Google's Quiet Privacy Policy Change Puts Web-Scraped AI Training Data Under Scrutiny
The update sits behind an embedded link, not on the page
Mark
Google says it updated its privacy policy. Why is that a story at all?
Mimi
Because the update says public web data feeds Bard and Cloud, and the magazine says the wording sits behind an embedded link in the Your Local Information tab. You have to click to find it.
Mark
Google says it was already transparent, though.
Mimi
Spokesperson Christa Muldoon says the policy already covered training models for things like Google Translate, and the update just extends it to newer services like Bard.
Luke
So the dispute is about prominence, not whether the disclosure exists. Calling it secret is the author's characterization. Google's position is that it was disclosed.
Mark
What's the bigger worry the piece raises?
Mimi
Quality. If the web fills with AI-written content, models trained on scraping may inherit flawed material and biases. NewsGuard found 50 sites almost entirely AI-written.
Luke
That 50 is a count of sites NewsGuard identified, not the whole web. And the Europol line about 90 percent by 2026 is an experts' estimate, not a measurement.
Mark
And OpenAI is the contrast?
Mimi
Deals with the AP, Shutterstock and Boston Consulting Group suggest it wants licensed, cleaner sources. The Shutterstock agreement runs six years.
Luke
But the article admits OpenAI is cagey about GPT-4's training data. Licensing deals don't tell us what was already used. The legal status of scraping is still unsettled, and the piece offers no evidence that Google's outputs are actually worse.
Il Polso
- A policy change that many observers see as buried beneath an embedded link has ignited suspicion about how Google treats the public web.
- Google's spokesperson insists the update merely extends long-standing transparency about training language models, a claim that clashes with the critics' sense of concealment.
- The web itself is being flooded with AI-generated content, and warnings that models trained on it may compound bias and error are growing louder.
- OpenAI is pursuing licensing deals with news agencies, image libraries and consultancies as a cleaner alternative, though its own secrecy about training sources leaves the question unsettled.
- Regulators and rivals are being watched closely, as no firm solution has yet emerged and the legal status of scraping for training remains uncertain.
Quietly, in the way that so much of the digital age's business is conducted, Google revised the words that describe what it does with the open web. Its privacy policy now states that publicly available material is mined to improve AI services such as Bard, and the disclosure sits where few readers would stumble upon it. The episode belongs to a long human argument about who owns the commons of shared knowledge, and whether a text once written for people can be quietly repurposed to teach machines. Underneath it lies a deeper worry: that the wells of information being drawn upon are themselves growing murky with machine-made words, so that tomorrow's systems may learn from the echoes of today's.
Two weeks before this piece appeared, Google updated its privacy policy to say that it mines publicly available information from the web to improve its AI services, naming Bard and Cloud among them. Analytics India Magazine described the change as made in secret, and Google's spokesperson, Christa Muldoon, rejected that framing. She said the company had already been open about using public open-web data to train language models for products such as Google Translate, and that the update simply extends the same language to newer services, guided by its AI principles.
The magazine's objection was less about the words than their placement. The disclosure does not appear on the page itself; it sits behind an embedded link within the Your Local Information tab, so a reader must go looking for it. The article places this inside a longer history of friction, pointing to the enormous volume of data Google handles each day, a lawsuit from a major American newspaper publisher, and complaints that its AI search tools draw away traffic from the sites they borrow from.
The heart of the piece is a worry about quality. The web, it argues, is already filling with AI-generated text, and models trained on scraped pages risk absorbing and amplifying that material's biases and errors. NewsGuard has flagged dozens of sites written almost entirely by software, and a Europol report suggests that much of online content could be synthetic by 2026.
Against this, the author holds up OpenAI's recent deals with the Associated Press, Shutterstock and Boston Consulting Group as an attempt to secure clean, first-source material, and suggests Google could learn from them. Yet the piece also admits that OpenAI has been guarded about the data behind GPT-4, and that the legal footing for scraping remains unclear. No concrete remedy is offered, though several countries have begun tightening AI rules. The closing warning is simple: big technology firms must take their own data seriously, or risk a kind of digital collapse. What to watch next is whether regulators and rivals push for firmer sourcing rules as licensing deals multiply.
Two weeks before this piece was published, Google changed its privacy policy. The revision disclosed that the company mines publicly available information from the web to improve its AI services, with Bard and Cloud named among them. Analytics India Magazine, in an opinion-leaning piece, describes the change as made in secret.
Google spokesperson Christa Muldoon pushed back on that framing. She said the company has been transparent in its privacy policy about using publicly available data from the open web to train language models for services such as Google Translate. The update, she said, extends that language to newer services like Bard. Google, she added, builds privacy principles and safeguards into its AI work, in line with its established AI principles.
The magazine says the placement of the change tells a different story. The revised wording on publicly accessible sources is not shown on the page. It sits behind an embedded link inside the Your Local Information tab, and a reader has to click through to find it.
The article places the change inside a longer history of friction. It says Google processes more than 20 petabytes of data a day and has faced legal fights over how it uses what it gathers. The largest newspaper publisher in the United States sued, arguing that advances in AI helped Google hold a monopoly over digital advertising. Google's AI search beta has been called a plagiarism engine, and the company has been accused of swallowing website traffic and leaving other sites starved for attention.
The piece's central argument is about quality. The internet, it says, is already contaminated with AI-generated junk, so models trained on scraped web data risk inheriting and compounding biases, producing flawed outputs. NewsGuard, an information-tracking site, has identified 50 websites as almost entirely written by AI software. The article also cites a Europol report saying experts estimate as much as 90 percent of online content may be synthetically generated by 2026.
Against that backdrop, the author presents OpenAI as a step ahead on clean data. The company has recently struck partnerships with the Associated Press, one of the largest news agencies in the US, with Shutterstock, and with Boston Consulting Group. The AP deal is said to explore how AI can support local news, and the article says it indirectly links OpenAI to 41 news agencies that the AP supports. The six-year Shutterstock agreement lets the Altman-run company train its large language model on images, videos and music from content creators.
The author reads these deals as an effort to obtain clean, first-source information, and suggests Google could learn from it. The article adds a caveat of its own: OpenAI has been very guarded about where it got the data used to train GPT-4, the model behind ChatGPT. Questions have been raised, but the legal status of scraping for training data remains a grey area. No concrete solution has been proposed, though several countries have moved toward stricter AI regulation.
The article closes with a warning. Telling people not to believe everything they see online has long been standard advice, it says, and big tech companies now need to take their own data seriously or risk a digital collapse. What to watch is whether regulators and rival companies press for firmer rules on how training data is sourced, as OpenAI keeps signing licensing deals with publishers and content providers.
Citazioni salienti
Google has been transparent about using publicly available web data to train language models for services like Google Translate, and the update extends this to newer services like Bard.— Christa Muldoon, Google spokesperson (paraphrased)
Experts estimate that as much as 90 percent of online content may be synthetically generated by 2026.— Europol report, as cited by Analytics India Magazine
Vuoi la storia completa?
Leggi l'originale su Analytics India Magazine ↗
Technology