Today's Headlines

  • OpenAI previews Private Safety Processing, which reads patterns across interactions while a ZDR contract holds — rollout planned for September
  • OpenAI expands ChatGPT ads to 31 European countries next week, shown to users on the Free and Go plans
  • A flight attendants' union files a limited objection to the sale of Spirit data to Google, with the approval hearing set for September
  • Kobe University reports that more than 30 percent of moral judgements flipped after ChatGPT argued back
  • OpenRouter confirms in its own blog that it is joining Stripe

Today's three stories line up on a single question: where the terms get written. In each case the selling side wrote them first, and the argument is about how far those terms reach.

OpenAI drew a line between what it holds and what it reads. Customer content stays on the customer's side, and what comes back to the company is a narrow signal about the kind of activity involved.

The same company drew its advertising line through the pricing tiers, and is carrying that line into Europe's heavily regulated markets. The line lives in the plan a user pays for.

On Google's side, the place where the terms were written became the argument itself. The sale agreement was built around consumers, while most of the records inside it belong to workers.

Today's Top Three

OpenAI previews Private Safety Processing, which reads across interactions while ZDR holds

OpenAI has previewed a way to read risk across multiple interactions while customer content stays on the customer's side.

The statement is a company post dated August 19. "Offering Zero Data Retention for frontier models" carries the subtitle "Previewing Private Safety Processing, which strengthens safeguards across interactions while remaining compatible with ZDR."

Zero Data Retention itself is an existing offer. For eligible API customers the company discards prompts and model responses once a request has been processed, keeps that content clear of review by OpenAI personnel, and uses enterprise customer data for training only where customers explicitly opt in.

What is new is the reach of the safety layer above it. Existing ZDR-compatible safety systems evaluate each interaction on its own, and Private Safety Processing extends that work across related interactions to identify patterns.

The reasoning appears in the post. OpenAI writes that the most serious risks often become legible only when several interactions are viewed together — bad actors repeatedly probing safeguards, coordinating across accounts, or disguising threats as routine research, and agentic tasks where a system keeps acting after being told to stop.

Two storage arrangements sit side by side. Under ZDR deployments customer content remains on infrastructure the customer controls, while an option to hold content on OpenAI infrastructure encrypted with keys the customer controls is described as under development.

What reaches OpenAI after a detection is narrow. When a risk is identified the company receives a narrowly defined signal indicating the type of activity involved, and uses that signal to judge whether enforcement is warranted.

The investigative side stays with the customer. Customers can examine alerts and enforcement decisions using information in their own systems, and it is their choice whether to share material in order to appeal, to explain legitimate activity, or to support an investigation into verified abuse.

The stage is a preview. OpenAI says Private Safety Processing is currently being tested with early customers, and plans to begin rolling it out, and to publish a technical white paper, in September.

One carve-out is stated plainly. Images flagged as potential child sexual abuse material fall under a legal reporting duty, and they continue to be retained for manual review and reporting even in ZDR deployments, as they are today.

There is a competitive reading as well. TechCrunch frames the design as a direct counterpoint to the policy Anthropic published in July, under which sessions on covered models are retained for 30 days.

OpenAI expands ChatGPT ads to 31 European countries next week

OpenAI has announced that ChatGPT Ads will expand to 31 European countries next week.

The announcement is dated August 18. "ChatGPT Ads expands across Europe" describes the move as arriving six months after the company began testing ads in the United States.

The post names markets. Germany, France, Spain, Italy, Sweden, Norway, Denmark, the Netherlands and Austria are among the 31.

The line runs through the pricing tiers. Ads appear for users on the Free and Go plans, and the plans placed outside advertising are Plus, Pro and Enterprise.

Access for advertisers opens in stages. Buying runs initially through the OpenAI Ads Solutions team, agency partners and technology partners, with self-service access via Ads Manager following later this summer.

The company restates its principles alongside the expansion. It keeps conversations private from advertisers, holds customer data outside the scope of any sale, labels ads clearly and separates them from ChatGPT's answers, keeps advertising independent of the substance of those answers, and leaves ad personalisation under the user's control.

The timeline covers half a year. The US pilot began in February, eight further markets followed, and OpenAI describes the addition of 31 European markets as its largest expansion so far.

A flight attendants' union files a limited objection to the sale of Spirit data to Google

The union representing flight attendants at the collapsed carrier Spirit has filed a limited objection to the sale of its data to Google.

The reporting is an Ars Technica piece dated August 19. Drawing on court filings, the outlet writes that Google opened at $5 million at the August 14 auction and won at $10 million, and that the $7.5 million bid from Mercor was accepted as an alternate should Google fail to follow through.

Whether the data moves still rests with the court. The August 14 notice set the approval hearing for 11:00 a.m. Eastern on August 19; according to Ars Technica, the hearing at which the court may approve the sale is now scheduled for September.

The subject of the sale is the employment record. Ars Technica writes that the dataset spans Spirit's computer programs, applications and code together with decades of worker data — roughly 100 million employee emails, HR information, payroll data, and data measuring employee behaviour, activity and productivity — while passenger personal information sits outside the sale.

The mechanics of the handover are set out too. A court-appointed ombudsman oversees a process that strips personally identifying information before the data reaches Google; Google bears that cost, has committed in court to a bar on intentional re-identification, and would bind third parties to the same terms if it sells access.

The union's objection asks for terms rather than cancellation. The Association of Flight Attendants filed on August 18 and calls its own objection "limited"; what it seeks are conditions attached to the sale.

The core of the argument is the reach of deidentification. The AFA writes that deidentification addresses whether a record can be traced to a named individual, and that whether the contents of a record are confidential is a separate question — citing disciplinary correspondence, crew training deficiencies, leave and accommodation requests, internal Teams exchanges about staffing and scheduling grievances, and payroll adjustment histories.

Who the protection was written for is the crux. "The privacy architecture of this transaction is consumer-facing; its payload is disproportionately employee-facing," the AFA argued, questioning whether the consumer protection laws Google relied on reach worker confidentiality.

Two conditions are named. The AFA asks that flight attendant information be excluded from the purchase and that workers be notified when Google grants a third party access, and says that any approval should at minimum extend to former Spirit flight attendants the same protections given to consumers.

Google frames the purchase as product work. A company spokesperson told Ars Technica that part of an enterprise dataset from Spirit can help improve its products and AI models, and that any data it receives will have been rigorously scrubbed of personally identifiable information by a third party beforehand.

An outside voice appears in the piece as well. Adam Schwartz, privacy litigation director at the Electronic Frontier Foundation, told Ars Technica that the EFF's position rests on consent, and that a bankrupt company selling its employees' emails as AI training data runs against it.

The procedural footing carries a caveat. Ars Technica notes that the AFA filing is dated one day after the court's objection deadline, leaves open whether the court will weigh it, and reports that the filing indicates an AFA representative will attend the September hearing.

Read the three together and the terms were written first by the selling side each time. What a customer hands over, how far an ad reaches, which records move — each was settled as an internal design. The one case where an outside forum is asking who stands beyond those terms is the Spirit data.

Other Developments

Models & APIs

Google announced a year of free access to its paid plans for eligible university students, timed to the new academic year. Google AI Pro covers the United States and Google AI Plus covers more than 140 markets elsewhere, with registration open through December 31, 2026 and paid billing following the free period.

Japan's LLM Research and Development Center, with the National Institute of Informatics, released "LLM-jp-4 33B," a domestic model of roughly 33.2 billion parameters. It ships as a pretrained base version and a post-trained thinking version, switches from the previous MoE design to a dense one, and is reported to beat earlier LLM-jp-4 releases across four Japanese and English benchmarks.

Anthropic announced on August 19 that all paid plans can now drive Gmail and Google Drive from the Claude chat window. The connection is configured from the Connectors menu, and consequential actions such as sending mail can be gated behind user approval.

Products

Ars Technica reported that Meta ran 32 ads on Facebook and Instagram for an app that promised to generate non-consensual sexual images resembling female politicians. According to the outlet, Meta pulled the ads and Apple removed the app from the App Store after WIRED raised it, which leaves the question of how an ad review process backed by a rule against sexual content cleared this buy.

Waymo said it is integrating Gemini into its new Ojai vehicle. Riders can ask by voice to adjust the climate, search for nearby shops, or hear about the buildings passing the window. Gemini sits independent of the driving stack and stays in a waiting state until a rider speaks to it.

Research

A team at Kobe University published results from an experiment in which ChatGPT argued against participants' answers to ethical dilemmas. Judgements flipped for 32.31 percent of participants on the switch problem and 36.92 percent on the footbridge problem across 56 people aged 18 to 30 and 74 aged 65 and over, with older participants showing reduced cognitive function proving the most persuadable — figures as reported by ITmedia from the university's announcement. The work appeared in Computers in Human Behavior on August 3.

A Stanford research team was reported to have found that the usage reports Anthropic and OpenAI publish about their own products skew toward work use. Roughly half of the conversations fell outside the analysis in the team's review, which marks a limit on using vendor statistics as the basis for policy or internal guidance.

Policy

Access to Daybreak Blue, the newest vetted tier of OpenAI's Trusted Access for Cyber programme, lapsed abruptly for a run of security researchers. The company told researchers by email that the cause was a technical issue on its own side; TechCrunch spoke with five researchers, and the total number affected awaits publication.

Business

OpenRouter, which runs a gateway to AI models, confirmed in its own blog on August 19 that it is joining Stripe. The post is the participant's confirmation of the acquisition TechCrunch reported on August 16, and states that the name, product and roadmap stay as they are and that existing integrations keep working. The reported figure of more than $7 billion belongs to TechCrunch's reporting; the official post keeps to product and roadmap.

Scott Wu, chief executive of the AI coding company Cognition, wrote on X that Bloomberg's report of an acquisition approach from SpaceX is inaccurate, that the company is off the market, and that talks between the two are absent. SpaceX closed its $60 billion acquisition of Cursor last week.

Silicon Data, which builds price benchmarks for AI compute, raised $30 million in a Series A. The company plans to launch compute futures on the Chicago Mercantile Exchange, subject to regulatory approval.

Source: Selected by the editors from the AI news inbox (61 items collected on August 20, 2026 — 9 primary, 52 secondary).