The takeaway: A quieter weekend, but the late-week news had a theme: who gets to use whose data. Brussels ordered Google to prise open Android and share its search data with rival AI assistants. A hack at the AI-music firm Suno appeared to reveal exactly how much copyrighted music it scraped to train on. OpenAI made its case for keeping teenagers on ChatGPT. Google's next Gemini slipped. And, away from the data fights, a Harvard model quietly showed AI doing something genuinely useful in cancer care. Here's Sunday's catch-up.

Brussels tells Google to open up — to its AI rivals

The European Commission handed down binding measures under the Digital Markets Act (reported 16 July) that force Google to let rival AI assistants plug into Android and to share some of its search data with competitors. Under the order, other assistants will be able to use voice activation and cross-app actions across roughly a dozen Android feature groups — subject to certification and user consent — and Google must share anonymised search data with rivals. Android interoperability is due by around July 2027; the search-data sharing starts January 2027, and non-compliance carries fines of up to 10% of global turnover.

The framing matters. This isn't a generic antitrust action — it is specifically about levelling the ground for AI assistants, at the exact moment every platform is racing to make its assistant the default. Google is opposing the measures on privacy and security grounds, which is the argument every gatekeeper reaches for. But the direction of travel in Europe is now unambiguous: the data and the entry points that made Google dominant in search are being pried open for the AI layer that comes next.

Sources: European Commission, CNBC, Euronews.

A hack at Suno appears to show what it trained on

A hacker accessed source code from the AI-music company Suno — via a supply-chain breach dating to November 2025 — and, according to 404 Media's reporting (rippling through the week), the code contained logs detailing large-scale scraping of copyrighted music for training. The numbers cited are specific: on the order of 113,000 hours from YouTube Music, tens of thousands more from Pond5, Genius and Deezer, with requests reportedly routed through a proxy provider to get around YouTube's defences. Suno's response was that this is "outdated source code no longer in use."

If the logs are what they appear to be, this is the kind of evidence that music-industry plaintiffs have been unable to get in the ongoing copyright suits: not a plausible inference that a model was trained on protected work, but an internal record of it. It doesn't decide anything on its own — provenance and context will be fought over — but "we can't know what's in the training set" has been the industry's shield, and a leak like this is exactly what erodes it.

Sources: TechCrunch (originating: 404 Media), Music Business Worldwide.

Google's next Gemini slips — but ignore the version numbers going around

Google is holding back Gemini 3.5 Pro to improve its coding performance, according to Bloomberg reporting (via 9to5Google, 16 July): a late-June training-data update reportedly produced "disappointing" results, and the company is now testing 3.5 Pro, an upgraded Flash model, and others with partners. That much is solidly sourced.

A word of caution on the rest. The versions of this story circulating on aggregator sites — a specific "third miss" of a 17 July deadline, and a named "Gemini 3.6 Flash stopgap" — do not appear in the reputable reporting, which mentions only "an upgraded Flash model," no 3.6, no stopgap framing, and no 17 July date. We're flagging the delay as real and the embellishments as unconfirmed. It's a small illustration of a recurring problem: AI-news aggregators inventing precise-sounding details that were never in the original.

Source: 9to5Google (citing Bloomberg).

OpenAI argues teenagers should stay on ChatGPT

OpenAI published a position piece, "Why teens deserve access to safe AI" (16 July), making the case that barring under-18s would leave them unprepared, and citing that around nine in ten of its teen users use ChatGPT for learning and productivity each week. Alongside the argument, it expanded its teen safeguards: more frequent break reminders during long sessions, and clearer pointers to parental controls — account linking, quiet hours, the ability to disable voice and image generation, and safety notifications — while stating parents don't get access to teens' conversations.

It's a notable stance to take publicly, and the timing isn't accidental: AI companies are under real pressure over minors and safety (see the classroom pushes from Anthropic and Google we covered this week). Framing continued teen access as the responsible choice — rather than the commercially convenient one — is a position OpenAI now has to keep defending with the safeguards to match.

Source: OpenAI.

Away from the data fights: AI reads a tumour

One genuinely encouraging item. A Harvard Medical School model called COMPASS, peer-reviewed in Nature Medicine, predicts how a patient's cancer will respond to immunotherapy — specifically immune-checkpoint inhibitors — from the gene-expression profile of their tumour. It was pretrained on more than 10,000 tumours across 33 cancer types and validated across 16 patient cohorts, and it reportedly improved on the best existing methods — raising precision by around 8.5% — while giving a biological rationale for each prediction rather than a black-box score.

That last part is the interesting bit. Immunotherapy works spectacularly for some patients and not at all for others, and knowing in advance who will benefit is a real, unsolved clinical problem. A model that not only predicts but explains, in terms a clinician can interrogate, is the shape of AI-in-medicine that's actually useful — and it's a reminder that while the headlines fixate on chatbots and lawsuits, some of the most consequential work is happening quietly in journals.

Source: Nature Medicine.

Also this weekend

• Credited Hollywood names are now openly making AI films. The Hollywood Reporter reports that established writers and directors — including a Doctor Who director and a Sharknado writer — are taking credited roles on films from the LA AI-studio Promise, a shift from quiet experimentation to public embrace. The credit is the story: it signals that working on AI-made film is no longer something the industry's professionals feel they must hide. (The Hollywood Reporter)

For Saturday's news, see yesterday's Daily Update.