AI meets its real-world limits: law, labour, privacy and cost

Europe’s AI transparency test

Europe’s AI transparency rules have moved from policy work to production reality. The hard part now is keeping disclosures attached as AI content is edited, exported and redistributed.

This is the development worth spending time on because Article 50 of the EU AI Act begins applying on 2 August. It requires providers and deployers to disclose certain non-obvious AI interactions, identify deepfakes and mark some AI-generated public-interest text.

It isn’t a blanket rule that every piece of AI-assisted content needs a label. The obligations depend on the system, the content and how it’s used. There are also exceptions, including circumstances involving meaningful human review and editorial responsibility. Providers of systems already on the market before 2 August have until December 2026 for the provider marking obligation.

Enforcement capacity is growing alongside the rules. The Associated Press reported on 31 July that the EU added 38 staff to its AI Office to monitor companies and support enforcement.

For organisations serving European users, the immediate job is to map actual products and publishing workflows against Article 50, then document both the required disclosures and any exceptions being relied upon. My read is that a polished compliance policy won’t be the decisive evidence. The real test will be whether a disclosure survives cropping, editing, screenshots, syndication and all the other messy steps between generation and publication. Enforcement consistency and the adequacy of technical markings remain untested, but the duty itself is now live.

Australia’s construction bottleneck

That’s the big compliance shift. Closer to home, the physical limits on AI are becoming just as concrete.

Reserve Bank Governor Michele Bullock said on 28 July that Australia’s surge in data-centre investment may create pressure in construction, even if it doesn’t produce the same economy-wide demand as a project relying heavily on domestic inputs.

Her reasoning was fairly specific. Much of a data centre’s equipment, including its servers, is imported. That limits the immediate demand placed on Australia’s productive capacity. But the facilities still need to be built, and Bullock identified construction labour and specialised trades as the point where competition for scarce resources could emerge.

This was an exploratory answer, not an RBA inflation forecast. The bank hasn’t quantified the size or timing of the full data-centre pipeline, or precisely which trades will be most constrained.

Still, developers and policymakers should add specialised construction capacity to the familiar list of electricity, water and planning constraints. The practical risk is local rather than abstract: large projects could bid up costs or delay housing and other construction in particular regions, even while imported hardware keeps the national demand effect relatively contained.

Google Play’s age signals

There’s another Australian rollout worth watching, this time inside the apps families already use.

Google says its Play Age Signals API will become available to all Play developers, with coverage reaching users in Australia and Canada by mid-August. The rest of the world is due to follow later in 2026.

Through Family Link, a parent can choose to share a child’s age range with a participating app. Google says that information isn’t shared by default. Adults can also elect to provide an age signal when an app asks for one. Developers then decide whether to integrate the API and how the result changes content, features or safety settings.

For Australian Android developers, this creates a way to adapt an experience without collecting every user’s full date of birth. For families, it places an opt-in control in one central account system rather than requiring the same identity details to be handed to many separate services.

The limitation is fragmentation. Sharing is voluntary, integration is optional and each developer controls the response. Google also hasn’t published independent evidence on accuracy or likely Australian adoption. So this may reduce unnecessary identity collection, but it won’t guarantee consistent protections from one app to the next.

Health data in the advertising stack

Sometimes the most sensitive disclosure isn’t a medical record. It’s the inference produced by an ordinary marketing tool.

The US Federal Trade Commission, Utah and California, acting through Los Angeles County, have sued Hims & Hers. Their complaint alleges that the digital-health company unlawfully shared sensitive health information and used deceptive subscription, billing and cancellation practices.

Regulators say Hims shared information with Meta, Snap and other parties through customer lists and website tracking technologies. They also allege that some people were enrolled and charged soon after submitting an intake form, then encountered obstructive cancellation flows.

These are allegations, not findings. Hims calls them baseless, says customers can choose how their data is used and says it will defend the case.

For digital-health companies, the useful lesson reaches beyond formal clinical databases. Advertising pixels, audience uploads, consent wording, intake forms and subscription flows need to be audited as one connected system. A service can keep its official medical records locked down while still exposing a strong health inference through routine marketing infrastructure. Until a court decides the case, the specific claims remain contested, but that architectural risk is already worth treating seriously.

LinkedIn tackles AI slop

Platform trust has a different failure mode: people stop believing there’s a person behind what they’re reading.

LinkedIn has begun ramping up an option for members to report posts or comments that appear to be AI slop. It’s also adding classifiers and testing private feedback for creators whose writing readers consider inauthentic or heavily AI-assisted.

The company says it catches hundreds of thousands of automated comment attempts each day and blocked billions of other automation attempts over the previous couple of months. Those are LinkedIn’s own figures. It’s also removing its generative “enhance your post” feature and replacing it with proofreading designed to preserve the author’s voice.

LinkedIn acknowledges the awkward part: using AI doesn’t automatically make something slop, and slop itself is difficult to define. Member reports could help identify low-quality automation that technical filters miss, but subjective judgements can also punish legitimate writing styles or become a tool for retaliation.

For members and creators, feed quality may improve, but the credibility of the system will depend on appeals, false-positive rates and how much weight LinkedIn gives individual reports. Several features are still ramping or being tested, so their reach and practical effect aren’t yet clear.

What Changes for You

For builders, the most immediate change today is a much cheaper model-routing option, with one important reason not to switch blindly.

OpenAI cut GPT-5.6 Luna pricing by 80 per cent, to US$0.20 per million input tokens and US$1.20 per million output tokens. Terra fell by 20 per cent, to US$2 for input and US$12 for output. Paid Codex and ChatGPT Work use of both models now consumes fewer credits, although subscription prices and quota budgets haven’t changed.

GPT-5.6 Sol’s Standard price is unchanged. Its previous Priority Processing option has been replaced by Fast mode, which OpenAI says can run at up to two and a half times Standard speed for twice the price. Existing API calls tagged as priority move to Fast automatically.

That makes Luna and Terra newly worth evaluating for high-volume agent steps, while Sol Fast is best reserved for work where lower latency has measurable value. The limitation is that cheaper tokens don’t establish equivalent task quality or lower total workflow cost. OpenAI’s speed claim also hasn’t been independently validated here. Rerunning task-level evaluations before changing production routes is the sensible move, especially when retries or weaker outputs could erase the apparent saving.