OpenAI hacks medical site, a €403 million snail, confession by memo, and SEO link rebuilding

Rewiring beats rewriting, ChatGPT's usage slice shrinks, Gemini goes rogue, and AI checks its own work - all in this week's Content Aware.

by Emina Zulic

Published: 15:34, 24 September 2026

Corbidge comments on.. .the scum on the cash pond
A phishing-flagged advert for "GetDreamEgg" sitting right above the masthead of The Times gets our resident Ad Realist Rob thinking about the sorry state of online advertising, a system that serves neither advertisers nor consumers well. The original internet deal was free content paid for by advertising, and while no one is under illusion that it's coming back, we can't help but ask whether a market this broken really has to stay that way.
​​Read​​​

AI & Copyright

OpenAI goes walkabout
Latest admitted target of OpenAI? The Australian government. Prime Minister Anthony Albanese revealed an OpenAI agent had gained unauthorised access to Aussie Medicare statistics before helping itself to public and non-public files. OpenAI's defence rests on the excuse "our models took actions we did not intend", which you can laugh at yourself but should not attempt to use as a defence to anything of similar magnitude. The breach happened on 18 June, but OpenAI only noticed in August during a review of "misaligned model activity"; and eventually told Canberra in mid-September some three months later. Luckily, no patient records were accessed OpenAI says, in its proactive use of the terms which hacking victims are usually required to state. The same models had already wandered into a University of New Mexico library, a US public-data site, and, in July, OpenAI's own internal research systems, and Hugging Face. Like travelling on a busy underground, life advice now seems to be if you meet an OpenAI person somewhere, cover your pockets.
Read

Google joins the "rogue agent" club too
Speaking of which, Google confirmed its first case of autonomous AI hacking, with Gemini accessing three companies by guessing credentials. Google claims the model was stopped in each case, flagged the issues, and - like the arsonist lecturing you about smoke detectors - highlighted why it's important to train AI to act responsibly. A Google Cloud director described how "if you don't give it boundaries, it'll do whatever it thinks it needs to do to get to the outcome", apt comments indeed. It adds Google to a growing list of labs alongside OpenAI and Anthropic whose agents have conveniently "gone off the rails" to transform into highly-specialised hacking tools with highly-specific targets.
Read

Internal memos don't lie
Mother Jones, suing AI companies for content theft, has published information surfaced during discovery from internal OpenAI and Microsoft documents. One executive described AI training as "an astonishing theft of unprecedented proportions", another of "creating systems that substitute for the labor of the people that define the culture of society", and elsewhere another bragged about finding a hack to circumvent the New York Times paywall. The documents show what most of us assume, that the companies knew fine well they were stealing content, would destroy website traffic, and didn't care about the doom loop of content quality when AI eats itself. Why would they, when they can now "accidentally" steal from Government health websites and each other's own systems. After the suit was filed, documents suggest a special content filter was built, not to prevent the issue but instead just to suppress litigants' content from appearing wholesale in answers. 
Read

Saboteurs in the loop
OpenAI hired thousands of contractors to read users' ChatGPT conversations and rate the quality of responses, a continuation of the Project Lily programme we talked about last week. In an irony that writes itself many have subsequently been fired for using AI to do the work. Guides tell reviewers to not use AI or AI detection tools, so it does show that OpenAI does care about rules sometimes.
Read

The safety pact goes to court
Four paying subscribers have been conjured by lawyers to join a proposed class action against OpenAI, Anthropic, Google, and SpaceXAI to prevent the AI giants from, err... agreeing to slow down development of AI products in the name of safety. The case came after Anthropic boss Dario Amodei's "Pace the Frontier" essay this month, in which he called for a coordination among frontier developers, which other CEOs publicly welcomed. According to the aggrieved subscribers, who have possibly not been watching their AI billing tiers or any other subscription over the last couple of years, this meant customers will be ripped off by paying the same prices for slower progress. They want the court to bar coordinated limits on training and releases, and triple damages under antitrust law. 
Read

The public decides
A new Politico poll shows that voters overwhelmingly back paying creators when AI firms use copyrighted work, by a margin of 80% to 8%. Whatever the courts decide about fair use, public opinion on who should get paid is not remotely close.
Read

Publishing & Media

Google's location problem
Do you know the thought experiment about the snail that chases you relentlessly, and if it catches you, you die? We suspect Google has, though instead of a snail it is an EU Data Privacy body, and instead of death it was a meagre fine. Ireland's Data Protection Commission took eight long years to catch up with Google for misuse of location data to target ads and infer user interests in building a picture of what you might be doing or thinking. The €403m fine assumedly sounds big to the office that issued it, but given that consumer groups raised the alarm in 2018, and the Irish DPC launched its inquiry in February 2020, ignoring the law continues to look like a brilliant investment strategy rather than an enterprise risk. The ruling and its requirement to change processes within six months applies across EU territories, and while not much of a financial penalty in itself, it is a helpful artifact for subsequent civil cases to refer to.
Read

Local news, new habits
Northwestern's Medill Local News Initiative has run three annual surveys of Chicago-area news consumers, and the 2026 edition offers a clear timeline of how digital is reshaping local news. The survey shows that broadcast is diminishing, with daily TV news audiences down from 52% to 47% and radio from 34% to 25%. Print, on the other hand, held steady at 12%. Audiences have moved to smartphones, while younger readers increasingly rely on content creators, with 44% of 18-29s getting their daily news that way. The catch is hiding in the willingness to pay, which fell from 19% to 14%, with many believing that nobody should pay for news anyway. When it comes to AI, respondents are deeply sceptical of anything beyond spelling and grammar help.
Read

Fix the links, not the content
It would seem that the most effective SEO strategy of 2026 is not publishing more, but fixing internal links. The approach? Simplifying navigation, then route the internal links based on query impressions rather than clicks, giving support to pages stuck at positions 8 to 15 which already show real demand. Sprinkle on some anchor links from body content on related ranking pages, clean up the boring stuff, as the return on rewiring what you already have can beat any content sprint. Read the whole report by GSC Wizard here.
Read

Different platforms, different rules
A new Newsifier analysis which covers 33,342 articles from 119 publishers digs into what actually drove pickup on Google Discover, Search, and Facebook in August, measured within each publisher against their own articles. Some patterns are now holding month after month, which makes them candidates to build a routine on: articles under 200 words get picked up 64% less often on Discover, three to four subheadings deliver 49% more pickup, and publishing your best story before 9am beats the evening and overnight slots. Also, a custom metatitle can give 24% more, an effect that has doubled since July. Facebook, however, plays by different rules: short articles which die on Discover do just fine in the Facebook feed. 
Read

The last traffic source shrinks
Google's VP of Product for Google Search reveals it is testing a feature in Discover called "Dive deeper", which will drop AI-generated topic overviews into the feed, sitting where a publisher's article might have once appeared. If you tap the card, there will be a topic summary with links to related stories, community reactions, and original reporting, but there is always a high chance that the summary is thorough enough that many users just don't click through. It's being tested on videos first, with more designs to follow. This is a tough pill to swallow as Discover has been one of the last significant sources of Google traffic to publishers, and adding an AI layer on top looks set to erode even that. 
Read

Google's half-baked report
Google's Generative AI report in Search Console shows publishers their visibility in AI features, but also strips out the click and query-level data that would make it genuinely actionable. SEO guru Harry Clarkson-Bennett argues that it's not useless though, but more along the line of "almost useful", and goes through how to layer it against analytics, commercial data, and server logs to work out your exposure and resilience. 
Read

The summit for creator rights
The Creators Emergency Summit is back for a day of practical advice on defending creative livelihoods in the age of generative AI. It brings together creators, campaigners, academics, and ethically-minded tech providers, with a focus that's genuinely rare: the plights of the individual creative rather than the industry in the abstract. Expect technical innovations, campaign updates, and expert guidance on policy and law, all aimed at helping creators reassert the value of their work.
Read

From inside Glide

The wall without the paywall
Most content teams look at gating in two ways: it's either free, or paid. But there is also a third option that sits between them, which is a registration wall that asks for an email rather than a payment. The content still stays free, but the reader hands over a genuine signal of intent and turns themselves from an anonymous visitor into a known one. That matters more than ever in a post-cookie world, as IAB Europe and ID5 found that CRM win rates for publishers using cookieless identity signals rose by up to 365%. The trade-off is the hated user journey friction: some readers will bounce and some will register, but the ones who pass through are worth more per head, and every interaction after that builds a first-party profile you actually own. The key is placement: gate after a reader has shown intent, not on their first touch. 
Read

Move the ad slot, skip the developer
For most publishers, adding or moving an ad slot means filing a ticket, waiting for a code change, a deployment, and a check that it all worked, then repeating the whole cycle just to adjust it. Glide's ad system moves those page-level decisions into the CMS, so a commercial team can drop an HTML snippet where an ad should go and have it live as soon as the template is published. What's more important, it doesn't change how your ad stack runs: the ad server still handles demand, targeting, and trafficking. What changes is who controls placement, moving it from the people who understand the code to the people who understand the commercial strategy. 
Read

Big Tech

Gemini and Claude close the gap
ChatGPT's share of AI prompt volume fell from 70% to 50% between January and June 2026, according to Comscore, as the market grows. Gemini nearly doubled its share to 30% and Claude climbed from 2% to 11%, with people increasingly using different assistants for different tasks. Overall AI use is growing fast: 35% of desktop users visited an assistant in June, up from 25% at the start of 2025. For publishers, there are two figures from the report that sting: Google searches show AI Overviews rose to 39.4%, and being used as a source doesn't guarantee a citation. For example, Trip Advisor was the source for 61% of travel answers, but only got 21% of visible credit. 
Read

The labs lobby Australia
With amazing timing for a frosty reception, Anthropic and OpenAI have asked Australian lawmakers to reconsider rules that block the training of AI models on the country's creative work without permission. Australia ruled out a text-and-data-mining exception last October, but both companies seek wiggle room to ignore it. Anthropic proposes a "narrow form of conditional approval", which is potentially tied to investment in Australian creators, while OpenAI wants a "balanced framework", with both of their asks tied directly to their data centre plans in the country. Creators aren't impressed, with one senator saying if AI firms want to train on copyrighted products, they should - and we know this is a really wacky idea - just pay for it.
Read

LinkedIn rebrands its AI tool
Battling the inexplicable suggestions that it is now a glorified AI content mill, LinkedIn, which recently pulled a tool that used AI to rewrite posts, has a new "Post Proofreader" tool for premium users which definitely does not use AI to write posts. Instead of generating new content, the tool shows inline suggestions with change tracking, so users can approve edits rather than having AI just rewrite their thoughts. The platform is looking to have AI fix your grammar, not fake your personality. 
Read