Fumbled maps and presidents, stealthy jailbreaks, and Perplexity's Chrome bid

Wikipedia's battle, the end of the dial-up era, and a social media for AIs - all in this week's Content Aware.

Published: 15:19, 14 August 2025

Corbidge comments on...byline blunders and bot trouble
In a move that is raising more eyebrows than a botched headline, Politico is now in hot waters with its staff union, PEN Guild, over two AI tools, LETO and Report Builder, which are claimed to be spitting out less-than-accurate content next to real journalists' bylines. Having the tools "sitting outside of the newsroom" and "not reflecting reality" isn't really a confidence booster for a newsroom built on facts. Our resident sceptic Rob discusses the thought of handing the pen to a glitchy algorithm.
Read

AI fumbles maps, presidents, but nails 007
Despite a grand fanfare, OpenAI's new GPT-5 can quickly be made to fail just like its predecessors. You want US States? Have "Onegon" and "Gelahbrin". Fancy a timeline of presidents? Meet Willian H. Brusen, in office circa 1991. The model will give correct answers in plain text, and even manages to generate visuals, but asking for images ends up with an alphabet soup. GPT-5 is not the only one at fault: Bing Image Creator thinks it's "United States of Ameriicca", and Gemini's state names read as if a cat sat on the keyboard. Oddly, the AI tools suddenly find their footing when asked about James Bond actors. Maybe secret agents are easier to track than US presidents.
Read

Cracking GPT-5: the rise of stealthy AI jailbreaks
Researchers uncovered a clever new way to jailbreak GPT-5 and sneak past safety guards using a mix of storytelling and subtle manipulation. The technique, called Echo Chamber, poisons the AI's context bit by bit, which leads to spilling of harmful info. Meanwhile, "zero-click" attacks such as AgentFlayer are quietly exploiting AI's ties to Google Drive, Jira, and Microsoft Copilot. As AI evolves and becomes more connected, these hacks expose how tricky it actually is to keep our digital assistants helpful but also safe.
Read

Is AI the mirror, or the issue?
Former Google X exec Mo Gawdat has seen where AI is heading, and he doesn't like it one bit. He warns of an inevitable 12-to-15-year dystopia starting in 2027, but not because AI is evil, but because it magnifies our flaws, instead of helping us emphasise our positive traits. Rather than being a helpful companion and easing our workloads, AI is fuelling layoffs, deepfakes, surveillance, and scams. According to Gawdat, the problem isn't the tech itself but the way we use it. "AI is just a hammer," he says, "but it's up to us whether we build with it or harm with it." There is a chance of a utopian AI future, but only if we sit down and regulate how we use AI, not just the design of it.
Read

End of an era
In a move that feels like it should've happened a decade ago, AOL will finally shut down its dial-up internet service on September 30, 2025, after long 34 years. That dial up sound that an entire generation became familiar with will sound no more. While most of us moved on a long time ago, a few thousand users, mostly in remote areas of the world, still relied on it.
Read

Reddit blocks archiving to thwart AI
Reddit is now blocking the Internet Archive's Wayback Machine from indexing most of its content. It came after spotting sneaky AI firms bypassing scraping bans by harvesting data from archived pages. In a swift "I don't think so" move which is citing privacy concerns and rule-breaking, Reddit now only allows homepage snapshots to be saved, crippling the archive's ability to preserve deleted posts or peek into user antics. Meanwhile, the Internet Archive remains mum, seemingly without a comeback in sight.
Read

Have your say: IPTC and the end of the Slopocalypse
Metadata masterminds the IPTC wants to stem chaos with a new proposal to bring order to AI-generated images and as members of the IPTC we at Glide are calling on friends and customers to help shape industry standards. With a little help in the metadata fields, the prompt entered, who wrote it, and any reference images involved, the overall tracking of AI imagery can improve. Optional, yes. Necessary? Absolutely, if we want transparency over mystery visuals. Drop your thoughts via the IPTC Contact Form or post to the public discussion list by Friday 29th August.
Read

Social media: is the tech the problem, or is it us
According to a new study from the University of Amsterdam, with or without algorithms, social media doesn't really change whether it's humans using it or chatbots. Researchers let 500 AI chatbots loose on a stripped-down, algorithm-free social media platform and sprinkled some political leanings into the mix, which led to the instant formation of echo chambers, the like-minded flocking together, and rewarding the most extreme voices. Sounds vaguely familiar? They did try to curb the chaos a bit with tricks such as chronological feeds and hiding follower counts but nothing really worked, if anything, they made it worse. The takeaway? When it comes to social media, maybe the problem isn't the tech, but us.
Read

Perplexity tries to snag Chrome
AI startup Perplexity just threw down a staggering $34.5 billion all-cash offer to buy Google's Chrome browser, which is nearly the double of its worth. Bold move or just a PR stunt? Only time will tell, since this offer lands amid a court fight over Google's alleged monopolistic ways. With Chrome owning over 60% of the browser world and acting as the gatekeeper to search, if Google is forced to sell, the entire online ad and search ecosystem could get flipped upside down. Even though we can't really tell whether this is a serious bid or if it is more about headlines than handshakes, the question who will own the browser throne is still there.
Read

GitHub CEO out, Microsoft in
There is a big change at GitHub: CEO Thomas Dohmke said bye-bye to go start a new venture, and Microsoft isn't replacing him. What's going to happen instead is now the leadership is directly reporting to Microsoft's CoreAI team, led by ex-Meta exec Jay Parikh. Will we now have an AI-powered Microsoft machine instead of the indie GitHub we all loved? Time will tell.
Read

Wikipedia vs UK's Online Safety Act
Wikipedia has lost a bid challenging its possible labelling as a "category 1 provider" under the UK's new Online Safety Act (which has already caused a commotion you wouldn't really expect). Such labelling would place some onerous duties on Wikipedia, such as verifying the identity of users, something that would fatally erode the core of what Wikipedia is. The UK High Court didn't side with the Wikimedia Foundation, but the judgement emphasised that regulators must avoid harming its operations. Looks like a badly written law if such ambiguity is apparent. While Wikipedia plans to continue seeking protection, the UK regulator Ofcom is expected to publish a list of categorised services soon.
Read