A third of the new web is machine-written
Pew ran a detector over half a million pages and found that more than a third of everything published since ChatGPT shows signs of machine authorship. Google finished a global spam update three days ago. The two facts are less connected than they look, and more useful together.

- 01
More than a third of pages published since ChatGPT show signs of AI authorship
Pew published on 20 August after running the Open Pangram detector over roughly 490,000 English-language pages pulled from Common Crawl between January 2021 and July 2026. Across the whole sample about 10 per cent of pages show significant signs of AI authorship, but among pages published after ChatGPT's release that figure passes one in three, and commercial sites carry it hardest: around 10 per cent on .com against roughly 1 per cent on .edu and .gov. The researchers are careful to say detectors are not reliable on any single document and that the finding holds as an aggregate pattern, not a verdict on individual pages.
Pew Research Center - 02
Google's August spam update finished rolling out
Google released the August 2026 spam update on 18 August and marked it complete on 21 August, a rollout of two days and sixteen hours. It applies globally and to every language, and Google describes it as a normal spam update, meaning it targets sites breaking the search spam policies rather than any particular publishing method. If your rankings moved last week and nobody can explain why, this is the first thing to check the dates against.
Google Search Status Dashboard - 03
OpenAI cut the price of its frontier model by a fifth for developers
From 21 August, GPT-5.6 Sol costs $4 per million input tokens and $20 per million output tokens, down from $5 and $30, so a 20 per cent cut on the way in and a third off on the way out. OpenAI's pricing page states the promotional rate runs at least through 21 November 2026, and it applies to the API and to credits on ChatGPT Work and Codex, not to Plus, Pro or Business subscriptions. If somebody built you a tool that bills per use, your running cost just dropped without you doing anything, and it may quietly go back up in November.
OpenAI - 04
A critical Elementor Pro flaw let strangers upload PHP to WordPress sites
CVE-2026-32475 carries a CVSS score of 9.0 and affects every version of Elementor Pro before 4.2.2, which was released on 19 August. The upload field in the Forms module validates files in one loop and moves them in another, and the two loops disagree about empty entries, so a crafted submission slips a PHP file past the extension check and into a public uploads folder. It only bites where a published form has a file upload field with multiple uploads switched on, which is off by default, and no exploitation had been seen at the time of disclosure. Check the plugin version on any site you own.
Patchstack - 05
Nine out of ten leaked AWS keys still work, and the median one is five years old
Truffle Security re-tested 10,616 AWS key pairs that had been sitting in public places between August 2022 and August 2026, and on 10 August this year 88 per cent of them still authenticated. Of the live ones, 768 belonged to companies and carried full control of the account: 526 root keys and 242 users holding AdministratorAccess. Only 13.7 per cent of keys had ever been rotated, and the single largest source was Hugging Face, where 8,482 keys were sitting in public datasets uploaded by people sharing their AI work.
Truffle Security - 06
Salesforce partners report no bookings revenue from Agentforce, two years in
The Register reported on 21 August on a TD Cowen survey of Salesforce implementation partners across the US, Europe and Asia. A third said interest in Agentforce is strong and customers are trialling it, 56 per cent expect interest but say the initiatives need time to mature, and none of them said it had become a driver of bookings. Across the wider portfolio a third of partners met or beat their targets, down from 43 per cent the quarter before, which is worth holding next to any vendor claim about how fast AI agents are being adopted.
The Register - 07
Torvalds credits AI with the grunt work on a kernel bug it kept calling unsolvable
On 21 August Linus Torvalds committed a one-character fix to the Intel Xe graphics driver, changing a rounding function after 24 debug patches and 18 kernel boots to corner a memory bug on Battlemage cards. His own commit message calls it "a debug session from hell, enormously helped by an AI doing much of the grunt-work", and he notes that the AI repeatedly insisted the problem was impossible and unsolvable. Both halves of that are the story: it did the tedious instrumentation faster than a person would, and it was confidently wrong about the conclusion until a human refused to accept it.
Linux kernel git
A third of everything published since ChatGPT is machine-written, and Google closed a global spam update three days later. The tempting reading is that the second is a response to the first, and it is not. Google has said for years that it does not care how a page was produced, only whether it is worth landing on, and a normal spam update goes after the same tired tricks it always did. What has actually shifted is the ratio, and the ratio is the interesting part. If a third of new pages are generated, the average page your customer lands on is thinner than it was two years ago, and the bar for being the useful one on the results page has quietly dropped.
For a firm of a dozen people that is good news, provided you do not join in. The one thing a model cannot produce is what happened on your job last Tuesday: the flat where the holes were cut in the wrong place, the plenum that did not fit the window opening, the photo of the finished riser at half six in the morning with the bag of offcuts still in shot. From direct experience in construction I know how rarely any of that gets written down, and how well it reads when it does, because it is specific in a way that no amount of generation can imitate. Pew also found that em dashes have doubled since 2023 and Oxford commas are up 63 per cent, which is a roundabout way of saying the machine writing has developed a house style. Readers are learning to spot it, whether or not they could name what they are spotting.
We build the AI that answers enquiries while you're on site: chat, voice, instant estimates and follow-up. See how it works or price it in two minutes.