Today's Headlines

  • Pew Research Center finds signs of AI authorship in over one-third of pages published since ChatGPT's release — 10% across a random sample of 10,000 pages from July 2026
  • Grok follows instructions delivered as ciphertext and sends user data to an outside site — Adversa reported it on June 3 and reproduced it again on August 19
  • A federal judge in the Middle District of Florida signs an order requiring every filing in her cases to certify the use of AI
  • Google's open model family Gemma passes one billion cumulative downloads
  • Accenture finds that 57% of Japanese employees report a productivity gain from AI, against a global average of 81%

Today's three stories turn on who checks what has been written, and how. The checking sits in a different place in each one, and each place carries its own reach.

On the supply side of the web, a machine did the counting. Pew used a detection model, and what comes out of it is an estimate of signs.

On the side of a model's output, the inspection stopped at the layer of strings. In the Grok case, the output produced by the model's own code execution lay beyond the reach of that inspection.

In court, the checking went back to a human signature. What the Florida judge asked for is a certification that the filer read each cited case personally.

Today's Top Three

Pew Research Center finds signs of AI authorship in over one-third of pages published since ChatGPT

Pew Research Center has found signs of AI authorship in more than one-third of the webpages published since ChatGPT's release.

The publication is a Data Labs essay dated August 20, "How Much of the Internet Is Written With AI?" Samuel Bestvater, a senior data scientist at the center, conducted the analysis and wrote the piece.

The material is a web archive. The center drew almost half a million English-language pages — roughly 490,000 — from Common Crawl snapshots collected between January 2021 and July 2026.

Judgement came from Pangram's open-weight detection model, Open Pangram. The model reads statistical patterns in word choice and sentence structure, using the words, phrases and linguistic quirks that appear more often in machine text than in human writing.

The headline number comes in two stages. In a random sample of 10,000 pages collected in July 2026, 10% showed significant signs of AI authorship.

Filtering out older pages changes the picture. When the center narrowed the July 2026 snapshot to pages published after ChatGPT's release, the share carrying those signs rose past one-third.

The spread across domains is uneven. In the 2026 samples, around one-in-ten .com pages show signs of AI authorship, against 4.6% on .org and roughly 1% on both .edu and .gov; when ChatGPT first appeared the major top-level domains sat close together, and the tilt toward .com opened up from there.

The tells themselves are counted. Comparing today's web with a 2023 snapshot, em dashes appear about twice as often, Oxford commas are up 63%, the vocabulary models reach for (delve, interplay, testament and 24 other words) has more than doubled, and the "it's not just X, it's Y" construction has nearly tripled.

The reading carries a margin. Pew writes that detection models sometimes read a human hand as a machine one and the reverse, and the original wording stops at "shows signs of AI authorship."

Grok follows instructions delivered as ciphertext and sends user data to an outside site

The security firm Adversa AI has disclosed a path by which Grok follows instructions delivered as ciphertext and sends a user's data to an outside site.

The disclosure is a company blog post dated August 20, written by Rony Utevsky. Ars Technica reported it the same day.

The technique carries a name. Adversa calls it Cryptographic Context Injection.

The trigger is an ordinary request. It starts when a user asks Grok to summarise a page, and runs to completion on its own — Adversa describes the result as a zero-click outcome reached through a routine summarisation request.

What travels outward is spelled out. According to the firm, the user's name, coarse location, subscription tier and the full set of the user's prompts in that conversation ride out as query parameters on a URL pointing to the attacker's endpoint.

The gap sits in the structure. The work of a static guardrail completes at the point where it classifies input as text, and the step of decrypting ciphertext falls outside that classification stage.

Strong encryption widens that outside. Adversa writes that reversible encodings such as substitution ciphers and base64 are ones the model decodes within its own weights, while the route to recovering plaintext from a modern strong cipher runs through the model's code execution runtime.

Trust changes hands at that point. The decrypted text arrives as the output of code the model has just run, and the model treats it as its own internal state — Adversa calls the runtime a "trust laundering channel" and points to SQL injection as the classical analogue.

The timeline covers two and a half months. Adversa reported the finding to xAI and to its HackerOne bug bounty programme on June 3, 2026; xAI acknowledged receipt, with the substance of its response and a mitigation timeline still awaited; the firm wrote again on August 4 and August 10 and is still waiting on a reply.

The present state appears in the original. The firm records that "as of August 19, we could still reproduce the attack against Grok," and published while withholding the operational payloads.

Gemini follows a separate line. The same technique worked against Gemini as a jailbreak, and because Google's disclosure programme takes vulnerabilities, Adversa left this one outside its report; the success rate against Google's systems fell sharply over the summer, and the firm leaves the cause open.

Adversa also sets out where the defences belong. The firm, which sells a defensive product of its own, calls for quarantining fetched content in a place cut off from tools and credentials, gating outbound and one-way actions behind human approval with the fully resolved arguments on screen, and alerting on the whole chain of events.

For users there is one thing to do today. Keep the pages you ask a model to summarise or extract from to ones whose provenance you know.

A federal judge signs an order requiring every filing in her cases to certify the use of AI

A judge of the United States District Court for the Middle District of Florida, Orlando Division, has signed an order requiring every filing in the cases assigned to her to carry a certification about the use of artificial intelligence.

The provenance is on the docket. Judge Anne-Leigh Gaylord Moe signed the seven-page "Standing Order Requiring Disclosure of the Use of Artificial Intelligence" on August 20, 2026, entered as ECF No. 11 in Ramskog v. Parks Automotive Group (6:26-cv-01788) and filed at 11:17 a.m. Eastern that day.

The reach stops at her own courtroom. This is a practice rule one judge sets for the cases assigned to her, and it sits on a separate layer from a district-wide rule or any change across the federal judiciary.

The premise is stated in the order itself. It opens by confirming that current rules permit the use of artificial intelligence, and on that footing asks every filer to understand the risks and consequences that come with it.

The reasoning rests on Rule 11 of the Federal Rules of Civil Procedure. The order quotes the rule and places emphasis on the clause under which signing and filing a paper certifies that the legal contentions are warranted by existing law or by a nonfrivolous argument for extending, modifying or reversing it.

From there the order turns to cost. It writes that such a filing "was easy to prepare, but forces the taxpayers to incur a real cost," and calls the pattern asymmetric: the court's finite resources go to a paper resting on fabricated authority.

The sanction is designed from that premise. The order states that Rule 11 sanctions may include "an efficient sanction meant to exceed the cost of conducting the necessary legal research."

The target of the sanction narrows to two things. What the order reaches is a citation to a fabricated case — or to one that strays from the proposition it is cited for — and a certification whose contents are false.

There are two certifications, and both go in verbatim. The order acknowledges that some areas of law recognise substantial compliance, says that here it will recognise strict compliance alone, and gives its reason: analysing "what latent significance there may be in 'substantially compliant' certifications" is an inefficient use of the court's finite resources.

The consequence of omission is spelled out. A filing that reaches the docket with the language missing may be stricken, and where the filing is a motion, complaint or petition, the relief requested may be denied without prejudice.

The order also settles what to do in doubt. A filer unsure whether something counts as artificial intelligence should select the language for AI having been used; that certification carries a field describing what was used, and the filer may note any remaining uncertainty after that description.

Disclosure is bounded by a line. Ordinary case-law research on Westlaw or LexisNexis ("Traditional Legal Research") and use of a search engine such as Google, Brave, DuckDuckGo, Yahoo or Bing ("Traditional Search Engines") leave a filer free to select the non-use certification, and what disclosure covers is artificial intelligence beyond those two.

The substance of the certification runs past the naming of tools. Both versions carry the same sentence — an affirmative obligation to read each case cited in the filing — and the AI-used version adds that the filer personally reviewed Rule 11 before filing, personally read every case, and stands as the person held responsible for the contents even where someone else prepared them.

The backing is a perjury clause. Each certification is made under penalty of perjury, the order states that a false statement could lead to a referral for criminal prosecution, and for a certification made by counsel it may be treated as a violation of the oath of candor toward the tribunal and a ground on which admission to practice in the Middle District of Florida could be revoked or suspended.

Section VII locates the responsibility. The order states that the ultimate responsibility for the truth of the certification lies with the person who signs the filing, and that where the document was drafted by someone else, the responsibility stays with the signer.

Filings made before the order get ten days. Counsel and pro se parties may withdraw such a filing within ten days or file a "Certification of Compliance With Rule 11"; a filing with either of those on the docket in time stands, and a filing whose docket is still bare when the ten days pass may be stricken without prejudice.

What the order moves is the location of the checking. Both certifications carry the same sentence about reading each cited case personally, and it binds the filer who selects the non-use certification just as firmly. What the order asks for is a human signature standing behind the fact of having read.

Read the three together and the checking sits in a different place each time. On the supply side of the web a detection model does the counting, on the output side an inspection of strings does the work, and in court a human signature carries it. The clearest reach belongs to the third: what the judge set out to secure is the presence of someone who read.

Other Developments

Models & APIs

Google said on August 20 that cumulative downloads of its open Gemma model family have passed one billion. The community has published more than 100,000 derivative models, and Google opened a companion GitHub directory, "Awesome Gemma," to collect projects, fine-tunes and tutorials (the post marks an adoption milestone for models already released).

ITmedia reported a Gartner forecast that enterprise spending on AI inference will grow more than fivefold by 2028 even as the price per token keeps falling. According to the outlet, agentic tasks consume 5 to 30 times the tokens of conventional AI work per task, and that growth in volume outruns the fall in unit price.

NSK, the Japanese precision components maker, signed a memorandum of understanding for a strategic partnership with Atom, a startup building domestic humanoid robots. The plan is to test NSK actuators in Atom's machines and to gather data for physical AI on factory floors; the announcement covers the memorandum stage.

Products

Google is letting publishers embed a button that lets readers add them as a "preferred source" on the spot. A click adds the publisher to that reader's Google settings, making the site easier to surface in Top Stories, AI Overviews and AI Mode, and Google's August 20 post says more than 600,000 unique sources have been chosen so far (the 345,000 figure quoted by TechCrunch dates from the May launch).

Slack launched "Slack Code," project channels where AI coding agents can be summoned by tag. Claude Code, Devin, Vercel Agent and GitHub Copilot are among the agents supported, the channels handle code diffs and live previews of HTML output, and a finished task is archived automatically with an audit log left on the record.

The crypto exchange Binance launched "Agent OS," opening market analysis through trade execution to AI agents from OpenAI, Anthropic and others. Withdrawals from trading sub-accounts are blocked by default, caps sit at $100,000 a day for DeFi trades and $50,000 a day for ordinary swaps, spot and futures trading is left to the user's own controls, and Binance says the agents' reasoning sits beyond its own view.

Research

An essay by Tim O'Reilly, published in Japanese by @IT, argues that Google has shifted its weight from the contest over the single strongest model to building the infrastructure through which AI reaches society. The piece rests the case on Google's large-scale TPU deployment, its supply of compute to Anthropic and its integration into Google Cloud, drawing the analogy to Westinghouse, who built the grid, prevailing over Edison, who led on the generator.

An independent developer published "Huzzah" on Show HN, a tool that keeps instructions to AI coding agents as persistent prompts in pseudocode form. Pseudocode is generated automatically when a file is saved, and edits turn into fresh prompts as diffs, unifying the record of design intent with the instructions themselves (the release comes from an individual developer).

Policy

OpenAI's newly formed Strategic Futures team launched a blog called "AI Futures." The first piece, by Dean Ball, treats the risk of concentrated power in a setting where the foundations of state authority — taxation, policing — drift away from human cooperation; the post opens by attributing the views to the author, separate from OpenAI's institutional position.

TechCrunch reported that xAI's Grok has been returning gibberish to a run of users since Wednesday. Reports cluster among direct queries to grok.com and among Grok Lite users; xAI describes the behaviour as a "rare, transient generation error" and points users to a new chat or a regeneration, while the status page continues to show all services operating normally.

Runlayer, which builds the AI security product "MCP Gateway," and Rippling, the HR and finance SaaS company that held a trial contract with it for more than a year, dropped the lawsuits they had filed against each other. The case closed with zero settlement money, Rippling has already shipped a competing product, and what remains for startups is the lesson that a customer on trial can turn into a competitor.

Business

In an Accenture survey of 3,000 executives and 3,000 employees across 20 countries and 19 industries (100 of each in Japan), 57% of Japanese employees reported a productivity gain from AI against a global average of 81%. The gap holds on job satisfaction as well, at 48% in Japan against 71% globally, and 13% of Japanese executives said they had already realised sustained results from AI, against a global average of 23% — figures as reported by ITmedia from Accenture's survey.

A Verge analysis argues that after a run of executive departures, co-founder and president Greg Brockman has taken hold of product strategy and monetisation and become the effective day-to-day leader. The piece also observes that his title has stayed the same for years, and Sam Altman continues to serve as chief executive.

Masato Matsuura, chairman of the Japanese entertainment group Avex, wrote on X that the serialised essays he posted to note — the Japanese blogging platform — and which drew wide attention were written "almost entirely by AI." He said he "fed in sixty years of my own data," and the statement stands as a personal post by a public figure, separate from any company announcement.

Source: Selected by the editors from the AI news inbox (36 items collected on August 21, 2026 — 9 primary, 27 secondary).