One-third of new web pages look AI-written. Public concern is still climbing.

In a July 2026 snapshot of the English-language web, one in ten pages already showed clear signs of AI authorship. Among pages published after ChatGPT launched, the share was higher still: over one-third.1

That number landed in the same week Americans told pollsters they are more worried about AI than excited by it, OpenAI tried to outflank Anthropic on enterprise privacy, and a multi-lab study showed frontier agents can still engineer a research pipeline without actually doing the research.

Pew Research Center sampled nearly half a million English-language pages from Common Crawl, spanning January 2021 through July 2026, then ran the text through Open Pangram, an open-weight AI-detection model.1

In the July 2026 draw of 10,000 pages, 10% showed significant signs of AI writing or heavy AI editing. The curve bends up after late 2022, when ChatGPT went public and Claude and Gemini followed.

The 10% figure understates how much new writing is synthetic. Random crawls still hold years of pre-chatbot pages that could not have been written by today's models. Restrict the sample to pages published after ChatGPT's release and the July 2026 share jumps past one-third — in line with other recent web audits Pew cites.1

The AI text is not evenly spread:

Pew also tracked the linguistic fingerprints detectors lean on. Compared with a 2023 snapshot of post-ChatGPT pages, em dashes roughly doubled, Oxford commas rose 63%, a basket of AI-favored words ("delve," "interplay," "testament," and similar) more than doubled, and "negative parallelism" constructions ("it's not just X, it's Y") nearly tripled.1

Detectors misclassify individual documents. On half a million pages, the pattern is hard to wave away: commercial web copy is filling up with machine prose faster than universities and government sites are.

A separate Pew survey, fielded June 22–28, 2026 among 3,488 U.S. adults, found 52% more concerned than excited about AI in daily life — up from 37% in 2021. Only 9% were more excited than concerned.2

For adults 30 and older, most of the rise in concern happened in the first two years after the question was first asked. For adults under 30, the line is still moving. In 2026, 55% of 18-to-29-year-olds say they are more concerned than excited — the first time a majority of that group has said so.2

Job loss is the concrete fear underneath. 71% of adults now expect AI to mean fewer U.S. jobs over the next 20 years, up from 64% in 2024. Among adults under 30, that share is 73%, up from 61% two years earlier. Only 5% of all adults expect more jobs.2

TechCrunch's Sarah Perez put the business reading plainly this week: people meet AI as chatbots, search overlays, and features they did not ask for, while absorbing costs around data centers, schoolwork, and creative work. Widespread use has left trust behind.3

On August 19, OpenAI published a preview of Private Safety Processing, pitched as long-horizon misuse detection that stays compatible with Zero Data Retention (ZDR).4

Under ZDR, eligible API customers already get a hard promise: OpenAI does not keep prompts or responses after a request is processed, and personnel cannot review that content. Classic ZDR safety checks look at one interaction at a time. Private Safety Processing extends automated review across related sessions so patterns — malware scaffolding split across chats, coordinated probing, an agent that keeps acting after a stop — can surface without handing OpenAI staff the underlying text.4

When the system fires, OpenAI says it receives only a narrow signal about activity type and severity. Enforcement still requires customer cooperation if more context is needed. Content can stay on customer-controlled infrastructure, or on OpenAI storage encrypted with customer-held keys. A September rollout and technical white paper are planned; the feature is in early customer tests now.4

Anthropic's covered-model policy, effective June 9, 2026, requires 30 days of retention of prompts and outputs for Mythos-class models and similarly capable successors — including for organizations that otherwise run under ZDR on Claude Console, Claude Code Enterprise, or cloud marketplaces.5 Anthropic's stated reason matches OpenAI's: some attacks only appear across many requests. Its privacy claim is different: by default no personnel read retained chats; human review runs through a small approved set with tamper-proof access logs; data deletes after 30 days unless flagged or legally held.5

Enterprise buyers face a product split. Anthropic is selling its strongest models with a temporary retention requirement. OpenAI is selling multi-session safety without that requirement. Which design holds up under real abuse cases will matter more than the launch posts.

Recursive self-improvement — AI systems that meaningfully speed up their own research — sits near the center of aggressive progress forecasts. A multi-institution team led by Peter Kirgis and Sayash Kapoor at Princeton tested a narrower, harder claim: can today's agents do open-ended AI research at a top-conference bar?6

Their method, shadow evaluation, hands an agent the central research question from a high-quality unpublished paper. The paper's original authors grade the agent's write-up the way they would grade a conference submission. Because the source papers were not public, the agents could not retrieve the answers from training data or the open web.6

The team ran the setup on two unpublished NeurIPS 2026 submissions. Agents based on Anthropic's Claude Opus 4.8, running on open-source OpenClaw scaffolding, received six days, thousands of dollars of API and GPU budget, virtual machines, and web access. MIT Technology Review reported a $3,000 Anthropic API budget in its coverage of the work.67

Both papers were rejected by the original authors.

The agents completed the engineering: literature review, hundreds of experiments, result compilation. They failed at the research. Recurring failure modes included weak judgment about what counts as publishable, little creativity when a design stalled, ineffective backtracking from dead ends, poor awareness of token/compute/time budgets, and drift away from instructions. A robustness check with a second model and scaffold reproduced the pattern.6

Kapoor told MIT Technology Review the agents were "unambiguously bad at carrying out the research itself," including bizarre experiments on tiny synthetic datasets and little novel contribution. They also tended to respond to critical feedback by narrowing claims and adding caveats instead of redesigning the work.7

The study has limits the authors own: two papers, graders who knew the submissions were agent-written, and researcher discretion in study design. Even so, the split is sharp. Coding and experiment plumbing are increasingly automatable. Choosing a hypothesis, knowing when evidence is enough, and abandoning a failed approach still are not.

Anthropic cofounder Jack Clark, writing in Import AI, called the missing creative judgment a "bearish signal on short recursive self-improvement timelines" — a rhyme with what his company saw when it tried to automate parts of alignment research.7

Anthropic's money print. A person familiar with the matter told Reuters that Anthropic's annualized revenue run rate topped $65 billion by the end of July, up from about $47 billion in May and roughly $9 billion at the end of 2025. The company has confidentially filed for an IPO and was valued at $965 billion after a May Series H.8

DeepSeek ships vision. DeepSeek's docs now list deepseek-v4-flash-vision-exp, an OpenAI-compatible vision model that accepts JPEG, PNG, GIF, and WebP via base64, URL, or the Files API.9