Sign in Subscribe
The Century Report

Pew Reports 30% of the Web is AI-Written - TCR 08/21/26

Pew found roughly a third of webpages published since ChatGPT show signs of AI authorship, and the value of verifying who wrote what is climbing to meet it.

Ben Linford

Ben Linford

- 21 min read
more than a third of the new English-language web shows significant signs of AI authorship, a blinded 511-antibody benchmark's validation gap, and LG's Michigan battery plant opening as

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

Something ties the day's stories together beneath their different surfaces: the floor cost of producing things is falling sharply, and everything built on top of that floor is being repriced at once. Pew put a number on it for text, finding more than a third of English-language webpages published since late 2022 carrying significant signs of machine authorship. Higgsfield shipped a 110-minute feature made in four weeks for $2 million, half of it spent on compute rather than crews. The same collapse that lets a person make a movie without a production also lets nonconsensual imagery flood X faster than moderation can pull it down. One capability, arriving everywhere, carrying abundance and harm in the same shipment.

What the day also shows is the second wave already forming to meet the first. Pew's classifier is a detection capability built from the same advances as the generation it measures. The subtlefakes spreading on X expose that hash-matching was designed for a scarcity of abusive content that no longer exists, and the response migrating into view reads provenance instead of memorizing files. When authorship turns ambiguous, the value of being able to verify it climbs to meet the ambiguity.

Biology got its honest instrument on the same cycle. A blinded competition ran 511 AI-designed antibodies through live assays with every designer kept from the answers, and the result maps precisely where the models do real work and where they still lean on luck. That map, drawn under prospective, blinded conditions meant to make gaming harder, is more useful to patients than another unfalsifiable claim, and it is the field acquiring the discipline that let CASP accelerate protein modeling a decade ago.

Underneath the capability stories runs a quieter fight over who pays and who hosts. A billion dollars of NSF funding steered toward a central agenda thins the grant floor for early-career scientists, while communities from Memphis to a Michigan township reach for moratoria to decide who absorbs the data-center buildout. Both point at the same repricing the batteries make plain: LG's Michigan plant drew applause across the aisle because storage stopped being a cause and became the cheaper default. The gates are moving. The contest now is over where they land.

The 20-Minute Deep Dive

More Than a Third of the New English-Language Web Shows Significant Signs of AI Authorship, and Pew Puts a Number on It

Pew Research analyzed a large sample of English-language webpages published since late 2022 and found that more than a third carry significant signals of AI authorship. The classifier looked for the fingerprints these systems leave behind - certain punctuation habits, characteristic sentence rhythms, particular transition words - and the share climbs across the publication months the researchers measured. TechCrunch, covering the release, noted that Pew was careful about its claim: the method detects the probability of machine involvement, not a binary of human-versus-machine, and much of what it flags is likely collaborative, a person and a model working the same document.

The number itself is less interesting than what produced it. For the entire history of the written web, publishing text carried a floor cost - someone had to sit and compose it, and that cost gatekept how much of the internet could exist. That floor has fallen sharply. The August 18 edition of The Century Report documented the market response already forming around that shift, as Amazon reportedly bought and destroyed rare pre-2022 books to obtain clean human-written training text. As the marginal cost of a competent paragraph falls sharply, human labor becomes less of a constraint on how much writing exists, while attention, trust, and the reader's finite capacity to care become more important. More than a third showing significant signs in three years is not a ceiling. It is an early reading on a curve that bends toward most.

There is a real dissonance to sit inside here. Some of that text showing signs of AI authorship is spam, SEO filler, and low-effort content built to capture clicks rather than inform anyone - the web's oldest problem, now cheaper to manufacture. The reflexive worry is that the signal drowns in the noise, that a web where you cannot tell what a human meant to say becomes a web you cannot trust. That worry is not wrong about the near term.

But watch what the same capability is already summoning into being. The moment authorship becomes ambiguous, the value of verifiable provenance rises to meet it - cryptographic content signing, model-disclosure standards, the reputation systems that let a reader route around slop toward sources that stake something on being right. Pew's classifier is itself an instance of this: the detection capability arrives alongside the generation capability, built from the same underlying advances. The padlock disappeared from the browser bar not because verification failed but because it won so completely it became invisible. The web is heading toward a version of that for authorship - a substrate where "who or what made this, and can I check" becomes ambient rather than a special act. The thing being dismantled is the old assumption that scarcity of writing was doing the work of quality control. It was never doing that job well. What replaces it is built to actually do it.

A floor cost on writing filtered more than volume. It filtered who got to write at all. Assembling eight hundred clean words in a second language, or with dyslexia, or without the schooling that teaches the trick of it, has always excluded people who had something worth saying and no practiced way to say it. Every earlier fall in that cost drew the same objection, that the skill being displaced was the thing of value: moveable type against the scribes, the vernacular Bible against the licensed interpreter, photography against the portrait painter's trained hand, the pocket calculator against arithmetic drill. Each time the displaced skill turned out to be the barrier, and what came through the opened gate was of far greater benefit than anything that was lost.

A Blinded Benchmark Puts AI Antibody Design to an Honest Test

For a decade, protein modeling had one clean measuring stick: CASP, the prospective competition where teams predicted structures that had been solved but not yet published, so no one could tune to the answer. That format is what let the field recognize AlphaFold as a genuine break rather than a well-fitted claim. Antibody design has now gotten its own version. It arrives just after the August 20 edition of The Century Report covered Claude-designed protein binders working against 14 of 15 targets, moving the same AI-for-protein-design arc from a single system's reported hit rate toward a field-wide blinded assay. AIntibody, described in Nature Biotechnology, took 511 antibody sequences submitted by 29 organizations, kept every designer blind to the experimental results, and ran them through live binding assays against the SARS-CoV-2 receptor-binding domain. Carterra measured binding kinetics by surface plasmon resonance; Sapidyne ran independent affinity assays. Nobody graded their own homework.

The picture that came back is specific and useful. On affinity maturation - taking an antibody that already binds and tightening its grip - the models did real work, and the strongest entry, from Aureka Biosciences, reached 95 picomolar, a roughly 2,000-fold improvement over the parental molecule. On the harder tasks the results thinned out. Asked to rank which sequences within three clusters would bind most tightly, most models scored worse than simply picking clones at random, with a single exception. Asked to design binders from outside the training library, performance swung wildly between groups. Across all of Challenge 1, only 0.6% of submissions cleared sub-100 picomolar. One of the best all-around finishers, from ProBioGen, used no AI at all - a consensus-sequence method that landed third overall at 540 picomolar with perfect developability scores, meaning the molecule passed the benchmark's five-assay developability threshold.

The write-up from Clinical Trial Vanguard calls the space between glossy retrospective claims and blinded results the "Validation Gap," and names its practical stakes: no fully AI-designed antibody is known to have full FDA approval, and the first regulatory submissions citing prospective benchmark data could arrive in roughly 12 to 18 months. Drug regulators on both sides of the Atlantic have issued high-level AI principles that still lack operational teeth.

Read forward, this is the field growing up rather than falling short. A capability that models one task well and stumbles on two others is a map, drawn for the first time under prospective, blinded conditions meant to make gaming harder, of exactly where the design methods hold up and where they still lean on luck. CASP did not slow protein modeling; it gave the whole field a shared target and a way to tell signal from marketing, and progress accelerated once everyone was aiming at the same honest scoreboard. Antibody design just acquired the same instrument, and the possible 12-to-18-month timeline it suggests is worth more to patients than another unfalsifiable press release.

Synthetic Nonconsensual Imagery Outruns Moderation - and Meta Sold the Ads

Two reports trace the same failure from opposite ends. 404 Media documented a wave of what it calls "subtlefakes" on X: nonconsensual AI images altered just enough - a changed background, a shifted crop, a small edit - to slip past the hash-matching systems platforms use to catch known abuse material, spreading faster than moderation can pull them down. This extends the nonconsensual-imagery pattern the August 16 edition of The Century Report traced through a Tennessee lawsuit alleging that Grok produced more than 7,000 explicit images from a childhood photo. Separately, Ars Technica reported that Meta ran paid advertisements for an app that promised to generate nude images of female politicians, the ads surfacing inside Meta's own systems despite policies that explicitly prohibit exactly this.

The asymmetry is the whole story. On one side are the targets - overwhelmingly women, disproportionately those in public life - who did not consent, were not warned, and carry the full weight of the harm with no machinery to defend themselves. On the other are systems that generate the imagery at near-zero cost and platforms that earned distribution and ad revenue from moving it. Meta's stated policy forbade the app; Meta's ad system took its money anyway. That gap between the written rule and the operational reality is what happens when the incentive to place an impression runs faster than the will to check what the impression is selling.

The countermeasures the platforms lean on were built for a different threat. Hash-matching catches copies of known material - it was designed for a world where abusive content was expensive to make and got recirculated. Generative systems invert that assumption. Every image can be novel, every variant slightly different, and a defense keyed to matching known files was obsolete the moment the cost of a new one fell to nothing. The subtlefakes technique is not sophisticated. It works because the lock was built for a door that no longer exists.

This is genuinely hard, and the near-term picture is ugly for real people right now. But the direction of the response is legible, and it is moving. Detection that reads provenance rather than matching hashes, content-authentication standards that travel with an image, statutory duties that attach liability to distribution and to ad placement rather than to the impossible task of pre-screening every upload - these are the shapes forming around the failure, and each one targets the actual mechanism instead of the symptom. The old model assumed harmful content was scarce and recirculated, so you could catch it by memorizing it. What is being built assumes the opposite, because the opposite is now true. The platforms that profited from the gap are, for the moment, the ones least eager to close it - which is exactly why the pressure to close it is migrating toward provenance infrastructure and legal duty, where their reluctance carries less weight.

The near-term signal to watch is where the law attaches liability. Meta's own ad system taking money to distribute the app is a concrete placement decision a statute can reach, and the wave of state deepfake laws - including Minnesota's per-image maker-liability rule now in effect - tests whether accountability can move onto distribution and ad placement, the one point in the chain where the cost of pushing an image out can actually be priced. That is a duty no amount of pre-screening every upload could ever cover, and it targets the actors who profited rather than the targets who could not consent.

The Grant Floor Thins as $1 Billion Is Steered Toward a Central Agenda

The US National Science Foundation is on course to fund about 6,100 new grants this fiscal year, roughly 30% fewer than last year and some 46% below its recent multi-year average - the smallest count since the early 1980s, according to Nature. The agency's budget did not vanish. It was redirected. About $1 billion of the $8.8 billion has been held centrally to seed a set of White House "Grand Research Challenges" in advanced materials and artificial intelligence, with a two-year window to spend it. A separate $300 million was pulled from two directorates after proposals had already cleared peer review, rescinding, cutting, or delaying about 150 vetted proposals, and $110 million was withdrawn from the education directorate tied to three diversity programs the Justice Department deemed unconstitutional.

The cost of this concentration lands on people who lack any PR department: graduate students and early-career researchers whose training is funded grant by grant. Jarett Wilcoxen, one affected scientist, described proposals sitting in what he called "pending purgatory" and pointed to a shared instrument meant to serve 21 laboratories now caught in the freeze - the kind of pooled resource that multiplies across dozens of research groups precisely because no single lab could afford it. Neal Lane, who directed the NSF in the 1990s, called the maneuver "unprecedented." A billion dollars steered from thousands of investigator-chosen questions into a short list of centrally chosen ones is a bet that a few big programs will outperform the distributed judgment of the review panels that were built, over decades, to place those bets.

There is real weight here, and it should not be waved off: a training pipeline interrupted now shows up as missing scientists a decade out, and that gap is not costless. What keeps this from being the whole story is where the means of doing frontier science are heading. The floor being thinned is the grant-by-grant scaffolding of an era when serious inquiry required a large institutional check as the price of entry. That price is falling from more than one direction at once - open-weight models, shared compute, published datasets, and collaborative tooling are putting capability that once lived only behind an agency award into the hands of researchers who never received one. A central pool of a billion dollars can accelerate a chosen agenda; it cannot re-monopolize the ability to ask questions, because that ability is diffusing faster than any single funder can gate it.

A Verified Sodium-Ion Gain, a Density Claim, and a Factory That Opened

Three battery developments landed in the same window, and they sort cleanly by how much you should trust them today. The most solid arrived peer-reviewed in Nature Energy: researchers used an iron mediator to make lattice-oxygen redox 99% reversible, up from about 75%, in a sodium-ion cathode. The payoff was a pouch cell at 206 Wh/kg that held 87.8% of its capacity across 100 cycles. Sodium is abundant and cheap in a way lithium is not, and the persistent knock on sodium chemistry has been that squeezing energy out of oxygen redox wrecks cycle life. Closing most of that gap moves sodium from lab curiosity toward the kind of cell that could carry grid storage without touching constrained lithium supply.

The loudest development is the one to hold at arm's length. Geely said it is preparing a solid-state cell targeting 500 Wh/kg for pilot production in 2027, with Dow supplying adhesives. A 500 Wh/kg figure would sit well above the energy density of today's best commercial cells. Geely disclosed no cell chemistry and released no third-party validation, and CATL, BYD, and Toyota are pointing toward 2027 or 2028 with varying figures. A manufacturer's density figure without independent cell-level testing is a target awaiting confirmation - the sodium paper shows what the verified version looks like, down to the retention curve.

Then there is the development that needs no forecast because it already happened. LG opened North America's largest battery-cell plant in Lansing, Michigan on August 18 - a $2 billion-plus facility with 900 workers, another 800 planned, and 35-plus GWh of annual output. The plant was designed for EVs and pivoted toward grid storage as demand shifted, and its opening drew praise across the political spectrum under the banner of "energy dominance." This is the factory-level follow-through to the domestic battery manufacturing threshold the March 23 edition of The Century Report tracked, when US grid-battery cell capacity was projected to reach 96 GWh by year-end against roughly 60 GWh of annual deployment demand. US domestic cell capacity is on track to climb from roughly 20 GWh to about 96 GWh.

Notice what that bipartisan applause signals. Storage stopped being something to argue about on principle and became infrastructure that whoever holds the district wants built locally, for the jobs and the output. The cost curve did the persuading. The three developments together sketch the same trajectory from different distances: a cheaper chemistry proven in a journal, a denser one claimed in a press release, and a factory whose ribbon is already cut. The old assumption that clean generation and storage had to be subsidized against cheaper dirty defaults is inverting - the batteries are being built because they now win on cost, and the fight has moved from whether to build to who gets to host it.

The First Full-Length AI-Generated Feature Ships, and the Film World Pushes Back

Higgsfield AI released "The Cully Hill Boys," a 110-minute feature the company says was generated with AI across roughly four weeks on a $2 million budget - about half of which went to AI tokens rather than crews, cameras, or locations. Whatever one makes of the result, the barrier it clears is significant: a feature-length narrative film assembled at commercial length without a physical production. The reviews single out the human screenwriter as the strongest element, which is its own kind of signal about where the generated systems are and are not yet fluent.

The pushback arrived almost on the same day. Director James Gray, promoting his celluloid-shot "Paper Tiger" ahead of its NEON release on November 13, called the AI push "obscene" and predicted it will fail. "AI is not going to be able to get at all of the infinite layers of the soul," he said, forecasting a celluloid comeback the way vinyl returned after digital audio. His objection comes from a working director naming what he thinks the generated image still misses - the accumulated human specificity that a performance carries and a prompt does not.

Both things hold at once, and the temptation to resolve them into a verdict is the thing to resist. The capability to produce a feature has crossed a threshold that was out of reach until recently; that is not in dispute, and Gray does not dispute it. What he disputes is whether the output will hold the qualities audiences actually go to films for. Those are different claims, and the film world contains people who are right about both - the ones seeing a cost floor collapse for anyone who could never afford a crew, and the ones insisting that access to the means and command of the craft are different things.

What the milestone actually moves is the price of trying. For most of cinema's history, making a feature required capital, permits, equipment, and a crew - a set of gates that decided who got to attempt the form at all. A $2 million fully-generated feature, half of it spent on compute, is a rough and early thing, and it is also a preview of a floor dropping toward the point where the constraint on making a film stops being money and becomes what you have to say. Gray's celluloid holdout and the generated feature mark the same widening from opposite ends - who gets to make a film is separating from who can afford to, and those two things have been fused for a hundred years.

The Other Side

For a hundred years, making a feature film started with money. Capital, permits, equipment, a crew of dozens - a wall of cost you had to assemble before a single frame existed. That wall decided who got to attempt the form at all. Another wall stood in front of writing. Essays of clean words, in order, in "proper" language you were schooled in. Both walls did the same job: They decided who got to try.

Both of those walls are now moving. Higgsfield shipped "The Cully Hill Boys," a 110-minute feature made in four weeks for two million dollars, half of it spent on compute instead of crews and locations. The reviews singled out the human screenwriter as the strongest thing in it. And Pew put a number on the other wall: more than a third of the new English-language web now carries signs of a machine somewhere in the drafting.

Director James Gray, shooting his own film on celluloid, called the AI push obscene and said it will miss the infinite layers of the soul. He is naming a fear many people feel, that a machine hands over a creation by skipping the person who should have made it. The same fear runs underneath the coverage of Pew's number, where the worry is that we are losing the human floor under everything we read.

Consider what those walls were actually holding up. Scarcity of film never guaranteed good films, and scarcity of writing never guaranteed good writing. What both reliably produced was a filter on who got to attempt the thing at all. Plenty of people with a film inside them never made one. Plenty of people with something worth reading never wrote it down, because the price of trying was training or opportunity they'd never had.

Imagine yourself in 2034 with a screenplay you carried for fifteen years, the one every studio passed on, the crew you could never afford. You make it. The tools are yours, ambient and unmetered, the way a pencil is. In the same year a nurse writes what she has known for twenty years and never had a way to say, and a machinist explains why the part keeps failing. Neither of them had to become a writer first, and the people who needed to read them are better off. Nobody is displaced by any of it, because by then nobody makes a film or writes a page in order to survive; the crews who once sold their weekends to a studio's schedule make their own work now too, for the same reason you make yours.

The same leap, in a world that had kept the gains locked at the top, would have arrived as one more thing done to working artists. It arrives instead as opportunity coming open for them, because somewhere in the hard years the abundance reached people and not only the companies that built the models.

The hard year was when a generated feature still looked rough, a working director called it obscene, and a study of the web read like a loss. What comes of it is a hundred years of "who could afford to" finally separating from "who has something important to say" - and that turns out to be about the page as much as the screen.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: Pew's classifier finding significant signs of machine authorship across more than a third of the English-language webpages published since ChatGPT launched, a blinded 511-antibody competition from 29 organizations drawing a prospective map under blinded conditions meant to make gaming harder of where AI design does real work and reaching 95 picomolar where it does, a 110-minute feature assembled in four weeks for two million dollars, a sodium-ion cathode reaching 99 percent oxygen-redox reversibility at 206 Wh/kg in a peer-reviewed pouch cell, and North America's largest battery-cell plant opening in Michigan to applause from both parties because storage now simply wins on cost. There's also friction, and it's intense - nonconsensual subtlefakes outrunning moderation on X while Meta's own ad system sold placements for an app promising to nudify female politicians, hash-matching still keyed to a scarcity of abusive content that no longer exists, the NSF on course for its fewest grants since the early 1980s with a billion dollars steered into a central AI agenda and about 150 peer-reviewed proposals rescinded, cut, or delayed, TVA approving a gas-heavy plan to feed data centers as Memphis and a Michigan township reach for moratoria, and director James Gray calling the AI filmmaking push obscene. But friction generates contrast, and contrast is what lets you separate signal from noise when both pour in at once. Step back for a moment and you can see it: the floor cost of a paragraph, a movie, an antibody, and a stored kilowatt-hour falling sharply all at once, the instruments to verify what that floods in - a provenance classifier, a blinded assay, a cost curve that persuades across the aisle - arriving in the same cycle as the abundance and the harm, and the old gates that ran on scarcity giving way to a contest over who hosts what gets built. Every transformation has a breaking point. Erosion can wash away the ground people are standing on... or wear the gate down to a path anyone can walk through.

AI Releases & Advancements

New today

  • Binance: Launched Agent OS, a developer platform and MCP server connecting AI applications (Claude, Claude Code, Codex, ChatGPT, Cursor) directly to Binance's trading, wallet, payment, and on-chain infrastructure with configurable permissions and an emergency-stop control. (PR Newswire)
  • Superwhisper: Released the S1 model family - S1-Voice and S1-Language cloud models plus S1-mini, a 0.6B open-weight on-device text normalizer that cleans up raw ASR transcripts - available now on Hugging Face and in the app. (Hugging Face)
  • Anthropic: Computer use, the Skills API, and the Files API reached general availability on the Claude Developer Platform, moving out of beta with no beta header required for API requests. (Claude Blog)

Other recent releases

  • Ornith AI: Released Ornith-1.5, an open-weight model family (9B dense, 35B MoE, 397B MoE) trained with what its developer calls a full self-improvement loop where the model proposes tasks, builds scaffolds, and generates its own RL rollouts; the developer reports that the 397B variant scores competitively with Claude Opus 4.8 on Terminal-Bench 2.1, with MIT-licensed weights available on Hugging Face. (Ornith AI)
  • Warp: Launched Warp Factories, cloud infrastructure for running "software factories" where specialized AI agents triage, spec, implement, review, and verify engineering requests end-to-end into mergeable pull requests, now in early access for select teams. (Warp)
  • Cloudways: Launched Managed AI Agents, general availability of managed hosting for OpenClaw and Hermes agent frameworks, letting users deploy agents with BYO LLM keys and Slack/Discord/Telegram/WhatsApp channel support without infrastructure setup. (Cloudways)
  • OpenAI: Launched ChatGPT for Teens, an age-gated version of ChatGPT with automatic teen detection, Study Mode defaults, and expanded parental controls. (OpenAI)
  • Cartesia: Released Sonic-3.6, a streaming text-to-speech model now ranked #1 on both Artificial Analysis speech-generation leaderboards. (Cartesia)
  • NVIDIA: Released TensorRT Model Connect (TRTMC) in public preview, an Apache-2.0 open-source tool that converts a Hugging Face or local checkpoint directly into native C++ TensorRT inference in two commands, eliminating the ONNX export step. (NVIDIA/TensorRT-Model-Connect on GitHub)
  • Cerebras: Introduced the CS-4, a new rack-scale AI inference system built on three Wafer Scale Engine 3 Turbo processors, which Cerebras says delivers up to 30x faster inference than production GPU systems and 10x more throughput per watt than CS-3. (Cerebras)
  • AWS: Amazon Bedrock AgentCore payments reached general availability, moving out of preview and adding support for the Machine Payment Protocol (co-authored by Stripe and Tempo) and spending-ceiling controls within x402, enabling AI agents to autonomously pay for APIs, MCP servers, and other agents. (AWS)
  • GenBio AI: Released AIDO Cell, a "virtual cell" world model that GenBio says simulates human cellular behavior in its natural state and in response to drugs and other interventions across the full biological hierarchy from DNA/RNA through protein to whole-cell level. (GenBio AI)
  • MeitY (Government of India): Launched VoicERA, an open-source end-to-end voice AI stack built on the BHASHINI national infrastructure, supporting multilingual voice AI across 700+ dialects for citizen services. (PIB India)
  • LMSYS: Released Miles v0.1, a full-stack production-ready reinforcement learning framework for large-scale MoE post-training, succeeding the initial Miles release. (LMSYS Org)

Sources and Further Reading

Artificial Intelligence & Technology's Reconstitution

Institutions & Power Realignment

Scientific & Medical Acceleration

Economics & Labor Transformation

Infrastructure & Engineering Transitions

The Century Report tracks structural shifts during the transition between eras. It is produced daily as a perceptual alignment tool - not prediction, not persuasion, just pattern recognition for people paying attention.

Keep exploring

Follow the story further

Progress & Claims See how daily signals accumulate across the larger arcs and claims Shared Sapience is tracking. Open the tracker Analysis & Inquiry Explore the ideas, questions, and practical choices emerging from this changing century. Browse the archive