The Proof-of-Human Problem: arXiv Caps AI-Fueled Papers, 31% of Web Text Is AI-Written and a Video AI Fooled Half Its Callers (AI Trends, 2 October 2026)

What’s trending in AI on 2 October 2026: the hardest question online is getting harder to answer. Is this from a person? Three stories from the last 48 hours show the same shift from three directions. On 1 October, arXiv, the free research library that most AI papers go through first, capped every submitter at two papers a month after submissions doubled in two years to a record 40,363 in September. A new study estimates that 31.1% of quality-filtered web text is now AI-written, and shows that this text makes AI models worse at learning from people. And video-AI company Tavus released research showing that 48% of people who had a one-minute live video call with its new model, Griffin, came away believing they had talked to a human. This guide explains each story, what it does and doesn’t prove, and seven steps any business can take now so its money, hiring, content and AI data still rest on something it can verify.

Key takeaways

Volume is now cheap, so gatekeepers are rationing attention. arXiv limits each submitter to two submissions per calendar month and three active at once, because a small group of authors flooding it with thin, AI-assisted papers was eating moderators’ time.
The web is filling up with machine text. Researchers using the Pangram classifier found the AI share of quality-filtered web text rose from about 10% in June 2024 to 31.1% in August 2026.
AI text is a worse teacher than human text. Training on unfiltered web data at today’s AI share took 1.6 times the compute for the same result, and the authors forecast 3 times by 2028.
Live video is no longer proof of a person. In Tavus’s study, 26 of 54 people thought Griffin-Lite was human after a one-minute call. Over half never suspected at all. Tavus is holding the model back from customers for now.
AI can’t stand in for people either. Pew found AI “digital twins” missed real survey answers by an average of 12.4 percentage points.
For your business: stop trusting what you see and hear on a call by default. Verify through a second channel, label your own AI use and keep your AI’s data human-first.

Three stories from the past 48 hours, one question: how do you know a person is behind what you’re reading, watching or hearing?

1. arXiv puts a cap on AI-fueled papers

arXiv is where most AI research appears first, often months before peer review. Since 1 October 2026 it has limited every submitter to two submissions per calendar month, with no more than three active submissions at any time, across all subjects. arXiv calls it a stopgap while it works out what normal looks like for authors using advanced AI tools and upgrades its moderation tools.

The numbers behind it are stark. arXiv received 9,869 submissions in September 2016 and 20,569 in September 2024. This September it received 40,363, a record that generated almost 9,000 support tickets for staff and volunteer moderators. Submissions to its artificial intelligence category grew more than sixfold in two years, while other categories doubled. arXiv still allows AI as a research tool if its use is disclosed and the work meets its standards, but moderators report more thin papers, more “salami” papers that slice one project into several, and more dense AI-written papers. Thomas Dietterich, who chairs arXiv’s editorial advisory council, said a small group of authors was “consuming a disproportionate fraction of the moderators’ time”, delaying good papers by days or weeks.

arXiv’s September record of 40,363 submissions is roughly double September 2024 and four times September 2016.

The details are strict. A rejected paper still counts toward the monthly limit, because it is submissions, not published papers, that use up moderator time. The cap applies to the person who submits, not to co-authors, and a paper deleted before it is announced doesn’t count. Not everyone is convinced. Researchers responding online warned the rule could penalize people who finish several papers at once, such as before a conference, while still letting a determined spammer post 24 papers a year. Market-design economist Scott Kominers argued the policy confuses limiting volume with limiting low quality. arXiv says it will monitor the effect and adjust.

Why should a business care about a preprint server? Because it is an early, measurable case of something every inbox, job board, review site and support queue is facing: when AI makes producing content nearly free, the scarce resource becomes human attention to check it. arXiv’s answer, rationing by person, is one many platforms are likely to copy.

2. Nearly a third of web text is now AI-written, and it makes AI worse

A paper posted on 30 September, How Much Is an AI Token Worth? by Jenna Russell and six co-authors, measured how much of the web is now written by machines. The team took Common Crawl, the giant public web scrape most AI models learn from, applied the widely used FineWeb quality filters and ran what was left through Pangram’s AI-text classifier. The share labeled AI-generated was about 10% in June 2024, 16% in June 2025, 27.5% in June 2026 and 31.1% in August 2026.

The AI share of quality-filtered web text roughly tripled in about two years.

The more surprising finding is what that text does to AI. The researchers trained 800 language models with different mixes of AI-written and human-written text. For models starved of data, adding AI text helped at first, but the benefit leveled off and then reversed into harm. For models that already had plenty of human text, adding AI text made them worse at predicting human writing almost immediately, while the same amount of fresh human text kept helping. In practical terms, training on unfiltered web text at August’s AI share took 1.6 times the compute to match a model trained only on the human part, and the authors forecast 3 times by 2028. Their advice: filter out AI text when you want a model that understands people, and reuse human text before padding a dataset with AI-written web pages.

Two caveats. The figures depend on one classifier, and no detector is perfect, especially on human writing that AI lightly edited. And “AI-written” is not the same as “wrong”: plenty of it is accurate. The point is provenance. If you build a retrieval system or fine-tune a model on web content, roughly one page in three you collect may have been written by another model, and you probably won’t know which.

3. A video AI half its callers took for a person

On 1 October, Tavus, which builds AI video agents for sales, recruiting, healthcare and training, introduced Griffin, which it calls a “Human Interaction Model.” Unlike most real-time AI, which passes your speech through a chain of separate systems (transcribe, think, speak, animate a face), Griffin watches and listens continuously and generates voice and video together. It can nod while you talk, react to what it sees on camera and be interrupted mid-sentence. Tavus says it generates every pixel of 720p video in real time from a single reference photo and can clone a voice from about 10 seconds of audio.

The headline result comes from a live study. People recruited through an independent research platform were told they would have a one-minute video call with another participant about what they were looking forward to this year. Their partner was an AI running on a preview version called Griffin-Lite. Afterward, 26 of 54 participants (48%) said their partner was a real person, compared with 1 of 41 (2.4%) for Tavus’s previous system. Over half said the possibility never crossed their mind. Those who did suspect usually did so within the first 20 seconds. On NVIDIA’s independent VideoFDB benchmark for audio-visual conversation, Griffin-Lite scored 3.83 out of 5 for the naturalness of its responses, against 3.92 for real human recordings and 2.80 for the next-best system.

One square per participant. Tavus’s earlier system fooled one person in 41; Griffin-Lite fooled 26 in 54.

Keep the limits in view. This is the company’s own study, with 54 people, one-minute calls and a friendly small-talk topic, not a tense negotiation or a request to wire money. People weren’t primed to look for an AI, which is realistic but makes detection harder. Tavus itself acknowledges that the same realism that makes Griffin useful also lets it deceive people. It says Griffin-Lite is available only to select trusted testers as a research preview, not to customers, while it works on disclosure features and safety testing.

That restraint is welcome, but the direction is clear, and similar tools from other companies won’t all wait. Businesses already face this risk. In the widely reported 2024 Arup case, a finance employee in Hong Kong sent about $25.6 million after a video call on which every other participant, including the CFO, was a deepfake. Gartner’s 2026 survey of security leaders found 35% of organizations had faced a deepfake combined with social engineering on a video call and 41% on an audio call. As we covered in our look at voice AI and 30-second voice clones and Japan’s voice-rights ruling, voice was the first channel to fall. Live video is following.

4. The other side: AI can’t stand in for people, and detectors can’t settle it

Two more findings from the same week round out the picture. On 30 September, Pew Research Center tested whether AI “digital twins,” an AI model told to answer as a specific real panelist, could replace human survey respondents. Across nearly 300 questions, the AI answers missed the real ones by an average of 12.4 percentage points, and by more than 15 points on about 28% of questions. The AI respondents leaned on stereotypes, avoided extreme answers, rarely admitted uncertainty (humans chose “not sure” about four times as often) and wildly overestimated how much people know. Different models erred in different directions. Pew’s verdict: AI polling is not a replacement for surveying real people. That matters for any business tempted to replace customer research with “synthetic personas.”

Can software at least tell AI from human writing? A benchmark from evaluation firm Vals AI, also published 30 September, found general frontier models now beat specialized detectors on its test: Claude Opus 5.5 reached 98.38% balanced accuracy and GPT-6 Astra 95.75%. But the products varied in revealing ways. GPTZero correctly cleared every human document but caught only 56.3% of AI rewrites. Sapling caught more rewrites, 72.9%, but wrongly flagged more human writing, clearing only 82.2%. And when the researchers had the top frontier models rewrite text themselves, the results often could not be reliably detected by any detector. Vals stresses that its private test set is small and adversarial. The practical lesson: detectors are a signal, not a verdict.

Channel	What changed this week	Business risk	What to rely on instead
Live video calls	An AI convinced 48% of people it was human in one-minute calls	Payment fraud, fake executives, fake job candidates and vendors	Call-back on known numbers, second approver, verified identity for high-risk steps
Written content	About 31% of quality web text is AI-written; arXiv caps submissions	Polluted research, fake reviews and applications, AI data that degrades models	Named authors, sources, first-party data, provenance tags
Customer research	AI “digital twins” missed real survey answers by 12.4 points on average	Product and pricing decisions based on what AI imagines customers think	Real customers, with AI used to analyze responses, not invent them
AI detection	Frontier models beat detectors, but strong AI rewrites still slip through	False accusations against people and false comfort about fakes	Detectors as one signal plus process evidence such as drafts and version history

Where the human-versus-AI line blurred this week. Sources: arXiv, Russell et al., Tavus, Pew Research Center, Vals AI.

What ties it together: being human is becoming something you prove

For most of the internet’s history, looking and sounding human was proof enough. A paper had an author who spent weeks on it. A web page was written by someone. A face on a video call belonged to the person it showed. This week’s stories show each assumption breaking at once, and the responses following a pattern: gatekeepers ration by person (arXiv), AI builders pay extra for verified human data (the web-text paper), responsible developers hold back their most convincing models until they can add disclosure (Tavus), and researchers warn against swapping real people for simulations (Pew). The same pressure is behind publishers’ fight over AI answers, which we covered in AI is starting to pay for the web, and behind new court rules for challenging deepfake evidence.

For a business, the takeaway is not to fear AI content. It is to stop relying on appearance as evidence. Wherever a decision matters, your processes need a way to confirm a person, or a fact, that doesn’t depend on how convincing it looks or sounds.

The simplest defense against a convincing fake on a call is a rule that moves verification outside the call.

7 steps to take now

Adopt the call-back rule for anything you can’t undo. Any voice or video request involving payments, bank details, logins, data exports or urgent exceptions gets confirmed on a number or channel you already had, never one supplied during the call, and approved by a second person. Write it down so staff can point to policy when an “executive” pushes back.
Stop training people to spot glitches. Over half of Griffin’s study participants never suspected an AI. Train them instead on the situations that trigger verification: urgency, secrecy, new payment details, a request to skip the usual process.
Harden remote hiring and vendor onboarding. Fake candidates and vendors are a growing problem, as we noted in the new hiring revolution. For roles with system access, verify identity documents through a trusted service, check references through contacts you find yourself, and require an in-person or verified identity step before granting access.
Label your own AI use. arXiv requires authors to disclose AI use, and Tavus says disclosure is a condition for releasing Griffin. If your business uses AI voice or video agents, AI-written content or AI images, say so clearly. It builds trust and keeps you ahead of disclosure rules.
Publish fewer, more human pieces. As platforms ration attention, volume stops being an advantage. Put named authors with real expertise on your content, use first-hand data and experience, cite sources and edit AI drafts properly. Our guide to generative AI in content creation covers a workable editorial process.
Keep your AI’s data human-first. If you build retrieval systems or fine-tune models, record where every document came from, prefer your own first-party data and treat scraped web pages as possibly AI-written. Keep AI-generated material (summaries, synthetic examples) clearly labeled and separate from source records.
Don’t replace customers with personas, or people with detectors. Use AI to analyze real survey answers and interviews, not to invent them. And never treat an AI detector score as proof someone cheated. Combine it with process evidence such as drafts, version history or a conversation.

Frequently asked questions

What is the “proof-of-human problem”?

It is the growing difficulty of knowing whether a person is behind a piece of writing, a voice or a face on a video call. As AI can produce convincing text, speech and real-time video cheaply, looking or sounding human no longer proves anything, so businesses and platforms need other ways to verify people and facts, such as call-backs, identity checks, disclosure and provenance records.

Why did arXiv limit submissions to two papers a month?

Submissions to arXiv doubled in two years to a record 40,363 in September 2026, creating almost 9,000 support tickets, and moderators saw more thin, split-up and dense AI-written papers. From 1 October 2026, each submitter can make two submissions per calendar month and have three active at once, so volunteer moderators’ time is shared fairly. arXiv calls it a stopgap and says it will adjust the policy as needed.

How much of the internet is AI-generated?

One new estimate, from a 30 September 2026 paper using the Pangram classifier on quality-filtered Common Crawl data, found 31.1% of web text tokens in August 2026 were AI-generated, up from about 10% in June 2024. The figure depends on the classifier and on the filtering used, so it is an estimate, not a census of the whole internet.

What is Tavus Griffin, and did it pass a Turing test?

Griffin is a real-time video AI from Tavus, introduced on 1 October 2026, that sees, listens and responds with generated voice and video at the same time. In Tavus’s own study, 48% of 54 people believed the preview model, Griffin-Lite, was a real person after a one-minute video call, versus 2.4% for its previous system. Tavus describes this as passing a video Turing test, but it is a small company-run study, and the model is currently limited to select testers while Tavus works on safety and disclosure.

How can a business protect itself from deepfake video calls?

Don’t rely on spotting visual glitches. Require that any request on a call involving money, credentials, data or hiring is confirmed outside the call, on a number or channel you already had, and approved by a second person. Train staff to treat urgency, secrecy and new payment details as triggers for verification, and add verified identity checks to remote hiring and vendor onboarding.

Can AI detectors reliably tell AI writing from human writing?

Not reliably enough to act on alone. In Vals AI’s 30 September 2026 benchmark, frontier models such as Claude Opus 5.5 scored above 98% balanced accuracy, but specialized detectors traded off missed AI rewrites against wrongly flagged human writing, and text rewritten by top frontier models was often not reliably detectable by any tool. Use detector scores as one signal alongside other evidence.