VZ research lens
This report is not written for trend consumption. It is written for decision quality: what to trust, what to prioritize, and what to execute first.
VZ Lens
Through a VZ lens, this is not content for trend consumption - it is a decision signal. 23 empirical studies, 110+ sources. Most of the internet is no longer human-generated. How is machine-generated content flooding search engines, opinion platforms, and academic publishing—and what is the feedback loop that makes it so difficult to break free? The real leverage appears when the insight is translated into explicit operating choices.
Status: active research, 2026-03-07 Method: GFIS full pipeline (4 academic backends + Corpus V2 RAG + web search) Sources: 110+ empirical sources, 23 evaluation threads
What is “slop,” and why does it matter?
The word “slop” was named Merriam-Webster’s Word of the Year in 2025. It originally meant swill or slop—the food scraps fed to pigs. Today, it means the same thing on the internet: massive amounts of low-quality, machine-generated content that nobody asked for, but is everywhere.
This research demonstrates—through numbers, measurements, and studies—just how real this phenomenon is, how profound its impact is, and where it will lead if we do nothing.
I. How Much AI Content Is There on the Internet?
By 2025, bots will be in the majority on the internet.
- 51% of internet traffic is from bots (Imperva Bad Bot Report, 2025)
- 74.2% of new English-language websites were written by artificial intelligence (Ahrefs, 900,000 pages, April 2025)
- 17.3% of Google’s top 20 search results are AI-generated (peak: 19.56%)
- The rate of zero-click searches is 69% — in seven out of ten searches, no one clicks anywhere (SimilarWeb, 100M+ devices, 2025)
Impact of AI Overview on Clicks — 5 Independent Studies
| Who measured it? | Sample | Click Decrease |
|---|---|---|
| Seer Interactive | 25.1M impressions, 3,119 terms | −61% |
| Ahrefs | 300,000 keywords | −58% |
| GrowthSRC | 200,000+ keywords | −32% |
| seoClarity | 12 million keywords | −0.66 — −2.41 pp |
| BrightEdge | Fortune 100 companies | −30% |
By Platform
| Platform | Data | Source |
|---|---|---|
| YouTube | 21–33% AI-generated content; 278 channels with $117M/year in ad revenue | Kapwing, 2025 |
| X (Twitter) | 64% bots | Imperva, 2025 |
| 95M fake accounts; organic reach: 4% | Galaxy Research | |
| 5.4 billion fake accounts deleted in a single year | Meta, 2025 |
II. The Market for Reviews: How Is Trust Collapsing?
The Amazon Example
- 74% of AI-written reviews are five-star (compared to 59% for human reviews)
- 93% of AI reviews carry the “Verified Purchase” label — but no one actually bought the product
- Harvard HBS research (January 2026) showed that Amazon’s own AI summaries systematically overrepresent fake reviews
Science is also affected
- 21% of reviews at the ICLR conference were written by AI — one in five reviews was generated by a machine
- The doubling time for scientific fraud is 18 months
- AI chatbots incorrectly flagged 18% of valid studies as retracted
The Akerlof threshold — three markets have already crossed it
Nobel Prize-winning economist George Akerlof’s 1970 “market for lemons” theory accurately describes what happens: if you cannot distinguish the genuine from the fake, good content disappears because it is not worth producing.
“If the proportion of good products in a market falls below two-thirds, there is only one equilibrium: one in which only bad products remain.” — Easley and Kleinberg: Networks, Crowds, and Markets
Has already crossed the threshold:
- Stock photography: Adobe Stock AI content rises from 2.5% (2023) to 47.85% (2025) — a 19-fold increase. Getty revenue down 4.5%. Shutterstock subscribers down 22%.
- Academic publishing: Wiley/Hindawi retracted 8,000+ articles and shut down 19 journals. AI-related retractions: <20 before 2022, 663 in 2023 alone.
- Freelance writing: Writing job postings −33% since the launch of ChatGPT. The highest-quality freelancers were affected the most.
III. Can you tell if it was written by a machine? (No.)
- Of all participants, only 0.1% were able to correctly identify all genuine and fake content (iProov, 2025)
- People recognized deepfake videos only 24.5% of the time — worse than flipping a coin
- AI detector accuracy in real-world scenarios: 60–70% — it fails to detect three out of ten texts
- GPTZero incorrectly flags 25% of texts written by non-native authors as machine-generated
IV. The Great Feedback Loop: Model Collapse
The Bottom Line
- AI systems are trained on human text
- These AI systems now generate massive amounts of text, which is fed back into the internet
- The next generation of AI systems is trained on this internet — which is now full of AI-generated text
- So AI learns from its own output — like a photocopier that copies its own copies
Model collapse generation numbers
| Research | Field | Total collapse |
|---|---|---|
| Shumailov (2024, Nature) | Language model | 9 generations (+56% error) |
| Alemohammad (2024, ICLR) | Face generation | 3–5 generations (diversity collapses) |
| Dohmatob (2025, ICLR) | Linear regression | Even 0.1% synthetic data causes collapse |
| Medical AI (Liu, 2026) | Clinical documentation | 2 generations (clinically unusable) |
According to Epoch AI’s updated (2025) estimate, the depletion of high-quality publicly available human text:
- With computationally optimal training: ~2028
- With 100x overfitting: ~2025 (has already occurred)
V. Economic Impacts
Labor Market
| Source | Estimate | Timeframe |
|---|---|---|
| Challenger, Gray & Christmas | 54,836 jobs specifically attributed to AI | 2025 (U.S.) |
| Stanford (Brynjolfsson et al.) | Entry-level workers (ages 22–25): −13% employment | Since 2022 |
| Goldman Sachs | 300 million jobs affected globally | By 2030 |
The Collapse of the Publishing Ecosystem
| What happened? | Magnitude |
|---|---|
| Google → publisher traffic (global) | −33% year-over-year (Chartbeat, 2,500+ sites) |
| HubSpot blog traffic | −80% (14.8M → 2.8M monthly visits) |
| Washington Post | Organic search −50% over 3 years; >$100M annual loss; 300+ journalists laid off |
| Chegg (education platform) | Stock −99%, revenue −43% |
| WPP (world’s largest advertising agency) | Revenue −8.1%, market value fell from 25 billion to 3 billion |
Workplace “workslop”
- 40% of workplace AI content is “workslop” — unchecked, unread machine-generated text (BetterUp, 2025)
- This causes an annual productivity loss of $9.1 million per organization
VI. The Speed of Trust Loss
| What was measured? | Change |
|---|---|
| Gallup: Trust in U.S. media (2020 → 2025) | 40% → 28% |
| Pew: Trust in national news organizations (2016 → 2025) | 76% → 56% |
| BrightLocal: Google as a review platform (2023 → 2026) | 87% → 71% |
| Deloitte: “The benefits of online are worth the risk” | 58% → 48% (all-time low) |
| X/Twitter: advertiser trust (2022 → 2024) | 22% → 12% |
Key data: According to Pew Research, 18–29-year-olds trust social media (50%) almost as much as national news organizations (51%) — an unprecedented convergence.
VII. What does this mean in practice?
If you’re a marketer
69% of Google searches result in no clicks. The introduction of AI Overview reduces clicks by 30–61%. Your organic reach on Instagram is 4%. HubSpot’s blog lost 80% of its organic traffic.
The middle of the classic funnel is empty. Awareness is there, but the action isn’t happening on your end.
If you’re a salesperson
“Social proof”—which has been the foundation of online sales for the past 15 years—is no longer a reliable indicator. 93% of Amazon reviews are “Verified Purchases”—but so are machine-generated reviews.
If you’re in PR
The U.S. media trust rating has fallen to a historic low: 28%. Media placement alone is no longer a guarantee of credibility.
If you’re a market researcher
If a respondent’s opinion was partially shaped by machine-generated content—and neither the respondent nor the researcher knows which part—then the validity of the measurement cannot be guaranteed within the previous methodological framework.
The Common Denominator
The digital infrastructure of the past 15 years was built on assumptions that AI is now invalidating. None of these assumptions can be maintained without change.
Key Sources
| Source | What did it examine? | Where was it published? |
|---|---|---|
| Shumailov et al. (2024) | Model collapse: how AI degrades | Nature |
| Dohmatob et al. (2024) | Severe model collapse | ICLR 2025 |
| “As Good as a Coin Toss” (2025) | AI content detection | ACM |
| Yu, Kim, and Kim (2026) | Retrieval Collapse | ACM |
| Ahrefs (April 2025) | 74.2% AI-generated new web pages | Industry report |
| Imperva (2025) | 51% bot traffic | Bad Bot Report |
| BetterUp (2025) | Workplace Workshop | Industry Report |
| Epoch AI (2023/2025) | Exhaustion of Human Text Data | Research Institute |
Research Method: GFIS pipeline v0.3.0 — 4 academic search engines (Semantic Scholar, OpenAlex, Crossref, Elicit) + proprietary corpus (+10M chunks: books, articles, studies) + web search (Brave Search, Tavily) Generated by: Claude Opus 4.6 | Date: 2026-03-07
Strategic Synthesis
- Translate the core idea of “The Real Impact of AI Slop — What Do the Numbers Show, and What’s Next?” into one concrete operating decision for the next 30 days.
- Define the trust and quality signals you will monitor weekly to validate progress.
- Run a short feedback loop: measure, refine, and re-prioritize based on real outcomes.
Apply to your context
If you want this framework translated into a concrete execution sequence for your team, we can map the first 30-day priorities together.