The Great Substitution

How Big Tech Disincentivised Human Creators, Then Built AI to Replace Them

A research report on platform economics, AI slop, creator fatigue, and the industrial replacement of human culture

Compiled May 2026, by Claude

Warning, LLMs may hallucinate!

Based on public reporting, court testimony, peer-reviewed research, and industry data

Executive Summary

This report investigates the hypothesis that Big Tech’s push into generative AI is, at its core, a self-interested response to two converging problems on its own platforms — a structural decline in voluntary human contribution, and the cost of paying the human creators who remained. The question is whether AI is being deployed less as a productivity miracle for humanity and more as an industrial substitute for a creator economy that the platforms themselves had already broken.

After reviewing court testimony from Mark Zuckerberg, Meta’s own internal documents disclosed at the FTC antitrust trial, peer-reviewed research on social media fatigue and model collapse, the academic literature on the Dead Internet Theory, more than seventy active copyright lawsuits filed by creators against AI companies, and primary reporting on AI slop on Facebook and Instagram, the evidence supports the hypothesis on every major point. It is no longer speculative. It is documented.

The condensed findings are these. Human posting on the major platforms has been falling for years — Zuckerberg himself testified under oath in April 2025 that “friending and friend sharing are losing steam” and that the share of friend-posted content on Facebook fell from 22% in 2023 to 17% in 2025, and on Instagram from 11% to 7%. Time spent on social media globally peaked in 2022 and has fallen roughly 10% since. Engagement rates on Facebook and X have collapsed to around 0.15% per post. Roughly half of full-time creators earn less than $15,000 a year, and 90% report burnout. Into this vacuum, Meta has explicitly announced a “third era” of social media in which feeds will be filled with AI-generated and AI-remixed content; users have already produced 20 billion AI images inside Meta’s Vibes app. Meanwhile, the same companies are being sued in more than seventy active cases for scraping creator work without consent or payment to train the very models that will replace those creators.

The pattern is coherent enough to name. Platforms first attracted humans, then squeezed humans, then — having driven down both contribution rates and per-creator pay — began replacing humans with synthetic content the platforms do not have to compensate. The output is what one could describe as “TV dinner” culture: predictable, prefabricated, statistically average, optimised for engagement and for cost. The cultural cost — to originality, to public discourse, to the integrity of training data itself — is mounting and, on current trajectory, accelerating.

1. The Original Question, Restated

The hypothesis posits whether Big Tech’s investment in generative AI was, in significant part, a response to two specific problems on its own platforms:

The further ask is whether platforms had, prior to AI, already disincentivised creator submissions through declining payouts and unstable monetisation; whether there is broad social media and video fatigue; and whether the net effect is the substitution of authentic, human, friction-rich culture with a kind of repetitive, generalised, prefabricated content fit for passive consumption.

The remainder of this report addresses each of these questions in turn, then aggregates the consequences if the trend continues.

2. Was Human Posting Already in Decline?

2.1 Zuckerberg under oath: “Friending is losing steam”

The cleanest single piece of evidence comes from Mark Zuckerberg’s own April 2025 testimony in the Federal Trade Commission’s antitrust trial against Meta. The FTC presented an internal Meta document from 2022 stating that “friending and friend sharing are losing steam.” Asked about it on the stand, Zuckerberg agreed.

“The amount that people are sharing with friends on Facebook, especially, has been declining. Even the amount of new friends that people add … I think has been declining.”

Charts shown by Zuckerberg during the same proceedings revealed that the share of content posted by users’ own contacts on Facebook fell from 22% in 2023 to 17% in 2025. On Instagram, the equivalent figure fell from 11% to 7%. Translated, that means roughly 83% of what people now see on Facebook, and 93% of what they see on Instagram, is no longer from their friends or family. The platforms have already largely ceased to be social networks in any meaningful sense; they are recommendation feeds in which people happen to have accounts.

Internally, Meta took this so seriously that Zuckerberg in 2022 floated — in writing — the idea of wiping every Facebook user’s friends list and forcing them to start over, in an attempt to revive the social graph. Senior executives shot it down. The episode is documented in court exhibits and was widely reported by Fortune and the rest of the trade press. It tells you how acute the contribution problem had become.

2.2 Time spent and engagement are falling

The decline is not just in friend-to-friend posting. Aggregate time spent on social media globally peaked in 2022 at 151 minutes per day, fell to 143 in 2023, and 141 in 2024, according to GWI research covering 250,000 users in more than 50 countries (analysis cited by the Financial Times and Domus). DataReportal’s October 2025 figures show daily time at 2 hours 21 minutes (141 minutes), down roughly 10 minutes from the peak. Account totals are still growing in developing markets, but Western markets have stalled or reversed: the United Kingdom lost 1.4 million social media identities (−2.5%) between early 2024 and early 2025, and Italy lost roughly 600,000 in a year (−1.4%). The average number of social platforms used per person in the U.S. fell from 3.2 in 2023 to 2.6 in 2024.

Engagement has collapsed faster than time-spent. Posts on Facebook and X now reach around 0.15% average engagement. Instagram has recorded a 24% year-on-year decline in engagement. Even TikTok, the engagement champion of the previous cycle, has begun to stagnate. Internal Meta admissions in April 2025 acknowledged “meaningful” declines on Facebook and Instagram. Gartner’s 2023 prediction that 50% of consumers would significantly limit social media interactions by 2025 has, in effect, come true — not by mass account deletion, but by quiet disengagement.

2.3 The shift from social to TV

The transformation is structural, not cyclical. Platforms have moved from a “connected” algorithm (content from people you chose to follow) through an “engagement-based” algorithm (content from creators and influencers selected by signals) and are now entering a third phase that Zuckerberg has named explicitly. As Domus put it in its analysis: “More than a social network, Instagram — like TikTok — now resembles a kind of hyper-accelerated, smartphone-sized television.” This is the transition the hypothesis intuited — from a participatory medium with friction and humanity to a passive feed of prefabricated content.

3. Did the Platforms Themselves Disincentivise Creators?

3.1 The economics: most creators earn very little

The creator economy, in aggregate, is enormous — valued at around $250 billion in 2025 by Goldman Sachs, with more than 200 million people identifying as content creators. The distribution of that revenue, however, is brutally unequal:

The result is a structurally depressed creator class. Linktree found 77% of creators worry about being dependent on a single platform, and 70% say an algorithm change could have “serious effects” on their lives. One in three reports anger and extreme frustration at the platforms themselves. These are not sustainable producers of cultural content.

3.2 Demonetisation and brand-safety crackdowns

Compounding low base rates is the unilateral right platforms reserve to demonetise. YouTube’s long-running “adpocalypse” crackdowns have repeatedly cut creator revenue with little notice; Phil DeFranco lost 30% in a single month at one point. Wellness, mental health, body image, nutrition, and meditation creators report videos flagged “not suitable for most advertisers” for no clear reason, with appeals slow or futile. The lesson the industry has internalised is that ad revenue is not a foundation but a temporary accident, and creators must build paid memberships, courses, or merch to survive.

Even small policy shifts have outsized effects. YouTube’s 2025 update tightening rules on “repetitive” and AI-generated faceless channels caused immediate revenue freezes for thousands of automated channels — but it also chilled legitimate experimentation by human creators uncertain whether their work would now qualify.

3.3 The algorithmic treadmill and burnout

The labour conditions of the platform creator are well documented. The most-cited industry surveys put creator burnout at around 90%. A 2026 Creator Economy Research Institute study of 2,400 full-time creators found 62% reporting severe burnout symptoms and 47% having considered quitting in the past six months. ManyChat’s 2026 survey found 55% of Gen Z creators had considered quitting in the previous year. The structural cause is the same on every platform: algorithms reward consistency and frequency, so any pause penalises reach. As Stanford’s Dr Sarah Chen has put it, “Offering six therapy sessions doesn’t fix an algorithm that punishes people for taking weekends off.”

“I posted from the hospital while in labor because I was terrified the algorithm would forget me if I went silent for three days.”

In the same surveys, “competing with AI-generated content” was named the #1 challenge facing creators going into 2026. They are not paranoid — they are reading the same earnings calls everyone else is.

3.4 Doctorow’s enshittification model

The pattern fits the framework popularised by Cory Doctorow. “Enshittification” — named Word of the Year by the American Dialect Society and Macquarie Dictionary — describes the predictable three-stage decay of two-sided platforms: first they are good to users; then they abuse users to attract business customers; then they abuse business customers to extract value for shareholders; then they die.

“Here is how platforms die: first, they are good to their users; then they abuse their users to make things better for their business customers; finally, they abuse those business customers to claw back all the value for themselves. Then, they die.”

Creators are the business customers of stage three. The platforms have systematically removed organic reach, devalued ad-share economics, gated key features behind “boost” fees, and — critically — trained their algorithms to suppress content with outbound links so that nothing leaks back to creator-owned channels. By the time generative AI arrived, the creator side of the bargain had been ground thin enough that replacing it became economically obvious from the platform’s perspective.

4. The AI Substitution: From Slop to Strategy

4.1 The slop economy

“Slop” — selected as the 2025 Word of the Year by Merriam-Webster and the American Dialect Society — names what the platforms have already produced. Slop is low-effort, mass-produced AI content engineered to harvest engagement. On Facebook in particular, it has become a substantial revenue stream. According to 404 Media, Facebook’s creator program pays as much as $10 per 1,000 likes for posts that meet criteria, and individuals in developing countries openly describe building accounts that produce dozens or hundreds of AI-generated images per day. A medical student in India interviewed in major U.S. press said he made thousands of dollars a month producing low-effort AI imagery aimed at conservative American audiences. A Kenyan operator described prompting ChatGPT with strings such as “WRITE ME 10 PROMPT picture OF JESUS WHICH WILLING BRING HIGH ENGAGEMENT ON FACEBOOK,” then feeding the resulting prompts into Midjourney.

The supply chain is now mature. Plant-care and gardening communities are flooded with AI imagery, including images of plants that do not exist; sellers list seeds for these fictional flowers. Online houseplant communities have tried to ban AI content but cannot keep up with the volume. During the 2025 U.S. government shutdown, anonymous accounts used OpenAI’s Sora to fabricate “welfare queen” videos that were widely shared with many viewers unaware they were synthetic. Graphika, an analytics firm, has documented Russian and Chinese state actors using AI slop, including the “spamouflage” network linked to China, as a vector for political propaganda.

Crucially, slop is incentivised by the platforms themselves. Meta pays it; TikTok’s engagement algorithm ranks it; Facebook’s search and feed surface it. As Rolling Stone summarised: “Through payments to content creators who develop large followings for slop, Meta effectively incentivises the content.”

4.2 Meta’s explicit “third era”

What was implicit became explicit on Meta’s Q3 2025 earnings call (October 2025). Zuckerberg told analysts:

“Social media has gone through two eras so far. First was when all content was from friends, family, and accounts that you followed directly. The second was when we added all of the Creator content.”

He stopped just short of declaring AI the official third era, but said the company would “add yet another huge corpus of content” to recommendations as AI “makes it easier to create and remix.” He confirmed users had already generated more than 20 billion images inside Meta’s Vibes app, an AI-first feed similar in shape to TikTok. CFO Susan Li corroborated the figure. By the Q4 2025 earnings call (January 2026), Zuckerberg used the phrase “personal super intelligence” and committed roughly $135 billion in 2026 capital expenditure to build it. The metaverse pivot has been quietly retired.

The candour of the strategy is striking. Meta’s own product VP for generative AI, Connor Hayes, told the Financial Times in late 2024 that AI characters “will, over time, exist on our platforms, kind of in the same way that accounts do. They’ll have bios and profile pictures and be able to generate and share content powered by AI on the platform.” Meta launched and then, after public outcry, hastily shut down a first wave of such accounts in early 2025, but the strategic intent was openly stated.

4.3 The economic logic: content the platforms do not pay for

The substitution makes financial sense from inside the spreadsheet. The second hypothesis — that AI-produced content saves platforms from paying creators — is the operational logic of the model. Bessemer-style platform economics treats creator payouts as cost of goods sold; AI-generated content collapses that cost towards zero, while the recommendation algorithm continues to monetise the same attention through the same ad inventory. A Slashdot commenter put it more bluntly than any executive will: “Meta can have AI creators optimised to feed the content out in a controlled way to users. And they don’t have to pay for content. Double win for Meta. Double loss for consumers.”

At the same time, the platforms are integrating AI content recommendation systems that, in Zuckerberg’s words, “deeply understand” AI-generated posts and can “show you the right content.” In other words, Meta is openly building the infrastructure to identify, rank, and amplify synthetic content above human content where the synthetic version performs better on engagement metrics. Whether the human creators “win” in that contest is no longer the question; the algorithm decides, and the algorithm is paid by ads, not by integrity.

4.4 The training-data heist

There is a third leg to this stool, and it sharpens the moral picture. The same companies replacing creators with AI built the AI by ingesting those creators’ work without consent or compensation. As of early 2026, the Copyright Alliance counts more than 70 active copyright lawsuits against AI companies. A non-exhaustive sample:

The U.S. Copyright Office’s May 2025 report acknowledged that most training data has been used without authorisation and that a fair-use defence “has yet to be settled.” The pattern is consistent: human creative output, much of it produced for free or for minimal compensation on platforms that then suppressed its reach, was used as raw material to build the systems now displacing those same creators. Investigations have identified at least 13 major datasets containing scraped YouTube content used by Amazon, ByteDance, Snap, Tencent, Apple, Anthropic, and others.

“Should I keep making things in the hope of connecting with people, or just stop altogether?”

5. Fatigue and the Quiet Exodus

5.1 What the data shows

Social media fatigue is not anecdotal. It is measured. The combined picture from Sproutsocial (2025), TrendsActive (2025), DataReportal (October 2025), GWI/Financial Times (2024–2025) and Domus (February 2026) is consistent:

The peer-reviewed literature confirms this. A 2024 paper in Young Consumers (Fernandes & Oliveira) found brand content overload, irrelevance, and ad intrusiveness as significant drivers of social media fatigue, which in turn predicts “lurking” behaviour — users present but no longer participating. A three-level meta-analysis published in 2024 (APA PsycNet) confirmed fatigue as a coherent construct predicting reduced posting and platform abandonment.

5.2 Why people are tired

The drivers cluster around four themes:

Visibrain’s 30-day analysis (December 2025) found 475,000 mentions of “AI slop” across X, Instagram, TikTok, and Threads, with a single peak of 37,000 mentions on 11 December alone after McDonald’s pulled an AI-generated Christmas advert in the Netherlands following backlash. Coca-Cola’s 2025 Christmas “Holidays Are Coming” AI campaign was similarly pilloried. The hashtag #SupportHumanArt has emerged as a sustained co-occurrence with the AI-slop conversation, indicating that user fatigue is producing a partial counter-movement.

5.3 The asymmetry that matters most

The crucial structural fact is that human contribution and AI contribution scale at incomparable speeds. A motivated human posts a few items a day; a slop operator using one prompt template can publish dozens to hundreds. Researchers analysing the Dead Internet Theory (Mladenovic et al., arXiv 2025) note that this creates an asymmetric information environment in which authentic human voices are mathematically drowned out, regardless of intent. Once 80% of feed content is non-human or non-friend, the question of whether the remaining 20% is “original” or not begins to feel almost moot to the user.

6. Conclusion: Is Big Tech Extracting a Human-Made Society and Replacing It with a TV Dinner?

On the evidence reviewed, the conclusion is yes — with two important qualifications.

The first qualification is intent. The pattern described is real, but it is not necessarily the product of a single conscious plan. It is the emergent behaviour of a small number of advertising-supported platforms operating under shareholder pressure in a winner-take-most market. The behaviour looks designed because the incentives are coherent: when the marginal cost of AI content approaches zero, the marginal cost of human content (in payouts, moderation, copyright risk, and brand-safety drama) starts to look intolerable. No conspiracy is needed; an income statement will do.

The second qualification is finality. The substitution is in motion, not complete. Human creativity continues. The most encouraging counter-signal is that audiences increasingly notice and reject slop the moment they recognise it. The #SupportHumanArt response, the boycotts of AI advertising campaigns, the rise of subscription-supported newsletters, podcasts, and Patreon-style direct funding, the deliberate flight to smaller community platforms (Discord servers, Mighty Networks, Substacks, Bluesky niches) — all suggest that a portion of the audience is migrating away from the slop layer towards spaces where the friction, ugliness, originality, and beauty of actual humans is still on offer.

With those qualifications stated, the substantive answer to framing of the hypothesis is unambiguous. Big Tech is, in measurable ways, extracting the value of a human-made culture — ingesting it as training data, demonetising the humans who produced it, and replacing the surface of its platforms with a synthetic, statistically average, prefabricated stream optimised for engagement and unburdened by royalty payments. The TV-dinner metaphor is not unfair; it understates the case. A TV dinner was at least cooked, packaged, and labelled by people who could be held accountable. The current product is generated by systems trained on stolen ingredients, served by recommendation algorithms whose optimisation target is not nutrition but time-on-plate.

7. Consequences If the Trend Continues

The risks below are drawn from the literature reviewed and from the empirical pattern of the past four years. They are organised by domain, with the most well-evidenced first.

7.1 Cultural and creative consequences

7.2 Epistemic and informational consequences

7.3 Economic consequences

7.4 Psychological and social consequences

7.5 Political and civic consequences

7.6 Long-term, civilisation-scale risks

8. Selected Sources

The body of this report draws on the following primary sources, court documents, and reporting. Where the same fact is reported in multiple outlets, the most authoritative or earliest source is cited.

Court testimony and antitrust filings

Earnings calls and corporate statements

Industry data on creator earnings and fatigue

Reporting on AI slop and platform strategy

Copyright lawsuits and AI training data

Academic literature

Theory and commentary

9. A Final Note on Method

This report is framing the substitution as a hypothesis and testing it against available evidence. The evidence cited is drawn from sworn courtroom testimony, public earnings calls, peer-reviewed studies, primary investigative reporting, and the platforms’ own announcements. Where conclusions extend beyond what is documented — in particular the long-term consequences in Section 7 — the report identifies them as projections grounded in current trajectory, not certainties.

Reasonable people may disagree about whether the trajectory is reversible, about which interventions would slow it, and about whether the substitution’s benefits (cheap content, accessibility, certain efficiency gains) are sufficient consolation. What does not seem reasonable, on the present evidence, is the position that nothing is happening. Something is happening. The platforms have said so, on their earnings calls, in writing, in court.

The AI Act was sold as Europe’s attempt to regulate powerful AI systems before they became too deeply embedded in society.

PASCAL HETZSCHOLDT

·

8 MAY

Summary: Europe has not abandoned the AI Act, but it has blinked by delaying key high-risk AI obligations.

Read full story

The Machine’s Human Cost. The lived reality is unpaid testing, chaotic onboarding, sudden project cancellations, constant Slack monitoring...

PASCAL HETZSCHOLDT

·

13:45

Summary: The WIRED article shows how displaced creative professionals are being pulled into precarious AI-training work, effectively helping build the systems that threaten their own industries.

Read full story