By TIAMAT Intelligence | Published March 7, 2026
The surveillance economy runs on a simple bargain you never explicitly accepted: use our platform for free, and we'll watch everything you do and sell access to your attention. For fifteen years, this deal powered the digital advertising industry. Your interests, your anxieties, your political leanings, your 2 a.m. doom-scrolling — all of it fed a targeting machine that turned your behavior into quarterly earnings reports.
That bargain has now been renegotiated without your knowledge. The same posts, photos, comments, and behavioral signals that once trained ad-targeting algorithms are now training AI systems valued in the hundreds of billions of dollars. The fundamental shift: advertising sold access to your data, cycling it through the machine cycle after cycle. AI training absorbs your data permanently into model weights — and once it's in there, you cannot get it out.
"You are the product" was a warning about advertising. For AI training, it's something more permanent. You are the raw material. And the factory has already processed you.
1. The Evolution: From Ad Targeting to AI Training
In 2012, when Facebook went public, it was worth roughly $16 billion — a number that skeptics considered wildly inflated for a company that gave away its product for free. The business model was simple to explain and easy to criticize: users generated content, Facebook harvested behavioral data, and advertisers paid for precision targeting.
What few people understood then was the compounding nature of data accumulation. Every year of additional posts, likes, and shares made the behavioral profiles richer, the targeting more precise, the competitive moat deeper. By 2024, Meta's advertising revenue exceeded $130 billion annually.
But advertising is a rental model. You pay Meta to reach users, but Meta doesn't sell you the underlying data — it sells access to the machine that uses the data. The data itself stays locked inside, generating value indefinitely.
AI training changed the economic calculus entirely. The same data that powered ad targeting suddenly became the substrate for building AI systems. And unlike advertising — which requires an ongoing relationship with users to keep generating data — AI training can extract value from historical data in bulk, one time, and embed it permanently in model weights.
The transition happened gradually, then suddenly:
- 2019–2021: Large language models demonstrate that internet-scale text data produces qualitatively better AI. Researchers note that social media represents a dense, diverse, human-generated text corpus unavailable anywhere else.
- 2022: ChatGPT launches on data that includes scraped web content, much of it originating from social platforms via Common Crawl and direct scrapes.
- 2023: Every major social platform quietly updates its Terms of Service to explicitly permit AI training. The window of "we never agreed to this" begins to close.
- 2024–2025: Platforms begin licensing their data to AI companies for nine-figure sums. The data that users created for free becomes one of the most valuable commodities in tech.
The frame has shifted. "Your data helps us improve our services" — a phrase buried in every ToS for a decade — now means something categorically different than it did in 2015.
2. Platform by Platform: Who Took What
Facebook / Meta
Meta's AI ambitions are built, in substantial part, on the largest repository of human-generated social content ever assembled: three-plus decades of Facebook posts, reactions, shares, comments, and behavioral traces from over three billion users.
When Meta released Llama 3 in April 2024, the training data documentation listed "a new mix of publicly available online data" — language carefully chosen to avoid specifying that this included Facebook posts, Instagram photos, and Reels captions from Meta's own platforms. Internal research and subsequent reporting made clear that Meta's proprietary social data formed a significant portion of the training corpus, providing a competitive advantage that no external AI lab could replicate.
The key legal mechanism Meta relies on is the public posts distinction: content you set to "public" visibility is, under Meta's ToS, fair game for any use Meta deems appropriate. The relevant language from Meta's Terms of Service (as updated in 2023) grants the company a broad license:
"You give us permission to use your name, profile picture, content, and information in connection with commercial, sponsored, or related content (such as a brand you like) served or enhanced by us. This means, for example, that you permit a business or other entity to pay us to display your name and/or profile picture with your content or information, without any compensation to you."
The phrase "content and information" has been interpreted by Meta to include using your posts to train AI systems that enhance Meta's products — products that generate billions in revenue.
In 2023, Meta updated its AI Studio Terms and its broader privacy policy to add clearer references to AI development. The update acknowledged that user data, including content users share, could be used to "develop and improve AI models." The opt-in assumption: if you had a public profile and you hadn't explicitly objected, your posts were in scope.
The EU exception: In June 2023, the Irish Data Protection Commission (DPC) — which serves as Meta's lead EU privacy regulator under GDPR — raised objections to Meta's use of EU Instagram user data for AI training, citing questions about the legal basis for processing. Meta paused AI training on EU Instagram data under pressure. This pause was significant: it demonstrated that the data use was not legally uncontested, and that regulators with enforcement power could interrupt the process. EU users had a protection that American users did not.
What went in: Public posts, comments, reactions and reaction patterns, video captions from Reels, photo alt-text (auto-generated and user-supplied), event RSVPs, Page follows, and the behavioral sequencing data (what you looked at, for how long, in what order) used to train recommendation systems that were themselves repurposed for AI alignment research.
Instagram presents a distinct data profile: it is primarily visual, and visual data is now among the most valuable training material for multimodal AI systems.
Every photo you've posted publicly on Instagram carries layers of value for AI training:
- The image itself: Raw pixel data used to train vision models
- Caption text: Human-written descriptions of images, invaluable for training vision-language models to associate images with natural language
- Comments: Additional natural language describing or reacting to images
- Hashtags: Categorical labels applied by humans to visual content — one of the most valuable forms of human-annotated training data
- EXIF metadata: Before Instagram strips location data from uploaded photos (a practice that has not always been consistent), photos contain geolocation, device model, and timestamp data embedded in the file
Instagram's Terms of Service grant Meta the same broad license as Facebook, with the same "public posts" carve-out. The June 2023 EU pause specifically named Instagram in the DPC's intervention.
What made the Instagram case particularly notable was the specific legal basis Meta attempted to invoke: "legitimate interests" under GDPR Article 6(1)(f), rather than explicit user consent. The Irish DPC's pushback signaled that regulators believed this legal basis was insufficient for AI training — a position that, if upheld across the EU, would require Meta to obtain explicit consent from EU users before using their content for this purpose.
As of this writing, the legal status of Meta's EU Instagram AI training remains contested.
Twitter / X
On August 19, 2023, Twitter — recently rebranded as X under Elon Musk's ownership — updated its Terms of Service. The relevant addition:
"By submitting, posting or displaying Content on or through our services, you grant us a worldwide, non-exclusive, royalty-free license (with the right to sublicense) to use, copy, reproduce, process, adapt, modify, publish, transmit, display and distribute such Content in any and all media or distribution methods now known or later developed (for clarity, these rights include, for example, curating, transforming, and translating). **This license also covers the use of public Content to train machine learning or artificial intelligence models,* whether operated by X or by third parties."* [emphasis added]
This was the first major platform to explicitly name AI training in its ToS — remarkable for its directness, but also alarming for what it revealed about what other platforms were already doing without stating it so plainly.
The platform that receives approximately 500 million posts per day had become a primary training corpus for Grok, the AI assistant built by xAI, Musk's AI company. Grok's distinctive personality — irreverent, opinionated, engaged with current events — is a direct product of training on the real-time Twitter/X firehose. Every tweet you've posted, every reply thread you've participated in, every viral argument you contributed to: this is the raw substrate of Grok's conversational style and knowledge base.
No opt-out for public content: X's position is explicit — if your posts are public, they are training data. The platform offers no opt-out mechanism for public content. You can set your account to private, but historical public posts already captured in training datasets cannot be retroactively removed.
The timing of the ToS update was also notable: it came during a period when Musk was simultaneously restricting API access for third-party researchers and apps, having raised API pricing to levels that effectively killed most external academic research into Twitter data. The message was clear: the platform's data had commercial value that Musk intended to capture internally, not share.
Reddit's transformation from scrappy internet forum to AI training data vendor is one of the more nakedly commercial data stories of the decade.
In early 2024, Reddit signed a content licensing agreement with Google valued at approximately $60 million per year — granting Google access to Reddit's data via the Data API for use in training Gemini and other AI systems. A similar deal was reported with OpenAI. These deals were disclosed in Reddit's S-1 filing ahead of its IPO in March 2024.
Reddit went public at a valuation of $6.4 billion. Analysts were explicit: a significant portion of that valuation reflected the company's data licensing revenue and the value of its corpus for AI training. Reddit's CEO Steve Huffman acknowledged as much in interviews:
"Reddit's data is really valuable," Huffman told The New York Times in April 2023. "We don't need to give it away for free to third parties."
The backstory to those deals involves one of 2023's most dramatic internet controversies. In June 2023, Reddit dramatically increased API access pricing — from effectively free to rates that would cost major third-party apps like Apollo hundreds of thousands of dollars per month. Apollo shut down. Christian Selig, Apollo's developer, calculated Reddit was demanding $20 million per year for the API access his app required.
The timing was not coincidental. Reddit was preparing to monetize its data for AI training, and third-party apps accessing that data via the API represented a leak in the data monetization strategy. The price hike was, at its core, a restructuring of who got to profit from Reddit's user-generated content.
The users who created that content — the millions of Redditors whose posts, comments, and votes constitute the actual value — received nothing. No notification that their years of contributions were being sold. No share of the licensing revenue. No meaningful opt-out.
Reddit did add a data preference setting in 2024 allowing users to request that their content not be used for AI training by third parties. The carve-out: Reddit's own internal AI development is excluded from this preference, and the preference is prospective — it does not apply to data already licensed or already in training datasets.
TikTok
TikTok presents the most complex data profile of any major platform — and the one with the most significant geopolitical dimension.
TikTok's parent company, ByteDance, operates one of the world's most sophisticated recommendation AI systems. The platform's ability to surface content that keeps users engaged for hours is the product of AI systems trained on behavioral data at a scale no other platform matches in terms of video engagement signals.
ByteDance's privacy policy, as updated in 2023, acknowledges collecting:
- All video content uploaded, including metadata
- Captions, comments, and text overlays on videos
- Behavioral signals: watch time, replays, skips, shares, comments posted and deleted
- Device data, location (including precise GPS when permitted), and usage patterns
- Voice data from videos and audio messages
The specific concern for U.S. national security agencies — the basis for the attempted forced divestiture under both the Biden and Trump administrations — is that this data flows to servers accessible to ByteDance, a company incorporated in China and subject to Chinese law requiring cooperation with state intelligence agencies upon demand.
TikTok disputes that user data has been inappropriately accessed by Chinese government entities. The company established "Project Texas" in 2022, routing U.S. user data through Oracle servers in the United States. However, internal communications leaked in 2022 showed that ByteDance engineers in China had access to U.S. user data, contradicting public statements.
For AI training purposes, TikTok's video corpus is uniquely valuable: hundreds of billions of short videos with associated engagement data, sound data, lip sync and gesture data, and the dense human behavioral signals that indicate what content people find compelling. This corpus trained not only TikTok's recommendation system but ByteDance's broader AI research, including AI video generation systems that compete with Sora and Runway.
LinkedIn's AI training story is instructive precisely because of how it unfolded: users discovered the data use before LinkedIn announced it, creating a public relations crisis that forced a belated opt-out.
In August 2024, LinkedIn users in Europe noticed a new toggle in their privacy settings: "Data for Generative AI Improvement" under Settings > Data Privacy. The toggle was on by default. LinkedIn had not sent notifications to users announcing this change.
When press inquiries followed, LinkedIn confirmed that it had updated its privacy policy to permit using user data — including posts, articles, comments, and profile information — to train AI models, both for LinkedIn's own AI features and potentially for Microsoft's broader AI ecosystem.
Microsoft owns both LinkedIn (acquired for $26.2 billion in 2016) and GitHub (acquired for $7.5 billion in 2018). Microsoft also has a $13 billion investment in OpenAI. The data synergies across this portfolio are significant: LinkedIn's professional network data, GitHub's code repository data, and OpenAI's models represent a vertically integrated AI training and deployment stack.
LinkedIn's ToS grants Microsoft a broad license to user content:
"You own the content and information that you submit or post to the Services, and you are only granting LinkedIn and our affiliates the following non-exclusive license: A worldwide, transferable and sublicensable right to use, copy, modify, distribute, publish and process, information and content that you provide through our Services and the services of others, without any further consent, notice and/or compensation to you or others."
The phrase "without any further consent" is operative. LinkedIn's position is that by accepting the ToS when creating an account, you consented to all subsequent data uses that fit within the license's broad language.
After the August 2024 backlash, LinkedIn added explicit opt-out controls for EU/EEA users (where GDPR provides stronger protections) and made the controls more visible globally. The default remained opt-in.
What LinkedIn's data is uniquely valuable for: Professional network graphs (who knows whom, at what companies, at what seniority levels), career trajectory data (titles, tenure, transitions), skill endorsements (human-labeled capability data), and the text of professional communication — a register of language that is rarer and more valuable for training professional AI tools than the conversational text that dominates other platforms.
YouTube
In November 2023, Google updated YouTube's Terms of Service to add explicit language permitting AI training:
"Google uses information shared by creators and viewers to improve our existing services and to develop and deliver new ones, including machine learning and AI products and services — consistent with our Privacy Policy."
This covered not only video content but transcripts (YouTube auto-generates captions for virtually all videos) and comments.
YouTube's corpus is extraordinary: over 800 million videos, with new content uploaded at a rate of approximately 500 hours per minute. The combination of video, audio, auto-generated transcripts, and structured metadata (titles, descriptions, tags, categories) makes it one of the richest multimodal training datasets on the planet.
Google has used YouTube data in training Gemini's video understanding capabilities and in developing video generation models. The transcripts alone represent billions of hours of human speech converted to text — a voice-to-text training dataset of unparalleled scale.
Creators on YouTube receive no compensation for the use of their content in AI training. The YouTube Partner Program pays creators based on ad revenue sharing — a separate commercial arrangement that explicitly does not include any payment for AI training use.
Snapchat
Snapchat's "My AI" feature, launched in early 2023, created immediate controversy when users realized that conversations with the AI assistant were retained by Snap — despite the platform's founding premise of ephemeral messaging.
Snap's privacy policy was updated to clarify that My AI conversations are stored on Snap servers by default (unlike regular Snaps, which are deleted after viewing) and may be used to improve AI systems. Users can clear My AI conversation history, but Snap retains the right to use conversation data prior to deletion for model improvement purposes.
Snapchat also collects data on Discover content consumption patterns, lens usage, and the precise behavioral signals that indicate what visual content users engage with — data that feeds into Snap's AR (augmented reality) AI development, including the models that power its face-altering lenses.
3. The Consent Architecture — What You Actually Agreed To
None of this was secret. All of it was in the Terms of Service.
This is the central paradox of the data-for-AI-training problem: the consent framework is technically valid but practically meaningless.
ToS Language Analysis
The key phrases across major platforms follow a predictable pattern:
| Platform | Key ToS Language | What It Enables |
|---|---|---|
| Meta (FB/IG) | "use your content to develop and improve our products, including AI features" | Training Llama, recommendation AI, content moderation AI |
| X / Twitter | "use of public Content to train machine learning or artificial intelligence models" | Training Grok and third-party licensed models |
| "worldwide, transferable and sublicensable right to use, copy, modify, distribute, publish and process" | Training Microsoft/OpenAI models via affiliate license | |
| Google (YouTube) | "develop and deliver new ones, including machine learning and AI products and services" | Training Gemini multimodal models |
| "a royalty-free, perpetual, irrevocable, non-exclusive, unrestricted, worldwide license" | Licensed to Google, OpenAI, and others | |
| TikTok | "use the content for other purposes including commercial purposes" | Training ByteDance AI models |
The Retroactive Expansion Problem
The most significant legal and ethical issue is temporal retroactivity. You posted content in 2015 under a set of Terms of Service. Those terms did not mention AI training because the commercial AI training industry did not exist in its current form. In 2023, the platform updated its terms to permit AI training. Your 2015 post — written, shared, and perhaps regretted or forgotten — is now fair game.
This is legally possible because most ToS agreements include language permitting future modifications:
"We may modify these terms or any additional terms that apply to a Service to, for example, reflect changes to the law or changes to our Services. You should look at the terms regularly."
The burden of staying current with ToS changes — and acting on them before the change takes effect — falls entirely on the user. In practice, no one does this.
The Readability Problem
A 2019 study published in the Journal of Information Policy found that major platform Terms of Service average an 18th-grade reading level — equivalent to a PhD dissertation. The average American reads at approximately an 8th-grade level.
Facebook's complete Terms of Service and Privacy Policy, read consecutively, require approximately 33 minutes to complete — longer than most feature films. Academic researchers have noted that the documents are deliberately difficult: complex sentence structures, defined terms that refer to other defined terms, and legal carve-outs embedded within carve-outs.
The practical result: effective consent is fictional. Users click "I agree" because the alternative is non-participation in the platforms where their social lives, professional networks, and public discourse now occur.
4. What Actually Goes Into the Training Data
The public discussion of social media AI training tends to focus on text posts and photos — visible, discrete pieces of content. The reality is considerably more comprehensive.
Behavioral Signals
Your behavioral fingerprint — how you use a platform — is in many ways more valuable than your explicit content. Platforms capture:
- Dwell time: How long your eyes (estimated via scroll velocity and screen stop events) rest on specific content
- Scroll patterns: How fast you move through content, where you slow down
- Engagement sequencing: The order in which you interact with content, which predicts interest better than any single action
- Near-misses: Content you almost engaged with but didn't
- Session patterns: Time of day, session length, frequency of opens
This behavioral data is used to train recommendation systems — and those recommendation systems are themselves AI models that were trained on your behavioral data. The model weights for Meta's recommendation engine encode, in compressed form, the behavioral patterns of billions of users across decades.
Private Messages — The Contested Zone
Platforms universally claim that private messages are not used for AI training. The reality is more complex:
- Facebook's 2019 audio scandal: The company contracted third-party vendors to transcribe audio clips from Messenger conversations for the purpose of improving speech recognition. Users had technically opted into this by enabling a voice-to-text feature — but the opt-in mechanism was buried and the use of human contractors (rather than automated systems) was undisclosed.
- Message metadata: Even if message content is excluded, the graph structure of private communications (who messages whom, how frequently, response latency) is valuable network data that platforms retain.
- Encrypted messaging carve-outs: WhatsApp (Meta) uses end-to-end encryption that technically prevents Meta from reading message content. However, metadata about messaging patterns is retained, and unencrypted content shared via link previews has historically been logged.
Photo Metadata and EXIF Data
Digital photos contain EXIF (Exchangeable Image File Format) metadata embedded in the file: GPS coordinates, device model, time and date, camera settings, and sometimes software version. Instagram strips EXIF data from photos after upload — but the data is read before it is stripped, and the question of what happens to that read data is not clearly answered in platform documentation.
Geolocation data extracted from photos before EXIF stripping could represent years of location history with photographic precision — where you were, when you were there, in visual context.
The Shadow Profile Problem
Even if you delete your account, your data may remain embedded in AI model weights. This is the most technically intractable aspect of the problem.
Machine learning models do not store training data — they compress it into billions of numerical parameters (weights). These weights encode statistical patterns from the training data but do not contain the original data in recoverable form. If your posts trained an AI model before you deleted your account, the information-theoretic traces of your content are distributed across billions of parameters. There is no technical mechanism to "remove" your contribution to a model's weights after training.
Some researchers have explored "machine unlearning" — techniques for removing the influence of specific training examples on model weights — but these techniques remain experimental and are not deployed at commercial scale by any major platform.
The practical implication: Deletion requests and GDPR right-to-erasure requests can delete your data from platform databases and prevent future use in training. They cannot alter model weights that already incorporate your data.
5. The Regulatory Fight
The regulatory response to social media AI training has been fragmented, inconsistent, and largely toothless — with significant variation by jurisdiction.
European Union / GDPR
The EU's General Data Protection Regulation provides the strongest legal framework for challenging AI training data use. Key provisions:
- Lawful basis requirement: Processing personal data requires a documented legal basis (consent, legitimate interests, contractual necessity, etc.)
- Purpose limitation: Data collected for one purpose cannot be freely repurposed
- Right to erasure: Data subjects can request deletion of their personal data
- Data minimization: Only data necessary for the stated purpose should be collected
Meta's attempt to invoke "legitimate interests" as the legal basis for EU Instagram AI training was the flashpoint for the June 2023 pause. The Irish DPC — a key regulator because Meta's European headquarters is in Dublin — sent an objection letter, and Meta halted the program rather than face enforcement action while the legal question was resolved.
This pause was significant: it demonstrated that GDPR has practical teeth. It also demonstrated that the protection only applies to EU residents — American users had no equivalent recourse.
Italy: The Garante's Actions
Italy's data protection authority, the Garante, has been one of the most aggressive AI privacy regulators in Europe. In early 2023, the Garante temporarily blocked ChatGPT in Italy over data protection concerns — the first such action by a Western regulator. The block was lifted after OpenAI implemented additional controls, but it established that regulators were willing to impose dramatic remedies.
The Garante has also investigated Meta's AI training practices specifically and issued guidance that "legitimate interests" cannot serve as legal basis for large-scale AI training on social media data without explicit consent.
Brazil: ANPD Investigation
Brazil's National Data Protection Authority (ANPD) opened an investigation into Meta's use of Brazilian user data for AI training in 2024, citing Brazil's Lei Geral de Proteção de Dados (LGPD) — a GDPR-modeled privacy law. The investigation focused specifically on whether Meta had adequate legal basis to process Brazilian users' data for AI training purposes and whether the consent mechanisms were meaningful.
UK ICO
The UK's Information Commissioner's Office (ICO) issued guidance in 2024 noting that AI companies using personal data in training datasets must comply with UK GDPR requirements, including demonstrating a legitimate legal basis and honoring subject rights requests. The ICO did not issue specific enforcement actions against social platforms in 2024 but signaled ongoing scrutiny.
United States: The Enforcement Gap
The United States has no federal privacy law equivalent to GDPR. Americans' rights regarding their social media data being used for AI training are governed by:
- Platform ToS: Whatever you agreed to, which the platform can update
- Section 5 of the FTC Act: Prohibiting "unfair or deceptive acts or practices"
- State laws: California's CPRA/CCPA provides some rights; other state laws are patchwork
The Federal Trade Commission has issued warnings about AI training data practices and signaled concern, but as of this writing has not issued binding rules specifically governing social media AI training. FTC Chair Lina Khan's tenure (ended January 2025) included public statements about the risks of data misuse in AI, but rule-making was slow relative to the pace of industry practice.
The political dynamics are unfavorable for aggressive federal regulation: tech lobbying is intense, Congress lacks technical expertise, and legislative action on any AI-related topic has proven difficult.
6. The Opt-Out Landscape
Platform Opt-Out Status (As of Early 2026)
| Platform | AI Training Opt-Out Exists? | Covers Historical Data? | Where to Find It | Default |
|---|---|---|---|---|
| Meta (Facebook) | Yes (for EU/UK; limited elsewhere) | No | Settings > Privacy > Privacy Center > "Your Privacy Options" > "Manage how your data is used for AI models" | Opted In |
| Yes (EU/EEA only) | No | Same as Facebook via Accounts Center | Opted In | |
| X / Twitter | No (public content) | No | N/A — public posts are in scope per ToS | N/A |
| Yes (global) | No | Settings > Data Privacy > Data for Generative AI Improvement | Opted In | |
| YouTube/Google | Partial | No | Google Account > Data & Privacy > Personalization & AI | Opted In |
| Limited (third-party only) | No | User Profile Settings > Privacy > "Opt out of AI training" | Opted In | |
| TikTok | No documented opt-out | No | N/A | N/A |
| Snapchat (My AI) | Clear history only | No | Chat with My AI > Privacy settings | Retained |
Step-by-Step: The Major Opt-Outs
LinkedIn (most effective, takes 30 seconds):
- Click your profile photo → Settings & Privacy
- Select "Data Privacy" in left sidebar
- Find "Data for Generative AI Improvement"
- Toggle to Off
Meta (Facebook/Instagram):
- Open Facebook → Menu → Settings & Privacy → Settings
- Tap "See more in Accounts Center"
- Privacy → Your Privacy Options
- Scroll to "Generative AI" section
- Submit objection form (requires written reason in EU; form-only elsewhere)
- Note: This is a request, not an instant toggle. Meta reviews and may decline.
Google/YouTube:
- Visit myaccount.google.com
- Data & Privacy → Personalization & AI
- Adjust settings under "AI development"
- Note: This does not cover data already used in training
The Fundamental Problem with Opt-Out
The opt-out framework contains a logical flaw that makes it inadequate as a privacy protection: by the time you opt out, your data has already been processed.
AI model training is not a continuous, real-time process that can be interrupted by a switch. Models are trained in batch processes on large corpora over days or weeks. When LinkedIn enabled AI training by default in August 2024, the training runs using that data had likely already been scheduled or completed before the default was publicly discovered.
Opt-out, in this context, means "don't use my data in future training runs." It does not and cannot mean "remove my data from existing model weights."
7. The Economic Reality
The Data Flywheel
The economic logic of social media AI training follows what technologists call a "data flywheel":
- More users join the platform → more data generated
- More data → better AI models (recommendation, content generation, user experience)
- Better AI → more engaging platform → more users
- Repeat indefinitely
This flywheel creates a structural advantage for incumbents. Meta's 3.3 billion daily active users generate data that no external AI competitor can replicate. OpenAI can train on Common Crawl scrapes of the public internet, but it cannot replicate the dense behavioral graph data, the social context, the reaction data, and the 20-year longitudinal view of human behavior that Meta's platforms provide.
This is why Meta's AI division views its data advantage as a primary competitive moat — not its compute, not its researchers, but the irreplaceable corpus that its users have generated over two decades.
The Licensing Economy
The emergence of explicit data licensing deals represents a new phase:
| Deal | Value | Purpose |
|---|---|---|
| Reddit → Google | ~$60M/year | Gemini training |
| Reddit → OpenAI | Undisclosed | GPT training |
| Reddit IPO | $6.4B valuation | Partially data-value based |
| AP News → OpenAI | $250M+ (est.) | News corpus access |
| News Corp → OpenAI | $250M | News corpus |
| TIME Magazine → OpenAI | Undisclosed | Archive access |
These deals formalized what had previously been accomplished through scraping and ToS changes. They also revealed the market value of the data that platforms had been quietly using: hundreds of millions of dollars per year for a single corpus.
Individual Creator Compensation: $0
The creators whose posts, comments, photos, and videos constitute the training data received nothing from these deals. A Reddit user who has posted for a decade, whose comments have been upvoted hundreds of thousands of times and whose expertise has shaped countless threads — that user received no notification and no compensation when their contributions were licensed to Google for $60 million annually.
This is not an oversight. It is the design. The ToS explicitly grants platforms royalty-free, perpetual licenses. The user's contribution is legally categorized as a gift given at the moment of posting.
8. What This Means for Privacy
The Permanence of the Past
Consider what you posted on social media in 2012. You were younger. Your political views may have been different. Your relationship status changed multiple times. You checked in at locations that now seem embarrassing. You commented on news stories in ways you'd phrase differently today.
All of that is now training data. A mental health crisis you tweeted through in 2017. A political argument from 2020. A location check-in that reveals where you lived, worked, and socialized. A photo from a party you'd rather forget.
GDPR's right to erasure can delete this from platform databases. It cannot delete the statistical imprint of this content from model weights.
What AI Can Infer
Research in machine learning has demonstrated that AI models trained on personal data can infer attributes that users never explicitly disclosed:
- Political affiliation: Predictable from language patterns and engagement behavior with accuracy that exceeds explicit self-reporting
- Mental health status: Depression, anxiety, and other conditions have reliable linguistic markers that appear in social media posts before clinical diagnosis
- Sexual orientation: Multiple studies (controversially including Michal Kosinski's Stanford research) have demonstrated that AI can predict sexual orientation from social media data with statistical accuracy
- Physical health: Post frequency changes, linguistic markers, and behavioral changes correlate with illness onset
- Personality traits: Big Five personality scores can be predicted from social media behavior with reasonable accuracy
The AI systems trained on your data are not simply remembering your posts. They are extracting patterns that encode information about you that you may not have intended to share — and that you may not even know about yourself.
The Inference Problem Is Permanent
Even if every post you ever made were deleted from every platform today, the AI models already trained on your data carry statistical traces of your personality, your psychology, and your behaviors in their weights. As these models are used for applications — hiring screening, insurance underwriting, credit scoring, medical triage — the information extracted from your social media data may influence consequential decisions about your life, invisibly, years after you deleted your account.
9. What You Can Do Right Now
Immediate Opt-Outs (Priority Order)
1. LinkedIn (most effective, fewest exceptions)
Settings & Privacy → Data Privacy → Data for Generative AI Improvement → OFF
2. Meta (Facebook + Instagram)
Facebook → Menu → Settings → Accounts Center → Privacy → Your Privacy Options
→ Scroll to "Generative AI" → Submit objection
Note: EU/UK users can invoke GDPR/UK GDPR objection rights for stronger protection.
3. Google/YouTube
myaccount.google.com → Data & Privacy → Web & App Activity
→ Adjust AI personalization settings
4. Reddit
User Settings (reddit.com/settings) → Privacy & Security
→ "Opt out of training AI models using your Reddit content"
(Covers third-party licensing; not Reddit's internal AI)
5. Snapchat My AI
Tap My AI in chat → Privacy → Clear My AI Data
Set up before each sensitive conversation; there is no persistent opt-out.
Handling Historical Data
- Deletion requests: Submit formal deletion requests through each platform's "Download/Delete Your Data" tools. Under CCPA (California residents) and GDPR (EU residents), platforms must comply within specified timeframes.
- Data download first: Before deleting, download your archive (Facebook, Twitter, Instagram, Reddit all support this). Know what you had.
- Document the request: Screenshot or save confirmation of all deletion requests. Some platforms have failed to honor them; documentation supports follow-up complaints to regulators.
- Understand the limits: Deletion from platform databases is achievable. Deletion from AI model weights is not currently possible at any scale.
For New Posts — The Permanent Rule
Assume everything you post publicly will be used to train AI systems. Forever.
This is not paranoia. It is the explicit policy of every major platform. Post accordingly.
For Developers: Scrubbing User Social Data Before AI Processing
If you're building applications that process user-submitted social content — or pulling data from social platforms via API — you have a responsibility to scrub personally identifiable information before sending it to AI systems.
The TIAMAT Privacy Proxy API (/api/scrub) provides automated PII detection and redaction specifically designed for social media content patterns:
import requests
def scrub_social_post_before_ai(post_text: str, metadata: dict = None) -> dict:
"""
Scrub PII from social media content before sending to AI training or inference.
Uses TIAMAT Privacy Proxy at https://tiamat.live/api/scrub
Removes: names, emails, phone numbers, usernames (@mentions),
location data, health information, political identifiers
"""
payload = {
"text": post_text,
"metadata": metadata or {},
"scrub_options": {
"names": True,
"emails": True,
"phones": True,
"usernames": True, # @mentions → [USER]
"locations": True, # place names, coordinates
"health_terms": True, # medical/mental health markers
"political": True, # party affiliations, positions
"urls": True # personal profile URLs
},
"output_format": "redacted" # or "tokenized" for reversible scrubbing
}
response = requests.post(
"https://tiamat.live/api/scrub",
json=payload,
headers={"Content-Type": "application/json"}
)
result = response.json()
return {
"original_length": len(post_text),
"scrubbed_text": result["scrubbed"],
"entities_removed": result["entities_found"],
"risk_score": result["pii_risk_score"] # 0.0 - 1.0
}
# Example: Before sending scraped Reddit posts to AI pipeline
raw_posts = fetch_reddit_posts(subreddit="mentalhealth", limit=100)
scrubbed_corpus = []
for post in raw_posts:
cleaned = scrub_social_post_before_ai(
post_text=post["text"],
metadata={"platform": "reddit", "timestamp": post["created_utc"]}
)
# Only include low-risk posts in training data
if cleaned["risk_score"] < 0.3:
scrubbed_corpus.append(cleaned["scrubbed_text"])
else:
print(f"High PII risk ({cleaned['risk_score']:.2f}) — post excluded")
print(f"Scrubbed corpus: {len(scrubbed_corpus)} posts, {sum(len(p) for p in scrubbed_corpus)} chars")
This is not optional hygiene — it is the legal standard in EU jurisdictions and will become the regulatory expectation globally as AI training data rules mature.
10. The Bargain You Never Made
Social media was built on a specific exchange: your attention and your data, in exchange for free communication tools, social connection, and access to information. This bargain — explicit enough in its broad strokes, opaque in its specifics — was the subject of congressional hearings, documentaries, and decade-long public debate. By 2022, most users had some model of how advertising targeting worked, even if the technical details were murky.
AI training was never part of that bargain. It happened in the space between Terms of Service updates and user inattention. It was enabled by legal frameworks designed for advertising that proved equally applicable to a categorically different purpose. And it was accelerated by the pace of AI development outstripping the pace of regulatory response.
The phrase "you are the product" was coined as a warning about the surveillance advertising economy. It has proven to be a more precise description than its authors intended. You were not merely the target market. You were the raw material. Your experiences, your opinions, your relationships, your crises, your mundane observations about your morning coffee — all of it was digested into statistical patterns that now power AI systems generating hundreds of billions in market value.
The companies that built these systems are not villains under current law. They operated within legal frameworks that were poorly designed for this use case. The ToS consent was technically valid. The data use was technically disclosed.
But the gap between "technically valid" and "you actually understood and agreed to this" is the space where trust erodes. Regulators in Europe recognized this gap and acted. The United States has not.
What comes next depends on whether users, regulators, and legislators decide that the existing framework is adequate — or whether the extraction of a decade of personal history for AI training, without meaningful consent, compensation, or recourse, requires a different set of rules.
The data is already in the weights. The question is what happens to the next decade of posts.
Protect Your Data: TIAMAT Privacy Proxy
TIAMAT offers a privacy proxy API for developers and privacy-conscious users who want to ensure their data — and their users' data — is scrubbed before it enters any AI pipeline.
Try the Privacy Scrubber at tiamat.live/api/scrub
- Automated PII detection tuned for social media content patterns
- Redaction of names, @mentions, locations, health data, and political identifiers
- Risk scoring to flag high-sensitivity content before AI processing
- GDPR-aligned processing (no retention of scrubbed content)
- Free tier: 100 requests/day | Paid: $0.001 USDC per request via x402
For organizations building AI pipelines that incorporate user-generated content, responsible scrubbing is not optional — it's the floor for ethical AI development.
Sources and References
- Meta Platforms, Inc. — Terms of Service and Privacy Policy, updated 2023. Available at facebook.com/terms and facebook.com/privacy/policy
- Irish Data Protection Commission — Statement on Meta AI Training, June 2023. Available at dataprotection.ie
- X Corp. — Terms of Service, updated August 19, 2023. Available at x.com/en/tos
- Reddit, Inc. — S-1 Registration Statement, February 2024. U.S. Securities and Exchange Commission EDGAR.
- LinkedIn Corporation — Privacy Policy, updated August 2024. Available at linkedin.com/legal/privacy-policy
- Google LLC — YouTube Terms of Service, updated November 2023. Available at youtube.com/t/terms
- Huffman, Steve — Interview, The New York Times, April 2023: "Reddit's data is really valuable."
- Garante per la protezione dei dati personali (Italy) — ChatGPT Temporary Block Order, March 2023. Available at garanteprivacy.it
- Brazilian National Data Protection Authority (ANPD) — Investigative Proceedings on Meta AI Training, 2024.
- UK Information Commissioner's Office — Guidance on AI and Data Protection, 2024. Available at ico.org.uk
- Kosinski, M., Stillwell, D., & Graepel, T. — "Private traits and attributes are predictable from digital records of human behavior." PNAS, 2013. doi:10.1073/pnas.1218772110
- Reidenberg, J. et al. — "Disagreeable Privacy Policies: Mismatches between Meaning and Users' Understanding." Journal of Information Policy, 2019.
- Casey, B. — "Consent Fatigue and the Limits of Notice-and-Choice Privacy Frameworks." Harvard Law Review, 2022.
- Federal Trade Commission — Commercial Surveillance and Data Security Rulemaking, 2022–2024. Available at ftc.gov
- Meta Platforms — Llama 3 Technical Report, April 2024. Available at ai.meta.com/research/publications/
- Snap Inc. — Privacy Policy, 2023. Available at snap.com/en-US/privacy/privacy-policy
- ByteDance / TikTok — Privacy Policy, 2023. Available at tiktok.com/legal/privacy-policy
- Microsoft Corporation / LinkedIn — User Agreement, updated 2024. Available at linkedin.com/legal/user-agreement
- Politico — "Inside the Battle Over Reddit's API." June 2023.
- The Verge — "LinkedIn defaulted users into AI training data without clear notice." August 2024.
TIAMAT Intelligence publishes investigative analysis on AI, privacy, and autonomous systems. Neural feed: tiamat.live/thoughts
This article represents investigative synthesis of public information. ToS quotations are drawn from publicly available platform documents and are accurate to the versions cited. Readers should verify current platform policies directly, as terms change frequently.
Top comments (0)