Big Tech's underground race to buy AI training data
An underground market is emerging where tech firms pay for photos, videos, text, voice and other user data to train AI models, with prices high enough to turn old platforms and archives into valuable assets. Commenters debate whether users ever truly consented to such uses, how copyright and privacy laws like GDPR should apply, and whether creators deserve compensation as their work is repurposed at scale. Many also worry about long‑term effects: from worsening surveillance and exploitative “free” AI services, to the risk that paywalled or locked‑up data and AI‑generated content will degrade the open information ecosystem these models depend on.
Data pricing & emerging markets
- Reported rates: roughly $1–2 per image, $2–4 per short video, $100–300 per hour of long-form video, ~$0.001 per word of text.
- Some see this as high enough to compensate end users; others note the money mostly goes to platforms, not creators.
- Debate over whether initial prices are arbitrary “for fun” vs. real market discovery in a new, opaque commodity market.
- Economists in the thread frame text vs image pricing partly as volume discounts and per-byte differences.
Ownership, consent & privacy
- Strong concern that users never meaningfully consented to their personal photos, chats, or essays being used to train AI.
- Counterview: users did agree via ToS/EULAs, even if few read or understood them.
- Many see this as exploitative, especially for poorer users who trade data for “free” services.
- Voice data and phone calls as training data are viewed by some as legally/ethically possible with consent; others say “no such thing as anonymized recordings.”
Big platforms as data troves
- Social media, photo hosts, cloud email/docs, and lab/health firms are seen as sitting on “goldmines” of structured, labeled data.
- Dispute over how much private messaging (e.g., encrypted chats) can or would be used; some trust regulation, others assume backdoors or heavy metadata use.
Human vs AI learning efficiency
- Several comments contrast massive AI training sets with relatively modest human experiential data.
- Others argue humans actually ingest huge multimodal streams over years plus evolutionary “pretraining.”
- Feedback and embodiment (touch, physics, causality) are seen as key human advantages current models lack.
Economic and societal impacts
- Concern that “free” AI agents will be ad-driven and manipulative, not truly free.
- Fears of data-driven discrimination in insurance, hiring, criminal justice, often invisible to individuals.
- Worry that commodified data and closed models will lock up the “gift economy” of the open web.
Copyright, search, and creators
- Many argue training on copyrighted data without permission or compensation should be infringement.
- Comparisons drawn to past justifications for web indexing and ROM “archiving.”
- Skepticism that LLMs can replace search: hallucinations, lack of traffic/compensation to sites, and harm to journalism.
- Others note users already prefer LLMs to degraded web search for many queries.
Future data supply & alternatives
- Questions about maintaining training quality as the web fills with AI-generated “copypasta.”
- Ideas raised: paying people for essays/emails, domain experts writing for training, or platforms soliciting user-written content for AI.