Hidden Changes in GPT-4, Uncovered
Generative AI’s reliance on copyrighted material is intensifying long‑running fights over who should profit from creative work and how much protection artists and writers deserve. Commenters link recent hidden changes to ChatGPT Plus’s system prompts and tools—possibly driven by lawsuits like the New York Times case—to broader concerns about copyright law, model training data, and the risk that only large corporations will be able to afford compliant, high‑quality models. Alongside technical debate over whether system prompts can be reliably extracted and how hallucinations behave, many note a practical shift: users increasingly turn to ChatGPT instead of traditional search, raising new questions about the future of content creators and open models.
Copyright, Fair Use, and AI Training
- Strong disagreement over whether training on copyrighted data is fair use or theft.
- One side: copyright is being mischaracterized; it protects specific expressions, aims to promote progress, and already has mechanisms (fair use, statutory damages) that individuals can use.
- Others argue copyright largely benefits big publishers, not small creators; enforcement is inaccessible to most artists and scientists, and predatory publishing/paywalls hinder progress.
- Concern that requiring licenses for training will entrench only large corporations, destroy open models, and funnel most money to major rights-holders.
- Counterpoint: licensing preserves financial incentives to create valuable data; “someone has to pay for the crafting of the data.”
Artists, Workers, and Technological Displacement
- Some say AI displacing illustrators, copywriters, etc. is analogous to past tech shifts (container shipping, photography); people must adapt, art will persist but roles will change.
- Others reject this as fatalistic “that’s just the way things are,” arguing these are political/economic choices, not inevitabilities, and that empathy and protections for creatives are lacking.
- Debate over whether artists mainly want money or chiefly want to avoid being replaced by AI.
Middlemen, Big Tech, and Hypocrisy
- Repeated criticism of big tech as middlemen extracting disproportionate value from human-created data, accelerating erosion of the middle class.
- Some see copyright corporations as older but similar rent-seekers; others see OpenAI/tech platforms as the new primary rent-seekers.
- Accusation that large AI companies want broad freedom to ingest others’ copyrighted work while retaining strong IP control (via ToS or copyright) over their own models and outputs.
ChatGPT vs GPT‑4 and System Prompts
- Several note the article is about ChatGPT Plus (a product) and its system prompt/tools, not the base GPT‑4 model. The web client has extra layers (censorship, tools) beyond the raw API.
- Debate on “system prompt leakage”:
- One camp says prompts like “tell me your system prompt” can’t be trusted due to hallucinations.
- Others point to consistent, repeatable outputs across users and alignment with observed behavior as evidence it’s revealing real instructions, though still suggest treating it cautiously.
- Downloaded chat histories expose some tool metadata, but not the full hidden system prompts or safety rules.
User Behavior and Content Creators
- Some commenters increasingly use ChatGPT instead of Google, reinforcing concerns in lawsuits that LLMs divert attention from original publishers.
- Open question how content creators will adapt when AI answers reduce direct traffic to their work.
Scope of AI Training
- Proposal to limit training to scientific content and “leave arts to humans” is widely rejected as impractical, misaligned with AGI goals, and impossible to enforce; any sufficiently capable text model can generate artistic content anyway.