As of November 30, 2025, neither the legal system nor the market has clearly settled in the way Jason predicted, but there are meaningful moves in that direction, so the prediction can’t be cleanly called right or wrong.

Legal requirements / permission for data owners

• In the U.S., there is still no specific statute that requires generative AI providers to get permission from all copyright holders whose works are used in training. Courts have issued early rulings treating some training on copyrighted books as fair use (e.g., Meta’s Llama case and Anthropic’s Claude case), confirming that permission is not categorically required, though the law remains unsettled and many lawsuits are ongoing. (theguardian.com)
• In the EU, the AI Act now imposes transparency and copyright-compliance duties on providers of general‑purpose AI models. They must publish a summary of training data and respect copyright, including honoring rights‑holders’ opt‑out from text and data mining; if a rightsholder opts out, authorization is required. But the regime is based on exceptions plus opt‑outs, not a blanket “permission for everything” rule. (hoganlovells.com)
• In China, interim measures for generative AI require use of “legitimate data sources,” non‑infringement of IP, and consent for personal data, again signaling stricter standards but not a globally uniform mandate that all training data be licensed. (staticpacific.blob.core.windows.net)
• In the U.S., the proposed Generative AI Copyright Disclosure Act would require companies to disclose which copyrighted works (with URLs) they used in training, but it explicitly does not ban use of copyrighted works for training and, as of late 2025, remains only proposed legislation, not enacted law. (en.wikipedia.org)

Market pressure, licensing deals, and citations

• Major AI vendors have voluntarily struck content‑licensing deals with publishers (e.g., OpenAI with the Associated Press, Axel Springer, and Condé Nast; Perplexity with Le Monde), which allow use of that content for training and display, including attribution and links in answers. This is clear evidence of market pressure toward licensing and attribution, but it is selective and far from universal. (cnbc.com)
• At the same time, many high‑scale models still appear to rely heavily on large web scrapes and fair‑use/text‑and‑data‑mining theories rather than comprehensive opt‑in licensing, as reflected in ongoing copyright lawsuits by news and reference publishers against OpenAI and Perplexity. (theverge.com)

Citation / linking behavior in products

• Some generative AI services, especially those positioned as search tools, now routinely provide citations and links. OpenAI’s ChatGPT Search integrates a search engine that “generates responses, including citations to external websites,” and Deep Research is built explicitly to create cited reports. (en.wikipedia.org)
• Competing systems like Perplexity and Bing Copilot similarly foreground linked sources, but empirical work on “attribution gaps” shows that web‑enabled LLMs frequently fail to credit many of the pages they read, and many responses contain few or no clickable citations. This underscores that citation practices are product‑design choices, not mandatory legal obligations. (arxiv.org)
• There is, as of 2025, no law in the U.S. or EU that generically requires chatbots like ChatGPT to attach per‑answer citations to their training data or even to all web sources consulted, though EU rules do require public summaries of training data for large models and respect for copyright and opt‑outs. (hoganlovells.com)

Why this is ambiguous rather than right or wrong

• Jason’s claim had two parts: (1) services like ChatGPT would be required to use citations and links, and (2) they would be required to get permission from data owners to operate at scale. As of late 2025:
  • Citations and links are common in some product modes (especially search) but not universally mandatory across all generative AI use cases.
  • Permission/licensing is becoming more common for certain high‑value content and is reinforced by EU and Chinese rules, but large‑scale models can and do still operate relying on fair use or text‑and‑data‑mining exceptions without comprehensive opt‑in permission.
• Because his statement is framed as an ultimate end state (“that’s where this is all going to shake out”) with no explicit time horizon, and regulatory and market trends are still evolving—some in the direction he predicts, others not—we cannot yet say his forecast has clearly succeeded or clearly failed.

Given this mix of partial convergence (more licensing, more attribution in some products, new transparency rules) and clear gaps (no global obligation for per‑answer citations or universal permissions, plus ongoing reliance on fair‑use/TDM), the fairest classification today is ambiguous rather than decisively right, wrong, or simply “too early” in a time‑bound sense.