← All pages

Fine-Tune

Definition: A clause disclosing that a platform fine-tunes — adapts an already-trained AI/ML model using additional, often user-derived data — rather than training a model from scratch, creating an ongoing feedback loop where user interactions keep reshaping the model. Risk level: This is more specific than generic “train AI/models” language: it confirms a live pipeline where a user’s own conversations, content, or profile data directly shape the model’s future behavior, not just a one-time historical training corpus. The risk grows further when that fine-tuning data flows to a third party — like a parent company — instead of staying with the platform that collected it.

Platforms Using This Clause

Common Wording

In addition, at this time, X prohibits any use of the X APIs and/or X Content to fine-tune or train a foundation or frontier model with the exception of Grok. — X, Other (Developer Policy, Restricted Use Rules), 2025-06-04 (the named “Grok” exception is only legible in the 2026-03-10 scrape — earlier scrapes captured this sentence with the exception name link-stripped out) — a restriction on developers, not a disclosure of X’s own fine-tuning practice: third parties are barred from fine-tuning models on X data via the API, while X’s own AI product is explicitly exempted from its own rule — a clear self-preferencing data moat.

We use your interactions with AI models and technologies like Bard to develop, train, fine-tune, and improve these models to better handle your requests, and update their classifiers and filters including for safety, language understanding, and factuality. — YouTube, Privacy Policy, 2023-11-16 (present in substance throughout, with “Bard” rebranded to “Gemini Apps” from 2024-11-20 onward) — the platform-fine-tunes-its-own-model direction: a user’s own AI-assistant conversations feed back into fine-tuning that same assistant, with no fine-tuning-specific opt-out shown in the captured snippets.

To train, fine-tune, evaluate and improve our Generative AI models used to create content (textual, audio, visual, and other media, or multimedia) on the LinkedIn platform (or elsewhere) and in LinkedIn’s lines of business. — LinkedIn, Privacy Policy, 2025-06-10 (present in substance throughout) — the same platform-fine-tunes-its-own-model direction as YouTube’s, naming concrete products (InBart, Collaborative Articles, Account IQ) built on the fine-tuned output. LinkedIn’s Privacy Policy additionally discloses (for UK members specifically) that data is shared with parent company Microsoft “so that it can train, fine-tune, and evaluate… Microsoft’s AI models” — the only confirmed cross-company fine-tuning data flow in this wiki.

…use or share Outputs that will be used to train, develop or fine tune models, services or other AI technologies… — Snapchat, Terms of Service, 2025-03-06 — a second confirmed restriction-direction occurrence alongside X’s: rather than disclosing its own fine-tuning practice, Snapchat bans its users from using or sharing AI Outputs to fine-tune other models or AI technologies, protecting its own AI Features’ outputs from being harvested into third-party training pipelines — framed as an end-user Acceptable Use rule rather than a developer-API restriction like X’s, and with no named self-exception comparable to X’s Grok carve-out.

…you are prohibited from using, accessing or aggregating data obtained from the Telegram platform to train, fine-tune or otherwise engage in the development, enhancement or deployment of artificial intelligence, machine learning models and similar technologies. — Telegram, Other (Bot/API Developer Terms), 2026-02-03 — a third restriction-direction occurrence, developer-API-facing like X’s: bans developers from fine-tuning AI/ML models on data obtained via the Telegram API, with no named self-exception.

Notes & Trends

X’s occurrence is structurally unique among the three: rather than disclosing that X fine-tunes its own model on user data, it’s a restriction on developers — banning them from fine-tuning any model on X data via the API — paired with an explicit carve-out for X’s own AI product (Grok). This is the wiki’s clearest “data moat” pattern tied to this keyword: X gets to use its own data for AI, but blocks competitors from doing the same. YouTube’s and LinkedIn’s occurrences both run in the platform-fine-tunes-its-own-model direction, with a feedback loop from individual user interactions (conversations, content) into live model updates — YouTube’s tied to its own AI assistant (Bard/Gemini), LinkedIn’s tied to its content-generation features. LinkedIn’s occurrence goes one step further than any other platform tracked here: it explicitly discloses that fine-tuning data is shared cross-company, with parent company Microsoft fine-tuning its own separate AI models on LinkedIn member data (UK members specifically) — a corporate-group data flow not disclosed on any other platform’s page in this wiki. Snapchat’s Terms of Service joins X as the second platform using this keyword in the restriction direction, but frames it as an end-user Acceptable Use Policy violation rather than a developer-API ban. Unlike X’s explicit Grok carve-out, it names no self-exception, leaving Snapchat’s own potential use of AI Outputs for fine-tuning unaddressed by this particular clause. Telegram’s “Other” page returns to the developer-API-facing framing, like X’s, and is the third platform in this wiki to use this keyword in the restriction direction rather than as a first-party disclosure.