이 페이지에서4분 읽기
개요
These measures mostly affect future collection by companies that choose to respect them; they do not remove work from models already trained.
심층 분석
Most large AI models are trained on material collected from the web. Common Crawl publishes huge crawls of the public internet, and image datasets such as LAION-5B, with about 5.85 billion image and caption pairs, were built by filtering them. LAION distributes links and captions rather than image files, but models trained with it downloaded the images from those links. To check whether your work appears, Spawning's Have I Been Trained lets you search LAION datasets by text or image. It covers only the datasets it indexes; there is no way to search the private data many companies use. The main prevention tools are: robots.txt. Many AI companies publish crawler names that sites can block, including GPTBot (OpenAI), ClaudeBot (Anthropic), CCBot (Common Crawl), Applebot-Extended (Apple) and Google-Extended, which controls use in Google's Gemini models without removing a site from Google Search. Some hosts and CDNs, such as Cloudflare, offer one-click AI crawler blocking. Platform settings. Services like LinkedIn and X have offered settings controlling training on user content, and Meta has offered objection forms in regions with stronger data protection law. Art sites such as DeviantArt and ArtStation introduced NoAI tags. Registries and legal reservations. Spawning's Do Not Train registry records opt-outs that some companies, including Stability AI for certain models, said they would honor. In the EU, the 2019 copyright directive allows rights holders to reserve text and data mining rights in machine-readable form, and the EU AI Act requires general-purpose model providers to respect such reservations. The limits are important. robots.txt is voluntary and not enforcement. Opt-outs are not retroactive, and trained models do not forget. Copies of your work reposted elsewhere are not covered by your site's rules. Blocking crawlers can also reduce visibility in AI-powered search. Opting out lowers exposure; it does not guarantee exclusion.
전략적 영향
위험과 안전
치명적인 AI 피해와 일상적인 AI 피해는 누가 위험을 이해하고 누가 조치를 취할 수 있는지에 따라 달라집니다.
더 명확한 결정들
공공 및 전문 지식은 강력한 안전 정책이 정치적으로 가능한지 여부를 결정합니다.
과장된 과장을 뚫고 나가기
명확한 설명은 과대광고, 연구실 홍보, 모호한 윤리 연극에 의한 포착을 줄입니다.
The Future of Opting Out of AI Training Datasets
Pressure is growing for opt-out signals that are standardized and legally meaningful rather than scattered across crawler names and platform menus. Standards bodies and industry groups are working on shared vocabularies for AI usage preferences, and EU rules are pushing providers to document how they respect reservations. Licensing deals between AI companies and publishers suggest a market for consented data is forming, though mostly for large rights holders. Lawsuits over training on copyrighted work may change the default rules in some countries. For now, individual creators should treat opt-outs as partial protection and check settings periodically, since platforms change them.
실제 구현
An illustrator searches her portfolio images on Have I Been Trained, finds several in LAION-5B, and adds them to Spawning's Do Not Train registry, knowing this only binds companies that honor it.
A photographer who runs his own site adds robots.txt rules disallowing GPTBot, CCBot, ClaudeBot and Google-Extended, while leaving Googlebot allowed so his pages still appear in search results.
A writer on LinkedIn switches off the setting that allows her data to be used to train generative AI models and notes that this does not undo past use.
A small publisher in the EU adds a machine-readable reservation of text and data mining rights to its site and terms of use, relying on the EU copyright exception that lets rights holders opt out.
위험 및 가드레일
실존적 위험을 공상과학처럼 다루면서 능력을 합성합니다.
높은 자율성 하에서 정렬과 표면 제품 안전성을 혼동합니다.
영어가 아니거나 전문가가 아닌 청중에게는 품질이 낮은 소스만 남겨 둡니다.
구현 로드맵
제품 손상, 오용, 통제력 상실/잘못 정렬 위험을 분리합니다.
일정과 심각도에 대한 귀하의 견해를 바꿀 수 있는 증거가 무엇인지 물어보십시오.
마케팅 주장보다 기본 소스와 구체적인 평가를 선호하세요.
인식뿐만 아니라 경력, 정책, 자금 조달 또는 기술 등 하나의 행동 경로를 식별하십시오.
계속 탐색하세요
Free newsletter
Get the daily AI briefing
Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.
One email each weekday. Unsubscribe in one click. We never sell or share your address.
Test yourself
Take the Opting Out of AI Training Datasets quiz
Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
자주 묻는 질문
What is Opting Out of AI Training Datasets?
Creators can reduce future use of their work for AI training by blocking AI crawlers in robots.txt, turning off training permissions in platform settings, and registering with opt-out services, and can check some public datasets with search tools. These measures mostly affect future collection by companies that choose to respect them; they do not remove work from models already trained.
What does the LAION-5B dataset actually distribute?
LAION provides URLs and captions. Model trainers downloaded the images from those links.
What does blocking Google-Extended in robots.txt do?
Google-Extended is a control token for AI training use. Blocking it leaves Search indexing untouched.
What is the main limitation of opting out after a model is trained?
Opt-outs affect future collection by companies that honor them. Past training is not undone.
Why does blocking crawlers in your own site's robots.txt not fully protect your work?
robots.txt is a request that well-behaved crawlers follow, and it only applies to the site that publishes it.
What does Spawning's Have I Been Trained tool let you do?
It searches the datasets it indexes. It cannot see private training data used by many companies.
계속 학습하세요
관련 가이드
이 주제에 대해 선택된 추가 가이드