We are also extremely proud of how rigorous and thorough our pre-release testing process is, including a full stack of frontier risk evaluations according to our Preparedness Framework and external red teaming. Read more safety work and evaluations in:
REGULATION
-
AI risks: misinformation, manipulation, and harm concerns
By
–
I hear you but I'm worried it will increase the risk of misunderstanding/misinformation/harm/manipulation
-
Web-Scale Data: Legal Risks and Compliance Requirements
By
–
Voilà! Overall, dealing with web-scale data is inherently risky and requires due care and good faith to reduce legal risk. In lawsuits, the burden of proof quickly shifts onto the defendants (because it's an affirmative defense) and they need to provide evidence of compliance.
-
Partner with IWF upfront to prevent illegal content in AI datasets
By
–
Partner with IWF upfront, not when problem arise! LAION knew about the risks of illegal content (e.g. child abuse) being in their dataset. It would have been sensible to partner with professional organizations like IWF to check files *before* publishing the dataset.
-
Fair Use Licensing for AI Datasets: Legal Compliance Strategy
By
–
Use an appropriate license compatible with Fair Use. LAION's creation of the dataset required one or more Copyright Exceptions with strict terms. Picking a license that respects the terms of these exceptions would reduce legal risk significantly, and show good faith.
-
User Generated Content Exclusion from AI Training Datasets
By
–
Exclude social media and UGC websites. Popular websites that host user generated content or personal profiles could have been excluded from the dataset. Platforms don't have full rights to their users private information, and avoiding that kind of content reduces legal risk.
-
LAION criticized for partnering with questionable data hosting site
By
–
Don't partner with dubious websites to host files! LAION shouldn't have requested TheEye, a questionable "archival" website, to host their data in order for them to link to it! Maybe instead get approval from platforms like HF so they share legal responsibility for hosting.
-
Respecting Manual Opt-Out Requests After Fact
By
–
Provide opt-out after the fact. To operate in good faith, manual opt-out requests should also be respected after the fact too. Because the dataset format allows users to bypass and override opt-out client side, not honoring request makes LAION look complicit in infringement.
-
Database Storage for LAION-5B: Removing Illegal Content Safely
By
–
Use a database storage method not version control. Storing the LAION-5B dataset on revision control means nothing can be removed (e.g. private data or illegal content) without leaving a trace or nuking history. Using a database storage + wrapper API would avoid this problem.
-
Copyright Compliance: LAION Must Prove File Deletion Records
By
–
Publish code & logs upfront show files were deleted. Complying with Copyright exceptions means the burden of proof is on LAION to show they qualify. They're required to delete the files so keeping records of this (or publishing them upfront) could help avoid court.