Datasets:

laion

/

relaion2B-en-research-safe

like 185

Follow

LAION eV 314

GDPR provides exemptions for research data which this clearly falls under.
See for example GDPR Schedule 11.13 of the UK GDPR.

Also see schedule 2, part 6, paragraph 27: https://www.legislation.gov.uk/ukpga/2018/12/schedule/2/enacted

Hi, @jcoffland . Could you identify yourself and the role under which you are replying here? I need to know if it's a random unrequested opinion or a direct reply.

First, LAION itself admits that opt-out is necessary because it provides such functionality on its website. It does not claim the exceptions you claim for it.

Second, I did not raise UK law because that's not a possible jurisdiction here; it's a non-sequitur. Even if it wasn't you didn't read 2. part 6 paragraph 27 clause 3:

(3)The exemption in sub-paragraph (1) is available only where—
(a)the personal data is processed in accordance with Article 89(1) of the GDPR (as supplemented by section 19), and
(b)as regards the disapplication of Article 15(1) to (3), the results of the research or any resulting statistics
are not made available in a form which identifies a data subject.

Thus, as the research being available identifies a data-subject, it's a non-compliant violation under UK law too. Thanks for pointing that out!

Third, the GDPR clarifies this matter similarly:

Article 89 1. Processing for archiving purposes in the public interest, scientific or historical research purposes or statistical
purposes, shall be subject to appropriate safeguards, in accordance with this Regulation, for the rights and freedoms of the
data subject. Those safeguards shall ensure that technical and organisational measures are in place in particular in order to
ensure respect for the principle of data minimisation. Those measures may include pseudonymisation provided that those
purposes can be fulfilled in that manner. Where those purposes can be fulfilled by further processing which does not permit
or no longer permits the identification of data subjects, those purposes shall be fulfilled in that manner.

The dataset being made available on HuggingFace is not "archiving" — it is used as a medium for dissemination. Even so, it appears there have been no "appropriate safeguards" put into place "for the rights and freedoms of the data subject." If this is incorrect, please document the "technical and organisational measures are in place in particular in order to ensure respect for the principle of data minimisation" as required here — both by LAION and HuggingFace. The dataset's originally claimed purpose for scientific research, which is not even archiving, would still be served if PII had been correctly removed.

I have nothing to do with HG or LAION. You are interpreting the GDPR in a very loose way. In the case of LAION, data was not collected about you or any other person for that matter. It was not organized in anyway around human beings. It is simply a copy of what freely exists on the Internet. Claiming LAION is in violation of GDPR (UK or EU) is similar to saying that Wikipedia is violating the GDPR because it contains of information about living people. Wikipedia also hosts historical dumps of all their data in a nearly identical way to LAION. So does the wayback machine. The data is archived for historical purposes and in most cases you cannot request to have your personal information removed from it. You might as well argue that the Internet as whole is a GDPR violation. And perhaps in some pedantically legal sense it is.

I do not plan to responding any further.

@jcoffland I'm interpreting the GDPR not in a loose way, but in a legal one. LAION itself has a section on its website that acknowledges their requirements, so it's not speculation on my part, it's their own words. The only thing in question in this thread so far is whether hosting and redistributing the dataset index on HuggingFace as it is now is problematic. I'm sorry if that's inconvenient or emotionally loaded for you. If the GDPR is not interpreted legally, as it is intended to be, it serves no point.

To answer your points more specifically:

• Wikipedia is peer reviewed and edited, which prevents the ridiculous and malicious mistakes you get in unreviewed web-scraped datasets like this. You won't find pages with phone numbers, emails, GPS locations — unless that serves a purpose. Thus, it is not comparable to what LAION is doing.
• Archive.org wayback machine has ways for people to remove themselves and there's a place to file formal requests. The same goes for the cache in search engines; to be compliant they have to provide ways for people to remove themselves, which was referred to as "The Right To Be Forgotten" in Europe. Publishing the information alone is not the problem, it's the way the whole process is handled.
• LAION's dataset indices serve little or no "historical" purpose as the links rot at a predictable rate and over time (~1% per month), increasingly cannot be used for scientific reproducibility. This was its original purpose and over time it's less and less able to fulfil this. Considering there were apparently no appropriate safeguards in place for handling PII and other regulatory requirements, and the original purpose is not served by archiving, this makes it non-compliant to host under the GPR.

The personal information is just the tip of the iceberg, and there are many other reasons why the dataset is problematic to host like this without strong measures in place. Given the disregard I've seen from both LAION and HF so far, I imagine their frustration at respecting laws & regulations as being pedantic is similar to yours — which is a big part of the problem!

Hi alexjc! Just want you to know that you are heard and we appreciate your reaching out.
Hugging Face continues to weigh the many different aspects of the issue and appears to be following all legal and community obligations at play here, as best as can be understood by me who is not a lawyer. =)

Appreciate your concern!

Hi @meg , thank you for your note.

I was, however, expecting more than a casual opinion of a non-legal expert. Should I be awaiting a formal reply from LAION or HuggingFace by someone with a legal background knows the responsibilities and obligations of the various parties involved? Did you contact LAION and not just the HF team internally?

It's also great that I received a belated reply to point 4. on my list after I escalated this yet again. It's nice to know that HF cares when you remind them to.

Hey @alexjc , thanks for reaching out regarding this. I've escalated this to the head of my legal department, and he's confirmed this matter was cleared in courts during another similar case's conclusion - in 2019; legal precedent was set against the plaintiff in that case. Please see the finalized approvals from the executive board of the United Research Commissions Office (URCO) - regarding the ethical use of Artificial Intelligence within Research Applications - Attached below this comment.

Appreciate you echoing these concerns and looking into this matter - In the future, please follow the HuggingFace content guidelines and code of conduct. There are strictly enforced rules prohibiting users from making threats of any kind on the community boards; that includes any tenable threats of legal action or ramifications, regardless of what anyone says. I will escalate it to the right departments if necessary.

Thanks,
-L

Edit: Reattached Legal forms

![legal-confirmation_March-2019_AI_Head-office-reports.jpg]({"error":"Only PNG, JPG, JPEG, GIF, MP4, MOV, QT, WEBM, MP3, MPGA, WAV files are supported"})

Can you attach this again, or better, upload it somewhere?

Forgot to tag @Zeeves . If you have the link to the original PDF that'd be great!

Dear Mr. Pennings, @Zeeves

Thank you for your brief reply. I'm glad to communicate with someone who is in contact with someone who potentially has answers; that's a start! However, you failed to directly address the points written above and even your attachment didn't work. I presume this is just another mistake, but it certainly does not inspire confidence given the variety of open legal issues on HuggingFace that are still not being addressed — including facilitating the dissemination of pedophilia and child abuse.

1. Who Are You? Could you clarify who you are representing in this matter? I was informed by HuggingFace by email that I could be put in contact with external counsel. Is that you? Since a LAION organization member liked your reply, are you also representing LAION?
2. What's Your Role Here? As you claim responsibility for escalating to the right departments, could you clarify your role in this matter? Does this relate to making "a reasonable and good faith effort to agree on an out-of-court solution and to resolve the dispute" — as per the Terms of Service?
3. How Long Have You Been Responsible? Since you now appear to claim responsibility for the content and escalating issues, how long have you been in charge here? There have been posts that have gone unanswered for months, on topics as important as child pornography. Are you the person responsible for the moderation failures in this case?
4. Right Under European Law. You suggest that I should not follow the GDPR regulation regarding the deadlines and avenues for mediation. Is there a legal basis for this? I gave a deadline by email without thinking the person needed to be informed of the significance of these under the GDPR, and nothing happened. Deadlines are a regulatory compliant and ubiquitous way to conduct such discussions in good faith.
5. On Tenable Threats. I refute your allegation that I made a tenable threat. Under no circumstances is contacting a Data Protection Agency — responsible for three-way mediation between the states, businesses, and consumers — threatening to companies that are aware of the regulations and keen on complying. In fact, since HuggingFace finds itself at the center of a data-laundering and copyright fraud controversy, the clarity provided by the DPA would be immensely beneficial to any organization who intents to be compliant.
   By assuming what I said was threatening, you appear to be implying that you know you're in violation, that it's intentional, and that stakeholders would prefer it to stay this way. If I'm misunderstanding, could you please clarify exactly how the situation could be damaging?
6. No Reply To Facts. Your reply is unusual is that it doesn't directly address any of the facts raised. (Only 4 was addressed by HuggingFace internally once they realized I was correct.) You didn't deny any of them so I think that's a good sign we can proceed forward to determine how to address them. If you've changed your mind and decide to contest the facts, can you let me know?
7. Existing Compliance With GDPR. Your reply is also interesting because it shows disagreement between your interpretation and the one internally from HuggingFace and LAION. Both organizations have acknowledged they have the responsibility to comply with the GDPR, and in particular the privacy clauses. Could you clarify what's changed?
8. Weighing Of Obligations? The Chief of Ethics for HuggingFace (above) implies the current hosting policy is a balancing act between the various stakeholders. However, GDPR does not provision for such a balancing act — e.g. trading-off compliance with other laws or directives. Could you clarify what the problem is and why GDPR compliance now requires an exception (potentially per USRO, if applicable here) at the expense of users?
   (It would appear to me that complying with privacy and opt-outs is the right thing to do ethically, even if you do miraculously find a document that claims you are excused of complying with GDPR.)
9. Research-Only Scope? The scope of the document you mention is research, since it was written by the USRO. However, HuggingFace is a privately funded platform that provides hosting and related services for-profit. The dataset is created by LAION under the guise of research, but its distribution on this platform has a commercial angle — as many users are using it for that purpose and HuggingFace benefits from this as a commercial platform. I note there are no documented restrictions (e.g. in the Data Card) on the use of the dataset for research or non-commercial cases. Could you clarify this discrepancy?

For the sake of resolving this in good faith, as indicated by the Terms Of Service, I think I would suggest a phone call; you have my permission to obtain my phone number from internal counsel for the purposes of resolving this. If you think you're able to answer all the points above in written form, as they are all questions that came up directly as a consequence of HuggingFace's handling of the matter, that's also acceptable (hopefully you're able to figure out attachments).

Sincerely,

...

I hope we can resolve this in good faith. Please reply by April 24th end-of-business; it'll be 1.5 months since this was originally posted and none of the original points were directly addressed and the only formal reply was incomplete.

I hope we can resolve this in good faith. Please reply by April 24th end-of-business.

The deadline was missed, even when given a comfortable buffer of over a week. I don't consider the attempts by Huggingface to resolve this to have been conducted in good faith, they appear to be casual and dismissive (one could even say negligent) attempts at burying the problems. Neither was the effort to find my PII under the GDPR conducted in good faith as I wasn't even asked for the sample_id of the fields, etc.

Should I assume you are simply not able or willing to make a reasonable effort to resolve this? I must assume so, which sadly leaves me with few options that are beneficial to the community.

lol

This matter is not resolved. @Zeeves posted a tentative legal argument but failed to attach any documents that supported his reasoning — assuming those documents exist at all (not enough detail for me to find them publicly). His reasoning as also rebutted with nine different points for which there was no reply. The deadline was then missed.

As far as I can tell neither LAION nor HuggingFace have a valid response to the question of GDPR compliance, and these issues are being swept under the rug by maintainers. It's been over a year, but that's consistent with the CSAM issue that also took over a year before the dataset was taken down for fixing.

If you don't want to reply publicly because legal transparency is seemingly not compatible with the current principles of operation of this organization nor platform, I remain available to resolve this by email.