Licensing “Open” AI Models

PLI’s Open Source Software - From Compliance to Cooperation

Kate Downing and Luis Villa

September 19, 2023

What are “open” AI models and why are they appealing?

WHAT IS A MODEL?

The “model” is the key artifact in machine learning:

• Giant million-dimensional matrix of weights and parameters
• i.e., lots of numbers – not a work of text or art in the traditional sense recognized by copyright law
• Can be iteratively improved— “fine tuning”
• May memorize some inputs; amount of memorization fact-specific

Paired, usually, with a “tokenizer”:

• Software that translates the training data into numbers
• Translates numeric outputs back into something human comprehensible

WHAT ARTIFACTS COMPLEMENT THE MODEL?

Important, protectable artifacts that may exist alongside the model include:

• Data
• used to create (“train”) the model
• Code
• used for training, inference
• Tokenizer
• Runtime
• Traditional software code (almost exclusively open source) that “executes” the model to create outputs

​

Note: weights, parameters, and runtime may shipped to consumers as one artifact

WHAT’S THE SECRET SAUCE?

The training process:

• Choice of datasets, filtering of datasets, deduplication
• Training software
• Model architecture
• Length of training
• Human review and feedback
• Sequence of training steps

AND

Do you have enough GPUs??

​

Only a select group of people at OpenAI know all the exact steps taken during a training process and they’re physically separated from everyone else at the company to maintain this secrecy.

OPTIONS FOR USE OF ML

• Hosted by someone else (via SaaS API)
• Create your own model from scratch
• Use a publicly available model (either directly or fine-tuned)

TODAY, WHAT DOES “OPEN” MEAN WHEN USED IN AI?

• Not “open” or “free” under the definitions of OSI or FSF
• Often contain field of use restrictions and sometimes discriminate against certain types of users - some of these relate to competitive products, but many of these relate to “unethical” uses
• Sometimes have triggers that require a user to upgrade to a for-fee commercial license
• “Open” in the sense that:
• They are freely available for download from the internet
• Use is $0
• Sometimes the license must be passed along to downstream users
• Have more in common with proprietary freeware licenses than open source licenses

Llama 2 Acceptable Use Policy - Prohibitions

​

iv. Sexual solicitation…

c. Engage in, promote, incite, or facilitate the harassment, abuse, threatening, or bullying of individuals or groups of individuals

d. Engage in, promote, incite, or facilitate discrimination or other unlawful or harmful conduct in the provision of employment, employment benefits, credit, housing, other economic benefits…,

g. Engage in or facilitate any action or generate any content that infringes, misappropriates, or otherwise violates any third-party rights, including the outputs or results of any products or services using the Llama 2 Materials

​

​

​

2. Engage in, promote, incite, facilitate, or assist in the planning or development of activities that present a risk of death or bodily harm to individuals, including use of Llama 2 related to the following:

a. Military, warfare, nuclear industries or applications...

c. …regulated/controlled substances

f. Any content intended to incite or promote violence, abuse, or any infliction of bodily harm to an individual

…

a. Generating, promoting, or furthering fraud or the creation or promotion of disinformation

​

WHY USE PUBLICLY AVAILABLE MODELS?

• It’s free
• No data leaves your company; no additional sub-processor to manage and disclose
• Potential for faster improvements
• Potentially more transparency around training data
• Some developers are honoring requests to exclude certain works from the training data
• Greater ability to fine-tune for your use case
• Eliminate risk of vendor limiting features or changing pricing model
• May be able to pick and choose compliance-related add-ons

Every large company adapted to the Internet eventually and now even Walmart has software engineers. The next evolution involves hiring people with AI expertise (though training may be done on cloud hardware).

​

POPULAR PUBLICLY AVAILABLE AI MODEL LICENSES

• Llama 2 license
• RAIL family of licenses
• Allen Institute for AI’s AI2 ImpACT License – Medium Risk Artifacts
• Various flavors of Creative Commons Licenses
• But this use not (yet?) endorsed by Creative Commons
• Various OSS licenses
• Unclear whether these actually apply! But being used anyway

The trillion dollar question: are AI models copyrightable?

OPEN QUESTIONS IN THE US

​

• Does training the model constitute copyright infringement?
• Is the model itself copyrightable?
• Is the model a derivative work of the training data?
• Is output from the model copyrightable?
• The US Copyright Office has said “no” unless it has been modified by a human and the human additions can be identified
• Is the output a derivative work of the training data?

​

COPYRIGHTABILITY OF AI MODELS

• Copyright law in the US and abroad protects the non-functional elements of software - areas of creative expression
• Models are the output of the training software and the various steps in the training process, but unlike other output like screen displays or GUIs, there is nothing there to qualify for a separate literary, pictorial or graphical copyright which such displays are typically eligible for (they are not themselves treated as “software”). Any specific weight or parameter cannot be said to be an encoding of any specific human expression
• In certain countries like Germany, there is a stricter interpretation of what is copyrightable in software and less material is eligible for copyright in those jurisdictions
• Even if models are copyrightable, what is the framework for deciding whether one is a derivative work of another or one infringes another?
• Substantial similarity and access are problematic since two people can independently create extremely similar models by using the same datasets

WHAT IF THE MODEL ISN’T COPYRIGHTABLE?

• At best, the license is actually a contract but:
• Without a consent mechanism such as an “I Agree” button, there may not be acceptance and contract formation
• What constitutes acceptance varies greatly by jurisdiction
• Unlike copyright law, which was unified by the Berne Convention, there is no uniform contract law. Every country has their own and even the states within those countries can and do have conflicting rules about contract formation and interpretation
• Licenses with no governing law provision are particularly problematic
• Copyright’s statutory damages are no longer applicable, meaning that to enforce a breach, the plaintiff would need to show real damages that are personal and in most cases, monetary
• The harm caused by ethical misuse is not suffered by the plaintiff
• Many jurisdictions will not recognize moral pain as a form of compensable damage

If AI models are copyrightable, can the open source approach work for AI?

APPLYING AN OPEN SOURCE-LIKE APPROACH TO AI MODELS

Pros:

• Ease of deployment—text file!
• Standardization simplifies choices for developers and can create less ambiguity for users
• High trust governance because no party can unilaterally change the license
• 25+ years of FOSS advocacy that can be leveraged
• Use of Berne → global(ish)

​

​

Cons:

• Static, so can’t react to:
• Changes in economics/tech (eg SaaS)
• New regulation (eg GDPR, CRA)
• Recalls(?)
• New IP laws/cases
• Intersection with regulation (eg privacy) not well-defined
• Weak limitation of liability
• OSS concepts like “derivative work,” “modification,” “copyleft,” “covered source code,” “build instructions, ” etc. have only loose analogs, if any
• If Berne isn’t applicable → licensing construct fails

​

Critical: these will mostly apply even to a “perfect” mature license

LEGAL PITFALLS OF ETHICAL RESTRICTIONS

• Can constitute copyright misuse if sufficiently unrelated to the copyright monopoly rights
• May not be interpreted as conditions of the license, and therefore their violation may be subject only to breach of contract claims and not copyright infringement claims
• The worst violations are likely to occur in places where the licenses cannot be enforced - no rule of law, courts uninterested in foreign plaintiffs, countries with laws and values that are not Western, etc.
• Violations of OSS license require proving relatively simple things, but violations of ethical licenses can require proving extremely complicated things involving knowledge, intent, local law, and a healthy dose of cultural relativism.

SOCIO-POLITICAL PITFALLS OF ETHICAL RESTRICTIONS

• ​
• ​

SOCIO-POLITICAL PITFALLS OF ETHICAL RESTRICTIONS

AI Model Licenses and Liability

ATTEMPTING TO LIMIT LIABILITY VIA LICENSING

• The EU’s AI Act will limit the types of limitations that can be enforceable in a contract
• Much of the potential harm from AI will be borne by third parties who are not parties to the contract and aren’t limited by its limitation of liability provision
• Regulatory violations such as violation of privacy rights, HIPPA, various financial regs, etc.
• Speech regulation—First Amendment probably shields an LLM’s output in the US, but other countries may apply regulation to “mere” LLM outputs
• Tort liability - fraud, defamation, libel, invasion of privacy (appropriating a person’s name or likeness, etc.)
• Product liability—coming next year in the EU, indeterminate in the US
• Criminal liability - certain speech can be criminally punished in some countries
• IP claims

Traditional OSS Licenses and AI Models?

USE OF TRADITIONAL OPEN SOURCE LICENSES FOR AI MODELS

OSS Interactions with AI Models:

• What happens when copyleft code is used to execute or train a model?
• What happens if a model is trained on copyleft code or content?
• What happens if a model outputs copyleft code?

​

Specific licenses:

• Licensing models under L/A/GPL:
• Terms are often specific to software—applicability to models unclear
• Licensing models under Creative Commons licenses
• Attempts to address database rights, but unclear to what extent a model is a derivative of the database

Conclusion

CONCLUSION

• Lots of unknowns, including very basic things like “what is the right framework for governance” and “what sources of liability might apply”
• Extreme humility is needed when drawing lessons from open source and applying them to AI – many apparent parallels are misleading at best, wrong at worst
• No one knows what they’re doing yet—don’t assume that, just because a big player did it, it’s actually correct or sensible
• Licenses aren’t a substitute for governance
• Licenses can’t be a substitute for legislation