ChatGPT and Other Generative AI Tools: Overview of Intellectual Property Issues
ChatGPT and other generative AIs have garnered global interest in recent months. According to Google Trends (see the image), the buzz is slowly fading away and now might be a good time to calmly talk about potential IP issues.
While the performance of ChatGPT and other generative tools (Midjourney, StableDiffusion, etc.) is impressive—everyone agrees on this—such tools create new risks, also in terms of intellectual property (IP).
Technical issues
The technical quirks of ChatGPT are well known. The tool tries to give the best guess response to a question, which can lead to unintended or wrong responses. In particular, the underlying model (GPT) suffers from “hallucinations”, like other large language models (LLMs). I.e., the responses can be nonsensical at times. In other cases, even if the confident-looking response appears plausible, it may in fact be biased, incorrect, or even toxic.
Remember that GPT was trained on a large portion of the internet, much of which contains non-curated and/or “bad” data. Still, most of the data sources can be assumed to be reasonably reliable. In this regard, we recently learned that our own website was used for training LLMs.
Beyond hallucinations, LLMs suffer from other alignment issues, which may result in a lack of helpfulness (they do not follow explicit instructions) or interpretability (you cannot understand how the model arrives at the response). Meaning risks in terms of transparency and traceability.
Data ownership and privacy
Data ownership and privacy are less often debated. Every posted question is managed by ChatGPT’s owner, OpenAI, which handles the data on the basis of contract law. Interestingly, you own ChatGPT’s outputs, including the right to reprint and sell (regardless of the free/paid plan used), and you can use them for any purpose, including commercial purposes, as long as you comply with the terms of use (https://openai.com/policies/terms-of-use).
That said, each generative tool has its own policy. For example, according to Midjourney’s Terms of Service (https://docs.midjourney.com/docs/terms-of-service), you don’t own the assets you create unless you are a paid member. If you are not a paid member, Midjourney grants you a license to the assets under the Creative Commons (CC) Noncommercial 4.0 Attribution International License, meaning that you can use the images provided that you don’t sell them (or otherwise make money off them) and give credit to Midjourney. Incidentally, such a legal construct may deserve further scrutiny as the CC license assumes copyrightable works, which is manifestly not the case with Midjourney outputs, see below.
IP issues
Beyond the above issues, there are numerous IP questions. For instance, all generative tools can cause copyright issues and text-generating tools can additionally harm patent and trade secrets.
Copyright
To start with, contents exclusively produced by ChatGPT or other generative tools are, in all likelihood, not copyrightable, at least not in jurisdictions emphasizing authorship (an AI tool is not a human author). However, individuals who use it in creating a work may try and claim copyright protection for their own contributions.
Similarly, you may use StableDiffusion or MidJourney to produce images, e.g., for a comic book. The selection and arrangement of the images that you make can be copyrighted (at least in some jurisdictions), not the images themselves.
领英推荐
A main issue about copyright is that works created by generative AI tools may be too close to the original works used to train such tools. In this case, the generated content could infringe the copyright of the original works. ?Decisive is the extent to which the AI tools tweak the original content, something that also depends on the user prompts.
Thus, AI-generated images and text should not be used lightly. In particular, “fanfiction” generated by means of AI tools can be problematic. E.g., publishing an alternate ending to Harry Potter would be risky, see, e.g., https://en.wikipedia.org/wiki/Legal_issues_with_fan_fiction.
Patents and trade secrets
There are other issues regarding patents and trade secrets, too. Interacting with ChatGPT may result in losses of IP rights. The reasons are that trade secrets require reasonable efforts to conceal the information from the public, while a claimed invention must be novel to be patentable.
Now, OpenAI tells you that anything posted into ChatGPT can be reviewed or used by them. Thus, your inputs may constitute a prior disclosure, something that could preclude patent protection and/or harm trade secret protection.
Accordingly, IP players (e.g., patent drafters and inventors) should be careful about the information provided as input to ChatGPT. That said, OpenAI is working on a new subscription for professionals who need more control over their data.
Incidentally, relying on ChatGPT to interpret a patent claim or description is a false good idea if the stakes are high. See, e.g., https://www.dhirubhai.net/feed/update/urn:li:activity:7044404873278181377?commentUrn=urn%3Ali%3Acomment%3A%28activity%3A7044404873278181377%2C7045084790705782785%29 for an illustration, where ChatGPT incorrectly interpreted the scope of a patent description. The same conclusions hold with other chatbots.
Takeaway messages
Let’s not spoil the fun. Generative AI is awesome. But there are a few things to keep in mind:
Further (very good) reading
?
Driving growth through Strategic Partnerships & Outstanding Customer Experience @ Ogre AI
1 年Thanks for sharing! Sonia, you may find relevant this post!