ChatGPT and Intellectual Property

During my reading to do with Alarm Flooding I used the ISA (International Society of Automation) web site. At the top of their page is the following notice:

Artificial Intelligence Notice: ISA prohibits the entry of ISA standards and related ISA intellectual property (“ISA IP”) into any form of Artificial Intelligence (AI) tools, such as ChatGPT. Additionally, creating derivatives of ISA IP using AI is also prohibited without express written permission from ISA’s CEO. In the case of such use, ISA will suspend a licensee’s access to ISA IP, and further legal action will be considered.

Increasingly, those who own copyrighted material are pushing back on programs such as ChatGPT. The ISA is not alone in this.

So I asked ChatGPT 3.5 the following question.

ChatGPT: There is increasing concern that programs such as ChatGPT are using intellectual property belonging to other people or organizations. What is your opinion on this, and are you changing your method of operation?

In response to the first question it said,

Concerns about the use of intellectual property (IP) in AI models like ChatGPT are valid and have been a topic of discussion among experts, users, and developers. The primary issues revolve around:

1. Data Sources: AI models are trained on vast amounts of data, which may include copyrighted material. This can raise questions about whether the use of such data constitutes fair use or infringement.

2. Attribution and Ownership: When an AI generates text, it's not always clear where specific pieces of information or particular phrasings originated, making it difficult to attribute content properly or respect the original creators' rights.

3. Ethical and Legal Implications: The use of IP without explicit permission can have legal consequences and ethical implications, especially for creators who may feel their work is being exploited without compensation.

In other words, it dodged the question.

I asked the follow-up question.

ChatGPT: How do we know that your replies are not using copyrighted information?

Two paragraphs in its reply to that question stand out.

3. Probabilistic Output: Each response is generated probabilistically, meaning the model combines and reformulates information in unique ways rather than recalling specific texts.

5. User Responsibility: Users are advised to use the generated content responsibly and to be aware of the potential need for further verification or adaptation, especially if the content is used for publication or commercial purposes.

Going back to the ISA notice, the following questions are in play.

• How can the owner of intellectual property even know that their material is being used? ChatGPT does not list sources or references.

• How can an organization control the use of derivative material? For example, someone may use an ISA standard as part of a design. If that design is published on the internet does it violate the ISA’s rights?

• How can a conscientious user of such information know that he or she is not inadvertently violating copyright issues?

My final question was,

ChatGPT: How do I know that your responses are not violating the intellectual property rights of the International Society of Automation?

Its reply:

Ensuring that responses from AI like ChatGPT do not violate the intellectual property rights of specific organizations, such as the International Society of Automation (ISA), involves several safeguards:

1. Generalized Knowledge: The AI model is designed to provide generalized knowledge based on patterns and information from a wide range of sources. It does not store or retrieve specific texts from proprietary sources verbatim.

2. Training Data Management: The training process includes data from various sources, but specific proprietary content from organizations like ISA would not be directly included unless it falls under publicly available information or is licensed for use.

3. Transformative Use: The model generates responses that are typically transformative in nature. It combines and reformulates information rather than copying it directly, which helps mitigate risks related to copyright infringement.

4. Filters and Guidelines: There are filters and guidelines in place to help ensure that the content generated does not directly reproduce large sections of text from any single source, especially proprietary ones.

5. Continuous Improvement: Developers continuously update and refine models to better handle issues related to IP rights, including improving the ability to differentiate between public domain and proprietary content.

The second paragraph is interesting. The ISA will not publish its information. But a secondary user may do so. How does ChatGPT know the difference?