Science fiction no more, artificial intelligence, or AI, is becoming ubiquitous in the workplace, gaining one million users in five days, attracting a multibillion-dollar investment from Microsoft Corp. and giving rise to expectations of future integration into Office 365 and Bing.
After launch, end users immediately recognized that ChatGPT, an AI text-generation tool, provides surprisingly well-written answers to their short prompts. Writing that once took days can be accomplished in minutes.
This article provides an overview of ChatGPT and other AI text-generation tools — or AI writing tools — and explores the five risks for patent attorneys related to using AI text-generation tools or patent drafting and editing tools, like ChatGPT.
Although this article primarily addresses concerns relevant to patent practitioners, many of the issues raised here apply to other attorneys as well.
Overview of ChatGPT
ChatGPT, a generative pretrained transformer, includes the "largest trained language model," according to TechTarget, having 175 billion machine learning parameters. ChatGPT generates real-time, on-demand, humanlike responses based on learning "how conversation works in a specific way," as Ethan William said in a recent article, adding, "[w]hen a user provides text input, the system analyzes the language and uses a text predictor to create the most likely output."
ChatGPT includes a content moderation application programming interface to filter out unsafe content, and instantaneously displays a humanlike, ad-free response consisting of the answer you are looking for.
AI text-generation tools, like ChatGPT, can help with act as a dictionary or thesaurus and can translate, provide a topical framework, and rephrase and/or expand boilerplate content — e.g., template patent application language, marketing copy.
AI writing tools hold the potential to simplify a patent practitioner's workload, allowing one to focus on the most substantive aspects of practice. An attorney is, of course, ultimately responsible for, and should review, all work product.
Caveat Emptor: Five Risks of AI Writing Tools for Patent Attorneys
1. Invalidating Self-Publication or Self-Disclosure of Invention
According to the U.S. Patent andTrademark Office's Manual of Patent Examining Procedure, a reference or document is considered to be a prior art publication if it is available and able to be located by an ordinarily skilled artisan.
The most serious risk of patent drafting tools is latent invalidation of a client's patent or application via premature self-publication or self-disclosure.
Can another person access information you enter into an AI writing tool, like ChatGPT? As cybersecurity blog Deploying Securely points out, "although ChatGPT is not currently trained in realtime and it says it doesn't store or train on questions you ask it … that could change in the future," and ChatGPT may "[a]ccidentally regurgitate sensitive information," which could be troublesome to patent applicants whose inventions were disclosed to ChatGPT.
Patent lawyers do not want to use a tool, no matter how powerful, when the usage could somehow cause the invention to be disseminated and made available to the public, i.e., a prior art printed publication, or end up in another's predated patent application.
Conceivably, providing such disclosure to an AI writing tool that is trained in real time could result in one's invention becoming part of the tool's store of knowledge and, in some circumstances, part of the public domain.
Fortunately, AI patent drafting and editing tools, including a version of ChatGPT, may provide a default or opt out provision to avoid sharing of data internally, e.g., used to improve our models, or externally, e.g., third parties.
Patent-specific generative AI drafting tools, Patent Pal and qatent are other known patent drafting and editing options using natural language processing or generative AI.
Patent drafting and editing tools, such as Patent Optimizer of LexisNexis, check for proper antecedent basis, lack of support in the disclosure, adequate support in the drawings and more.
Another issue to be considered is the risk of using a single generative AI model that is fed input information from multiple patent applicant clients. Conceivably, such a model could generate inappropriate responses based on the commingling of client data, e.g., invention disclosures.
A specification generated for one client might contain information the AI learned from another client. To avoid such risks, the client may keep their own patent drafting model — and consult the patent attorney for preliminary claim drafting and final review — or the patent law firm may consider whether separate, client-personalized AI models may be used, each one trained on an individual client's data — e.g., patent disclosures.
Verdict: Although an attorney should be apprised of the risks of dissemination of client data, this risk appears mitigable by restricting the usage of the input data. For example, the vendor agreement could have, at a minimum, a "no sharing" data privacy agreement between the client or law firm and the AI service provider, or at the very least no external, or third-party, sharing.
If possible, the models could be personalized and potentially provide a reasonably safe level of data security like cloud computing. The terms of service may include the standard cloud terms, as well as AI-specific terms — e.g., potentially no cross-training of client data, separate models for each client.
2. Breach of Ethical Duty of Confidentiality
Title 37 of the Code of Federal Regulations, Section 11.106, requires that USPTO practitioners "not reveal information relating to the representation of a client" without the client's informed consent or applicability of another exception — e.g., USPTO duty of disclosure compliance or to prevent death, harm or crime.
Nonetheless, according to the American Bar Association's 2022 Cloud Computing survey, 70% of lawyers already use cloud computing. Of course, not all the attorneys using cloud computing are in violation of USPTO Rule 11.106.
Cloud computing is the post-pandemic norm, and there is an expectation that the cloud data remains confidential and not shared with others. With AI that could be constantly learning based on inputs and outputs, the risk of revealing client information is greater, but it does not have to be.
With relation to cloud computing, the American Bar Association's model Rule 1.6(c) recites: "[a] lawyer shall make reasonable efforts to prevent the inadvertent or unauthorized disclosure of, or unauthorized access to, information relating to the representation of a client."
Although bar ethics rules may vary jurisdiction by jurisdiction, practitioners need to be conscientious about the cloud tools they employ, including AI tools, and the risks they present to client information.
According to IP ethics attorney Emil J. Ali of McCabe & Ali LLP, while attorneys may seemingly rely on informed consent to use cloud computing, many engagement letters still warn clients of the use of smart devices, cloud computing, and other third party, hosted services that may store data — both due to the fact that such data is being shared with a third party, but also that it might be saved and accessed by others, including law enforcement.
Confidentiality and reliance on the terms of service could follow a cloud approach, as AI may be an extension of a cloud service. Acquiring informed consent to use a generative AI tool — for cost-effective patent drafting — may be a wise practice for ethics compliance, and marketing, purposes.
Alternatively, as discussed above, it may make more sense for the AI patent drafting tool to be kept on site with the client and/or provided virtually directly to the client — and the patent attorney provides guidance along the way.
3. Copyright Infringement
As the USPTO has noted, "[w]hile an AI machine cannot currently own intellectual property rights, it may be able to infringe others' rights."
Suppose that an attorney uses text generated by AI, which was trained on copyrighted material. The model learned from copyrighted material and incorporates text learned from copyrighted material into its answers.
Might this violate the reproduction right of a copyright owner under Title 17 of the U.S. Code, Section 106(1)?
Or would this be a noninfringing fair use or derivative work? Is training an AI text tool based on the Internet considered to be copyright infringement?
Should the patent lawyer or client train the model on only the client's currently owned patent applications and/or the public domain?
There are at least two pending generative AI copyright litigation cases where the alleged copyright infringement was using artistic works to train an AI model to generate new works in the style of that artist.
As in the repurposing of artistic works, ChatGPT can be adapted to various styles or impersonations. But let's assume that the patent attorney will not ask the AI writing tool to write in a specific style, or like a specific person — e.g., act as Steve Jobs when drafting patent application.
In the pending Pearson Education Inc. v. Chegg Inc. litigation regarding generated text, the plaintiff Pearson sued Chegg in the U.S. District Court for the District of New Jersey because Chegg's textbook answers to "end-of-chapter questions" allegedly violated Pearson's copyrights.
Chegg may argue the textbook answers are derivative works and/or fair use. Although this case is not necessarily about AI, because ChatGPT is a proposed search engine replacement, the discussion of how questions (inputs) and answers (outputs, e.g., of search engines) are treated under U.S. copyright law will be closely followed.
Moreover, according to the USPTO's 2020 AI report, mass digitization and text and data mining, or TDM, are activities with
copyright implications, [and] may be considered copyright infringement or fair use, depending on the facts and circumstances at issue ... Many publishers now include TDM terms in their contracts and expressly set a licensing fee for for-profit entities or permit licensing at no additional cost for researchers and public research organizations, while ensuring that the licensed content is machine-readable and searchable.
Verdict: To be determined. Copyright infringement based on training AI models using copyrighted materials is a developing area of law. A patent lawyer or technology company, or client, could inquire about the training data used and attempt to limit the amount of copyright material used by asking about mass digitization and TDM terms in their vendor agreements.
Although in each of the above training data copyright cases the end user was not the target of litigation, patent lawyers — end users — may have a desire to avoid using vendors that infringe others intellectual property.
4. Inadvertent Inventorship by AI
Up to now, we have envisioned AI as primarily a writing/editing tool but suppose a patent attorney enters the inventors patent claims into ChatGPT, and the model further invents. Does the AI deserve inventorship credit?
Under U.S. patent law today, as the U.S. Court of Appeals for the Federal Circuit held in the August Thaler v. Vidal decision, AI cannot be an inventor because an inventor must be a natural person under the Patent Act.
The USPTO follows the Federal Circuit approach. According to the USPTO, based on patent statutes and Federal Circuit case law, "inventorship requires that an inventor must be a natural person [which] is reflected in the numerous references to the inventor as a 'person' in Title 37 of the Code of Federal Regulations." Likewise, there are no U.S. copyrights for the machines.
Verdict: Inadvertent AI inventorship is a red herring because the machine is not a natural person and hence does not qualify for inventorship for U.S. patents or U.S. copyright authorship.
5. Confidently Wrong
Is it safe to rely on ChatGPT or the Internet as a source of truthful information? Can ChatGPT draft a patent application? Yes. Will it be accurate, valid and enforceable? That is a tricky question to answer.
At least in its current form, according to an article in The Verge, "the answers which ChatGPT produces have a high rate of being incorrect," although "they typically look like they might be good." Even ChatGPT's creator, Open AI, has acknowledged that it "sometimes writes plausible-sounding but incorrect or nonsensical answers."
There have been noted instances of inaccuracy. However, according to The Indian Express, "ChatGPT 4 could have as many as 1 trillion parameters, capable of delivering more accurate responses."
Verdict: As subsequent iterations of AI writing tools, are made available, the risks of inaccuracy will be far outweighed by the benefit of the robust, accurate outputs. Nonetheless, attorneys are should review and are ultimately responsible for firm work product.
Other Issues: Attribution, Duty to be Technologically Competent
This article lists the five risks for patent attorneys but is not completely comprehensive. For example, the terms of service of ChatGPT require attribution of any published content to both the source and to AI generation. Is that attribution credit something you want to include in your patent application? Maybe a different, patent specific drafting service (that does not require attribution credit) is preferable?
As Open AI notes, generative AI programming code "may be subject to third party licenses, including, without limitation, open source licenses."
Open-source software terms may also need to be reviewed and complied with for patent law firms creating their own software tools. If a law firm or technology company decides to develop its own AI writing tool, using for example, OSS Python or Apache, and the code is edited by ChatGPT, the code may be subject to open source software licenses.
Additionally, the ABA Model Rules include a comment that lawyers shall maintain competence, and a statement that "[t]o maintain the requisite knowledge and skill, a lawyer should keep abreast of changes in the law and its practice, including the benefits and risks associated with relevant technology."
What if ChatGPT becomes the prevailing standard?
If in the future, like cloud computing today, 70% of lawyers are utilizing ChatGPT or similar AI writing tools in their practice, will a lawyer be deemed not to live up to the guidance of remaining technologically compete by refraining from doing the same?
How the inputs and outputs are treated is critical. The inputs (and potentially outputs) may include sensitive client information. The vendor agreement (for the generative AI vendor) preferably addresses how the inputs (questions/commands) and outputs (answers/responses) are treated, and the treatment excludes sharing with third parties and limits or prohibits internal viewing by the generative AI vendor except for basic maintenance.
 Mollman, Steve. ChatGPT gained 1 million users in under a week. Here's why the AI chatbot is primed to disrupt search as we know it. Yahoo! Finance. December 9, 2022. Available at https://finance.yahoo.com/news/chatgpt-gained-1-million-followers-224523258.html quoting OpenAI founder Greg Brockman's Twitter Handle @gdb's tweet on Dec. 5, 2022 at 3:32 AM ("ChatGPT just crossed 1 million users; it's been 5 days since launch."). Available at https://twitter.com/gdb/status/1599683104142430208.
 Capoot, Ashley. Microsoft announces new multibillion-dollar investment in ChatGPT- maker OpenAI. CNBC. January 23, 2023. Available at https://www.cnbc.com/2023/01/23/microsoft-announces-multibillion-dollar-investment- in-chatgpt-maker-openai.html.
 Walia, Priya. Microsoft Plans To Add ChatGPT To Office 365. How Would It Benefit Small Businesses? International Business Times. January 17, 2023. Available at https://www.ibtimes.com/microsoft-plans-add-chatgpt-office-365-how-would-it-benefit- small-businesses-3657754.
 Schemeizer, Ronald. Definition GPT-3. TechTarget. June 2021. Available at https://www.techtarget.com/searchenterpriseai/definition/GPT-3.
 William, Ethan. How ChatGPT is able to generate human-like responses in conversation. Medium. January 5, 2023. Available at https://medium.com/@sahayababumanickam/how-chatgpt-is-able-to-generate-human-like-responses-in-conversation-37af9ade9aca.
 OpenAI. ChatGPT: Optimizing Language Models for Dialogue.OpenAI November 30, 2022. Available at https://openai.com/blog/chatgpt/.
 Manual of Patent Examining Procedure (MPEP) at 2128(I). United States Patent & Trademark Office. Available at https://www.uspto.gov/web/offices/pac/mpep/s2128.html.
 Haydock, Walter. ChatGPT's implications for cybersecurity. December 16, 2022. ChatGPT's implications for cybersecurity. Available at https://www.blog.deploy- securely.com/p/chatgpt-implications-for-cybersecurity.
 OpenAI. Service Terms. November 3, 2022. OpenAI. Available at https://openai.com/api/policies/service-terms/. (a version of ChatGPT provides an opt out option to avoid data sharing with an organization ID). See also Simens, Maddie. How your data is used to improve model performance. OpenAI Help Center. January 2023. Available at https://help.openai.com/en/articles/5722486-how-your-data-is-used-to- improve-model-performance.
 PatentPal. Terms of Service. PatentPal August 1, 2022. ("PatentPal will not share the contents of your patent documents or any output generated using our services with any third parties outside of processing the patent data on Amazon Web Services (AWS)"). Available at https://patentpal.com/terms/. See also qatent at https://qatent.com/.
 LexisNexis. PatentOptimizer. LexisNexis. Undated. Available at https://www.lexisnexisip.com/solutions/patent-drafting/patentoptimizer/.
 United States Patent & Trademark Office (USPTO). Appendix R – Patent Rules. USPTO. October 31, 2019. Available at https://www.uspto.gov/web/offices/pac/mpep/mpep-9020-appx-r.html.
 Dennis Kennedy. 2022 Cloud Computing. American Bar Association. November 17, 2022. Available (by subscription) at https://www.americanbar.org/groups/law_practice/publications/techreport/2022/cloud- computing/.
 American Bar Association. Model Rules of Professional Conduct, Rule 1.6: Confidentiality of Information. Undated. Available (by subscription) at https://www.americanbar.org/groups/professional_responsibility/publications/model_rule s_of_professional_conduct/rule_1_6_confidentiality_of_information/.
 USPTO. Public Views on Artificial Intelligence and Intellectual Property Policy. USPTO. October 2020. Available at https://www.uspto.gov/sites/default/files/documents/USPTO_AI- Report_2020-10-07.pdf
 Copyright Law of the United States and Related Laws Contained in Title 17 of the United States Code – October 2022 (Circular 92). October 2022. Available at https://www.copyright.gov/title17/title17.pdf.
 GitHub Copilot Litigation. We've filed a lawsuit challenging GitHub Copilot, an AI product that relies on unprecedented open source software piracy. Because AI needs to be fair & ethical for everyone. GitHub Copilot Litigation. November 3, 2022. Available at https://githubcopilotlitigation.com/. See also https://stablediffusionlitigation.com/pdf/00201/1-1-stable-diffusion-complaint.pdf.
 Pearson Education, Inc. v. Chegg, Inc. (D. N.J. 2022) Inc. Case Number 2:21-cv- 16866, in the U.S. District Court for the District of New Jersey.
Public Views on Artificial Intelligence and Intellectual Property Policy at 23.
 Id. quoting STM. Text and Data Mining: Building a healthy and sustainable knowledge ecosystem for Europe. December 2017.
 Thaler v. Vidal. 43 F.4 th 1207 (Federal Circuit 2022) Appeal Number 21-2347. Available at https://cafc.uscourts.gov/opinions-orders/21-2347.OPINION.8-5- 2022_1988142.pdf.
 Public Views on Artificial Intelligence and Intellectual Property Policy at 4.
 Id. at 19 ("Under current U.S. law, a work created without human involvement would not qualify for copyright protection").
 Vincent, James. AI-generated answers temporarily banned on coding Q&A site Stack Overflow. The Verge. December 5, 2022. Available at https://www.theverge.com/2022/12/5/23493932/chatgpt-ai-generated-answers- temporarily-banned-stack-overflow-llms-dangers.
 ChatGPT: Optimizing Language Models for Dialogue.
 Tech Desk. Open AI's GPT 4 could support up to 1 trillion parameters, will be bigger than ChatGPT 3. The Indian Express. January 23, 2023. Available at https://indianexpress.com/article/technology/tech-news-technology/chatgpt-4-release- features-specifications-parameters-8344149/.
 OpenAI. Sharing & Publication Policy. OpenAI. November 14, 2022. Available at https://openai.com/api/policies/sharing-publication/.
 Siegel, Daniel. Hot Buttons: Revise the Rules of Professional Conduct for Technology. American Bar Association. November/December 2021. Available (subscription) at https://www.americanbar.org/groups/law_practice/publications/law_practice_magazine/2021/nd21/hb/.
See also Macaulay, Don. What Is A Lawyer's Duty Of Technology Competence? February 2, 2018. SmartLawyer. Available at https://nationaljurist.com/smartlawyer/what-lawyers-duty-technology-competence/.