Lawyers Need Wisdom, Not to Call GPT

LLMs as writing editors, existential thoughts on the role of lawyers, and a case study of corporate product promotion disguised as academic research + OpenAI lawsuits

Note 05-03-2024

As I have said before, OpenAI’s legal and ethical problems are mounting. This past week a lot has been going on in its legal department. Let me give you a quick recap with sources for further reading.

• OpenAI filed a motion in court on Monday last week, seeking to dismiss key claims from The New York Times (NYT) lawsuit. See my coverage of the lawsuit here.

  NYT attached an appendix to its original complaint, Exhibit J, that shows one hundred examples of GPT-4 memorizing long passages from NYT articles. OpenAI states in its motion to dismiss that NYT was only able to produce these “anomalous results” “after tens of thousands of attempts” and “ by targeting and exploiting a bug (which OpenAI has committed to addressing) by using deceptive prompts that blatantly violate OpenAI’s terms of use.”

  Furthermore, OpenAI argues that it is not competing with NYT since no one uses GPT-4 to read articles. This is technically correct. However, as I see it, it remains concerning in and of itself that GPT-4 is capable of reproducing so much copyrighted text, even if no one does so. It clearly indicates that GPT-4 was built on the creative work of others without permission or acknowledgement.

  Interestingly enough, OpenAI does not ask the court to dismiss NYT’s core claim that GPT-4 has been built on millions of NYT articles without permission, nor does OpenAI argue why “the fair use” exemption applies to AI training. Hopefully, these matters will be settled in court.

• On Wednesday, three other news organizations, The Intercept, Raw Story, and AlterNet filed separate lawsuits against OpenAI and Microsoft in the Southern District of New York for alleged copyright infringements. All plaintiffs are represented by the same attorneys, Loevy & Loevy.
  
  To understand the two big unanswered questions about the intersection of generative AI and copyright law, see my posts on “the input phase question” and “the output phase question”, written in non-legalese.

• Last Tuesday, a resident of Florida filed a class action lawsuit “on behalf of all others similarly situated” against OpenAI in the Northern District of California, for using stolen private information “from hundreds of millions of internet users, including children of all ages” to built its products. The lawsuit also goes into depth on how the unregulated development of AI can lead to existential threats and many malicious applications of the technology, see the complaint’s pg. 29ff.

• On Thursday, Elon Musk sued OpenAI (complaint is here), and its co-founders Sam Altman and Greg Brockman, for not staying true to OpenAIs non-profit mission of developing AGI to the benefit of humanity. A mission Mr. Musk had invested a lot of money in before OpenAI partnered up with Microsoft and took the commercial route with its products. I recommend Matt Levine’s coverage of the case here.

• Finally, The Wall Street Journal reported on Wednesday that the Securities and Exchange Commission (SEC) is examining internal communications of CEO Sam Altman as part of an investigation into whether the company’s investors were misled, following the OpenAI board’s decision to fire and then re-hire Altman in November.

---

Lawyers Need Wisdom, Not to Call GPT

AI as Editors 

For fun, I have routinely asked ChatGPT or Claude 2 to play the role of editor on my finished article drafts. 

I have been doing it for many months now and experimented with different prompts. Usually something simple like: "Please provide feedback to the following post". If I had room in my budget to hire a savvy human editor, I would do that instead, but it isn't cost-efficient at this stage in my “newsletter writing journey”. For now, I have attempted to rely on the machine-wisdom of large language models. 

At first, the AI feedback seemed pretty neat. It often resembled something my teachers could have told me in school. But time went on and I have now come to realize that the models have been fooling me with the same phrases and same edit suggestions over and over again. Very rarely, if ever, have I received a valuable pointer from the language models that seriously made me rethink something I had written. In fact, now that I have become accustomed to the usual feedback it’s consistently dull and uninspiring. 

I consider this as a good thing. The jobs of writers and editors are probably safe for now. Substituting human writing is a challenge that for AI companies seems very achievable, yet it may be far beyond reach. AI is excellent at bland writing without substance which unfortunately is a resource in no short supply on the internet. But longer write-ups with some level of depth and quality that people want to read for the sheer enjoyment of reading is not something I have seen LLMs produce. 

With that anecdote in mind, let’s consider the job of a lawyer. A lawyer’s job is to bridge the gap between symbolism and reality. Let me explain what I mean by that. 

Symbolism and Shared Reality

Words are commonly understood symbols that represent thoughts, feelings, actions, qualities, material and immaterial objects (even hyperobjects), etc. The human ability to communicate complex and abstract concepts to each other through language has led to all the ideologies and systems we have today on earth. That includes money, religions, and political systems (the best books I have read on this topic is Sapiens and Homo Deus by Yuval Noah Harari). 

Let’s consider the effect of language. It’s trippy to think about how I can write this text here in this time and space, and you can be in another time and space, read it, and cognitively make sense of it. This telepathy-like ability humans have developed with language is unique to the human species. Common languages is what have enabled humans to collaborate and innovate much more than our congeners like gorillas or chimpanzees. Human’s advanced abilities to collaborate and communicate have enabled us to create amazing science, knowledge, art, literature, design, music, cuisine, architecture, etc. 

Beyond language, we share a reality. For example, when a group of people walks through a forest, they share the experience of walking through a forest. But our shared reality has its limits. After all, humans are not a hive mind. You may perceive something differently than I do and have other opinions on things than I have.  That is of course understandable since we have different upbringings, wirings, and we each possess information that others don’t.  

Still, I would argue that such a thing as an objective reality do exist. I will never be able to prove it but my hunch is that something is right and wrong in ultimate terms - like the world’s major religions have always claimed. We just can’t see it from where we are standing. 

As we as individuals face resistance in life, grow, learn, and become older, we see with more and more clarity what objective reality is and what is right and wrong. This increasing clarity is not an intellectual knowledge that a scientist would be able to quantify or measure. It’s more of a non-verbal, intuitive knowledge that can only be acquired through lived experience. We call it wisdom = knowledge about objective reality. 

Wisdom is the number one skill lawyers need to have.  The clarity to see reality for what it is and then describe it in language, that is the finest job of a lawyer. The point of a court case is to reach truth. Not the ultimate, objective truth of reality - such a thing can never be proven – but to come as close to this ideal as possible. In other words, bridging the gap between our shared reality and language, or objective reality and symbolism. 

Can a large language model really take on a job like this? A new preprinted study by researchers from the AI Center of Excellence, Onit Inc. in New Zealand suggests: yes, it may be time for lawyers to “call GPT”. However, the study has some serious credibility issues. Let’s take a closer look. 

Better Call GPT, Comparing Large Language Models Against Lawyers