It seems ChatGPT can replace me for some patent tasks

Today I did an experiment.  The goal was to find out whether ChatGPT can replace me for some patent tasks.  It seems the answer is “yes”. 

My situation was that I needed to prepare a response to an Office Action in a patent case in the USPTO.  This particular patent case is extremely difficult subject matter.  The subject matter is very software intensive, with lots of deep neural networks, machine learning, and AI.  Against all odds the Examiner did not impose a 101 rejection.  The Examiner rejected claim 1 as supposedly obvious over a two-way combination of two NPLs.

It took quite a lot of work, over the course of several days, to prepare a response to the Office Action.  This is the kind of case in which I work on it for some hours, review the patent application as filed, study the Office Action, study each of the NPLs, and write part of a response.  Then I set it aside for a day or so, then resume the effort.

Part of the work included doing OCR on each of the NPLs, so that I could do text searches in the NPLs.  (Of course in the Office Action the USPTO had flattened each NPL to a pure image-based PDF, so there was no choice but to do OCR.)

Eventually after many days of work I had arrived at a proposed response to the Office Action.

Then I decided to do an experiment, asking ChatGPT to come up with arguments against the 103 rejection.

This is a published case, so I did not need to worry about revealing client secrets.  I did click around in the settings of ChatGPT to say “do not use my session to train the AI”.

The last time I used ChatGPT was maybe a year ago, and I asked it to describe the neighborhood where I live.  ChatGPT said the neighborhood had tennis courts and a swimming pool, neither of which was true.

And we have all read the many news stories where a lawyer gets sanctioned for filing a brief in court that turns out to have hallucinated court cases in it.

Here (with redactions) is the query that I gave to ChatGPT:

Please see U.S. application nn/nnn,nnn which I call “the Patent Application”.

And please see Plinge et al., “redacted title” (redacted citation to professional conference volume), which I call “Plinge”.

And please see “redacted title” by Spelvin et al, which I call “Spelvin”.

A patent examiner has devised a combination of Plinge and Spelvin.  The patent Examiner has rejected claim 1 of the Patent Application as supposedly being obvious over Plinge and Spelvin.

Please find arguments to convince the Examiner that claim 1 is not obvious over Plinge and Spelvin.

About ten seconds passed and what appear on the page was a lot of text.  The text, when printed on paper, amounted to about six pages at 1½ spacing.

The first few pages were an analysis of the two NPLs and a detailed discussion of reasons why the two NPLs don’t really add up to a good 103 rejection.

There was a place where the chatbot said “oh and on this point about X you do not want to overstate the argument because if you say that Plinge completely lacks X it is not true and you would lose credibility.”  It said I need to acknowledge that there is a certain amount of X in Plinge in a particular way.

Then there was a page detailing reasons why one skilled in the art would not combine the two references.

The chatbot then said “here is suggested language for a response to the Office Action”.   This was about two pages.

At this point I had two goals:

• • compare my proposed response with the proposed response from the chatbot, and
  • scrutinize the proposed response from the chatbot from the top to the bottom, checking carefully for anything that might be a hallucination or untruth.

The second goal was, of course, quite a lot of work.  This is kind of like in law school when I did tech-citing and sub-citing for a law review, meaning careful checking that the citation was technically accurate, and careful checking that the cited reference really said what it was claimed to say.  This included checking to see if a quotation was scrupulously accurate, and it included working through the reason and logic of each line of argument.

This was also kind of like being a partner in my law firm and checking the work of a young associate.

Unlike a year ago when ChatGPT hallucinated a tennis court and swimming pool, I was pleasantly surprised that I could not find any hallucination or untruth in this document.  So far as I could discern, it was a document to which I could safely affix my signature and USPTO patent registration number.

Now we turn to the first goal — comparing my proposed response with the proposed response from the chatbot.

Each argument that I had devised for my proposed response could be easily found in the proposed response from the chatbot.  I am rather chagrined to report that in several places, the chatbot was more concise and clearer than what I had written.

Not only that, in the “no motivation to combine” area, the chatbot identified a line of attack that I had missed.  I had found one reason why the two NPLs ought not to be combined, and the chatbot found that reason too — but the chatbot found a second reason.

It had taken perhaps a dozen hours for me to arrive at my proposed response.  It took the chatbot about ten seconds to arrive at its proposed response.

It took me well over an hour to fact-check and cite-check the chatbot’s response.

Now I need to quote the last few sentences of the response from the chatbot:

One important caveat

I would not yet file the above verbatim without seeing the examiner’s actual rejection.  The precise wording matters enormously in an obviousness response.  In particular, the examiner may have identified particular pages/figures of Plinge and Spelvin for each limitation and may have supplied a specific rationale for combining them.

If you provide the Office Action (PDF or pasted §103 rejection for claim 1), I can do the more valuable next step: map the examiner’s exact Plinge/Spelvin citations limitation-by-limitation, identify precisely where each citation fails, and draft a point-by-point response designed to answer the examiner’s stated rationale rather than a generic §103 response.

So indeed I pasted the text of the Office Action into the chat.

About ten seconds passed, and the chatbot spewed about ten printed pages’ worth of text onto the screen.  The chatbot had fact-checked every citation and quotation made by the Examiner, and found a place where the Examiner made a mistake when saying where a certain quoted passage actually came from.  The chatbot then offered a revised proposed response.

Of course I then fact-checked and cite-checked the chatbot’s second proposed response, which took about another hour.  Again, despite scrupulous checking, I was not able to find any hallucinations or mistakes.

I then spent around an hour splicing together various selections from the two documents to arrive at a single document that I could confidently sign and file at the USPTO.

This is, of course, a single anecdote.  Past performance is no guarantee of future performance.

But to the extent that this anecdote ends up being an indication of things to come, it is clear that law firms will gradually find that they do not need so many associates.  A law firm partner that until now supervised the work of, say, six associates is going to find that three of the associates can be let go.  The remaining four lawyers, armed with ChatGPT or some other AI, will be able to do the work previously done by seven lawyers.

Another observation is that in this particular case, the client will be better served by the “Carl plus chatbot” approach than by the “Carl alone” approach.

In this case the total amount of professional time for “Carl alone” was around a dozen hours.

In this case the total amount of professional time for “Carl plus chatbot” was around fifteen hours (plus two staring-at-the-screen intervals of about ten seconds each).

As mentioned above, there is no doubt in my mind that the client will be better served by my filing the “Carl plus chatbot” response” instead of the “Carl alone” response.  I believe the chances of overcoming the rejection are either the same or better with the inputs from the chatbot.  As I say, there are places where the chatbot was more concise or clearer than me.  And there are places where the chatbot thought of something helpful to say that I had not thought of.

Might as well acknowledge the elephant in the room.  If I were to eventually arrive at some level of comfort and confidence using the chatbot to help prepare a response to an Office Action, it would permit me to get work done faster, and with higher quality of output, than before.  Maybe it would even save money for the client.

Oh and this was the free-of-charge version of ChatGPT.  Some people pay for their ChatGPT, but in this case they did not receive one cent from me.

Dear reader, please post a comment below.