GPT-3 Can Talk Like the Philosopher Daniel Dennett Without Parroting His Words

Earlier this year, Anna Strasser, Matthew Crosby, and I fined-tuned the large language model GPT-3 on the philosophical writings of Daniel Dennett.  Basically, this amounts to training a chat-bot to talk like Dennett.  We then posed ten philosophical questions to Dennett and our to Dennett model, "digi-Dan".  Regular readers of this blog will recall that we then tested ordinary research participants, blog readers, and experts in the work of Daniel Dennett, to see if they could distinguish Dennett's actual answers from those of digi-Dan.

The results were, to us, surprising.  When asked to select Dennett's answer to a philosophical question from a set of five possible answers, with the other four being digi-Dan outputs, Dennett experts got only about half right -- significantly better than the 20% chance rate, but also significantly below the 80% we had hypothesized.  Experts often chose digi-Dan's answer over actual Dan's answer.  In fact, on two questions, at least one of the four digi-Dan outputs was selected by more experts than was Dennett's own response.  (Blog readers performed similarly to the experts, while ordinary research participants were at about chance.)

Anna Strasser and I then brought my son David Schwitzgebel into the collaboration (Matthew Crosby unfortunately had to withdraw, given a change in career direction), and we wrote up the results in a new paper in draft "Creating a Large Language Model of a Philosopher".  Comments welcome, as always!

Presenting our initial results to audiences, we sometimes heard the following objection: Could digi-Dan be doing so well because it's just parroting Dennett's words?  That is, might we have "over-trained" the model, so that it produces long strings of text more or less word-for-word directly from Dennett's corpus?  If so, then the Dennett experts aren't really mistaking a computer program for Dennett.  Rather, they're mistaking Dennett for Dennett.  They're just picking out something Dennett said previously rather than what he happened to say when asked most recently, and nothing particularly interesting follows.

That's an good and important concern.  We addressed it in two ways.

First, we used the Turnitin plagiarism checker to check for "plagiarism" between the digi-Dan outputs and the Turnitin corpus, supplemented with the texts we had used as training data.  Turnitin checks for matches between unusual strings of words in the target document and in the comparison corpora, using a proprietary method that attempts to capture paraphrasing even when the words don't exactly match.  We found only a 5% overall similarity between digi-Dan's answers and the comparison corpora.  Generally speaking, similarity thresholds below 10%-15% are considered ordinary for non-plagiarized work.  Importantly for our purposes, none of the passages were marked as similar to the training corpus we used in fine-tuning.

However, since the Turnitin plagiarism checking process is non-transparent, we chose also to employ the more transparent process of searching for matching strings of text between digi-Dan's answers and the training corpus.  We found only five matching strings of seven words or longer, plus another sixteen strings of six words or longer.  None of these strings has distinctive philosophical content.  A few are book titles.  The rest are stock phrases of the type favored by analytic philosophers.  If you want to see the details, I've pasted a table at the end of this post containing every matching string of six or more words.

Digi-Dan is thus more sophisticated than the objector supposes.  Somehow, it creates textual outputs that Dennett experts often mistake for Dennett's own writing without parroting Dennett's exact words.  It can synthesize new strings of Dennett-like prose.

It by no means follows that digi-Dan thinks like a philosopher, and we emphasize that some of its outputs are unlike what Dennett would say.  But we still find the results quite interesting.  Maybe humanity is on the cusp of creating machines capable of producing texts that seem to sparkle with philosophical cleverness, insight, or common sense, potentially triggering new philosophical ideas in the reader, and perhaps also paving the way for the eventual creation of artificial entities who are genuinely capable of philosophical thought.

-------------------------------------------------------------

Table

Strings of six or more words that match between the GPT-3 outputs and the Dennett training corpus.  The occurrences column indicates the number of separate training data segments in the training corpus in which that phrase appears.  The occurrences total for shorter strings excludes the occurrences in larger matching strings.  (Therefore, if any n-gram that is a subset of a larger n-gram appears in the table, that means that it appeared independently in the text, rather than appearing only within the larger n-gram.  For example, “intuition pumps and other tools for thinking” occurs once outside of “in my new book intuition pumps and other tools for thinking.”)

String	# of words	occurrences
in my new book intuition pumps and other tools for thinking	11	1
is organized in such a way that it	8	1
there is no such thing as a	7	10
figure out what it ought to do	7	1
intuition pumps and other tools for thinking	7	1
there is no such thing as	6	14
i have learned a great deal	6	2
organized in such a way that	6	2
a capacity to learn from experience	6	1
but if you want to get	6	1
capacity to learn from experience we	6	1
in my book breaking the spell	6	1
in such a way that it	6	1
is organized in such a way	6	1
my book breaking the spell i	6	1
of course it begs the question	6	1
that is to say there is	6	1
that it is not obvious that	6	1
the more room there is for	6	1
to fall into the trap of	6	1
what it ought to do given	6	1