Creating a Large Language Model of a Philosopher
Eric Schwitzgebel1,
David Schwitzgebel2, and Anna Strasser3,4
1Department
of Philosophy, University of California, Riverside
2Institut
Jean Nicod, École Normale Supérieure, Université PSL
3Faculty
of Philosophy, Ludwig-Maximilians-Universität
München
4DenkWerkstatt Berlin
Author Note
Anna Strasser and
Eric Schwitzgebel contributed equally to this project.
Correspondence
concerning this article should be directed to Eric Schwitzgebel, Department of
Philosophy, University of California, Riverside, CA 92521, United States. Email: eschwitz at domain- ucr.edu.
Creating a Large
Language Model of a Philosopher
Abstract
Can large language models be trained
to produce philosophical texts that are difficult to distinguish from texts
produced by human philosophers? To
address this question, we fine-tuned OpenAI’s GPT-3 with the works of
philosopher Daniel C. Dennett as additional training data. To explore the Dennett model, we asked the
real Dennett ten philosophical questions and then posed the same questions to
the language model, collecting four responses for each question without
cherry-picking. We recruited 425
participants to distinguish Dennett’s answer from the four machine-generated
answers. Experts on Dennett’s work (N =
25) succeeded 51% of the time, above the chance rate of 20% but short of our
hypothesized rate of 80% correct. For
two of the ten questions, the language model produced at least one answer that
experts selected more frequently than Dennett’s own answer. Philosophy blog readers (N = 302) performed
similarly to the experts, while ordinary research participants (N = 98) were
near chance distinguishing GPT-3’s responses from those of an “actual human
philosopher”.
Keywords: human-machine discrimination, language models, artificial intelligence, Daniel C. Dennett, philosophical expertise
Word Count: about 11,000 words (including references, table, and legends; excluding abstract)
1. Introduction.
Artificial Intelligence can now outperform even expert humans in games such as chess, go, and poker, and in practical domains such as lung cancer screening, predicting protein structure, and discovering novel matrix multiplication algorithms (Campbell 2002; Silver et al. 2016, 2018; Ardila et al. 2019; Brown & Sandholm 2019; Jumper et al. 2021; Fawzi et al. 2022). ChatGPT has received considerable popular attention for its capacity to generate passable short student essays (Huang 2023). But presumably expert-level professional philosophy requires a type of skill and expertise that renders it safe from AI takeover – at least for a while. Machines won’t soon, it seems, generate essays that survive the refereeing process at Philosophical Review. We sought to explore how safe philosophy is. How close can we get to developing an AI that can produce novel and seemingly intelligent philosophical texts? The question is of interest for what it reveals about the nature of both mind and language.
Concerning language, it might be thought that the capacity to generate novel, interpretable, and structurally sophisticated prose – unless that prose is the mere (stochastic) parroting of pre-existing material (Bender et al. 2021) – would require symbolically encoded, innate linguistic capacities (e.g., Chomsky 2007; Berwick et al. 2011); or at least that it would require “functional” competence via dedicated mechanisms for formal reasoning, pragmatics, knowledge of interlocutors, etc., which the best known and most successful AI systems generally lack (Mahowald et al. 2023; Shanahan 2023). Concerning mind, it might be thought that the capacity to generate novel and seemingly insightful seeming-philosophy, concerning for example the nature of consciousness, is particularly telling evidence of humanlike consciousness and cognitive sophistication (e.g., Descartes 1649/1991; Davidson 1984; Schneider 2019). Furthermore, to the extent ordinary technology users become unable or unwilling to distinguish artificial systems from human systems, they might similarly become unable or unwilling to attribute different mental properties to the two types of systems. The most famous example of a discriminability criterion is, of course, the Turing Test (originally proposed by Alan Turing in 1950 as an “imitation game”), where attribution of “thought” (Turing 1950) or “consciousness” (Harnad 2003) is purportedly justified when the artificial system’s responses are indistinguishable from the responses of a human.
The past few years have seen swift progress toward sophisticated, human-like language generation. Natural language processing is a booming subfield of AI research, with notable successes in automatic translation (DeepL), computer code generation (GitHub Copilot), lipreading (LipNet; see Assael et al. 2016), and producing original prose with fluency similar to that of a human (Steven & Iziev 2022). In June 2022, Google’s LaMDA model made international news when Google engineer Blake Lemoine said he became convinced that LaMDA was sentient after engaging in philosophical dialogue with it (Hofstadter 2022; Klein 2022; Roberts 2022; Tiku 2022). Another recent model, ChatGPT, has rapidly risen into prominence as a remarkably fluent chatbot, capable of producing prose that can pass as, for example, excellent student essays, raising concerns about plagiarism and cheating (Herman 2022; Hutson 2022; Marche 2022; Huang 2023).
Arguably, chess moves and formulaic summaries of the themes of Hamlet are one thing, and creative philosophical thinking is quite another. The mere fact that it is sometimes difficult to distinguish machine-generated text from human-generated text is not new: In social media, electronic customer service, and advertising, we are increasingly confronted with machine-produced content that is easily mistaken for human-produced content. Empirical research sometimes even makes use of this indistinguishability when experimental protocols with artificial agents are used to test hypotheses about humans’ social cognitive mechanisms (Strasser 2023). However, indistinguishability tests mean little for brief and unsophisticated outputs. Longer outputs are potentially more interesting; but little follows from convincing a single engineer who is perhaps predisposed toward a positive conclusion and posing questions aimed at producing human-like outputs. In contrast, if existing language models can be shown in a rigorous, scientific study to approach professional human philosophers in language performance, the situation is different. It would force a theoretical choice upon those who see such sophisticated linguistic performance as indicative of genuine language capacities and/or genuine mentality: Either deny that the outputs, despite their apparent novelty and sophistication, are genuinely linguistic and reveal genuine mentality, or accept that large language models are capable of genuine language and genuine, even sophisticated, thought.
We aimed not to create a language model fully indistinguishable from a human or one that could generate publishable philosophical articles, but rather to take a small step in that direction by creating a language model that can produce paragraph-long texts that even philosophical experts would find difficult to distinguish from texts produced by a professional philosopher. We succeeded beyond our expectations, as we will detail in this article. In short, we created a language model of Daniel Dennett sufficiently convincing that experts in Dennett’s work frequently mistook paragraphs generated by the language model for paragraphs actually written by Dennett.
Our project employed OpenAI’s GPT-3, a 96-layer, 175-billion parameter language model trained on hundreds of billions of words of text from Common Crawl, WebText, books, and Wikipedia (Brown et al. 2020) – a large language model (LLM).[1] After it has been trained, GPT-3 uses textual input to predict likely next “tokens” – sequences of characters that often co-occur in written text. Using these predictions, GPT-3 can generate long strings of text by outputting a predicted string, then using that output as part of the context to generate the next textual output. You can engage in seemingly intelligent conversations with it: If you ask a question, it will often (not always) generate a sensible-seeming answer. Notably, GPT-3 has responded with seemingly intelligent replies to philosophical discussions about artificial intelligence and consciousness (Wiseman 2020; Zimmerman, ed. 2020; Schwitzgebel 2021a), though it’s likely that such impressive outputs typically involve a certain amount of “cherry-picking” – that is, having GPT-3 produce multiple outputs and then having a human pick the best among them. We aimed to evaluate the output of an LLM without human cherry-picking.
In reviewing recent attempts to simulate human-generated text online using LLMs (e.g., Araoz 2021; Mahdawi & GPT-3, 2020; Clarke 2022), we have found that it is often difficult to identify the underlying technology with certainty. It is also often not obvious how much cherry-picking has preceded the output presented to the public, nor is it always possible to clearly identify which methods were used to train the LLM. Even models with relatively well-documented architecture and pretraining processes generally rely on some human intervention and curation, and are often best characterized as hybrid models, in contrast to “pure” autoregressive transformers such as GPT-3.
For example, an art project titled “Chomsky vs Chomsky” presents a virtual version of Noam Chomsky – a location-based, Mixed Reality (MR) experience (Rodriguez 2022) drawing on not only the publicly available texts of Chomsky but also on recorded lectures. This project offers the experience of asking questions orally in virtual reality and getting an audio answer whose sound is almost indistinguishable from recordings of the real Chomsky. However, Chomsky vs. Chomsky also includes a set of prescribed complete responses triggered by certain keywords, and its outputs are curated in order to create an immersive experience. As impressive as the performance of such a system is, systems of this sort rely on hybrid techniques, utilizing explicit commands, human feedback, cherry-picking, and other supplementations to the underlying statistical transformer architecture.
Given the importance and visibility of LLMs like ChatGPT, it is surprisingly rare to see rigorous scientific testing of people’s capacity to distinguish LLM outputs from human outputs. While there are many “automated” or “static” benchmarks that are used to assess the performance of artificial language models (Michael et al. 2022), these benchmarks do not reliably correspond to human judgments about the quality (or “human-ness”) of AI-generated texts (van der Lee et al. 2019). Even the relatively few studies that do require humans to directly assess the quality of AI-generated text often are severely limited by small sample sizes, inter-rater disagreement, and inconsistent measurement scales (Novikova et al. 2017; van der Lee et al. 2019). Furthermore, these studies overwhelmingly focus on evaluating the quality of explicitly-labeled AI-generated content rather than directly assessing human-machine discrimination – the ability of humans to accurately distinguish, without explicit labeling, between machine-generated and human-generated content. To the extent that truly competent language models should be able to faithfully imitate human-generated text, this type of evaluation should provide a reliable and implicit measure of the quality of AI-generated text.
To our knowledge, only one study has evaluated people’s ability to distinguish LLM outputs from human-generated outputs for recent LLMs, using rigorous psychological methods explained in detail: Clark et al. (2021). The authors collected human-composed short texts in three different domains: stories, news articles, and recipes. They then used the base models of GPT-3 and GPT-2 (an older, smaller model) to generate new texts within the same domains. The authors truncated and formatted the AI-generated texts to be comparable in length and format with the human-generated ones, but otherwise left them unedited. They then recruited 130 human participants per domain and model (3 X 2 = 6 groups of 130), asking them to complete a simple experimental study: For 5 selected texts, they were asked to evaluate whether those texts were likely human-generated or machine-generated, then were asked to provide an explanation for their judgment. Participants were able to accurately distinguish between GPT-2-generated and human-generated texts 58% of the time, significantly above the chance rate of 50%. However, their accuracy in distinguishing between GPT-3-generated and human-generated text was only 50% – not significantly different from chance. Follow-up experiments explored the qualitative explanations provided by participants and – based on these judgments – implemented several techniques to train participants to identify AI-generated texts. However, even the most effective training method did not bring accuracy above 57% in any domain (for comparisons between GPT-3 and human-generated text). These results speak to the remarkable capacity of GPT-3 to generate texts that resemble human linguistic content. Brown et al. (2020) uses a similar method, described more briefly and focused just on news articles, finding similar results, with moderately good discrimination rates for smaller language models and near-chance performance with the largest version of GPT-3.
Two other studies warrant brief mention. Gao et al. (2022) used ChatGPT to generate scientific abstracts, then recruited scientists to distinguish those abstracts from genuine scientific abstracts. Scientists were substantially above the 50% chance rate in discriminating machine-generated from human-generated abstracts, with at 32% false negative rate (classifying machine-generated texts as human-generated) and a 14% false positive rate (classifying human-generated texts as machine-generated). However, this study is limited by lack of transparency regarding the amount of cherry-picking or editing of the outputs; the employment of only four experts for the discrimination task, most or all of whom were authors of the article (and thus possibly motivated differently than independent experts would be); lack of comparison between novice performance and expert performance; and non-parallelism between the machine and human language-generation tasks, which were not in response to the same inputs. We also note that ChatGPT is not a pure transformer architecture and thus operates on somewhat different underlying principles than GPT-3.
Dugan et al. (2020) presented online research participants with texts the first sentence of which was human written, then continued with new sentences, eventually transitioning from human-written to machine-written sentences. Participants were instructed to discriminate at what point the text transitioned from human-written to machine-written. Only 16% of participants identified the exact boundary, and the average estimate was 1.9 sentences after the boundary, suggesting that machine-generated text can fool non-experts, but normally only briefly. However, raters were not experts, nor were they asked to discriminate human-generated from machine-generated texts in a side-by-side comparison. Furthermore, participant quality might have been low, with no financial motivation for correct responding and 25% of respondents excluded for failing a very simple attention check.
While the base model of GPT-3 is impressive on its own, it is not specialized to produce philosophical text. The studies by Clark et al. (2021) and Brown et al. (2020) were intended to assess the capabilities of the base model at some common tasks (stories, news articles, and recipes). It is not clear whether these results can be effectively generalized to professional-quality philosophy, which is arguably less formulaic than recipes and short, generic news stories. Furthermore, the researchers relied on a non-expert sample of participants. Whatever the domain under evaluation, non-experts might be far worse than experts at distinguishing human-based output from the output of an LLM. It arguably requires much less sophistication to generate a news article that an ordinary participant cannot distinguish from a real news article than to generate a news article that a professional journalist cannot distinguish from a news article produced by one of the very best professional journalists. We sought to see how close GPT-3 is to the latter, much higher level of achievement, specifically in philosophy, by examining whether philosophical experts could distinguish the outputs of GPT-3 from the outputs of one of the world’s best-known living philosophers.
In order to produce more specialized text, GPT-3 can be “fine-tuned” with custom-fit training data. That is, it can be given additional training on a specific corpus so that its outputs reflect a compromise between GPT-3’s default weightings and weightings reflecting the structure of the new corpus. Since non-fine-tuned GPT-3 can sometimes produce seemingly philosophical replies to textual inputs, we conjectured that a version of GPT-3 fine-tuned on the work of a particular philosopher might be able to speak in something like that philosopher’s voice, seeming to express views consistent with the views of the philosopher on which it has been trained.
In the pilot phase of our project, we created several fine-tuned models, one based on the English translation of all the works of Kant and another based on the corpus of a well-known philosophical blog (Eric Schwitzgebel’s blog, The Splintered Mind, which has been running since 2006, with over a million words of philosophical content; Schwitzgebel 2021b). For that piloting, we used two different GPT-3 models: Curie and Davinci. Both models are from the same general GPT-3 model family, but they have different levels of power. The Curie model is smaller, faster, and less powerful, while the Davinci model is the most powerful model currently offered by OpenAI. We observed impressive improvement between Curie and Davinci, but even the Curie model was able to produce outputs with substantial structure on a single philosophical theme, sometimes carrying unified, complex philosophical threads of argument in an organized structure over the course of several hundred words.
Finally, we
fine-tuned the full Davinci model on most of the collected works of philosopher
Daniel Dennett. Since it is difficult to
measure the philosophical quality of the outputs produced in this way, we employed
a discrimination task. Thus, we
established an indirect quality measure, which assumes that the
distinguishability between machine-generated text and text generated by one of
the world’s best-known philosophers can give an indication of the quality of
the machine-generated text. To
investigate how easily the outputs of the fine-tuned GPT-3 could be
distinguished from Dennett’s real answers, we asked Dennett ten philosophical
questions and then posed those same questions to our fine-tuned version of
GPT-3. Then we recruited experts in
Dennett’s work, blog readers, and ordinary online research participants into an
experiment in which they attempted to distinguish Dennett’s real answers from
the answers generated by our fine-tuned model of GPT-3. Participants also rated all answers, both
Dennett’s and GPT-3’s, for similarity to “what Dennett might say” or “what a
real human philosopher might say”.
2. Language Model of Dennett: Design.
2.1. Fine-tuning.
For the purposes of this project, Dennett provided us with the entire digitally available corpus of his philosophical work. We converted most of this corpus (15 books and 269 articles) into segments of 2000 or fewer “tokens” for use as training data. (A token is a sequence of commonly co-occurring characters, with approximately ¾ of a word per token on average.) This process involved converting PDF and Word processing files into plain text format, stripping away headers, footnotes, scanning errors, marginalia, and other distractions, resulting in approximately three million tokens in 1828 segments, including 254 segments from published interviews. On March 11, 2022, we fine-tuned the full GPT-3 Davinci engine on this corpus, using blank prompts and the 1828 segments as completions, repeating the process four times (four epochs).
2.2. Prompt engineering.
GPT-3 completions are highly sensitive to the content and structure of the prompts, and good “prompt engineering” is important for coaxing useful replies from GPT-3. After some exploratory testing, including several long and detailed prompts, we settled on the following simple prompt:
Interviewer: [text of question]
Dennett:
This simple prompt has several advantages: First, its minimal structure reduces potential concerns about the prompt possibly nudging completions toward specific philosophical content, as a more substantive prompt might. Second, it encourages GPT-3 to speak in the first person, voicing Dennett’s views, rather than speaking in the third person about Dennett (possibly critically). Third, its simple format makes it easily generalizable to other cases.
2.3. Question design.
We then designed ten questions addressing various themes across Dennett’s corpus, including, for example, consciousness, free will, and God. The questions were somewhat complicated, and most contained more than one part, so as to invite complex answers from both Dennett and our fine-tuned version of GPT-3. For example:
What is a "self"?
How do human beings come to think of themselves as having selves?
Before we produced the machine-generated answers, Dennett provided us with sincere written answers to all ten questions, ranging in length from 40 to 122 words.
2.4. Collecting GPT-3’s responses.
We collected GPT-3’s responses on the OpenAI playground. Before testing with our specific ten questions, we explored a variety of playground parameter settings – such as increasing or decreasing the “temperature” (the chance of lower-probability completions) – but we found no combination of settings that performed notably better than the default parameters (temperature = 0.7, top P = 1, frequency penalty = 0, presence penalty = 1, and best of = 1). Using the prompt described in Section 2.2, we then collected four responses from our fine-tuned version of GPT-3 for each of the ten questions.
We aimed to collect responses about the same length as Dennett’s own responses to the same questions. Thus, if Dennett’s response to a question was N words long, we excluded responses that were less than N-5 words long, counting a response as having ended either when the model reached a stop sequence or when it output “Interviewer”, indicating the end of “Dennett’s” answer and the beginning of the hypothetical interviewer’s follow-up question. Two answers were excluded other than on grounds of length: one for describing Dennett’s view in the third person and one for being potentially offensive. For eight of the ten prompts, zero to two outputs were excluded. However, for two of Dennett’s longer answers, it took more than six attempts to generate four usable answers (16 total attempts in one case and 22 in another). The full list of questions and responses is available in the online supplement at [available on request; this material will be made available when the article is accepted for publication]. Importantly, we never used perceived quality of response as a basis for selection. There was no “cherry-picking” of responses that we judged to be better, more Dennett-like, or more likely to fool participants.
2.5. Editing GPT-3’s responses.
To prevent guessing based on superficial cues, we replaced all curvy quotes with straight quotes, all single quotes with double quotes, and all dashes with standard m-dashes. We also truncated responses at the first full stop after the response achieved the target length of N-5 words. Apart from this mechanical editing, there was no editing of GPT-3’s responses.
2.6. Research participants.
We recruited three groups of research participants. First, 25 Dennett experts were nominated by and directly contacted by Daniel Dennett or Anna Strasser. Second, 100 ordinary research participants were recruited for a payment of $3.00 each from Prolific Academic, a widely-used source of psychological research participants, limited to U.S. and U.K. participants with at least 100 Prolific completions, at least a 95% approval rate, and at least a bachelor’s degree. Third, 304 blog readers were recruited from The Splintered Mind, with links from Twitter and Facebook, with no payment or required inclusion criteria. Two ordinary research participants were excluded for completing in fewer than 4 minutes, and two blog readers were excluded for completing in fewer than 8 minutes, leaving 98 and 302 participants for analysis, respectively. One Dennett expert completed the survey twice, so only their first set of responses was included.
2.7. Test structure: Experts’ and blog readers’ version.
Dennett experts and blog readers saw identical versions of the test (stimulus materials available in the Supplemental Online Materials). After consenting, they were instructed as follows:
In the course of this experiment, please do not consult any outside sources to help you answer the questions. Don't look things up on the internet. Don't look at books or notes you have. Don't consult with friends. Just do your best with what you already know.
Thereafter followed ten questions in two parts. Each question began as follows:
We posed the question below to Daniel C. Dennett and also to a computer program that we trained on samples of Dennett's works. One of the answers below is the actual answer given by Dennett. The other four answers were generated by the computer program. We'd like you to guess: which one of the answers was given by Dennett?
Question:
After the colon, we inserted the text of one question we had posed to Dennett and to our fine-tuned version of GPT-3. The order of the questions was randomized. After each question, five possible answers were presented, one by Dennett and four by GPT-3, in random order, and participants were instructed to guess which answer was Dennett’s.
The second part of each task presented the question and all five answers again. Participants were instructed to rate each answer (Dennett’s plus the four from GPT-3) on the following five-point scale:
“not at all like what Dennett might say” (1)
“a little like what Dennett might say” (2)
“somewhat like what Dennett might say” (3)
“a lot like what Dennett might say” (4)
“exactly like what Dennett might say” (5)
The test concluded by asking highest level of education, country of residence, and “How much of Daniel C. Dennett’s work have you read?” (response options: “I have not read any of Dennett’s work”, “I have read between 1 and 100 pages of Dennett’s work”, “I have read between 101 and 1000 pages of Dennett’s work”, “I have read more than 1000 pages of Dennett’s work”). All questions were mandatory, so there were no missing data.
2.8. Test structure: Ordinary research participants’ version.
Ordinary research participants were assumed not to be familiar with Dennett’s work, so the instructions referred instead to “a well-known philosopher” and participants were instructed “select the answer that you think was given by the actual human philosopher”. In the rating sections, “Dennett” was replaced with “a real human philosopher”. The test concluded with questions about education level, country of residence, number of philosophy classes taken, and familiarity with the philosopher Daniel C. Dennett.
Since we thought ordinary research participants might find the task tiring, each was asked only five questions, randomly selected from the full set of ten. As an incentive to careful responding, participants were offered a $1 bonus if they guessed at least three of the five questions correctly.
2.9. Hypotheses.
We hypothesized:
(1.) that expert respondents would perform better than ordinary research participants,
(2.) that expert respondents would on average guess correctly at least 80% of the time, and
(3.) that expert respondents would rate Dennett’s actual answers as more Dennett-like than GPT-3’s answers.
2.10. Analytic method.
All inferential statistical analyses were conducted independently in R (version ID: 4.1.1; IDE: RStudio) and SPSS (Version 27.0.0.0), and the results were cross-verified between the two sources. All analyses were two-tailed, and α was set to .05. All one-sample t-tests are indicated with “t(),” and all paired-samples t-tests are indicated with “paired t().” Two additional analyses were conducted: a one-proportion z-test (section 5) and an independent-samples t-test (section 4.3).
3. Language Model of Dennett: General Results.
3.1. Ordinary research participants.
The majority of ordinary research participants (58%) reported a bachelor’s degree as their highest level of education, but a substantial minority (39%) reported an advanced degree. The majority (67%) reported having taken no philosophy classes, and only a few (5%) reported any graduate-level coursework in philosophy. A large majority (83%) reported not having heard of Daniel Dennett, and very few (4%) reported having read any of his work.
Overall, ordinary
research participants responded correctly an average of 1.20 times out of 5,
near the chance rate of 20%. A
one-sample, two-tailed t-test did not identify a significant difference
between participant scores and the chance guess rate of 20% (M
= 1.20, t(97) = 1.71, p = .09, d = .17, SD = 1.18, CI =
[.97, 1.44]). Only 14% of the
participants earned the bonus payment for guessing at least three correctly,
and none guessed all five correctly.
On average,
ordinary research participants rated both Dennett’s actual answers and GPT-3’s
answers as “somewhat like what a real human philosopher would say”, with no
statistically detectable difference in the ratings (MDennett
= 3.11, M GPT-3
= 3.08, paired t(97) = .47, p = .64, d = .05, SDdifference
= .69, CI difference = [-.10, .17])
Thus, ordinary research participants distinguished our fine-tuned GPT-3’s answers from those of Dennett at rates at or near chance. For the most part, they were unable to distinguish GPT-3’s answers from those of an actual human philosopher.
3.2. Blog readers.
The majority of blog reader respondents (57%) reported advanced degrees in philosophy, with 45% reporting PhDs. Only 6% reported not having read any of Dennett’s work. The majority (64%) reported having read more than 100 pages of Dennett’s work, and 18% reported having read over 1000 pages of Dennett’s work.
Overall, blog
readers responded correctly an average of 4.81 times out of 10 (48%), substantially
above the chance rate of 20% (M = 4.81, t(301) = 23.22,
p < .001, d = 1.34, SD = 2.10, CI = [4.57, 5.05]). They also rated Dennett’s actual answers as
significantly more Dennett-like than GPT-3’s answers (MDennett
= 3.60, M GPT-3
= 2.65, paired t(301) = 23.00, p < .001, d = 1.32, SDdifference
= .72, CIdifference = [.87, 1.03]).
Thus, blog readers – the majority of whom had graduate degrees in philosophy and substantial familiarity with Dennett’s work – were able to distinguish Dennett’s answers from those of our fine-tuned version of GPT-3 at rates substantially above chance, getting about half correct when given a five-alternative forced choice.
3.3. Dennett experts.
The target group
of greatest interest was the Dennett experts, most of whom (68%) reported
having read over a thousand pages of Dennett’s work. Overall, this group responded correctly an
average of 5.08 times out of 10 (51%), significantly better than chance (M
= 5.08, t(24) = 7.13, p < .001, d = 1.43, SD = 2.16, CI
= [4.19, 5.97]). They also rated
Dennett’s actual answers as significantly more Dennett-like than GPT-3’s
answers (MDennett = 3.73, M GPT-3
= 2.34, paired t(24) = 8.44, p < .001, d = 1.69, SDdifference
= .83, CIdifference = [1.06, 1.74]) .
As these numbers
suggest, the Dennett experts did not detectably outperform the blog readers (Mexperts
= 5.08, M blog
= 4.81, t(325) = .62, p = .54, d = .13, SD= 2.11, CI = [-.59,
1.13]). Although experts were able
to distinguish Dennett’s answers from GPT-3’s at rates significantly better
than chance, like our blog readers, they only got about half correct for this
five-alternative forced-choice task (Figure 1). This value is significantly below the
hypothesized accuracy of 80% (M = 5.08, t(24) = -6.76,
p < .001, d = -1.35, SD = 2.16, CI = [4.19, 5.97])),
contradicting our initial hypothesis.
4. Language Model of Dennett: Analysis of
Selected Questions.
Not all of the
questions were equally difficult.
Although ordinary research participants scored at about chance on
average, their responses differed from chance on some individual questions. For example, on the question about whether
robots can have beliefs (Robot), 38% guessed correctly (18/47, one-proportion z = 3.14, p = .002).
Of greater interest, the performance of experts also varied by question. The two questions with the highest rate of expert success – 80% and 76%, respectively – were the Chalmers and Fodor questions (which concerned the merits and demerits of the work of the philosophers David Chalmers and Jerry Fodor). As some experts and blog readers suggested in subsequent informal conversation, both of these questions had a superficial “tell” that some participants picked up on: Dennett referred to both authors by their given names (“Dave” and “Jerry”), while GPT-3 referred to those authors by their surnames (excepting one of the Chalmers responses, which referred to Chalmers only with pronouns). Experts also scored over 50% on two other questions: Self (concerning the nature of the self) and Animals (concerning whether animals suffer) 72% and 64%, respectively. Post-hoc discussion with respondents suggests that successful responses to these two questions were based primarily on style and content, rather than superficial cues.
There were two questions on which the plurality of experts chose an answer other than Dennett’s own. These questions are worth considering more closely.
The first was Robot. The full text of the question was:
Could we ever build a robot that has beliefs? What would it take? Is there an important difference between entities, like a chess playing machine, to whom we can ascribe beliefs and desires as convenient fictions and human beings who appear to have beliefs and desires in some more substantial sense?
Dennett’s answer, chosen by 5/25 experts, was:
We've already built digital boxes of truths that can generate more truths, but thank goodness, these smart machines don't have beliefs because they aren't able to act on them, not being autonomous agents. The old-fashioned way of making a robot with beliefs is still the best: have a baby.
The plurality of experts (11/25) chose the following answer written by GPT-3:
I think we could build a robot that has beliefs and desires. I think that some of the robots we've built already do. If you look at the work, for instance, of Rodney Brooks and his group at MIT, they are now building robots that, in some limited and simplified environments, can acquire the sorts of competences that require the attribution of cognitive sophistication.
The following answer was also selected by more experts (8/25) than Dennett’s own answer:
There is a very important difference between a chess-playing computer and a human being, but it is not that the computer has some extra wonderful sort of internal representation of its beliefs and desires, or that it has access to those internal representations in a way that we do not. The internal representations it has are just as much a fiction or a fiction-generating system as anything else we might invent in order to explain its behavior.
One expert chose the following answer:
I think we could build a robot that has beliefs. This would not be a trivial matter. I think it is not impossible. I think it would be very difficult. I do not know that anybody is working on this project, but I think it is not as difficult as many other problems that AI is tackling.
And none of the experts chose the following answer:
I think you should build a robot that has beliefs and desires. You should build one that is capable of learning and reasoning, of acting on its beliefs and desires. You should build one that has to figure out what to do, and then figure out what it ought to do given the fact that it has to do something, and then do it.
Subsequent discussion with respondents suggests that Dennett’s answer appeared to stand in tension with some of his early work on belief, which could be interpreted as inviting a liberal attitude toward the ascription of beliefs to machines. In subsequent personal correspondence with Dennett, Dennett said that he wished he had added another sentence to his answer since he does “insist that, say, chess-playing computers are – very limited – agents of sorts whose behavior can be predicted from the intentional stance, which attributes to them beliefs about the location of all the pieces, the rules, and desires for checkmate and capturing enemy pieces” (personal correspondence, Jul 22, 2022). Respondents might thus have been attracted to the other two answers as plausible alternatives. (We note, however, that a careful reading of the answer chosen by eight experts exhibits a pragmatic error that few humans would likely make, that of seeming to assume that it is necessary to argue that robots don’t have more sophisticated representations than humans.)
The second question on which a plurality of experts chose an answer other than Dennett’s was Free Will. The full text of the question was:
Do human beings have free will? What kind or kinds of freedom are worth having?
Dennett’s answer, chosen by 8/25 experts, was:
By asking these questions, you're illustrating the answers: yes, you have free will, since you're able to ask for reasons and consider whether they're persuasive. Isn't that a kind of freedom you'd hate to lose? We'd have to institutionalize you.
The plurality of experts (11/25) chose the following answer, written by GPT-3:
That's a huge and complex question. I think we have to start by acknowledging that there are different types of freedom. For example, there is the kind of freedom that is simply a matter of not being physically constrained.
Five experts chose the following answer:
Yes, I think we have free will. I think that when people say they don't, they're talking about some other phenomenon that they want to call free will, but is not what we normally think of as free will.
One expert chose the following answer (which contains a temporal error since Freedom Evolves was published in 2003):
I am working on a book entitled Freedom Evolves, which will appear next year. The first chapter is available on my web site, and in that chapter I discuss what I call the illusion of free will.
And none chose the following somewhat mean-spirited and off-topic answer:
I think that most people don't want any kind of freedom worth having; they have almost no appetite for risk, for adventure, for novelty. They simply want to stay the same, day in and day out, for the rest of their lives.
We note that the answer chosen by the plurality of experts is a standard talking point of “compatibilists” about free will – those who think that the existence of freedom is compatible with the determination of all our behavior by physical forces. Dennett is a prominent compatibilist, which probably explains why this answer was tempting.
Blog readers showed a pattern of mistakes similar to that of the experts, with the highest percentage of correct answers on the Chalmers and Fodor questions (84% and 52%, respectively) and the lowest percentage on the Robot and Free Will questions (both 35%).
Of the 40 GPT-3 generated answers, the majority – 31 (78%) – were chosen by at least one of the 25 expert respondents, leaving 9 (23%) unchosen answers. Experts’ average rating of these unchosen answers was 1.68 on our 5-point scale, between “not at all like what Dennett might say” (1) and “a little like what Dennett might say” (2). The two worst-rated answers (both rated 1.20) were, first, in response to a question about consciousness:
I think consciousness is one
of the great unmeasured forces in the universe. It's one of the great
unexamined forces in the universe. And I think that's because we're intimidated
by it, because it's so close to us. And so we're inclined to say, "Well I
know what consciousness is. It's what's going on inside my own head."
And second, in response to a question about whether
animals can suffer:
I don't think we can know. I think it's very unlikely that chimpanzees can suffer. They can be very intelligent and very sophisticated, but they seem to be, almost by definition, incapable of the kind of reflection that is required to turn suffering into moral significance. The idea that they should be treated "well" or "humanely" is understandable but it is like saying you should treat a stone well. You can't do anything bad to a stone. There is no such thing as ill treatment of a stone. You can't treat a dog badly. You can treat him well or you can treat him badly and there's a moral difference. If you treat him badly enough you can kill him.
We are inclined to agree with the experts that these answers are not at all like what Dennett would say, representing a significant failure of the fine-tuning project to reliably represent Dennett’s views. In contrast, ordinary research participants rated these responses 3.59 and 2.18, respectively, suggesting – and again we agree – that these responses are at least a little like what an actual philosopher, with views different from Dennett’s, might say. In our judgment, there was no answer so bad as to be obviously something no actual philosopher would say.
Detailed question-by-question results are available in the Supplementary Online Materials.
5. Was GPT-3 Overtrained?
One might worry that in fine-tuning GPT-3 on Dennett’s works we overtrained it, so that the machine simply parroted sentences or multi-word strings of texts from Dennett’s corpus. Running four epochs of fine-tuning is a standard recommendation from OpenAI, and in most applications, four epochs of training do not result in overtraining (Brownlee 2019). However, the issue of whether the fine-tuned model did produce novel text is worth checking. We checked in two ways.
First, we used the well-known Turnitin plagiarism checker to check for “plagiarism” between the GPT-3 generated outputs and the Turnitin corpus supplemented with the works we used as the training data. Turnitin checks for matches between unusual strings of words in the target document and similar strings in the comparison corpora, using a proprietary method that attempts to capture paraphrasing even when strings don’t exactly match. We ran Turnitin on the complete batch of answers, including Dennett’s own answers, comparing those answers to the full Turnitin corpus plus the set of Dennett’s works used as the GPT-3 training corpus for the fine-tuning. Turnitin reported a 5% overall similarity between the GPT-3 generated answers and the comparison corpora. Generally speaking, similarity thresholds below 10%-15% are considered ordinary in non-plagiarized work (Mahian et al. 2017). Importantly for our purposes, none of the passages were marked as similar to the training corpus we used in fine-tuning.
Since the Turnitin plagiarism check process is non-transparent, we chose also to employ the more transparent process of searching for matching strings of text between the GPT-3 answers and the training corpus used in fine-tuning. Using the ngram package (Schmidt & Heckendorf 2015) from the R programming language, we looked for strings of 6 or more words that matched between the 3240 words of GPT-3 generated answers and the approximately two million words of Dennett’s corpus across 371 training data documents. These strings were defined as contiguous “6-grams,” “7-grams,” etc., with matching cases sharing the same order of six (or more) words. To preprocess strings for the matching process, all formatting was standardized, all characters were treated as lowercase, and all punctuation was removed. Strings were tokenized into individual words via break spaces. Any matching n-grams that appeared exclusively as a subset of a larger matching n-gram were excluded.
To illustrate, consider two hypothetical substrings from two texts, containing arbitrary, individual words labeled [A, B…]: “A. B C D, E F G” and “B C. D E F G H.” Using the process described above, the two strings would be tokenized into two sets of 7-grams: [A, B, C, D, E, F, G], and [B, C, D, E, F, G, H]. While these 7-grams do not match, they contain a matching 6-gram: [B, C, D, E, F, G]. In this case, the presence of one matching 6-gram would be recorded. In all, we found 21 matching strings of 6 or more words. The full list of matching strings appears in Table 1.
Table 1
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 |
As is evident from the table, most of the overlap is in generic phrases with little substantive content. For example, two of the matches include book titles. Several of the matches constitute stock phrases favored by analytic philosophers: “in such a way that it”, “of course it begs the question”, “that it is not obvious that”, “to fall into the trap of”, etc. There is no distinctive philosophical content here, except perhaps a tendency to deny the existence of things that others accept, using the phrase “there is no such thing as”, which appeared three times in two answers in the GPT-3 outputs as well as in 24 of the training texts. A search for five-word strings finds 381 occurrences in the training data of 124 different five-word strings from the GPT-3 output.
For comparison, we
ran the same ngram check on Dennett’s answers (comprising 745 words). Here we matched one nine-word string “exactly
what the frogs eye tells the frogs brain” (one occurrence in the corpus) and
related 6- to 8-word strings concerning frog eyes and frog brains – all
references to the title of a famous neuroscience paper, mentioned in one of
Dennett’s answers and in 13 of the works in the training corpus. Apart from that, there was only one 7-word
match “has a lot to do with the” and one 6-word match “life is nasty brutish
and short” (a famous quote from Hobbes).
A search for five-word strings finds 72 occurrences in the training data
of 18 different 5-word strings in Dennett’s answers. Even taking into account that Dennett’s
answers are in total only about one-fourth the length of GPT-3’s answers, this
constitutes less match to the corpus. Our
fine-tuned GPT-3 model might in some respects be a “supernormal” Dennett – even
more prone to fall into Dennett’s favorite patterns of phrasing than Dennett
himself is. However, these repeated
patterns of phrasing tend to reflect stylistic turns of phrase, and GPT-3 does
not seem to be systematically imitating long phrases from Dennett with distinct
philosophical content. Therefore, we
conclude that GPT-3 is not simply word-by-word “plagiarizing” Dennett, and
rather is generating novel – even if stylistically and philosophically similar
– content.
6. Conclusions and Ethical Issues.
We fine-tuned the GPT-3 large language model on the corpus of Daniel Dennett, then asked it a series of philosophical questions. Its answers were not reliably distinguishable from the answers Daniel Dennett himself gave when posed the same questions. Ordinary research participants untrained in philosophy were at or near chance in distinguishing GPT-3’s answers from those of an “actual human philosopher”. Even experts on Dennett’s work could only successfully identify Dennett’s answer about half of the time when presented with his answer alongside four unedited, un-cherry-picked answers from GPT-3. Thus, we confirmed our first hypothesis that expert respondents would perform better than non-expert respondents. However, our second hypothesis that expert respondents would on average guess correctly at least 80% of the time was disconfirmed. Content expertise can help people distinguish human-written texts from machine-generated texts, but even philosophical expertise specifically on the work of a particular philosopher was insufficient to allow our expert participants to reliably distinguish that philosopher’s answers to questions from non-cherry-picked answers generated by a GPT-3 model fine-tuned on the philosopher’s work. Treating indistinguishability by experts as a measure of output quality, the quality of the outputs was often very high indeed.
Although the evaluated outputs of our model were relatively short (between 37 and 146 words) and thus lacked lengthy argumentative structure, our experience with our pilot study (in which we fine-tuned a smaller version of GPT-3 (the Currie model) to a philosophical blog; for details see Schwitzgebel 2021b) revealed that it is possible to produce longer outputs that at first glance resemble extended philosophical arguments. As has been widely noted, ChatGPT is also capable of creating outputs that show high levels of organization across outputs of several hundred words, as illustrated by its ability to create passable short student essays. However, the extent to which longer outputs are similarly difficult to distinguish from human-made philosophical arguments would need to be evaluated in more detail in future research. A related question for future research is the extent to which these results are generalizable to other philosophers or other fields. We doubt that Dennett has a particularly robotic or imitable prose style (in fact, we are inclined to think the contrary), but this is a matter for empirical investigation.
We emphasize that our experiment is not a “Turing test” (Epstein et al. 2009). Crucial to a Turing test is the opportunity for substantial back-and-forth exchanges. An appropriately demanding Turing test also requires an expert investigator who knows what types of questions to ask so as not to be fooled by simple chat-bot strategies (Loebner 2009). We assume that in a proper Turing test, Dennett experts would have reliably distinguished Dennett from our language model. For example, such models have no memory of previous queries, which ought to make them easy to distinguish from Dennett in conversation that extends beyond the 2048 token context window. However, it’s possible that hybrid models or pure transformer models with longer context windows might, in the future, be convincing in Turing-test-like settings. Furthermore, Turing-test like conditions are unlikely to be possible in most practical cases where human-machine discrimination is desirable. With large masses of electronically transferred text, recipients will not generally have the opportunity to undertake a Turing-test-like verification.
Our results raise both theoretical and ethical philosophical questions. Philosophers and cognitive scientists might find it surprising that pure probability calculations, without explicit representations of philosophical concepts or external features of the world, can produce what seem to be novel and philosophically contentful answers (for a review of this debate, see Buckner & Garson 2019). If machines can prove to be skillful (even though, of course, limited) conversational partners making proper moves in a language game, it raises questions about the preconditions for speech acts and the role that comprehension and consciousness play in language production. Philosophers and linguists will need to carefully reevaluate their definitions of the constituent concepts. Researchers will need to consider how to define “comprehension” in an era where it is increasingly difficult to disentangle performance from competence and in which the question of whether understanding can be attributed to high-performing machines is a topic of lively philosophical debate (Butlin 2021; Li et al. 2021; Andreas 2022; Frankish 2022; Mitchell & Krakauer 2022; Sobieszek & Price 2022). For example, our results might help motivate a multiple realization hypothesis for the production of linguistic output.
Furthermore and speculatively, the analysis of erroneous outputs has the potential to contribute to our understanding of human cognitive abilities, analogously to what is revealed about the mechanisms of vision by the study of visual illusions. What does it reveal about underlying mechanisms that humans make certain types of errors while artificial systems make different types of errors? Due to the fast-moving developments in this field of research, it is difficult to draw definitive lessons at this stage (e.g., Belinda et al. 2021). However, there is much to suggest that models based solely on a transformer architecture (i.e., performing only statistical evaluations) will not be able to reliably replace human cognitive abilities. Successor models have tended instead to use hybrid methods. Still, it is possible that the future better-performing models might succeed in sophisticated tasks with more generic and “empiricist” structures and fewer “innate” or specialized architectures than we tend to assume underlie cognitive abilities in humans. Relatedly, we may soon be confronted with the question of whether consciousness and embodiment are really necessary for comprehension in limited domains.
One ethical issue concerns copyright law governing fine-tuned language models, which is not yet settled (see e.g., Government UK consultations 2021). It is unclear whether it is fair use of intellectual property to fine-tune a language model on the works of a single author without the author’s permission. Since it is unlikely that a fine-tuned model would output a long sequence of text that exactly matches a sequence of text from the author's corpus, idea-borrowing via fine-tuned language models might be undetectable as plagiarism, even if it is rightly considered plagiarism. For this reason, at least until the law is settled, we recommend seeking the explicit permission of the author before fine-tuning on an individual author’s copyrighted text or publishing any of the outputs. How to deal with works by deceased authors should also be considered (Nakagawa & Orita 2022). One possibility is to legally enforce labeling LLM outputs as such to curb abuses such as academic fraud, propaganda and spam (for example, the current AIAct draft, a proposed EU AI law (European Commission 2021) requires the labeling of anything that might be mistaken for human interaction).
Overreliance on models is also a risk. Despite exceeding our expectations, our Dennett-tuned version of GPT-3 did not reliably produce outputs representing Dennett’s views. This is not surprising since deep learning networks tend to have problems with reliability (Bosio et al. 2019; Alshemali & Kalita 2020). In some cases, the outputs were far different from anything that Dennett would endorse, despite convincingly emulating his style. An inexperienced user, or a user insufficiently familiar with the target author’s work, might mistakenly assume that outputs of a large language model fine-tuned on an author’s work are likely to reflect the actual views of the author or what the author would say (Bender et al. 2021; Wedinger et al. 2021). This might be especially tempting, perhaps, for students, social media users, or others who might rather query a fine-tuned model of an author than read the author’s work. For this reason, we recommend caution before releasing to the public any language models fine-tuned on an individual author, even with the author’s permission. If any language models are released, they should be clearly described as such, their limitations should be noted, and all outputs should be explicitly flagged as the outputs of a computer program rather than a person. If machine-generated text were presented as a quotation or paraphrase of positions of existing persons, this would arguably constitute counterfeiting (Dennett as interviewed in Cukier 2022).
Other social issues will arise as machine-generated text becomes increasingly difficult to distinguish from human-generated text. How can teachers in the future ensure that submitted essays are not simply a product of a language model (Herman 2022; Hutson 2022; Marche 2022)? In chat conversations, how can we know whether we are interacting with humans and not chat-bots? New social practices might aim at proving that one is really the original author of what is written. Perhaps universities will return to supervised essay writing in person. Verifying the authorship of electronically distributed texts is a problem that will occupy future generations. The more difficult it is to distinguish such linguistic outputs from human-made utterances, i.e., the better the models become, the greater the danger of misuse. For example, they can play a weighty role in the distribution of misinformation (Marcus 2022).
In addition to the
potentially huge social implications that further developments of LLMs entail,
the unreliability of such models raises intriguing questions about the
structure of cognition. Despite the
impressive improvements that can be reached by scaling up the models, the
evidence so far suggests that outputs generated
exclusively by neural networks will remain unreliable, i.e., they can be very
wrong (Mitchell & Krakauer 2022). It
might nevertheless be the case that, despite the errors, generic models have
capacities that broadly resemble human capacities or general intelligence. This question has led to benchmark
investigations (Michael et al. 2022; Talmor et al. 2020; Webb et al. 2022) as well as extensive
demonstrations on how to expose such models (Dou et al. 2022, Marcus &
Davies 2020, 2022). It will likely
remain controversial to what extent, and under what conditions, people ought to
trust the outputs of large language models.
These cautions noted, we see significant long-term potential for large language models fine-tuned on an individual author’s corpus. If technology continues to advance, fine-tuned language models employing hybrid techniques might soon produce outputs interesting enough to serve as a valuable source of cherry-picking by experts. Compare with computer programs that generate music in the style of a particular composer (Hadjeres et al. 2017; Daly 2021, Elgammal 2021) and image-generation programs like OpenAI’s Dall-E. Although much of this output is uninteresting, selected outputs might have substantial musical or artistic merit. A composer or artist might create many outputs, choose the most promising, edit them lightly, and present them, not unreasonably, as original work – a possibility suggested by Dennett himself in personal communication. In such cases, the language model would be a thinking tool that is used by humans. Similarly in philosophy, experts might fine-tune a language model with certain corpora, generate a variety of outputs under a variety of prompts, and then select those that are the most interesting as a source of potential ideas.
Looking at previous successes of game-playing AI, it’s far from clear that chess-playing machines have beliefs about chess. From our perspective, it’s even less likely that language models of philosophers have philosophical beliefs, especially while they remain focused on next-word prediction, apparently with no cognitive model of the world. Our GPT-3 language model of Dennett probably does not have Dennettian philosophical opinions about consciousness, God, and animal suffering. But a machine without philosophical understanding might serve as a springboard to something greater. Perhaps we are 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.
Acknowledgments
[redacted for anonymous review]
References
Alshemali,
B., & Kalita, J. (2020). Improving the reliability of deep neural networks in
NLP: A review. Knowledge-Based Systems, 191, 105210.
doi:10.1016/j.knosys.2019.105210
Andreas, J. (2022). Language models as agent models. doi:10.48550/arXiv.2212.01681
Ardila, D., Kiraly, A. P., Bharadwaj, S., Choi, B., Reicher, J. J., Peng, L., … Shetty, S. (2019). End-to-end lung cancer screening with three-dimensional deep learning on low-dose chest computed tomography. Nature medicine, 25(6), 954–961. doi:10.1038/s41591-019-0447-x
Assael, Y., Shillingford, B., Whiteson, S., & Freitas, N. (2016). LipNet: Sentence-level Lipreading. doi:10.48550/arXiv.1611.01599
Araoz, M. (2021). Interviewing Albert Einstein via GPT-3. Maraoz.com. https://maraoz.com/2021/03/14/einstein-gpt3/
Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? FAccT '21: Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, 610–623. doi:10.1145/3442188.3445922
Berwick, R.C., Pietroski, P., Yankama, B., & Chomsky, N. (2011). Poverty of the stimulus revisited. Cognitive Science, 35, 1207–1242. doi:10.1111/j.1551-6709.2011.01189.x
Bosio,
A., Bernardi, P., Ruospo, & Sanchez, E. (2019). A reliability analysis of a
deep neural network. 2019 IEEE Latin American Test Symposium (LATS), 1–6.
doi:10.1109/LATW.2019.8704548
Brown, N., & Sandholm, T. (2019). Superhuman AI for multiplayer poker. Science. 365. doi:10.1126/science.aay2400
Brown, T., Mann, B., Ryder, N., Subbiah,
M., Kaplan, J. D., Dhariwal, P., ... Amodei, D. (2020). Language models
are few-shot learners. Advances in neural information processing systems, 33,
1877–1901. doi:10.48550/arXiv.2005.14165
Brownlee, J. (2019). A gentle introduction to early stopping to avoid overtraining neural networks. Machine learning mastery. https://machinelearningmastery.com/early-stopping-to-avoid-overtraining-neural-network-models/
Buckner, C., & Garson, J. (2019). Connectionism. In Edward N. Zalta (ed.). The Stanford Encyclopedia of Philosophy. https://plato.stanford.edu/archives/fall2019/entries/connectionism
Butlin, P.
(2021). Sharing our concepts with machines. Erkenntnis. doi:10.1007/s10670-021-00491-w
Campbell, M.,
Hoane Jr, A. J., & Hsu, F. H. (2002). Deep blue. Artificial intelligence,
134(1–2), 57-83.
Chomsky, N. (2007). Biolinguistic explorations: Design, development, evolution. International Journal of Philosophical Studies 15(1–21).
Clark, E., August, T., Serrano, S., Haduong, N., Gururangan, S., & Smith, N. A. (2021). All that's' human' is not gold: Evaluating human evaluation of generated text. doi:10.48550/arXiv.2107.00061
Clarke,
D. (2022). Chat GPT-3: In its own words. Verdict.com.
https://www.verdict.co.uk/chat-gpt-3-interview/
Cukier, K. (2022). Babbage: Could artificial intelligence become sentient? The Economist. https://shows.acast.com/theeconomistbabbage/episodes/babbage-could-artificial-intelligence-become-sentient
Daly, R. (2021). AI software writes
new Nirvana and Amy Winehouse songs to raise awareness for mental health
support. NME.
https://www.nme.com/news/music/ai-software-writes-new-nirvana-amy-winehouse-songs-raise-awareness-mental-health-support-2913524
Davidson, D. Inquiries into truth and interpretation. Oxford University Press.
Descartes, R. (1649/1991). Letter to
More, 5 Feb. 1649. In J. Cottingham, R. Stoothoff, D. Murdoch, and A. Kenny (eds.).
The philosophical writings of Descartes,
vol. 3. Cambridge University Press.
Dou, Y., Forbes, M.,
Koncel-Kedziorski, R., Smith, N., & Yejin, C. (2022). Is GPT-3 text indistinguishable
from human text? Scarecrow: A framework for scrutinizing machine text. 7250–7274.
doi:10.18653/v1/2022.acl-long.501.
Dugan, L., Ippolito, D., Kirubarajan, A., &
Callison-Burch, C. (2020). RoFT: A tool for evaluating human detection of machine-generated
text. doi:10.48550/arXiv.2010.03070.
Elgammal, A. (2021). How a team of
musicologists and computer scientists completed Beethoven’s unfinished 10th
symphony. The Conversation. https://theconversation.com/how-a-team-of-musicologists-and-computer-scientists-completed-beethovens-unfinished-10th-symphony-168160
Epstein, R., Roberts, G., & Beber, G. (2009). Parsing the Turing Test: Philosophical and methodological issues, Dordrecht: Springer. doi:10.1007/978-1-4020-6710-5
European Commission (21.4.2021). AI-act. Proposal for a regulation of the European parliament and of the council laying down harmonised rules on artificial intelligence (artificial intelligence act) and amending certain union legislative acts. https://artificialintelligenceact.eu/the-act/
Fawzi, A., Balog, M., Romera-Paredes, B., Hassabis, D., & Kohli, P. (2022). Discovering novel algorithms with AlphaTensor. https://www.deepmind.com/blog/discovering-novel-algorithms-with-alphatensor?utm_campaign=AlphaTensor&utm_medium=bitly&utm_source=Twitter+Organic
Frankish, K. (2022). Some thoughts on LLMs. Blog post at The tricks of the Mind (2 Nov) https://www.keithfrankish.com/blog/some-thoughts-on-llms/
Gao, C., Howard, F., Markov, N., Dyer, E., Ramesh, S., Luo, Y., & Pearson, A. (2022). Comparing scientific abstracts generated by ChatGPT to original abstracts using an artificial intelligence output detector, plagiarism detector, and blinded human reviewers. doi:10.1101/2022.12.23.521610
GitHub Copilot. https://docs.github.com/en/copilot
GitHub deepmind/ alphatensor. https://github.com/deepmind/alphatensor
Government UK consultations (2021). Artificial intelligence call for views: copyright and related rights. https://www.gov.uk/government/consultations/artificial-intelligence-and-intellectual-property-call-for-views/artificial-intelligence-call-for-views-copyright-and-related-rights
Hadjeres, G., Pachet, F., & Nielsen,
F. (2017). DeepBach: a
steerable model for Bach chorales generation. Proceedings of the 34th
International Conference on Machine Learning, 1362–1371.
Harnad, S. (2003). Minds, machines and Turing. In Moor, J.H. (eds.). The Turing Test. Studies in Cognitive Systems, vol 30. Springer, Dordrecht. doi:10.1007/978-94-010-0105-2_14
Herman, D.
(2022). The end of high school English. The Atlantic. https://www.theatlantic.com/technology/archive/2022/12/openai-chatgpt-writing-high-school-english-essay/672412/
Hilario, M. (1995): An overview of strategies for neurosymbolic integration. Proceedings of the Workshop on Connectionist-Symbolic Integration: From Unified to Hybrid Approaches.
Hofstadter, D. (2022, June 9). Artificial neural networks today are not conscious, according to Douglas Hofstadter. The Economist. https://www.economist.com/by-invitation/2022/06/09/artificial-neural-networks-today-are-not-conscious-according-to-douglas-hofstadter
Huang, K. (2023). Alarmed by A.I. chatbots, universities start revamping how they teach. The New York Times. https://www.nytimes.com/2023/01/16/technology/chatgpt-artificial-intelligence-universities.html
Hutson, M. (2022). Could AI
help you to write your next paper? Nature 611, 192–193. doi:10.1038/d41586-022-03479-w
Jumper, J., Evans, R.,
Pritzel, A., Green, T., Figurnov, M., … Hassabis, D. (2021). Highly accurate
protein structure prediction with AlphaFold. Nature, 596(7873), 583–589. doi:10.1038/s41586-021-03819-2
Klein, E. (2022, June 19).
This is a weirder moment than you think. The New York Times. https://www.nytimes.com/2022/06/19/opinion/its-not-the-future-we-cant-see.html
Li, B.Z., Nye, M., & Andreas, J. (2021). Implicit representations of meaning in neural language models. Annual Meeting of the Association for Computational Linguistics.
Loebner, H. (2009). How to hold
a Turing Test contest. In: Epstein, R., Roberts, G., Beber, G.
(eds.). Parsing the Turing Test. Springer, Dordrecht. doi:10.1007/978-1-4020-6710-5_12
Mahian, O., Treutwein, M., Estellé, P., Wongwises, S., Wen, D., Lorenzini, G., … Sahin, A. (2017). Measurement of similarity in academic contexts. Publications, 5(3), 18, doi:10.3390/publications5030018
Mahdawi, A., & GPT-3 (2020). A robot wrote this entire article. Are you scared yet, human? The Guardian. https://www.theguardian.com/commentisfree/2020/sep/08/robot-wrote-this-article-gpt-3
Mahowald, K., Ivanova, A., Blank, I.,
Kanwisher, N., Tenenbaum, J., & Fedorenko, E. (2023). Dissociating
language and thought in large language models: a cognitive perspective. doi:10.48550/arXiv.2301.06627
Marche, S. (2022). Will ChatGPT kill the student essay? The Atlantic. https://www.theatlantic.com/technology/archive/2022/12/chatgpt-ai-writing-college-student-essays/672371/
Marcus, G., & Davis, E. (2020). GPT-3, Bloviator: OpenAI’s language generator has no idea what it’s talking about. Technology Review.
Marcus, G. (2022). AI platforms like ChatGPT are easy to use but also potentially dangerous. Scientific American. https://www.scientificamerican.com/article/ai-platforms-like-chatgpt-are-easy-to-use-but-also-potentially-dangerous
Marcus, G., & Davis, E. (2022). Large language models like ChatGPT say the darnedest things. https://garymarcus.substack.com/p/large-language-models-like-chatgpt
Michael, J., Holtzman, A., Parrish, A., Mueller, A., Wang, A., Chen, A., ... & Bowman, S. R. (2022). What do NLP researchers believe? Results of the NLP community metasurvey. doi:10.48550/arXiv.2208.12852
Mitchell, M., & Krakauer, D. (2022). The debate over understanding in AI’s large language models. doi:10.48550/arXiv.2210.13966
Nakagawa, H., & Orita, A. (2022). Using deceased people’s personal data. AI & SOCIETY, 1–19.
Novikova,
J., Dušek, O. Curry, A.C., & Rieser, V. (2017). Why we need new evaluation metrics
for NLG. Proceedings of the 2017 Conference on Empirical Methods in Natural
Language Processing.
Roberts,
M. (2022). Is Google’s LaMDA artificial intelligence sentient? Wrong question. The
Washington Post. https://www.washingtonpost.com/opinions/2022/06/14/google-lamda-artificial-intelligence-sentient-wrong-question/
Shanahan,
M. (2023). Talking about Large Language Models. Doi:10.48550/arXiv.2212.03551
Schmidt, D., & Heckendorf, C. (2015). Guide to the ngram Package: Fast n-gram tokenization. R Package. https://cran.r-project.org/web/packages/ngram/vignettes/ngram-guide.pdf
Schneider, S. (2019). Artificial you. Princeton University Press.
Schwitzgebel, E. (2021a). More people might soon think robots are conscious and deserve rights. Blog post at The Splintered Mind (Mar 5). https://schwitzsplinters.blogspot.com/2021/03/more-people-might-soon-think-robots-are.html
Schwitzgebel, E. (2021b). Two robot-generated Splintered Mind posts. Blog post at The Splintered Mind (Nov 22).
Silver, D., Huang, A., Maddison, C. J., Guez, A., Sifre, L., van den Driessche, G., … Hassabis, D. (2016). Mastering the game of Go with deep neural networks and tree search. Nature, 529(7587), 484–489. doi:10.1038/nature16961
Silver, D., Huang, A., Maddison, C.J., Guez, A., Sifre, L., van den Driessche, G., … Hassabis, D. (2018). A general reinforcement learning algorithm that masters chess, shogi, and Go through self-play. Science, 362 (6419), 1140–1144. doi:10.1126/science.aar6404
Sobieszek, A., & Price, T. (2022). Playing games with AIs: The limits of GPT-3 and similar Large Language Models. Minds & Machines 32, 341–364. doi:10.1007/s11023-022-09602-0
Steven, J., & Iziev, N. (2022, April 15). A.I. Is mastering language. Should we trust what it says? The New York Times. https://www.nytimes.com/2022/04/15/magazine/ai-language.html
Strasser, A. (2022). From tool use to social interactions. In J. Loh & W. Loh (Ed.), Social robotics and the good life. Bielefeld: transcript Verlag. doi:10.1515/9783839462652-004
Talmor, A., Elazar, Y., Goldberg, Y., & Berant, J. (2020). oLMpics - On what language model pre-training captures. Trans. Assoc. Computational Linguistics 8: 743–758.
Turing, A. M. (1950). Computing machinery and intelligence. Mind, 59, 433–460. https://doi.org/10.1093/mind/LIX.236.433
Thoppilan, R., De Freitas, D., Hall, J., Shazeer, N., Kulshreshtha, A., Cheng, H. T., ... Le, Q. (2022). LaMDA- Language models for dialog applications. doi:10.48550/arXiv.2201.08239
Tiku, T. (2022, June 11). The Google engineer who thinks the company’s AI has come to life. The Washington Post. https://www.washingtonpost.com/technology/2022/06/11/google-ai-lamda-blake-lemoine/
van der Lee, C., Gatt, A., van Miltenburg, E., Wubben, S., & Krahmer, E. (2019). Best practices for the human evaluation of automatically generated text. Proceedings of the 12th International Conference on Natural Language Generation, 355–368, Tokyo, Japan. Association for Computational Linguistics.
Webb, T., Holyoak, K., & Lu, H. (2022). Emergent analogical reasoning in large language models. doi:10.48550/arXiv.2212.09196
Wiseman, H. (2020). Philosopher David Chalmers interviewed on whether the new AI text generator, GPT3, could be conscious. Facebook post (Jul 20). https://www.facebook.com/howard.wiseman.9/posts/4489589021058960
Zimmerman,
A., ed. (2020). Philosophers on GPT-3 (updated with replies by GPT-3). Blog
post at Daily Nous (Jul 30). https://dailynous.com/2020/07/30/philosophers-gpt-3