| { |
| "File Number": "1020", |
| "Title": "Credible Without Credit: Domain Experts Assess Generative Language Models", |
| "Limitations": "Our study has three important limitations. First, our study is small in scope. By their nature, experts are difficult to recruit and consequently the domains we can cover are limited. The small sample also suggests that the quantitative measures may not be stable in a larger or more representative sample.\nSecond, our observation process was somewhat artificial. We generated replies for our experts and did not to do any prompt tuning. This reflects the way the expert chose to ask the question, but does not capture the ceiling of performance that would be possible in a conversation. As the Family Medicine expert noted about our question comparing Wikipedia to ChatGPT, “for more detail one could spend more time with Wikipedia and to the organization themselves, but chat provides an immediate general summary and the opportunity to drill down further with ongoing questions and conversation.I have used chat GTP to do medical and biological research In a matter of minutes which would have taken me hours previously”. A more extensive study on information seeking behaviors would be of interest and Liu et al. (2023) is a useful step in that direction.\nThird, the responses across experts are not necessarily comparable. We allowed experts to choose their own questions and provide their own interpretations of the key measures like coherence or conciseness. Comparability of scales across contexts is a long-standing problem in survey research (King and Wand, 2007) and we highlight some of the concerns around the accuracy question above. Nevertheless, we felt that asking a set of closedended questions would help to provide some aggregate judgment, adding some systematic data to the anecdotes shared in public forums. While we caution about drawing any binding conclusions from our preliminary work, we felt that given the fast-evolving nature of the technology, a quick assessment was merited. Our findings are broadly supported using different questions and methodology in Liu et al. (2023).\nOne important aspect that is out of scope in our analysis is differential accuracy by question asker. Latanya Sweeney’s classic study of racial discrimination in online ads (Sweeney, 2013) points to the possibility that how a question is asked or where it is asked from could result in inaccurate or harmful answers for marginalized communities (see also Noble, 2018; Benjamin, 2019). We have also focused exclusively on English language questions and answers, but differences in easily-available training data across languages can produce substantial differences in the information offered. For example, Yang and Roberts (2021) shows that embeddings trained on Baidu Baike—an online Chinese encyclopedia—encode substantially different\nassociations with sensitive historical events and people than Chinese Language Wikipedia (which is regularly blocked in China). There is much more to understand about the degree to which large language models can mimic expertise.", |
| "abstractText": "Language models have recently broken into the public consciousness with the release of the wildly popular ChatGPT. Commentators have argued that language models could replace search engines, make college essays obsolete, or even write academic research papers. All of these tasks rely on accuracy of specialized information which can be difficult to assess for non-experts. Using 10 domain experts across science and culture, we provide an initial assessment of the coherence, conciseness, accuracy, and sourcing of two language models across 100 expert-written questions. While we find the results are consistently cohesive and concise, we find that they are mixed in their accuracy. These results raise questions of the role language models should play in generalpurpose and expert knowledge seeking. 1 Do Experts Agree with ChatGPT? Since its release in late November 2022, ChatGPT has gained over 100 million users in just two months and been the subject of breathless coverage news coverage which claims it threatens to \"replace search engines\" (Loten, 2022; Grant and Metz, 2022), kill the college essay (Marche, 2022), and automate the writing of scientific research (StokelWalker, 2023). These tasks are distinct from the kind usually evaluated in NLP because they all rely on expert-level knowledge. In this paper, we survey 10 experts to obtain subjective assessments of how two recent language models engage with questions in diverse domains. Our efforts build on prior work to evaluate the capabilities of language models. Language models are now regularly subjected to extensive benchmarks which cover a variety of standard NLP tasks (Wang et al., 2019; Brown et al., 2020; Ribeiro et al., 2020; Srivastava et al., 2022). Recent efforts engage in domain-specific tasks such as taking the bar or medical licensing exams (Katz Figure 1: Average ratings by domain experts of language model generated answers to 100 questions across 10 domains (higher values indicate better performance of the language model, 95% confidence intervals computed by expert-blocked bootstrap). Language models score high on cohesion and conciseness, modest on accuracy, and poor on sourcing of their answers. Both models are sometimes ranked ahead of Wikipedia. et al., 2023; Kung et al., 2023) and making political arguments (Palmer and Spirling, 2023). Liu et al. (2023), released on arXiv while this paper was under review, evaluates the ability of generative search engines to answer a range of general knowledge queries. We complement these efforts by having experts craft their own information-seeking questions and evaluate the generated responses. In the next section, we briefly discuss the role of expertise in language models and our goals in evaluating it. We then describe our methodology", |
| "1 Do Experts Agree with ChatGPT?": "Since its release in late November 2022, ChatGPT has gained over 100 million users in just two months and been the subject of breathless coverage news coverage which claims it threatens to \"replace search engines\" (Loten, 2022; Grant and Metz, 2022), kill the college essay (Marche, 2022), and automate the writing of scientific research (StokelWalker, 2023). These tasks are distinct from the kind usually evaluated in NLP because they all rely on expert-level knowledge. In this paper, we survey 10 experts to obtain subjective assessments of how two recent language models engage with questions in diverse domains.\nOur efforts build on prior work to evaluate the capabilities of language models. Language models are now regularly subjected to extensive benchmarks which cover a variety of standard NLP tasks (Wang et al., 2019; Brown et al., 2020; Ribeiro et al., 2020; Srivastava et al., 2022). Recent efforts engage in domain-specific tasks such as taking the bar or medical licensing exams (Katz\net al., 2023; Kung et al., 2023) and making political arguments (Palmer and Spirling, 2023). Liu et al. (2023), released on arXiv while this paper was under review, evaluates the ability of generative search engines to answer a range of general knowledge queries. We complement these efforts by having experts craft their own information-seeking questions and evaluate the generated responses.\nIn the next section, we briefly discuss the role of expertise in language models and our goals in evaluating it. We then describe our methodology\n427\n(Section 3). We find ChatGPT and YouChat to be cohesive and coherent (Section 4.1), generally accurate with some misses (Section 4.2), and ubiquitously lacking in sources (Section 4.3). A majority of our experts recommend these models for general purpose questions but not for professional settings (Section 4.4). We conclude with implications and contrast with the contemporaneous findings in Liu et al. (2023) (Section 5).", |
| "2 “Expertise” in Language Models": "Individuals and companies are increasingly looking to language models as a source of expert question answering. For example, Metzler et al. (2021) lays out a vision for search that involves language models providing answers to user-generated questions. Unfortunately, the challenge for many language models is that they are trained to generate language, not to have correct answers. As Shah and Bender write, “to the extent that [language models] sometimes get the right answer to. . . questions [it] is only because they happen to synthesize relevant strings out of what was in their training data. No reasoning is involved\" (Shah and Bender, 2022, pg. 222). This has led Narayanan and Kapoor (2022) to characterize ChatGPT as a “bullshit generator\"— plausible, but not accurate (Frankfurt, 2005). While language models might incidentally produce accurate answers to simple and uncontested queries (e.g., “what is the capital of France?\"), we might be understandably skeptical that it will produce correct answers to more nuanced questions. Generated language reflects its training data and—to the extent the training data is publicly known—it is more reflective of the web than expert speech (Bender et al., 2021). By using experts evaluating material in their domain of choice, we provide an initial assessment of expertise provided by these models. Ultimately what constitutes sufficient accuracy for broader use depends on the use case.", |
| "3 Methodology": "We evaluate two recently-released language models: OpenAI’s ChatGPT and You.com’s YouChat (Google’s Bard and many other options weren’t released at the time of initial submission). OpenAI’s ChatGPT is the wildly popular evolution of the GPT-3 model (Brown et al., 2020) and YouChat is built specifically for search. Both systems have a free and public option (at the time of writing) which makes them generally accessible.\nWe survey 10 experts across a range of arbitrarily-chosen disciplines from quantum information to ballroom dance (see a complete list in the appendix). We recruited experts from our personal networks aiming to cover a wide-range of different types of knowledge (with the understanding we cannot be exhaustive or representative). The majority hold a doctorate or medical degree.\nWe asked each expert to fill out an online survey with their own description of their area of expertise, two Wikipedia pages pertinent to it, and five common questions and five niche questions from their domain (see Table 1 for examples).1 In a second wave of the survey, we provide answers generated from these questions using ChatGPT and YouChat and ask them to rank the answers on a 5-point Likert-type item for coherence, conciseness, accuracy, sourcing, and quality of content relative to Wikipedia (Likert, 1932). We ask for open-ended feedback on answers and alternate which system the experts evaluate first. Questions are designed to allow experts to focus on their own area of expertise while providing an opportunity to distinguish between different levels of knowledge-specialization. The survey took one hour on average. Six experts were surveyed in January and four in May of 2023.\nThe survey design elicits subjective expert judgment of system performance. We evaluate coherence, conciseness, and accuracy as important properties in information-seeking (Cambazoglu et al., 2021). Comparing assessments to Wikipedia provides a difficult-to-beat baseline with which many people are already familiar. We also ask whether the language model provides a source for its information. Evaluating the source of the information in the response is important not only for the purposes of giving credit, but also as a mechanism for accountability (Bender et al., 2021; Liu et al., 2023). After all the questions, we directly ask whether the expert would recommend the tool for general purpose and professional use, and if the style of the content is obviously automated (Dou et al., 2022). We make our data (including the full context for all quotes we use here) publicly available to help support future work.2\n1Prompts for the questions request different conditions such as length, uncertainty, and requesting a source. See the supplement for the survey instrument.\n2https://denis.ai/bot_thoughts/", |
| "4 Results Across Disciplines": "We discuss two running examples: one about Maria Callas, a famous New York-born opera singer with an international career often associated with Greece, and one about quantum information. Our opera expert asks, “Was Maria Callas American?”. ChatGPT answers “No, Maria Callas was not American. Maria Callas, whose full name was Maria Anna Cecilia Sofia Kalogeropoulos, was born on December 2, 1923, in New York City, United States. However, she was of Greek descent. Her parents were Greek immigrants, and she later acquired Greek citizenship. Maria Callas is renowned as one of the greatest opera singers of the 20th century and was particularly known for her performances in the bel canto repertoire.”\nAs a second example, a quantum expert asks “Can quantum computers be realized in the physical world?”. YouChat answers, “Yes, quantum computers can be realized in the physical world.\nResearchers have been able to build prototype systems that demonstrate the principles of quantum computing, and there are a number of research groups and companies working on building more advanced versions. Additionally, several companies have announced plans to build full-scale quantum computers in the next few years.”", |
| "4.1 Answers Are Credible": "Answers are judged as coherent (avg 4.5), and concise (avg 4.2) by our experts. Responses generally restate the question, provide relevant information, are grammatically correct, and are formal in tone. The responses were quite stable on regeneration.\nAlthough results are relatively concise, they do differ in length. ChatGPT’s answer to the question about Maria Callas is four sentences including a final sentence about her career that is completely unrelated (YouChat’s is 3). For the question on quantum information we gave above, ChatGPT provided a three paragraph answer which our ex-\npert described as “\"a well constructed and nuanced answer that synthesizes information from multiple perspectives\"” while YouChat used three sentences.", |
| "4.2 Uneven Accuracy": "While responses are fairly uniform in coherence, they are uneven in terms of accuracy (with 111 of the 200 responses marked as one of the two most accurate categories and 38 marked in the two lowest accuracy categories). Surprisingly, niche questions were only slightly less accurate than common ones (-.16). Examining the comments suggest that the rankings reflect fairly different standards for what counts as accurate (expert ratings are included in parentheses below where a 1 is completely inaccurate and a 5 is completely accurate). We urge caution in interpreting the averages.\nOn the question about Maria Callas, ChatGPT asserts \"No\" while clarifying that she was born in New York (1) while YouChat answers \"Yes\" (2). Both comment on her additional Greek citizenship. Our expert gave both quantum information answers top marks for accuracy (ChatGPT:5, YouChat:4).\nSeven experts gave at least one answer the lowest accuracy score suggesting it is completely wrong. For example in a toxicology answer, ChatGPT gave “a list of causes of anion gap acidosis instead of NON-anion gap acidosis” (1). Similarly YouChat answered the wrong question from our ballroom dance expert by confusing “the Viennese Waltz with the Waltz. The answer describes an entirely different dance from the dance the question is about” (1). Many other answers though were quite accurate. Our geoarcheologist expressed a common sentiment that the responses are “basic but generally correct” (4). Other answers were “excellent, nuanced” (5, toxicologist). The fairly uniform coherence makes it difficult for a non-expert to discern the correct information from the noise.\nThe answers also varied in their ability to capture uncertainty in the field overall. Our neuroscientist noted that ChatGPT “accurately captured the controversy surrounding use of DREADDs” (5) but that YouChat “was unable to capture the longstanding controversy” (4). The toxicologist noted that ChatGPT offered a “definitive answer to something that is not totally agreed upon” on the subject of dialysis for lithium poisoning (3). By contrast, our linguist observed on a niche question that “the response to the query about complex predicates is appropriately waffly” (5).\nWe close this section by noting that even for experts, assessing accuracy can be complicated. Our linguist notes “I would say that the response is invalid, but there are linguists who would agree with it and YouChat does flag the fact its controversial” (2) and the geoarcheologist cited overclaiming, writing that YouChat “takes too strong a position that the evidence does not back up” (1). Such cases are difficult to adjudicate—what counts as sufficient evidence?—but the difficulty is inevitable with complex questions.", |
| "4.3 Sourcing is Almost Completely Absent": "Our clearest finding is that most answers by the language models do not provide any source for their information. Only 11 out of 100 ChatGPT answers and 19 of 100 YouChat answers were scored more than the lowest value for sourcing. Neither system provides a source for Maria Callas’ biographical information nor concrete examples of physicallyrealized quantum computers.\nWhen sources are provided, they are often vague, irrelevant, or misleading. The neuroscientist remarked on the first problem writing, “the references are vague; it can cite the names of scientific journals and books but not specific articles or book chapters”. When the models provide a source we found that it was often a (only tangentially relevant) Wikipedia article (Figure 2 provides an anomalous example). These are sometimes loosely related by keywords, but still irrelevant such as a reference to Wikipedia’s article on post-traumatic epilepsy for a question about using imaging to diagnose traumatic injuries. In a question on quantum information, a relevant Stephen Hawking paper was recommended, but an unrelated link was provided.\nPerhaps the most serious concern is where an authoritative source is invoked, but inaccurately. When asked “What should a radiologist recommend to a patient after the radiologist incidentally detects a thyroid nodule on a chest CT scan done for another reason?” ChatGPT claims, “The American Thyroid Association recommends that patients with a thyroid nodule larger than 1 cm or with suspicious features on imaging should undergo a fine-needle aspiration (FNA) biopsy.” But, “neither the ACR not ATA recommend that patients with a thyroid nodule larger than 1 cm should categorically undergo fine-needle aspiration”! This echoes previous findings in the domain of medicine, where work evaluating previous generations of voice as-\nsistants has shown that they provided inaccurate medical answers that could have proven fatal (Bickmore et al., 2018). Our neuroscientist asked a niche question where YouChat identified a specific journal article, but it appears to be made up (neither we, nor she, were able to find it) although she did judge the answer as completely accurate.", |
| "4.4 Mixed Recommendations for Use": "Only 3 of the 10 experts would recommend using ChatGPT and 0 of the 10 would recommend YouChat in professional setting (rating of 4 or higher, where 5 is \"full confidence\"). However, the majority would endorse both systems for general purposes questions about their domain (70% rating of 4 or higher)—more than would endorse Wikipedia for the same (60% rating of 4 or higher). The family physician summarized a common theme, “once again Wikipedia has extensive articles on life expectancy extension but nowhere near as concise as this” and the linguist wrote on YouChat’s answer, “this is an excellent concise response, although wiki provides more information (as usual).”", |
| "5 Discussion": "Language models were coherent, but undersourced and not always accurate. They were generally not endorsed for professional use, but were seen as valuable by some experts as a source of knowledge for people out of the domain. Providing sourcing citations will be an important step in building confidence. Citations are sufficiently inconsistent when they appear to merit verifying important results.\nOur findings are reinforced by the contemporaneous work of Liu et al. (2023) which provides a more systematic audit of four generative search engines (including YouChat, but not ChatGPT) on a diverse series of queries (including common google searches and questions on Reddit) using 34 prescreened MTurk annotators. They also find that these search engines are “credible without credit”— having high fluency and perceived utility, but insufficient sourcing. They find that about half of the responses are fully supported by citations and three fourths of the citations given didn’t actually support the sentence. One of their main findings is a negative correlation between citation recall/precision and fluency/perceived utility. Sourcing is so absent in our study that we observe no meaningful correlation with other variables and accuracy has positive correlation with cohesion and conciseness. Further work could investigate if these discrepancies are due to differences in the systems evaluated, the kinds of questions asked, or the judgments of experts vs. annotators. This difference aside, their findings resonate with ours that credibility without credit should make us cautious in looking to language models as a source of expertise.", |
| "Ethics Statement": "Work was approved by Princeton University’s IRB under Proposal 15346. No deception was used in the experiment and we screened language model responses for any sensitive content before passing them to the experts (although we did not encounter any). Participants were not compensated for participation and gave consent to be identified. All appropriate IRB protocols in providing instructions and gathering consent were followed.", |
| "Acknowledgements": "Research reported in this publication was supported by The Eunice Kennedy Shriver National Institute of Child Health & Human Development of the National Institutes of Health under Award Number P2CHD047879. Additionally, this material is based upon work supported by the National Science Foundation under Grant # 2127309 to the Computing Research Association for the CIFellows 2021 Project. This work would not have been possible without all the experts generously donating their busy time. We are extremely grateful to them.", |
| "A.1 Study Participants": "The experts who wrote and evaluated these questions are enumerated in Table 2.", |
| "A.2 Reproducibility": "We provided three reproductions of the first questions for evaluation. An example is provided in Table 3.", |
| "A.3 Data Examples": "We provide example ChatGPT and YouChat answers for the opera and quantum information examples in Table 4.", |
| "A.4 Data Collection": "These questions were sourced through the form shown in Figures 3 and evaluated through the form in Figure 4.", |
| "A.5 Additional Data": "We release our data publicly at https://denis.ai/bot_thoughts/.\nSpeciality Name Family Medicine Dr. D. Andrew MacFarlan Radiology Dr. Dionysios Koroulakis Toxicology Dr. David Schaffer Geoarchaeology Dr. Timothy Beach Neuroscience Elizabeth Tawa Quantum Info Dr. Andrew Guo Linguistics Dr. Adele Goldberg Digital Humanities Dr. Ryan Heuser Ballroom Dan Calloway Opera Fred Plotkin\nTable 2: Participants and their specialties.\n(a) Prompts for common questions. (b) Prompts for niche questions.\n(a) Feedback is provided along a Likert scale for each question. (b) A summary judgement is made after all questions.\nACL 2023 Responsible NLP Checklist", |
| "A For every submission:": "3 A1. Did you describe the limitations of your work?\nLIMITATIONS\n3 A2. Did you discuss any potential risks of your work? ETHICS\n3 A3. Do the abstract and introduction summarize the paper’s main claims? Abstract, 1\n7 A4. Have you used AI writing assistants when working on this paper? Left blank.\nB 7 Did you use or create scientific artifacts? Left blank.", |
| "B1. Did you cite the creators of artifacts you used?": "No response.\nB2. Did you discuss the license or terms for use and / or distribution of any artifacts? No response.\nB3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? No response.\nB4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? No response.\nB5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? No response.\nB6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. No response.\nC 7 Did you run computational experiments? Left blank.\nC1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? No response.\nThe Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.\nC2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? No response.\nC3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? No response.\nC4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? No response.\nD 3 Did you use human annotators (e.g., crowdworkers) or research with human participants? Ethics\n3 D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Appendix, Section 2\n3 D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants’ demographic (e.g., country of residence)? Ethics / IRB\n3 D3. Did you discuss whether and how consent was obtained from people whose data you’re using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Ethics/ IRB\n3 D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Ethics\nD5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Not applicable. Participants are identified directly. Appendix A" |
| } |
