Abstract

Large language models can generate language that is often indistinguishable from language generated by humans, but they lack human motivations, beliefs, and accountability. These systems pose risks to our information ecosystem: bias amplification, fabrication, and misinformation are some of the forecasted negative consequences of mass adoption of the technology. The recurring regulatory proposal to mitigate such risks is obligatory source disclosure. The underlying assumption is that if people know the origin of AI-generated language, they can exercise appropriate caution when engaging with it. Here we apply concepts from linguistics and cognitive science to ask what appropriate caution means when engaging with AI-generated linguistic content. We discuss an idealized model of human communication as a motivated activity aimed at increasing the mutually shared beliefs among conversation partners. Building on this model, we develop a set of conceptual tools and empirical signatures to evaluate whether humans engage with AI-generated linguistic messages in the same way as they would with a fellow human or if they approach it differently. Our preliminary empirical investigation implies that even when humans know that some language was generated by AI, they nevertheless treat it on par with human language, albeit with somewhat diminished trust. The main implication of this finding is that source disclosure is not a sufficient regulatory strategy to manage risks that will arise with the proliferation of synthetic language.

Keywords: source disclosure, human communication, linguistics, human–AI interaction

1. An Unprecedented Shift

Throughout history, whenever humans have encountered language, they have known that it was created by a fellow human being. Humans are the only species with a capacity for language,1 and both creating and understanding language has been inextricably entwined with human beliefs, desires, and accountability.2 After all, understanding a sentence requires not only an understanding of the meanings of words and how they are combined but also an understanding of what the person who utters it tries to accomplish. When someone asks, “Can you promise me?” it is not questioning one’s ability to make a promise; rather, it is a more polite way to request a commitment. Likewise, someone who answers “yes” is not just affirming their ability to make a promise but the promise itself.

The monopoly that humans had over language generation persisted despite technological developments like industrialized printing, sound recording, and widespread computing. All these technologies made it so that what was produced could be altered, edited, and copied, but they did not change the need for a human originator: one with characteristically human motivations that humans know how to reason about. Recent technological advances in generative language modeling have disrupted this status quo. Now, large language models (LLMs) can produce language that is practically indistinguishable from human language.3 This means that for the first time in our species’ history, we have to understand and reason about language that has no human originator and no human motivations or accountability.

We have yet to learn how this technology will affect society at large. What is clear, however, is that our world is increasingly being flooded with AI-generated text, and not all of it harmless. AI-generated content threatens to pollute and erode our trust in the information ecosystem with highly convincing fabrications,4 amplification of biases.5 social engineering, and election manipulation campaigns.6 Industry leaders, governing bodies, and researchers have acknowledged these risks and the urgent need for ways to differentiate between human and AI-written text. The leading proposals to this end include provenance disclosure strategies, which make the source transparent to the audience.7

Of course, enhanced transparency is necessary from an ethical point of view. But whether or not such transparency is sufficient to mitigate the risks that might emerge from the proliferation of AI-generated linguistic content depends largely on whether people can use disclosure information to the fullest degree. The relevant question is this: if people know that some piece of language was machine-generated, does their thinking and reasoning about it change? Can they leave at the door the type of reasoning about language that trades in human minds and motivations? Or is it the case that irrespective of the source of language, people cannot help but understand language as if it were generated by humans? 

1.1. Understanding Human–AI Communication: A Cognitive Science and Linguistics Approach

The assumption behind AI disclosure proposals is that if people were to know the source of AI-generated content, they would exercise appropriate caution. But it is unknown whether such strategies are on the right path to mitigating the risks of widespread AI-written text online. The effectiveness of AI disclosure depends on whether knowledge about the source of the text can help humans approach AI-generated language differently from human-generated ones. But this is an empirical question that requires a systematic investigation into how humans process such artificially created language.

To start, we also should specify what it would mean to exercise appropriate caution with AI-generated content. While humans are certainly able to be cautious about content communicated to them, the notion of caution is closely tied to understanding human minds.8 For instance, we are cautious of email scams that promise to make us incredibly rich for a small advance payment because we can reason about the motivations and pay-off structure of the person who wrote it. We are not cautious of our pocket calculator in the same way—even if it could output the wrong number—because we know that it is not motivated at all. This may seem like a trivial difference, but the latter scenario becomes especially problematic when “calculators” can output reasonable-sounding, well-formed sentences. Therefore, to explore what it means to be cautious in human–AI interaction, a new conceptual framework is needed.

This article is an attempt to draw attention to the issues and questions that arise once we approach this problem through the lens of scientific research on human language and communication, from the fields of linguistics and cognitive science. We will take a three-pronged approach: (1) first, we will  review key concepts from linguistic research on language and communication in humans; (2) then, we will discuss some questions this research poses for an empirical exploration of human–AI interactions; (3) finally, we will summarize our preliminary experimental research, which suggests that humans can only use source information in a limited way, and they reason about AI-generated text similarly to how they reason about human-generated text.

2. Human Communication as Common Ground Management

To understand the effectiveness of any intervention strategy, we need to understand the system that is being intervened on. What do humans do when they use language to communicate? Although there are many models of communication, we focus here on one with its roots in philosophy of language and linguistics.9 At its core, this model is delightfully simple. It captures human communication as the interplay between two kinds of content. The first type is simply what we know: our beliefs, our goals, what we want to accomplish from the conversation, and so forth. The second type of content is what we take our conversational partner to know: their beliefs and goals—or at least our best guesses about them. When we look at the two content types simultaneously, what emerges is a special category of beliefs—the ones held by all conversation participants—that is called the common ground.  The common ground is whatever constitutes the shared space of information among people having a conversation; it sets the background of the conversation. 

2.1. Adding to the Common Ground

The common ground might be the single most critical concept for the analysis of human communication. Equipped with this concept, we can think about any single piece of communicated language as both shaped by, and attempting to reshape, what is in this shared space of knowledge. Information exchange can be seen as a cooperative endeavor where people aim to increase the common ground. The main way of doing so is by saying something, that is, by making an assertion. An assertion can be understood as a proposal to the listener to accept or reject the information. If the listener chooses to accept the information put forth, the body of mutually shared beliefs between conversation partners increases.

To see the model at work, consider a line that Donald Trump delivered in his infamous January 6, 2021, speech10:

(1)  We must ensure that such outrageous election fraud never happens again.

What exactly does (1) assert? Or, put differently, how does the speaker of this sentence intend to influence the common ground? Evidently, the sentence conveys an urgent need to prevent future election fraud from occurring. And, upon hearing it, the listener is faced with a choice: they can either accept the information it conveys and come to share the belief that it is true—thus making it common ground—or reject it and keep it out of the common ground. This opportunity to reject the assertion is what is generally at stake in discussions of trust and caution. What it means to be cautious is to be able to resist assertions from sources that are unreliable or have dubious motivations, especially if the assertion is implausible. However, speaker reliability and belief change can dissociate. We modulate our trust in the assertion based on a whole host of situational factors, besides general unreliability of the source.11 Not believing an unreliable speaker’s assertion becomes much easier when we understand why they are trying to manipulate us—that is, when we perceive their hidden agenda for changing our beliefs. Believing an unreliable speaker is also easier when what is expressed is plausible and we do not see an ulterior motive (e.g., a notorious liar admitting to some lies). 

2.2. Taking Things for Granted

There is a second, less obvious, component to analyzing language use in this model: a specification of the common ground that the assertion intends to modify. If every piece of language aims to reshape what conversational participants take as shared beliefs, what are the beliefs that a listener has to already share with a speaker—using the technical term presuppose—so that a sentence like (1) makes sense? Here is the critical presupposition: that election fraud has already happened. Otherwise, what could the speaker mean by “never happens, again”? A pronouncement like (1) is clearly aimed at an audience who takes for granted that there has been election fraud, hence the use of a word (again) tied to this presupposition.

With presuppositions, issues of trust and caution become more complicated. In the common ground management model, presuppositions relate to the conversational background differently from assertions: whereas assertions add to the common ground, presuppositions express old information that is already part of it. The presupposed content sets the grounds for the asserted information to make sense. This aspect of presupposed content becomes clearer when you imagine someone responding to (1) with “I disagree.” Someone responding this way seems to convey that they do not think it is crucial to prevent election fraud, that they are unconcerned about election integrity. We interpret their disagreement this way because direct disagreements are generally taken to be about the asserted content of a sentence, leaving its presuppositions largely unaffected.

But what if you did not share the speaker’s presuppositions? Let us imagine three cases: Avery, Brook, and Cecil, three individuals who each have a different set of prior beliefs (see Figure 1). Avery is a MAGA Republican and strongly believes that the only reason Donald Trump could have lost is electoral fraud. For them, the sentence in (1) makes perfect sense. Avery thinks that there was election fraud, and they could agree with the assertion that election fraud should be prevented going forward. This is where the core model works most smoothly. The presupposition of (1) was already part of the shared common ground between Trump and Avery, which would then evolve to contain the newly asserted information.

Turning now to our second case study, Brook considers themselves a “centrist Democrat” and has “unwavering trust in all American institutions.” Brook finds it inconceivable that election fraud could ever happen. The issue that Brook has with (1) is not just a disagreement with its content, but something more foundational. Because the existence of election fraud is not common ground, it is not clear what Brook should even do with the assertion. Our model here predicts a communication breakdown. There might not be a successful exchange between Trump and Brook because they fail to share enough common ground. Brook might even disregard the assertion altogether once the conflicting presupposition has been noticed.

Our final case study is Cecil. Cecil is apolitical, has few relevant beliefs about the election and trusts whatever the president of the country says. Cecil is different from Avery, because they do not presuppose that there has been election fraud, but differs from Brook too, because they do not presuppose that the election was clean. What happens when Cecil hears (1)? Does smooth communication break down in this case, as it did with Brook? The answer is likely no. Though the president has presupposed something that Cecil did not previously share with him, the sentence itself clues Cecil in as to what kind of common ground Trump wants with his audience. And motivated by a desire to make sense of what the president says, they would simply accept the presupposition and add it into the common ground. That is, Cecil could come to believe that there was election fraud without much deliberation, because it was never put forth as something to deliberate in the first place; it was presupposed.

This phenomenon—whereby presupposing something as common ground can actually make it common ground—is called presupposition accommodation.12 Accommodation is one way human communication is fundamentally cooperative. Trusting the speaker not to speak falsely, and wanting to avoid conversational hiccups, a listener agnostic about some content can tacitly adjust their own beliefs to meet the speaker’s expectations about the common ground. In this way, accommodation makes human communication fluid and more efficient, though it involves more sophisticated reasoning about others’ minds.13 A speaker can just say, “My spouse works at the post office,” presupposing that they have a spouse, rather than asserting two things, “I have a spouse; she works at the post office,” anticipating that their listener is likely to adjust their own beliefs without fuss as needed to preserve smooth conversation. But accommodation is frequently used by politicians, advertisers, and others who are in the business of convincing people. Why would they not? It makes their audience believe something without having to assert it. Precisely because people are generally willing to accommodate other speakers, presupposing novel content lets one “sneak in” information that otherwise might be more debatable.

3. Managing “Common Ground” with an AI

To recap, our model conceptualizes human-to-human communication as an evolving process whereby communicators increase what is shared common ground among them. Listeners can entertain the notion that a speaker might be unreliable or even malicious, but they still consider each sentence as if it aims to change their beliefs. We discussed two properties of this system. First, even if a listener knows that the speaker might be misleading, it might be effortful for them to resist believing what was asserted, especially so if the content is otherwise plausible and the speaker’s motivation for misrepresentation is unclear. Thus, speaker reliability and belief change can dissociate. Second, the directness with which belief change is brought about can vary. What is presupposed might not get the same level of scrutiny it would were that information directly asserted. Thus, presupposing new information can lead listeners to accommodate unreliable information.

Using this conceptual framework, we can formulate key questions about human–AI communication, including the effectiveness of disclosure strategies. To start, we can ask whether humans treat AI-generated language as being fundamentally different from human-generated language in communicative settings. Because AI is not motivated the same way as human speakers are, human listeners should not be relying on the usual things such as situation-specific speaker goals and beliefs in order to decide when to trust and not to trust. Thus, reliability from belief change should not dissociate in the way we expect with humans. When told that an AI-generated sentence is unreliable, people should simply be more cautious in how that sentence influences their beliefs, and judgments of low reliability should go hand in hand with resisting the corresponding belief. Conversely, if the human mind is wired to think about any linguistic communication as human-to-human communication, with all the involuntary inferences that go along with it, reliability and belief change might still dissociate, just as with human communication.  

Second, if people are able to reason about AI-generated language differently from human-generated language, they should treat any new information in the same way, irrespective of whether it is presented as presupposed or asserted. In human-to-human communication, presuppositions can get “snuck in” because of people’s tendency to accommodate the speaker: rather than challenge presuppositions they do not share with a speaker, listeners reason about what the speaker might want the common ground to be like and try to make that a reality by tacitly adjusting their own beliefs. With AI, such accommodating behavior is unwarranted, as it is unclear what it means to reason about what common ground a machine “wants” or “intends” to share with the interlocutor. What we might find, then, is that new information is seen as new information, no matter how it is “packaged.” People should be ready to challenge presuppositions at least as much as assertions. On the other hand, if people differentiate between AI presuppositions and AI assertions in how they contribute to their beliefs, and especially if they come to believe presuppositions more than assertions, that would suggest that people tend to accommodate presuppositions without explicitly evaluating or challenging them. This would be a sign that humans interpret AI-generated language just as they would in a human-to-human communicative setting, adopting the same cooperative stance with machines as they do with a motivated human.

4. Preliminary Experimental Investigation

As a first step to investigating these issues, we carried out an experiment with 205 contributing participants.14 Their task was to read social media posts created for the experiment and then evaluate how much they believed a follow-up sentence using a slider scale from “strongly disbelieve” to “strongly believe” (see Figure 2). Our main manipulation was the source of the social media posts. For every participant, we presented some posts as created by a Canadian news organization, while others as created by an AI-news generator that “may provide unreliable information.” 

 4.1. Trust in Humans vs. AI

Our first key finding is that participants trusted the human posts more than the AI posts (Figure 3). When they saw a human post, they believed statements expressing their content (follow-up sentences conveying the same asserted or presupposed content as the post) 23 percent more than when the post was generated by AI. When it comes to the perceived reliability of the post, this difference was even more pronounced. On average, participants found human posts to be 33 percent more reliable than AI ones. This finding shows that people can limit their trust in AI-generated content to some degree, at least when it is explicitly marked as potentially misleading. While this is promising, it is important to note that ours was a highly artificial experimental context that might be capturing an upper bound for this effect. Our participants likely understood the experiment to be about their evaluation of AI-generated information and tried to modulate their responses based on source to a greater degree than they might in more naturalistic contexts.

4.2. Reliability and Belief Change

The second finding concerns the relationship between belief change and reliability. As just discussed, we found that when the content was generated by AI, participants did not perceive it as reliable. We also found that these low-reliability judgments were not fully reflected in participants’ evaluation of the information contained in AI-generated posts. When the follow-up sentence conveyed information that was asserted or presupposed in an AI-generated post, participants rated it 12 percent more believable than when the follow-up stated, “The above post is reliable.” Put differently, when explicitly asked, participants report distrust in the AI-generated content, but they still allow that content to influence their beliefs. This, of course, makes sense in human-to-human communication where evaluating how to communicate with a potentially unreliable person is generally modulated by reasoning about what they are trying to achieve in the conversation and why. Human lies are motivated and figuring out these motivations is essential for separating truth from falsehoods. But this modulation does not make sense when it comes to AI-to-human communication, where an AI does not have the same human motivations. Without a solid understanding of why LLMs generate false (or true) sentences, we just do not know when to expect them.

4.3. Assertions and Presuppositions

Thus far, when we discussed participants' beliefs in the content of a post, we did not distinguish based on whether that content reflected its presupposition or assertion. When we examine how participants treated each type of content, we in fact observe a statistically significant difference: participants tended to believe the presupposition of AI posts more than they believed their assertions. While the size of this effect is small (a 3.2 percent difference), it still indicates that participants reason about presuppositions and assertions of AI sentences differently, just as they do in human-to-human communication. This makes sense if people, upon reading an AI-generated sentence, opt for implicitly accommodating its presuppositions, differently—and perhaps less cautiously—from how they deal with assertions.

 5. Discussion

Our preliminary experimental investigation yields two key findings. First, in an idealized experimental context, humans can indeed modulate their trust when language is marked as AI-generated. This means that if source disclosure is salient enough and clearly highlights reliability issues, then it might help mitigate some AI risks, particularly ones that would arise from overestimating the reliability of information due to not knowing its origin. It is an open question what factors affect the size of this modulation effect, and how it translates to more naturalistic scenarios.

         Our second key finding was that participants displayed clear signatures of human communication even when engaging with AI content. They changed their beliefs in response to sentences to a higher degree than they reported trust in them. This finding implies that a general mistrust is insufficient to counteract information uptake from AI sources, just as it would be in human-to-human communication.15 In line with this finding, participants did not treat asserted and presupposed information on par. To a large degree, they were more willing to accommodate the presuppositions of machines—even though these arguably lack the kind of beliefs and goals that people have when exchanging new information. This suggests that a critical part of interpreting machine-generated language involves an implicit assumption that it was produced by something with a humanlike mind and motivation structure. Thus, if the underlying policy goal is to help people treat the AI-generated linguistic content as qualitatively different from human language, source disclosure seems insufficient.

         This insufficiency reveals the underlying tension between the engineering goal of LLMs and the societal goal of people understanding LLM content as machine rather than human generated. The ease of usage of LLMs is derived from their ability to mimic human-to-human communication.16 It is precisely because LLMs pick up on human conversational patterns that they are so easy to use. Nevertheless, these models’ ability for fluid mimicry is outpacing their ability to be reliable communicative partners for humans.17 This is especially problematic because human communication has blind spots that can be exploited in AI–human interactions. The relative ease of accepting information compared to suspending belief, the willingness to accommodate presuppositions without much explicit deliberation, all persist even when interacting with text generated by machines.

         It is important to note that our empirical research is in multiple respects preliminary. First, there are a host of other linguistic phenomena to be explored in trying to pin down how humans understand machine-generated language. Finding differences among them might reveal when and how humans could use disclosure information better. Similarly, differences in the format in which AI content is presented, for example, news pieces vs. back-and-forth conversation, may have distinct consequences for uptake of new information. Second, our experiment paints with a wide brush, and assumes that humans understand machine-generated text uniformly. Of course, there might be a variety of possible ways in which individual differences could influence how people think about machine-generated text. Two critical ones are demographic factors that might result in unequal burden when it comes to bearing the negative consequences of AI-generated text proliferation, and factors that relate to familiarity with these systems. All these factors undoubtedly need to be further explored.

In terms of solutions to protect our information ecosystems from AI, we find that source disclosure might be necessary but insufficient. While not ineffective altogether, it fails as a method to guard against the dissemination of harmful information, because humans approach communication with machines with the same cooperative and accommodating stance as they do human communication. It is an open question whether more widespread understanding of scientific models of communication, as well as increased AI literacy, provides people with tools to identify the ways synthetic text can be used to manipulate their beliefs. As for policy, our preliminary conclusion is that mitigating AI risks might require more robust and transformative regulations than just disclosure.

We are moving to an era where machines have the ability to generate fluid, humanlike language, but it is not clear whether their creators are ready for this change. Here we made some steps to explore humans’ cognitive preparedness, starting from a model of human-to-human communication. Can we, at least under optimal circumstances, treat machine-generated text differently from human-generated language? Are we capable of conceiving of language differently when we know that we should? Our initial report is mixed but not encouraging. What is certain is that more research is needed about the humans that will communicate with machines. After all, making the best of this new era will require not only understanding the machines that brought it about, but also the humans that live in it.

Data/Code Availability

https://osf.io/rwj87/

Bibliography

 Abid, Abubakar, Maheen Farooqi, and James Zou. "Persistent Anti-Muslim Bias in Large Language Models." Paper presented at the 2021 AAAI/ACM Conference on AI, Ethics, and Society, Virtual Event, May 19–21, 2021.

Aravind, Athulya, Danny Fox, and Martin Hackl. "Principles of Presupposition in Development." Linguistics and Philosophy 46, no. 2 (2023): 291–332.

Bang, Yejin, Samuel Cahyawijaya, Nayeon Lee, Wenliang Dai, Dan Su, Bryan Wilie, Holy Lovenia et al. "A Multitask, Multilingual, Multimodal Evaluation of ChatGPT on Reasoning, Hallucination, and Interactivity." Preprint submitted February 8, 2023. https://arxiv.org/abs/2302.04023.

Beaver, David, and Henk Zeevat. "Accommodation." In Oxford Handbook of Linguistic Interfaces, edited by Gillian Ramchand and Charles Reiss, 503–38. Oxford: Oxford University Press, 2007.

Oxford Handbook of Linguistic Interfaces. Oxford: Oxford University Press, 2007.

Bolukbasi, Tolga, Kai-Wei Chang, James Y. Zou, Venkatesh Saligrama, and Adam T Kalai. "Man Is to Computer Programmer as Woman Is to Homemaker? Debiasing Word Embeddings." In Advances in Neural Information Processing Systems 29, edited by D. Lee et al., 4349–57. San Diego, CA: Neural Information Processing Systems Foundation, 2016.

Clark, Herbert H., and Susan E Brennan. "Grounding in Communication."  In Perspectives on Socially Shared Cognition, edited by L. B. Resnick, J. M. Levine, and S. D. Teasley, 127–49. Washington, DC: American Psychological Association, 1991.

Dash, Debadutta, Rahul Thapa, Juan M. Banda, Akshay Swaminathan, Morgan Cheatham, Mehr Kashyap, Nikesh Kotecha et al. "Evaluation of GPT-3.5 and GPT-4 for Supporting Real-World Information Needs in Healthcare Delivery." Preprint submitted April 26, 2023, https://arxiv.org/abs/2304.1371.

Dugan, Liam, Daphne Ippolito, Arun Kirubarajan, Sherry Shi, and Chris Callison-Burch. "Real or Fake Text?: Investigating Human Ability to Detect Boundaries between Human-Written and Machine-Generated Text." Proceedings of the AAAI Conference on Artificial Intelligence 37, no. 11 (2023): 12763–71. https://doi.org/10.1609/aaai.v37i11.26501.

Gilbert, Daniel T. "How Mental Systems Believe." American Psychologist 46, no. 2 (1991): 107–19.

Grice, H. Paul. "Meaning." The Philosophical Review 66, no. 3 (1957): 377–88.

Grice, Herbert P. "Logic and Conversation." In Speech Acts, 41–58. Leiden: Brill, 1975.

Gupta, Maanak, CharanKumar Akiri, Kshitiz Aryal, Eli Parker, and Lopamudra Praharaj. "From ChatGPT to ThreatGPT: Impact of Generative AI in Cybersecurity and Privacy." IEEE Access 11 (2023): 80218–45.

Hauser, Marc D., Noam Chomsky, and W. Tecumseh Fitch. "The Faculty of Language: What Is It, Who Has It, and How Did It Evolve?" Science 298, no. 5598 (2002): 1569–79.

Heim, Irene Roswitha. The Semantics of Definite and Indefinite Noun Phrases. Amherst: University of Massachusetts Amherst, 1982.

Karttunen, Lauri. "Presupposition and Linguistic Context." Theoretical Linguistics 1, nos. 1–3 (1974): 181–94.

Kotek, Hadas, Rikker Dockum, and David Sun. "Gender Bias and Stereotypes in Large Language Models." Paper presented at the ACM Collective Intelligence Conference, Delft, Netherlands, November 6–9, 2023.

Madiega, Tambiama. Artificial Intelligence Act. European Parliament: European Parliamentary Research Service (2021).

McGowan, Alessia, Yunlai Gui, Matthew Dobbs, Sophia Shuster, Matthew Cotter, Alexandria Selloni, Marianne Goodman et al. "ChatGPT and Bard Exhibit Spontaneous Citation Fabrication during Psychiatry Literature Search." Psychiatry Research 326 (2023): 115334.

Mercier, Hugo. Not Born Yesterday: The Science of Who We Trust and What We Believe. Princeton, NJ: Princeton University Press, 2020.

Nikolinakos, Nikos Th. "A European Approach to Excellence and Trust: The 2020 White Paper on Artificial Intelligence." In Eu Policy and Legal Framework for Artificial Intelligence, Robotics and Related Technologies-the AI Act, 211–80. New York: Springer, 2023.

Sadasivan, Vinu Sankar, Aounon Kumar, Sriram Balasubramanian, Wenxiao Wang, and Soheil Feizi. "Can AI-Generated Text Be Reliably Detected?" Preprint submitted March 17, 2023. https://arxiv.org/abs/2303.11156.

Salvagno, Michele, Fabio Silvio Taccone, and Alberto Giovanni Gerli. "Can Artificial Intelligence Help for Scientific Writing?" Critical Care 27, no. 1 (2023): 1–5.

Schulman, John, Barret Zoph, Christina Kim, Jacob Hilton, Jacob Menick, Jiayi Weng, Juan Felipe Ceron Uribe et al. "ChatGPT: Optimizing Language Models for Dialogue." OpenAI (blog)  (2022).

Shah, Chirag, and Emily Bender. "Envisioning Information Access Systems: What Makes for Good Tools and a Healthy Web." Unpublished manuscript, September 1, 2023. https://faculty.washington.edu/ebender/papers/Envisioning_IAS_preprint.pdf.

Shen, Yiqiu, Laura Heacock, Jonathan Elias, Keith D. Hentel, Beatriu Reig, George Shih, and Linda Moy. "ChatGPT and Other Large Language Models Are Double-Edged Swords." Radiology 307, no. 2 (2023): e230163.

Simchon, Almog, Adam Sutton, Matthew Edwards, and Stephan Lewandowsky. "Online Reading Habits Can Reveal Personality Traits: Towards Detecting Psychological Microtargeting." PNAS Nexus 2, no. 6 (2023): pgad191.

Sperber, Dan, Fabrice Clément, Christophe Heintz, Olivier Mascaro, Hugo Mercier, Gloria Origgi, and Deirdre Wilson. "Epistemic Vigilance." Mind & Language 25, no. 4 (2010): 359–93.

Stalnaker, Robert. "Assertion." In Pragmatics, 315–32. Leiden: Brill, 1978.

———. "Common Ground." Linguistics and Philosophy 25, no. 5/6 (2002): 701–21.

———. Context. Oxford: Oxford University Press, 2014.

Von Fintel, Kai. "What Is Presupposition Accommodation, Again?" Philosophical Perspectives 22 (2008): 137–70.