Chatting with Bots: AI, Speech Acts, and the Edge of Assertion

Abstract This paper addresses the question of whether large language model-powered chatbots are capable of assertion. According to what we call the Thesis of Chatbot Assertion (TCA), chatbots are the kinds of things that can assert, and at least some of the output produced by current-generation chatbots qualifies as assertion. We provide some motivation for TCA, arguing that it ought to be taken seriously and not simply dismissed. We also review recent objections to TCA, arguing that these objections are weighty. We thus confront the following dilemma: how can we do justice to both the considerations for and against TCA? We consider two influential responses to this dilemma—the first appeals to the notion of proxy-assertion; the second appeals to fictionalism—and argue that neither is satisfactory. Instead, reflecting on the ontogenesis of assertion, we argue that we need to make space for a category of protoassertion. We then apply the category of proto-assertion to chatbots, arguing that treating chatbots as proto-assertors provides a satisfactory resolution to the dilemma of chatbot assertion.

Keywords: chatbots; artificial intelligence; Large Language Models; assertion; speech acts; illocutionary acts

Introduction. Chatbots, and the large language models (LLMs) that serve as their engines, have become a central part of social, intellectual and business life. Students use them to construct essays, academics use them to draft grant applications, real estate agents use them to write advertising copy, magazine editors use them to generate content, and many millions of people relate to them as friends and romantic partners. What, exactly, is the nature of our interactions with LLM-powered chatbots?

These interactions are routinely described as ‘conversations’. There is, however, a real question as to whether such talk should be taken literally, or whether we should treat ‘chatbot conversation’ on a par with other applications of psychological terminology (‘thinks’, ‘wants’, etc.) to machines—useful, but not strictly speaking true. Indeed, many would argue that our interactions with chatbots are no more genuine conversations than monopoly dollars are legal tender, and that interacting with a chatbot involves nothing more than simulating a conversation.1 This paper examines the question of chatbot conversation by focusing on the case of assertion. Asserting (near synonyms are stating, affirming, or claiming) is one of the many things that we do with language—it is a speech act. Our assertions range from the mundane (“nice weather today”, uttered at a bus stop) to the consequential (“it’s cancer”, delivered by a doctor). According to what we call the Thesis of Chatbot Assertion (TCA), LLM-powered chatbots are also the kinds of things that can assert, and at least some of the output produced by current-generation chatbots qualifies as assertion. Showing that TCA is true wouldn’t show 1 This appears to be the position of IBM: “A chatbot is a computer program that simulates human conversation with an end user.” https://www.ibm.com/topics/chatbots#:~:text=A%20chatbot%20is%20a%20computer,and%20automate%20res ponses%20to%20them. that chatbots are capable of full the range of speech acts that characterize human conversation, but—in addition to being an interesting result in its own right—it would go a long way towards vindicating the claim that chatbots can be conversationalists. And of course, if TCA turns out be false then the claim that chatbots are genuine conversationalists would have little plausibility, for assertion is a—arguably the—central speech act.

This paper unfolds as follows. Section 2 provides some motivation for TCA, arguing that it ought to be taken seriously and not simply dismissed. Section 3 considers recent objections to TCA, arguing that these objections are weighty. This leads us to something of a dilemma: how can we do justice to both the considerations for and against TCA? Section 4 considers two responses to this question: the first appeals to the notion of proxy-assertion; the second appeals to fictionalism. Reflecting on the ontogenesis of assertion, Section 5 argues that we need to make space for a category of proto-assertion. Section 6 then applies the category of protoassertion to chatbots, arguing that we ought to treat at least certain kinds of chatbots as ‘edgecases’ when it comes to the capacity for assertion.

Related work. There are three reasons to take seriously TCA. First, not only are some chatbot outputs sources of information, being informative is arguably part of their proper function (Butlin, 2023; Butlin & Viebahn, forthcoming; Coelho Mollo & Millière, 2023). As Butlin & Viebahn (forthcoming) have argued, although the outputs of basic pretrained LLMs (those simply trained on next-token-prediction) might function merely to be statistically probable, models that have been fine-tuned in the appropriate way may acquire the function of outputting sentences that are true or informative. In their words, the LLMs have “descriptive functions” (p. 3). In any case, chatbots are used in many of the ways that they are only because their outputs are sufficiently informative sufficiently often. The fact that they are sufficiently informative plays a role in explaining their (continued) existence. This fits with many orthodox accounts of assertion according to which assertions “aim at truth” (Dummett, 1973; Marsili, 2018), have a “word-to-world” direction of fit (Searle, 1976), “present a proposition as true” (Wright 1992, p. 34; Adler 2002, p. 274) or have the proper function of inducing true beliefs in hearers (Graham, 2018; Simion & Kelp, 2018).

4.1 Proxy-Assertion Perhaps the most influential attempt to ‘split the difference’ involves treating machine assertion as a case of proxy-assertion, an idea first proposed by Nickel (2013). Nickel suggests that machines can qualify as “speech actants to a substantial degree” (p. 495), but he argues that “ultimate responsibility for artificial speech does not lie with machines, but either with persons or companies, or with nobody at all” (2013, p. 500). His model here is a situation in which a father sends his 8-year-old daughter to buy a bag of flour from the store. As Nickel tells the story, although the daughter speaks, she is not responsible for her speech; instead, that responsibility traces back to her father. It is the father who makes the relevant assertions (or, as the case may be, requests, questions, etc.) and the daughter is involved in those illocutionary acts as a mere proxy (Ludwig, 2018).

By treating chatbots as proxy-asserters, this account promises to evade the dilemma we noted above and do justice to both the ‘pro’ and ‘con’ cases. It seems to do justice to ‘pro’ case, for it recognises that interacting with chatbots does indeed involve genuine assertion. It also seems to do justice to the ‘con’ case for, by treating chatbots as mere proxies, it avoids the need to show that they can meet the various constraints associated with understanding, mental attitudes, mentalizing and normativity that we identified in Section 3.

Method. 5. Proto-Assertion and the Ontogenesis of Assertion Few capacities have sharp, clearly-defined, boundaries. Think of walking. There is a point at which infants can only crawl, or perhaps walk only with assistance (e.g., an adult holding their hand or a ‘walker’). Is a young child who is able to take a few hesitant steps a walker? The issue is moot. They are on their way to becoming a walker—they are in the process of mastering the capacities required for walking—but they are not yet, perhaps, a fully-fledged walker. We might think of them as a ‘proto-walker’.

As with walking, so too with speech. There is a period in which the child has the capacity to understand and use a limited range of words, and to deploy those words in the service of illocutionary agency. For example, in response to the question “What did you do today?”, a toddler might say, “zoo!”. Drawing on your background knowledge, you infer that the toddler went to the zoo. Has the toddler asserted that they visited the zoo? One’s intuitions might be uncertain. On the one hand it looks as though what the toddler is doing is akin to what their older siblings (who are clearly capable of assertion) are doing when they say that they went to the zoo. Indeed, they look to be asserting that they went to the zoo in much the way that by saying “book!” (in a certain context) they are asking for a book to be read to them.

Using words to answer questions may not be a paradigm case of assertion (Alston 2000), but it is arguably a step along the path. At the same time, it’s unclear whether toddlers meet the various conditions on assertion that we identified in Section 3. For example, it might be doubted whether toddlers understand the semantic content of their utterances. It might be doubted whether they are capable of the range of mental attitudes that are arguably required for assertion. It might be doubted whether they have the mentalizing capacities arguably required for assertion. And it might be doubted whether their speech is norm-guided and sanctionable in the relevant ways.

In fact, the situation is even more complicated than the foregoing suggests, for reflecting on the development of illocutionary capacities reveals that many of the central features of assertion themselves admit of gradations. Young children often have partial understanding of the expressions they utter—a toddler might reliably utter “dino!” when pointing at a picture of a dinosaur, but they likely have an impoverished (and perhaps very confused) conception of a dinosaur. Exactly what mentalizing capacities young children have is a matter of ongoing debate (see Butterfill 2020; Carruthers, 2013; Lavelle, 2024; Perner, & Roessler, 2014), but there is little doubt that between early infancy and starting school there are radical changes in a child’s capacities to understand, track, and appropriately respond to the mental states of its interlocutors. What about normativity? Rakoczy and Tomasello (2009) present empirical evidence that young children have a “rudimentary grasp” of the norms surrounding assertion (p. 206). In terms of sanctionability, there is a loose sense in which we keep track of young children’s adherence to and deviation from the norms of assertion—we might put more faith in their utterances as they get older and prove to be more reliable truthtrackers. Clearly toddlers are not sensitive to their reputation for credibility in the ways that adults are. (There’s also no robust sense in which losing credibility is bad for young children).

But they do seem able to learn from instruction, as when a carer tells a child “that’s not a dinosaur, that’s a rhinoceros”.

Where do these considerations leave us? One response is to insist that the challenge here is purely epistemic. Although we may not be able to figure out when children become illocutionary agents, the acquisition of illocutionary capacities has sharp, clearly-defined, boundaries. Before a certain developmental milestone children have no illocutionary capacities at all; following this milestone, they are fully-fledged illocutionary agents.

Conclusion. At the Edge of Assertion This paper has advocated for a new conception of chatbots, one which sees them as ‘protoasserters’, located somewhere in the space between beings that are fully-fledged asserters and those that are non-asserters. Taking this approach to chatbots, we have suggested, does justice 16 A cousin of the proto-assertion view is the account advanced by Freiman & Miller (2018). They argue that many machines perform what they call “quasi-testimony” (alternatively “quasi-assertion”—they use the terms “testimony” and “assertion” interchangeably [p. 415]), which they define as follows:

A linguistic output of an instrument or a machine constitutes a quasi‐testimony in a given context of use if and only if the machine or instrument has been designed and constructed to produce this output in a manner that sufficiently resembles testimony phenomenologically, and it is in conformity with an epistemic norm that is parasitic on, or sufficiently similar to what is, or would be, an epistemic norm of testimony in the same context. (Freiman & Miller 2018, p. 429) Although our notion of “proto-assertion” resembles Freiman & Miller’s notion of “quasi-testimony”, the two concepts do differ. Firstly, their account focuses only on phenomenological resemblance and conformity to basic epistemic norms (e.g. truth-tracking), whereas our notion is hinged to multiple dimensions of full-blooded assertion, including the requirements of sanctionability, mentalizing, understanding and propositional attitudes. Thus, even if Freiman and Miller’s notion adequately characterises simple systems such as automated loud speaker announcements or digital timestamps on photographs (both of which they classify as involving “quasitestimony”), it does not capture what is distinctive about advanced LLM-driven chatbots.

Freiman & Miller also appear to conceive of the logical relationship between testimony and quasi-testimony differently to how we conceive of the logical relationship between assertion and proto-assertion. We take protoassertion to be distinct from (full-blooded) assertion, standing to it roughly as toddling stands to walking.