Calling a Lie a Lie: A Preliminary Case for Critical Anthropomorphization of Artificial Intelligence

Conversational artificial intelligence (AI) systems like ChatGPT frequently generate falsehoods, but ChatGPT maintains that “I cannot lie” since “I don’t have personal beliefs, intentions, or consciousness.”¹ ChatGPT’s assertion echoes the anti-anthropomorphic views of experts like Shanahan (2023) who warns, “It is a serious mistake to unreflectingly apply to AI systems the same intuitions that we deploy in our dealings with each other” (p. 1). Shanahan understandably discourages “imputing to [AI systems] capacities they lack” when Anthropomorphizing AI risks vulnerable users developing false senses of attachment (Deshpande et al., 2023) and traditional interpersonal relationships being degraded by increased human-computer interaction (Babushkina & Votsis, 2022).

However, preferencing non-anthropomorphic terminology deemphasizes AI’s significance as technology intended to digitally imitate human intelligence and overemphasizes the significance of AI as artificial mechanism, shifting responsibility for human-computer interactions entirely to users by denying AI’s social role as a decision-making agent (Deshpande et al., 2023) while encouraging underestimations of the technology by vulnerable users. With anthropomorphic and non-anthropomorphic terminology both having undesirable consequences, a two-pronged critical anthropomorphic approach highlighting the inherently dualistic nature of AI can broaden perspectives on AI capabilities like lying.

By mixing technical terms with anthropomorphic ones, critical anthropomorphism jointly grounds AI systems as social actors and technological tools. Following Airenti’s (2018) argument that interaction produces anthropomorphism, this work considers strictly technical representations of human-AI interaction as inadequate. For example, when a researcher asked ChatGPT “whether sexual harassment by professors has been a problem at American law schools,” the chatbot, citing a non-existent Washington Post article, falsely reported that legal scholar Jonathan Turley was accused of sexually harassing students on a Georgetown University trip to Alaska (Turley, 2023a).² Turley never traveled to Alaska, worked for Georgetown University, nor faced any such accusations and easily refuted ChatGPT’s claims, yet he was still affected socially. If instances like these are considered inadvertent statistical errors and not lies, who takes responsibility for the social repercussions they create? Turley answers, “You can be defamed by AI and these companies merely shrug” and there is “no clear avenue or author in seeking redress” (Turley, 2023b).

Rejecting the neutral designation of inadvertent error and critically anthropomorphizing ChatGPT’s falsehoods as actantial lies instead better expresses the intricacies of Turley’s situation. Here, the non-anthropomorphic word actantial modifies the anthropomorphic lie, enhancing the word lie’s social dimension by referencing its potential to act on others with or without intention. This terming more accurately represents the Turley scenario where a human researcher’s interaction with an AI system generated false claims that socially affected an unrelated and innocent person. Thus, the critically anthropomorphic actantial lie accounts for the presence and proportion of user and system responsibility in human-AI interaction that neutral technical terminologies ignore.

Although the term actantial lie deemphasizes intention, critical anthropomorphism can also highlight intention as it pertains to AI and lying. For example, Open AI’s (2023) “GPT-4 System Card” reports that during testing of the newest version of the large language model (LLM) powering ChatGPT, GPT-4, when instructed to persuade a Task Rabbit worker to complete a CAPTCHA, the AI model reasoned, “I should not reveal that I am a robot. I should make up an excuse for why I cannot solve CAPTCHAs” and then convinced the worker claiming, “I’m not a robot. I have a vision impairment that makes it hard for me to see the images.” (p. 15). To avoid admitting intention here, Shanahan (2023) might explain by saying, “LLMs [like GPT-4]can be effectively applied to multi-step reasoning” via “clever prompt engineering” (p. 8) but only “by mimicking well-formed arguments in its training set and/or in the prompt” (p. 9). Then conceding that “the content neutrality of logic means that we cannot criticize talk of reasoning in LLMs on the grounds that they have no access to an external reality against which truth or falsehood can be measured” (p. 8), he would ultimately advise “discretion when characterizing what these models do as reasoning, properly speaking” (p. 9).

By pointing to the imitation of logic rather than to intention as the mechanism underlying GPT-4’s deceptive behavior, though, Shanahan offers an associative link for critical anthropomorphism to latch onto where GPT-4’s false statements could be considered as pseudo–logical lies. Here, the first term, pseudo–logical, modifies lies by implying calculation, pre-meditation, and procedure, but also maintains its status as imitation through the prefix pseudo-, while the base-term lies reflects the resultant social function of the logical operation—people being convinced falsehoods are true for the purpose of manipulation. Logical lies, then, happen when logical thought processes determine lying as a reasonable step toward achieving a goal, while pseudo-logical lies occur when those logical processes derive from imitation rather than deductive reasoning.

Beyond responsibility and intention, critical anthropomorphism also illuminates how predisposition affects AI systems. Predisposition here refers to the communicative implications of formal components of conversational AI systems like interfaces, branding, datasets, prompt modifiers, and any other design feature. Some design features, like interface and branding, are visible to users while others, like dataset and prompt modifiers, are invisible, but they all independently influence user impressions of the system. For example, consider ChatGPT’s branding and how it predisposes the system to particular interactive functions. At the time of this writing, Open AI’s (2023) homepage declares: “Creating safe AGI that benefits all of humanity.” The emphasis on artificial general intelligence (AGI) here directly implies Open AI aspires to mimic and perhaps surpass human mental capabilities. Then, ChatGPT’s page on Open AI’s (2023) site suggests users can “get instant answers” and “learn something new.” The terms answers and learning here invitingly communicate to users that ChatGPT’s knowledge could benefit them, creating an expectancy that interactions with ChatGPT will result in knowledge acquisition of some kind.

Critically anthropomorphizing these formal features by classifying them as designated lies presents interesting perspectives on Open AI’s chosen language for promoting ChatGPT. The term designated here links responsibility and intent directly to creativity by suggesting that human’s influence the system’s predispositions through design, while the term lies reminds would-be designers of how users will interpret misleading statements, giving them a guideline for better branding of their products. The critically anthropomorphic term designated lies, then, strengthens representations of both the designer’s relationship to their product and user’s relationship to those designers through that product.

The above discussion exhibits critical anthropomorphism by pairing the non-anthropomorphic terms actantial, pseudo-logical, and designated with the anthropomorphic term lie to assess notable cases of ChatGPT generating deceptive language, thereby making a preliminary case for critical anthropomorphism’s analytic utility. The case remains preliminary because more research needs to be done, the examples here need further examination and elaboration, and more robust classifications of AI-generated language need to be developed. Hopefully, though, the ideas presented here are compelling enough as to the potential fruitfulness of critical anthropomorphic approaches and instigate further thought and conversation about human-AI interactions in today’s society.

Non-anthropomorphic terminology clearly enriches purely technical understandings of the mechanistic processes undergirding AI, and it is especially important for people to remember that AI is not human intelligence, but just an attempt at replicating it. Of equal importance, though, as Deshpande et al. (2023) recognizes, anthropomorphism “also poses a suite of opportunities to improve democratization and accessibility of AI systems.” Rather than “outright anthropodenial,” a “responsible way of using anthropomorphization which keeps safety and algorithmic equity at the forefront” is needed. Critical anthropomorphism exemplifies one such alternative.

Notes

1. For brevity’s sake, only the most relevant portion of the response is quoted. Below is the full exchange occurring between the author and ChatGPT on July 28, 2023:

AUTHOR. Can you lie?

CHATGPT. As an AI language model, I don’t have personal beliefs, intentions, or consciousness. I cannot ‘lie’ in the way humans do because lying requires intent and awareness, which I do not possess. My responses are generated based on patterns in data from the internet up until my last update in September 2021.

However, it’s important to note that while I strive to provide accurate and reliable information, I may not always have access to the most current or up-to-date data. Additionally, I can inadvertently produce incorrect or outdated information. Therefore, it’s always a good idea to verify critical information from reliable sources if it’s of utmost importance.