--- title: "The AI Could Have Said Otherwise" description: "Suppose that two people are placed in a controlled environment. They are given access to the same AI model, use exactly the same computer and internet…" url: https://www.independentpress.com/article/the-ai-could-have-said-otherwise date: 2026-10-08 categories: ["Philosophy","Technology"] author: "Bailin Lu" --- # The AI Could Have Said Otherwise ![Bailin Lu AI Could have Said Otherwise](https://images.ctfassets.net/ewtdlsoyixc1/45PBZm0SvEbv2DU8x23HaV/63fb7a60f860321de35ee498739d24d2/Bailin_Lu_AI_Could_have_Said_Otherwise.jpeg) **** **The AI Could Have Said Otherwise** Suppose that two people are placed in a controlled environment. They are given access to the same AI model, use exactly the same computer and internet settings, and type in exactly the same prompt. There are no previous cookies or browsing history and neither of them has had any previous conversation with the AI model. They press enter. Will they receive the same answer? The intuitive answer may be yes. If two people type the same equation into the same calculator, we expect the same result. However, using the same controlled settings does not necessarily fix every step in an AI generation process. Large language models (LLMs) generate text in small units called tokens, which may be words or parts of words. At each step, the model assigns probabilities to possible next tokens based on the text already provided or generated. A decoding process then selects one. Some decoding methods use sampling, allowing variation among possible continuations rather than always selecting the most probable one (Holtzman et al., 2020). Consider a simple sentence input into an AI model: _“The greatest challenge posed by artificial intelligence is…”_ There are many reasonable ways to continue the sentence. The model might continue with “misinformation,” “automation,” “control,” or “bias.” That word choice becomes part of the context for the next step of the generation process. If the AI model chooses misinformation as the next word in the sentence, the generation process may develop into a discussion about democracy, while the selection of automation as the next word in the sentence may develop into a discussion on employment. A difference near the beginning of the generation process can therefore lead to substantially different outputs at the end of the generation process. The amount of variation is not completely outside human control. LLM temperature, in this context has nothing to do with how hot the computer is, it is a numerical setting that adjusts the probabilities and creativity involved in the AI model’s text generation process. Higher values make less likely tokens more available for selection, while lower values concentrate selection on more likely tokens (OpenAI, n.d.). Some systems also provide a random seed: a starting value for the process that generates the numbers used in sampling. Reusing it, by passing the exact same value into the LLM’s code every time you run a text generation process, can help reproduce a similar sequence of selections when other conditions remain fixed. OpenAI nevertheless describes outputs produced with the same seed and parameters as only “mostly” deterministic, rather than guaranteeing identical responses every time (Anadkat, 2023). Two people providing what appears to be the same input can therefore receive different outputs. This does not mean that every AI response must vary, or that the process is wholly random. It means that an ordinary input instruction need not guarantee one unique output result. However, a more interesting question follows from this conclusion: **If the input is the same, but the output varies, what exactly did the input determine?** Suppose a person asks an AI: _“Write an article explaining why social media harms democracy.”_ The first response may focus on political polarization. The second may argue that misinformation is the major problem. The third may focus on algorithms and the economics of attention. In all three cases, the user has clearly contributed something to the output. The topic was chosen by the user. The general direction of the argument was also chosen by the user, since the instruction already assumes that social media harms democracy. The form of the output was chosen as well, because the user asked for an article rather than a list, speech, or poem. But what about the supporting arguments, examples, structure, and wording? The user asked for an article against social media but did not specify political polarization as its supporting argument. Nor did they specify the structure of the second paragraph or the objections it should consider. These output features appear in the finished article randomly. The input request in the LLM leaves room for the text generation process to produce several different output versions. That is, an instruction can constrain an artifact without completely determining its properties. This distinction matters for AI-assisted authorship. I have previously argued that an authorship claim should rest on a creative intention realized through the person’s contribution to the content or expression of the work (Lu, 2025). This draws on Hilpinen’s account of artifacts, in which an artifact’s existence and some of its properties depend on its author’s productive intention (Hilpinen, 1992). Intention and contribution should not be used interchangeably here. Intention concerns what a person aims to create; contribution concerns what they actually supply or do in producing it. Intending to write an excellent article does not supply its argument. Correcting a typing error contributes to an article without necessarily being a creative contribution. The question for authorship is how the person’s creative intention is realized through what they actually do. There is an important connection here to the philosophy of causation. David Lewis approached causation through counterfactual dependence: roughly, whether an effect would still have occurred if its supposed cause had not occurred (Lewis, 1973). His later account of influence examines how alterations in a cause correspond to alterations in an effect (Lewis, 2000). This suggests a way to investigate how instructions shape AI outputs, although the application proposed here is not itself Lewis’s account of authorship. A similar method can be applied to AI generation. First, hold the submitted instruction constant, along with the model, conversation history, and relevant generation settings. Suppose the same prompt is submitted fifty times: “_Write a 1,000-word article defending universal basic income using an economic argument._” Suppose that across these fifty AI-processed text generations, universal basic income remains the topic, the articles remain supportive, and economic reasoning remains central. These are hypothetical results, not findings from an experiment conducted for this article. Their agreement with the instruction gives us reason to connect these features to the user’s contribution. Consistency alone, however, is insufficient: a feature could also recur because of the model’s usual tendencies. Meanwhile, the examples may vary between unemployment, automation, and administrative efficiency. The structure, objections, and wording may change as well. The same input instruction is compatible with these different outputs. Their variation shows what the instruction did not uniquely fix. It does not give us direct access to everything the user intended. Now consider the opposite experiment. Instead of keeping the prompt the same, the user changes one part of it. Suppose the instruction is changed from: _“Write a 1,000-word article defending universal basic income using an economic argument.”_ to: _“Write a 1,000-word article defending universal basic income using an argument about personal freedom.”_ The question is not simply whether the output changes. We already know that it can change without any alteration to the prompt. Instead, we ask whether this particular alteration reliably shifts the reasoning from economics toward personal freedom across repeated generations. If it does, that is evidence that the user’s instruction influences this output feature. It does not show that the user supplied every argument through which the new approach is expressed. If we need a term for this method, it can be called a counterfactual variation test. Hold the instruction constant and examine which features vary. Then alter one instruction and examine whether the relevant feature changes reliably in the requested direction. Taken together, these comparisons can help distinguish what the prompt specifies from what it leaves open. What must be noted is that this is a test of the prompt’s influence, not a complete test of authorship. It cannot establish that a contribution is creative merely because it affects the output. Nor can it establish that a person lacks authorship merely because the result could have been different. It may be tempting to describe this as a “quantum” characteristic of AI. But ordinary language generation does not require quantum computation. At most, the comparison draws attention to several possible outputs and one result actually received; that description also applies to ordinary processes involving chance. The philosophical issue is what follows when a person’s instruction leaves several outcomes open. It is not whether the model is somehow quantum. Consider a writer who asks an LLM for several possible endings to a short story they are writing, selects one of the outputs the LLM provides because it resolves a conflict they have developed within their writing, and modifies the ending to fit the story. The initial prompt did not determine the exact ending incorporated into the story; it only contributed to the writer’s subsequent selection and revision. Here, the writer’s subsequent actions seem to express a creative intention. These actions also seem to add to the writer’s contribution to the finished product. Human creativity does not require complete control over every detail. An unexpected result can become part of an intentional work. Conversely, merely accepting an LLM’s output does not establish that that output’s reasoning and wording were creatively shaped by the user. We need to examine the actual process, including the grounds for selection and the work of revision. This is why causing an article to be produced through an LLM input does not automatically establish authorship over every part of the output. A prompt plays a causal role, but an authorship claim still needs to explain how the person’s creative decisions shaped the work. The variation test helps us investigate the instruction’s role in producing an article; but it does not define authorship or explain a person’s creative contribution. Nor does a limit on the user’s claim automatically make the AI an author. Randomness does not establish intention. Under an account that requires creative intention, some generated features may fall outside the authorship claims of both the user and a system that lacks such intention. None of this establishes that unpredictable work is less creative, or that work produced at scale has less artistic value. Those are further questions. The narrower question here is which features a person can reasonably claim to have authored. Repeated generation lets us compare the instruction with features of the output: which features recur, which vary, and which respond reliably when an instruction changes. That makes the prompt’s influence open to examination, while leaving room to consider the person’s later selection and revision. What varies tells us what the instruction left open. What responds reliably to a changed instruction gives us evidence of the user’s influence. Neither observation settles authorship by itself. The fact that AI could have said otherwise therefore matters beyond the question of whether two responses will be identical. It helps us identify the limits of what the initial request specified. From there, we can ask how the person developed, selected, or revised the final output. If authorship rests on a relationship between creative intention and the resulting work, this is the relationship we need to explain. The question is not whether the human controlled every possibility, but how their creative decisions shaped the artifact that was ultimately produced. **** **Descriptive Credit:** Concept, central argument, philosophical analysis, drafting, writing, and final editorial decisions by Bailin Lu. **AI Usage Disclosure:** AI assisted with technical research, locating and formatting citations, and article structure.