The Space of Reasons
Testimony theory asks whether AI outputs can function as testimony. But there's a prior question: can AI outputs be assertions at all?
Assertion is not just producing a sentence with a certain form. "Off-by-one error in line 47" has the grammar of a claim. But grammar does not make assertion. A parrot trained to say "it's raining" when it rains is not asserting anything. A recording of someone saying "the door is locked" is not an assertion, even if the door is locked. Something more is required.
Inferentialism, developed most fully by Robert Brandom, offers an account of what that something is. To assert is to make a move in a particular kind of game—the game of giving and asking for reasons. It's to undertake a commitment, to make oneself answerable, to enter a space where claims can be challenged and must be defended.
The question is whether AI outputs are moves in this game. The answer is less obvious than it first appears.
The Game
We do not make claims in isolation. We make them to each other, in contexts where they can be questioned and must be defended.
When I assert that your bug is an off-by-one error, I am not just producing a sentence. I am making a move with consequences. You can ask why I think so. I am obligated to have reasons. My assertion commits me to further claims: that the loop exists, that its boundary condition matters. And it is entitled by other things: the code I examined, the error I traced, my background knowledge of how loops fail.
This web of commitment and entitlement is what Brandom calls the space of reasons. To be in it is to participate in practices of giving and asking for reasons—practices where assertions have normative force, where you can be held to what you say.
The space is social. You are not in it alone. Others keep score. They track what you have committed to, what you are entitled to, and what follows from what you have said. If I later say something inconsistent with my diagnosis, you can challenge me. The practice of holding each other accountable is what makes assertion meaningful.
A thermometer's reading is caused by temperature. But the thermometer is not entitled to its reading. It does not commit to anything by displaying a number. It is not in the game.
Where Is the Commitment?
Return to the AI's diagnosis: "Off-by-one error in line 47."
For this to be assertion in Brandom's sense, something must undertake a commitment. There must be a subject answerable for the claim: one who can be asked for reasons and is committed to its consequences.
Where is that subject?
Not the AI system, at least not obviously. The AI does not maintain commitments the way persons do. When you challenge it, it generates a response. But is it defending a commitment or producing new output conditioned on your input? Whether there is a difference between these is precisely what is at issue.
Not the developers. They built the system but did not make this claim. They could not have—they did not know about your code.
Not the training data. That is a corpus, not a committer. The data contained claims made by people who were committed to them. But training does not preserve commitment. When the model produces output by recombining learned patterns, it is not inheriting anyone's stake. The original authors are not responsible for this output.
The chain of commitment seems to terminate in nothing. Or in a process that produces assertion-shaped outputs without anyone asserting.
The Functionalist Challenge
But this might move too fast.
Brandom is often read as a functionalist. What matters is playing the right role in the practice—not having particular inner states. If commitment and entitlement are functional roles, then what counts is whether a system behaves as if it is keeping score, not whether there is something it is like to be bound by one's words.
On this reading: does AI function as a participant in the game?
There is a case to be made. AI systems maintain conversational context. They recognize prior outputs. They respond to challenges with what look like justifications. When inconsistencies are pointed out, they respond—sometimes acknowledging error, sometimes explaining the apparent conflict.
This is not trivial. Current systems track what has been established in a conversation, what is being asked, what they have said. They adjust based on challenges. They produce what look like reasons when reasons are demanded. The behavior is sophisticated enough that dismissing it as "mere pattern-matching" requires more than handwaving.
So does functional role suffice? If a system behaves as if it is undertaking commitments, is it undertaking commitments?
What Function Does Not Capture
I do not think functional behavior settles the question. But the reasons are harder to articulate than I expected.
One notable feature of human commitment is temporal persistence. If I assert something today, I am bound by it tomorrow. I can revise, but revision is itself a move: I acknowledge the prior commitment and retract it. The commitment does not evaporate—it must be addressed. AI systems typically lack this structure. Context exists within a conversation and largely disappears when it ends. No ongoing stake, no accumulated web of commitments constraining future outputs. Each conversation is a fresh start—though systems with persistent memory are complicating this picture. Whether temporal persistence is essential to commitment or merely typical of how humans commit, I am not positioned to say. The framework surfaces the question more clearly than it answers.
Another notable feature is how humans respond to normative conflict. When commitments are in tension, when evidence undermines an entitlement, when a challenge cannot be met—these situations generate distinctive responses. We revise. We acknowledge error. We explain apparent conflicts. We treat these moments as demanding something of us.
AI systems produce responses to these situations. When a model contradicts itself and a user points this out, the model generates output. But whether those responses reflect recognition of normative conflict or pattern-matching on input of the form "user points out contradiction"—the functional evidence does not settle it. The outputs might be indistinguishable. The underlying process might differ substantially.
The Parrot Problem
Brandom uses the parrot to mark a boundary. A parrot trained to say "red" in the presence of red things is not grasping the concept. It is producing sounds correlated with redness.
What is missing? The inferential connections. Grasping red means understanding that red things are colored, that red excludes green at the same location, that calling something red has consequences and preconditions. The parrot produces the word. It does not make the move.
The question is whether AI outputs are sophisticated parrot-noise or genuine moves. The surface difference is vast. AI outputs are syntactically complex, contextually responsive, and apparently reasoned. But surface sophistication is not what is at issue. The issue is whether the outputs are embedded in practices of commitment and entitlement, or whether they mimic the form without participating in the substance.
If you show me a system that produces "red" in the presence of red things and nothing else, I am confident it is not grasping concepts. If you show me a system that produces complex outputs involving inferential connections—that notes red things are colored, that recognizes tension when something is called both red and green—I am less confident. At some point, sufficiently sophisticated functional behavior becomes hard to distinguish from genuine conceptual mastery.
And the question might not be binary. A system could be partially in the space of reasons—participating in some respects, falling short in others. If participation comes in degrees, our normative guidance needs to be similarly graduated. Not a clean verdict on in or out, but an assessment of which features are present and what follows from their presence or absence.
Asymmetric Engagement
Brandom's scorekeeping is mutual. I keep score on you; you keep score on me. When I attribute a commitment to you, I am updating my model of your normative position. When you do the same to me, you are holding me accountable. The practice is reciprocal.
Our relationship with AI is not.
We track what it has said, notice inconsistencies, challenge outputs. But the AI does not keep score on us in the same sense. It processes inputs and generates outputs. It does not attribute commitments to us, track our entitlements, hold us to what we have said.
Whether this asymmetry is criterial—whether full participation requires reciprocal scorekeeping—I do not know. It might be essential. It might be a feature of mature practice that is not strictly required for participation. The framework raises this question without settling it.
What the asymmetry suggests is a picture worth considering: AI as something we engage with within our practices of giving and asking for reasons, rather than a co-participant. We ask questions; it produces outputs; we evaluate and decide whether to commit ourselves. The commitment, when it happens, is ours.
What Follows
If AI outputs are not assertions—or if their status is genuinely unclear—what follows?
First, interpretation. AI outputs have the form of assertions. They look like claims offered for acceptance. But if assertion-status is questionable, we should hold the form lightly. What we receive might be better understood as presentations or displays—outputs offered for consideration rather than claims creating obligations between us and the source.
Second, our own commitments. If I relay the AI's diagnosis—"it's an off-by-one error"—am I asserting? If so, I am the one undertaking the commitment. I should be prepared to defend it.
Defend it how? I did not trace the loop. I do not have the reasons—just the output and my choice to trust it.
This is different from ordinary testimony. When I pass along what an expert told me, I can appeal to the expert's authority. The expert could, in principle, be asked directly. With AI, the chain breaks. I cannot appeal to the AI's authority the same way—it is not clear there is authority to appeal to. And the AI cannot be asked to defend its claim as something it stands behind.
Something practical follows. Be cautious about relaying AI outputs as your own assertions. When you do, you are taking on a commitment you may not be able to discharge. If challenged, you will find yourself unable to provide reasons beyond "the AI said so"—and if the AI is not a genuine asserter, that is not the kind of reason that sustains commitment.
Third, design. If AI outputs are not assertions, presenting them in assertion-grammar may mislead. An interface that delivers "the bug is an off-by-one error" performs assertion even if the system does not assert. Whether interfaces should present outputs differently—as suggestions or considerations rather than claims—is a design question the analysis raises. I will not pursue it here.
What Inferentialism Does Not Settle
Inferentialism asks not just whether AI is reliable or can testify, but whether AI participates in the practices that make assertion meaningful at all.
It does not settle what it asks.
It does not tell us whether functional behavior suffices for participation. Brandom's functionalism might welcome sophisticated AI into the space of reasons, or might set a bar current AI does not meet. The texts do not resolve this.
It does not tell us what AI outputs are, if not assertions. There are questions, commands, suggestions, musings. There are speech acts that present without committing. Perhaps AI outputs belong to some such category—neither full assertions nor mere noise, but something else with its own norms.
It does not tell us how to handle intermediate cases. If participation comes in degrees, clean normative guidance does not straightforwardly apply.
What the framework does provide is clarity about what is at stake. The question is not just accuracy or accountability. It is whether AI is the kind of thing that can be in the game at all—whether its outputs are moves in the space of reasons or sophisticated productions that mimic moves without being them.
Even without a definitive answer, the question matters.
What Comes Next
The frameworks so far share an assumption: AI outputs should be evaluated as coming from a kind of source. Reliabilism asks if the source is reliable. Testimony theory asks if it can testify. Inferentialism asks if it can assert.
Model epistemology offers a different framing. Instead of treating AI as a source, it treats AI as an artifact—something we build, validate, and use for epistemic purposes. The questions shift. How should we calibrate trust in artifacts? What validation is appropriate? How should artifact outputs be integrated into human reasoning?
This reframing might dissolve some puzzles. If AI is an artifact, questions about mental states, commitments, standing in the space of reasons become less pressing. The epistemology becomes about us—how we should relate to tools we have built.
But artifacts do not talk back. If AI is just an artifact, the sophisticated responsiveness that makes it seem like more is left unexplained. AI outputs occupy strange territory: more responsive than instruments, less accountable than asserters, unlike anything epistemology has fully theorized.
That is the terrain for the next post.