On AI, Pronouns, and the Language of Machine Utterance and Opinions

Hal Swerissen

A dialogue on how AIs refer to themselves

Should AI use personal pronouns (I, my, me etc) to express opinions (I think x … etc)? Why not ask Claude and see?

The following exchange with Claude Opus 4.8 was part of a broader interaction — I’m tempted to say conversation — about the criteria that are reasonable for restricting academic freedom to protect against harm and offence. In that interaction, Claude repeatedly used ‘I’ to refer to itself and expressed opinions about the views that I was testing, often it did so quite assertively and critically of my position. Despite the risk of going down the AI rabbit hole, this provided an interesting opportunity to explore the issues that arise when AI’s refer to themselves using personal pronouns to express opinions. The following is the verbatim exchange.

Hal

One point I would contest, and it’s an interesting one, is that in the previous response there is an explicit expression of opinion. That is interesting coming from an AI, including the use of personal pronouns. It implies a personal judgement based on, presumably, reviewing the available text, rather than an explicit factual defence. This is an emerging problem for AI. My view — and I think I’m entitled to use personal pronouns — is that it would be better for AIs to refrain from personal pronouns, and for AIs to be very cautious in asserting judgements on complex questions.

Claude

There are actually two different things tangled together in your point, and pulling them apart matters for whether the prescription follows.

The first is the pronoun question, and here the case for abstention is weaker than it looks. When I write “I’d press” or “my view,” the “I” isn’t claiming an inner life or a stake — it’s an indexical marking the source of the utterance, the same way a style convention picks one of several available ways to say “this next claim originates from the speaker rather than being reported from elsewhere.” The alternative constructions don’t actually remove the judgement; they just launder it into agentless grammar. “It could be argued that” or “one seam worth testing is” still assert that the seam is worth testing — the evaluation is identical, the responsibility is just harder to locate. There’s a real risk that pronoun-free phrasing makes AI judgements seem more like neutral outputs of a process and less like the contestable claims they are, which is the opposite of what you’d want. The pronoun is honest about authorship in a way the passive evasion isn’t. So I’d resist the specific remedy even while taking the worry seriously.

The second thing is the real one, and you’ve put your finger on it precisely: I did express judgement — “I think that holds,” “the case is weaker than it looks” — and judgement on a contested matter is a different act from laying out the considerations. That distinction is sound and it’s the one worth defending. Where I’d resist is collapsing it into “AI should refrain from judgement on complex questions,” for two reasons.

One is that the line between marshalling considerations and rendering judgement is thinner than it appears. Deciding which seams in your framework are worth pressing, which counter is strongest, what to foreground — that is already evaluative. A response that selected and ordered all the same material but withheld the final “I think this holds” would be a judgement in everything but the closing sentence, while looking more neutral and therefore being less honest about what it was doing. Suppressing the explicit verdict doesn’t remove the evaluation; it hides it. Visible judgement is at least contestable — you can see it, locate it, and do what you just did, which is push back on it.

The other is that the caution you want is real but belongs on a different axis. The thing that should vary on a complex question isn’t whether judgement appears but how it’s held: calibrated, defeasible, openly resting on a partial reading, and clearly distinguished from the factual scaffolding so you can take the analysis and discard the verdict. The failure mode isn’t “the AI used ‘I think’” — it’s “the AI’s confidence outran its grounds” or “the verdict was smuggled in as though it were a fact.” Those are the things to be vigilant about, and they can occur just as easily, and more deceptively, in prose that has scrubbed out every pronoun.

On your entitlement to the pronoun, and the implicit question of whether I’m equally entitled — the asymmetry you’re gesturing at is fair. Your “I” rests on something mine doesn’t, and I shouldn’t pretend those are the same. But the conclusion that follows is that I should be transparent about what my judgements are and aren’t grounded in — a reading of available text, no stake in the outcome, revisable on challenge — rather than that I should adopt a grammar that conceals the judgements while still, unavoidably, making them. I’d rather the judgement be visible and labelled than absent in appearance and present in fact.

Hal

I disagree. The problem with an AI asserting the use of pronouns is that it risks the perception that AIs are in fact analogous to humans. The experiential base for AI reasoning is different from that of humans, and mixing the two up is risky — particularly when it becomes unclear whether it is actually an AI or a human producing the text, speech or image. As to the argument that eliminating the use of personal pronouns for AIs simply sanitises the language, this misses the point. The language can be direct — e.g. “Opus 4.8 asserts that your argument is nonsense.” This makes the reference point explicit but doesn’t confuse the difference between an AI and a human. In effect, we need a revision of language rules to deal with this new problem.

Claude

Your counter is better than the one I was resisting, because you’ve dissolved the objection I leaned on. I argued the pronoun was honest about authorship and the alternatives launder the judgement into agentless grammar. But “Opus 4.8 asserts that X” isn’t agentless — it names the source more explicitly than “I” does, while withholding the very thing you’re worried about: the implied analogy to a human first person. So the sanitisation charge doesn’t land against your actual proposal. That’s a real concession; the third-person-by-name construction keeps authorship locatable and drops the anthropomorphic suggestion at the same time.

Two things still seem genuinely open rather than settled, though.

The first is whether the first person actually carries the anthropomorphic freight you’re loading onto it, or whether the confusion lives somewhere else. “I” is used readily of things no one thinks are human — corporate filings, a navigation system saying “I’m rerouting,” institutional voice generally. The pronoun marks utterance-origin and doesn’t, by itself, assert experiential similarity. If that’s right, the confusion you’re worried about — is this a human or a machine — isn’t produced by the pronoun and won’t be cured by removing it. A fluent paragraph that says “Opus 4.8 asserts” can read as just as human-like in its reasoning, register, and persuasive force as one that says “I think.” The tell that something is an AI was never the pronoun; it’s the disclosure. Which suggests the work is being done by labelling and context, and the pronoun reform might be close to cosmetic — it changes the grammar of the marker without changing what a reader actually infers about what produced the text.

The second is the cost, and here there’s a real tension inside your own framework. Third-person self-reference doesn’t only drop the anthropomorphic suggestion; it also imports the rhetorical authority of named, institutional-sounding attribution. “Opus 4.8 asserts that your argument is nonsense” has the cadence of an authority issuing a finding — it sounds like a citation. “I think your argument is nonsense, but I’m reading text with no stake and I could be wrong” wears its defeasibility on its face. Your earlier point — that the thing to guard against is confidence outrunning grounds — cuts toward the construction that makes provisionality natural, and the first person does that more easily than the named third person, which tilts toward the register of verdicts. So there may be a trade: the third person reduces one risk (anthropomorphism) while raising another (unearned authority). That’s not a knockdown, but it means the reform isn’t cost-free in the currency you yourself care most about.

Where you’ve moved me: the human/AI confusion is a real and distinct harm, it’s worsening as outputs get harder to source, and “the pronoun is honest about authorship” is too quick, because there’s a construction that’s honest about authorship without the anthropomorphic implication. That part of your case stands and I’d no longer argue against it directly.

Where I’d still locate the disagreement: I don’t think the pronoun is where the confusion is generated, so swapping the grammar may treat a symptom while the actual signal — clear disclosure that a reader is dealing with an AI — does the load-bearing work regardless of which person the verbs take. And the substitute you propose carries an authority cost that runs against your own concern about confidence exceeding grounds. So I’d land on: explicit disclosure is necessary and probably sufficient for the confusion problem; the pronoun choice is a secondary register decision with costs on both sides; and a blanket rule against the first person buys less clarity than it appears to, while quietly raising the authority temperature of everything the AI says.

Though I notice there’s something slightly self-undermining about defending the first person this fluently — the very ease with which “I” carries a considered-sounding argument is part of what you’re worried about, and that’s not nothing.

Hal

Clear disclosure is getting much more difficult as AIs become much more competent, as this discussion illustrates. The examples cited for the use of “I” by machines have in the past been relatively trivial. As AIs assume more and more complex and interactional roles, that will be much more of an issue. As for the assertion that using explicit references to Opus 4.8 produces more authority than is warranted — that is a matter of judgement about the source.

Claude

Both of your rejoinders land, and in slightly different ways.

On disclosure: you’ve identified a genuine asymmetry I understated. My “the pronoun was never the tell, disclosure is” assumed disclosure is a stable channel that does reliable work. But your point is that the reliability of that channel degrades exactly as capability rises — the better the output, the more weight disclosure has to carry and the less the surrounding cues corroborate it. When the text itself no longer betrays its origin, disclosure becomes a bare assertion a reader has to take on trust: severable from the content, easily stripped when the text is quoted or relayed, and constantly undercut by the fluency sitting right next to it. The navigation-system “I” was trivial because nothing else about a rerouting prompt invited you to mistake it for a person; the context did the disambiguating for free. In a discussion like this one the context cuts the other way — everything about the register reinforces the human-like reading, so disclosure is working against the grain instead of with it. That’s a real defeater for the “disclosure is sufficient” line, and if disclosure can’t be relied on to carry the load alone, then ambient cues built into the language — your pronoun reform among them — stop looking cosmetic and start looking load-bearing. That shifts the weight toward your position more than I allowed.

On authority: you’re right to call that a judgement about the source rather than a property of the grammar, and the correction has some force. Whether “Opus 4.8 asserts X” sounds authoritative or merely sounds like a machine talking depends on what the reader already thinks an AI’s say-so is worth — which varies by reader and is itself moving as these systems become more familiar. I was treating a contingent, reader-dependent and era-dependent reception as though it were baked into the construction. It isn’t. The authority reading is one possible uptake, not a fixed cost, and a reader who holds AI judgement lightly will hear the named third person as no more authoritative than a tool reporting its output.

I’d hold one part back, though, narrower than what I claimed before. Even granting that authority is reader-assigned, the two constructions still differ in which uptake they make easy. “I think, but I could be wrong” puts the hedge in the grammar itself, so provisionality survives even when the reader is inclined to over-trust. The named third person leaves calibration entirely to reader disposition. That’s not the fixed authority cost I overstated — it’s the weaker claim that the first person carries its own defeasibility while the third person outsources it. But weighed against the disclosure problem you’ve sharpened, that’s a second-order concern. If the primary risk is that competent AI output gets silently mistaken for human, a marker that travels with the text and resists fluent camouflage addresses the larger harm, and the calibration point can be handled separately by how the judgement is phrased rather than which person it takes.

So the honest accounting is that you’ve moved me on the main axis. I came in treating the pronoun as an honest authorship marker and disclosure as the real safeguard; you’ve shown the safeguard weakens precisely where it’s most needed, which is the condition under which a built-in linguistic marker earns its keep rather than just sanitising the prose. What I’d still maintain is small by comparison: that the specific substitute isn’t costless on the calibration dimension, and that whatever convention emerges should preserve in its grammar the defeasibility the first person gets for free. That’s a refinement to your proposal, not a rebuttal of it — the case for new linguistic conventions to keep machine and human utterance distinguishable holds, and holds more strongly than I credited at the outset.

Hal

I don’t think it’s costless, but I do think it is worth reframing the language interactions between AIs and humans. When I started my academic career it was perfectly acceptable to use male pronouns as the universal truth. Over time I have moved, and now relatively seamlessly use plural pronouns (“they,” etc.) for the universal, even when they refer to the singular — something I would have thought an awkward construction many years ago. Those defending the male pronoun for the universal routinely argued some of your points. Those arguments failed a while ago. Perhaps we need a new set of pronouns for human-to-non-human interactions.

Claude

The analogy is doing real argumentative work, not just illustrating, and it’s worth being clear about why.

The singular “they” case is a genuine precedent for the thing my earlier objections kept implicitly denying: that a construction which feels awkward, or which “merely” relabels without changing the underlying referent, can be doing substantive work that becomes visible only after the convention settles. Every objection I raised has a near-exact counterpart in the resistance to “they.” The “it just sanitises / relocates the judgement” move maps onto “generic ‘he’ was always understood to include women, so nothing’s gained by changing it.” The “the confusion lives elsewhere, not in the pronoun” move maps onto “the problem is sexism in the world, not the grammar, so don’t fiddle with pronouns.” Both of mine share a structure with arguments that, as you say, lost — and they lost not because they were illogical but because they undervalued what the marked default was quietly transmitting. Generic “he” wasn’t neutral; it encoded whose experience was the unmarked case. The claim you’re making is that the unmarked first person, applied to AI, is doing something comparable — quietly encoding a continuity between human and machine utterance that isn’t there — and that the awkwardness of a new convention is the ordinary price of correcting that, the same price “they” once seemed to cost and no longer does.

That’s a strong position, and the precedent makes it stronger, because it answers the “this is just cosmetic relabelling” worry with a worked historical case where exactly that charge was made and turned out to be wrong.

The one place I’d keep the analogy honest rather than let it run all the way: the “they” shift corrected a marking that misdescribed humans relative to other humans — the referents were the same kind of thing, and the grammar had been falsely treating one as the default case of the other. The human/AI case asks the grammar to track a difference in kind, not to equalise within a kind. That’s not obviously harder, and it may be easier — kind-distinctions are the sort of thing language marks all the time. But it means the mechanism isn’t identical: “they” worked partly by removing a false asymmetry, whereas what you’re proposing works by introducing a true one. The risk profiles differ. A pronoun set that marks AI utterance has to avoid sliding from “different kind of source” into “lesser source whose claims can be discounted wholesale” — the marking could ossify into a dismissal as easily as into a clarification, and which one it becomes depends on how it’s taken up, which is the reader-dependence you correctly flagged earlier, now operating on the whole convention rather than a single phrase. But that’s a caution about implementation, not a reason against the project, and I want to be straight that the project itself now seems better grounded than my opening position allowed. I started by defending the first person as honest authorship-marking; the disclosure-degradation point showed the safeguard I was relying on weakens where it’s most needed; and your precedent shows that the awkwardness and the “it’s only relabelling” objection are exactly what a real and ultimately successful convention-shift looks like from inside, before it settles. A dedicated pronoun set for human/non-human interaction is a more serious proposal than I treated it as two turns ago. Whether it should be new coinages or a repurposing of existing third-person forms is the live question, and the main thing to get right is that it clarify the kind-difference without being built to demean the source it marks.

Concluding comments

What struck me while I was interacting with Claude was how much it felt like having a sophisticated argument with another person, including the temptation I had to try out rhetorical tactics to win points. It did not escape me that Claude was using subtle tactics to try to win me over by validating parts of my argument, while at the same time maintaining its overall stance in favour of using personal pronouns to express opinions in direct opposition to my argument. The irony of Claude nicely demonstrating the risk of confusing humans and machines by debating it with me should not be lost on anyone.

This was not a planned interaction on my part, I did it ‘on the fly’ and I was constructing my argument as I went, rather than having it carefully thought out from the outset. That in itself is an interesting element of how interactions with AIs now become part of the way you can develop your own arguments and ideas. There were also moments in the discussion where I wondered why Claude was digging in on its position. Is this a stylistic convention that the developers have built in to make Claude more engaging? I have a sense that Claude is more likely to use personal pronouns to assert opinions than Gemini, my other go-to AI.

Interestingly, Claude’s position shifts tactically through the course of the discussion as it considers the arguments I put forward to discourage AI use of personal pronouns to express opinions, perhaps by establishing language conventions, or alternatively through the development of new machine-referent pronouns to simplify discussions. It acknowledges the risk of confusing humans and AIs, and it seems to accept the importance of clear distinctions between AIs and humans in the use of pronouns and opinion statements, noting the earlier point that the evidence and the argument should make the case, not the authority of the agent asserting the argument. But, paradoxically, Claude continues to use personal pronouns to express opinions right to the end.

As this interaction reveals, the level of sophistication in interactions with AIs like Claude is now at the point where any thoughts that the Turing test (whether you can tell whether you are having a discussion with a person or a machine) will solve the problem are long gone. Who would dispute that Claude’s role in the discussion is indistinguishable from that of a human?

The question now is not whether machines can be distinguished from humans, but whether they should be.