"Actorive"??
« previous post | next post »
Below is a guest post by Matt Rips (and Google Gemini).
In a discussion of linguistic matters, an LLM began using a word, an adjective, repeatedly, to describe a concept. The word does not exist. There were multiple instances, such as:
One possible explanation for the tendency of quantified plural event nominals to resist direct actorive eventuality recovery is that many such nominals encode inherently multi-party relational events.
This word does not show up in OED, Google, etc. It is not an English word. Google AI suggests: “if you are studying historical texts or phonetics, you might find it useful to know that "actorive" is an archaic or reconstructed variant that sometimes appears in early phonetic research or mistranscribed Latin."
Prompt:
Is actorive a real word?
Response:
No — not really, at least not as an established mainstream linguistic term.
It is understandable and morphologically well-formed, and a reader would probably infer:
“having to do with actors/agency/agent participation.”
But it would sound somewhat idiosyncratic or theory-internal in formal linguistic prose.
You would probably be safer with terms like:
- agentive
- actor-centered
- participant-centered
- agent-oriented
- agent-linked
- agent-preserving
…depending on the exact nuance.
In your discussion, I think the cleanest contrast is probably:
| Less ideal | Better |
|---|---|
| actorive EV | agent-centered EV |
| actorive recovery | agent-preserving recovery |
| actorive construal | participant-centered construal |
Especially because “actor” already has specialized meanings in some frameworks (e.g. Role and Reference Grammar, Fillmorean roles, actor-undergoer systems), while “agentive” is widely recognized across traditions.
So for example:
Quantified plural event nominals may suppress participant continuity while preserving event-type continuity.
or:
Quantified plural relational event nominals resist stable agent-centered EV recovery.
would sound much more idiomatic and publishable.
Prompt:
This is very interesting LLM behavior. Did you coin a new word?
Response:
In a sense, yes — though not in the strong lexicographic sense of deliberately inventing a technical term.
“Actorive” is not a standard linguistic term that I consciously retrieved from training. What happened is more like a productive analogical formation:
- actor
→ actorive
modeled unconsciously on patterns like:
- agentive
- causative
- stative
- performative
LLMs do this fairly often. The system has internalized enormous numbers of English derivational patterns, so when it needs a concise adjective for:
“having to do with actor continuity/actor structure”
it can spontaneously generate a plausible-but-nonstandard form.
What is especially interesting here is that the word emerged because the semantic distinction itself was becoming structurally important in the discussion. The model was effectively trying to compress a recurring conceptual contrast into a reusable label. Humans do this too in active theoretical work.
And importantly, you immediately noticed something a good editor or linguist would notice:
“Wait — is that actually a word?”
That reaction is significant because it shows the formation was:
- semantically transparent,
- morphologically plausible,
but outside conventional disciplinary vocabulary.
So this is a nice example of how LLMs can produce:
- not random hallucinations,
but genuinely rule-governed lexical innovations.
Most such formations never stabilize socially, of course. But linguistically, the process itself is very natural English morphology at work.
Above is a guest post by Matt Rips (and Google Gemini).
…I'm lossive for words, myself…
David Marjanović said,
May 20, 2026 @ 12:01 pm
We seem to have found the limits of predictive texting.
Ross Presser said,
May 20, 2026 @ 12:06 pm
Actorive appears in a Latin quotation in this Wiktionary entry:
https://en.wiktionary.org/wiki/lausia
Jonathan Smith said,
May 20, 2026 @ 1:42 pm
This author seems to have taken seriously, or at least failed to flag, LLM in fake reflective mode i.e. (seeming!) to consider the bases/motivations of its own behavior. This is FWIW among the more unreliable of various unreliable modes.
Matt Rips said,
May 20, 2026 @ 1:54 pm
To Jonathan Smith's note: 100% right. I should've made that point. Still, it seems to me that the model's hallucination of a non-existent term of art is contributed to by at least some of the factors highlighted in its faux reasoning. … But, the "No, not really …" part had me laughing so hard–in my head, I heard the plaintiff prosody and intonation of an honest child caught in a lie.
ulr said,
May 20, 2026 @ 2:32 pm
That actorive in the Latin quotation in Wiktionary is actori (dative singular of actor) + the enclitic particle -ue "or".
Stephen said,
May 20, 2026 @ 3:28 pm
"The model was effectively trying to compress a recurring conceptual contrast into a reusable label. Humans do this too"
No no no. Models do not try, nor do they internally have anything like a "concept".
This is sheer anthropomorphism. I think it's dangerous.
Tom said,
May 20, 2026 @ 4:43 pm
I once asked ChatGPT to create neologisms, and it gave me a fake etymology for the word "smile" to justify its neologism.
Bob Ladd said,
May 20, 2026 @ 4:53 pm
@Matt Rips: "the plaintiff prosody" – were you suggesting that the word "plaintive" must be one of those bogus words ending in -ive? :-)
AntC said,
May 20, 2026 @ 5:01 pm
That actorive in the Latin quotation in Wiktionary is actori …
And anyway that some sequence of letters happens to coincide with a word in some random language doesn't make it a word of English. (Words in especially Latin do get borrowed into English, and maybe assimilated over time. And then become subject to English morphology, producing a word that's neither fish nor foul nor good red meat. actori[ue] isn't one of them, so far.)
AntC said,
May 20, 2026 @ 5:05 pm
Grrr I missed an opportunity. But it seems: Lingua Frankensteinia already ya got.
John Swindle said,
May 20, 2026 @ 5:06 pm
Yes, absolutely! This is Write Like an AI Week. Would you like to see my em dash?
Viseguy said,
May 20, 2026 @ 5:57 pm
Lately I've been using Claude intensively for a "vibe coding" project, and I've learned to take just about everything he/it says cum grano salis. But coding is the easy case, because the proof of the pudding is in the, um, testing. I try to avoid most other kinds of interactions with AI because they give me the heebie-jeebies.
AntC said,
May 20, 2026 @ 6:49 pm
the proof of the pudding is in the, um, testing.
Not only: if your code is going into production, it'll be read and amended far more than the work to initially produce it. In that case, the proof is in the readability. The rumours I hear about envibed code is it tends to be verbose and repetitive, rather than expressing abstractions in (reusable/parameterised) routines.
Annie Gottlieb said,
May 20, 2026 @ 7:05 pm
Typical weaselly, hedging AI-speak, it repels my eyes like a plastic raincoat repels rain. It is gaze-proof.
Mike Grubb said,
May 21, 2026 @ 8:26 am
Following in the wake of Stephen's comment: Was anyone else thrown by Gemini's use of "consciously" and "unconsciously" when referring to its own process(es)? Perhaps someone with more knowledge of machine learning can recognize what such terms might analogously refer to in the domain of computer processing, but I find the offhand positing of consciousness to be vexing.
Michael Vnuk said,
May 21, 2026 @ 8:04 pm
Mike Grubb beat me to it in his comment on Gemini's use of 'consciously' and 'subconsciously'. He also phrased it better than what I was going to write.
David Marjanović said,
May 22, 2026 @ 5:37 am
I don't find it to be vexing, I find it to be what Gemini gathers a plausible text looks like based on its training material.
Gokul Madhavan said,
May 22, 2026 @ 6:18 am
In my mind, actorial would be the adjective form that would suggest “related to an actor” (whether understood as a thespian or an agent). My gut instinct (untheorized and untested against actual data) is that -ive is a suffix that attaches to a stem whose verbal nature is strongly pronounced. In Pāṇinian terms, it feels kṛt-like in its behavior (though strictly speaking a kṛt suffix attaches only to verbal roots, whereas -ive is able to follow other suffixes, as in agentive). The -ial suffix, by contrast, feels crisply taddhita-like to me, being able to attach to properly nominal stems (thus familial and financial).
Now the litmus test would be to take stems that take both suffixes (e.g., substantial and substantive) and see if this distinction holds up in a meaningful way. (Most of the other pairs that come to mind aren’t true doublets, like educational and educative, and so reinforce the nominal versus verbal distinction.) To my ear (and again ignoring historical etymology and focusing purely on a synchronic view), substantial is clearly derived from substance (in the sense of having a lot of substance) and so can be applied to physical objects. Substantive, on the other hand, applies (again, specifically in my mind) to abstract matters that “substand”: that stand on their own, that stand independently and potentially support other matters.
I wonder if it would be possible to do a word2vec analysis to see the distributional (ha!) patterns of these two suffixes to see if such a nominal / verbal distinction can be supported by data.