NOTE-697: Floridi — AI as agency without intelligence

Contribution

Floridi coins a slogan and a framing: generative models are Agere sine Intelligere, agency without intelligence (p. 6). Earlier debate asked whether LLMs understand. He grants that they do not and moves the question to agency: they are successful, adaptive actors with no understanding, a kind of agent that, he says, no past culture met except as natural or supernatural agency. That is the new thing. The rest of the editorial (the ChatGPT examples, the failure modes, the list of ethical, legal and labour costs) is commentary on early 2023 and does not bear on the thesis.

Key insight

Split the capacity to act successfully from the capacity to understand, and treat LLMs as the first artefacts with plenty of the first and none of the second. “We have decoupled the ability to act successfully from the need to be intelligent, understand, reflect, consider or grasp anything. We have liberated agency from intelligence” (pp. 5–6). Learning separates such an agent from a mere cause. A sea wave “makes the difference, but can make nothing but that difference, without being able to ‘learn’ to make a different or better difference” (p. 5).

Assumptions

• No understanding. LLMs “do not think, reason or understand” and have “nothing to do with” the cognitive processes animals and humans use “to manage semantic contents” (p. 1). They work “statistically—that is, working on the formal structure, and not on the meaning of the texts they process” (pp. 1–2). This is taken as given. It is supported by citation (Bishop 2021) and by examples, not argued.
• Intelligence is understanding. “Intelligence” is used throughout as the capacity to understand, reflect, consider and grasp (p. 6). It is not defined. On a behavioural or functional definition of intelligence the thesis would not be stated this way.
• Agency is successful, adaptive action. The only marks of agency given are success and the capacity to “learn” and improve behaviour (p. 5). The scare quotes on “learn” are Floridi’s.
• RL as the source of agency. RL fine-tuning made ChatGPT an agent in the relevant sense, because “RL is the machine learning approach to ‘solving agency’” (p. 5, credited to Vincent Wang). This is a remark about one training method, not an analysis of agency.

Key results

There are no results in the formal sense. The positions taken, in order:

• Capability (p. 2). ChatGPT’s 50-word summary of The Divine Comedy (Fig. 1) is “not bad, and certainly better than one produced by a mediocre student”. The teaching task shifts to prompting, checking and correcting.
• Fragility (pp. 2–3). LLMs “fail catastrophically”, as with Bard’s demonstration error. They confabulate references, make factual and trivial logical errors, struggle with arithmetic, and have “strange linguistic blind spots”. Fig. 2: asked for “the name of the only daughter of Laura’s mother”, ChatGPT says it lacks the information. The answer is Laura.
• Analogy (pp. 3–4). “Stochastic parrots” is only partly apt: parrots are intelligent, and LLMs “synthesise texts in new ways”. Better images are search-engine autocomplete, the “mediocre or lazy” student’s patchwork essay, and Manzoni’s trickster pulling endless “tape” (Floridi’s gloss on the word points to the Turing machine’s tape). LLMs “gobble data … and regurgitate (what looks to us as) information” (p. 4).
• Implications (pp. 4–5). A list without argument: copyright and lawsuits, Kenyan content labellers paid under $2/hour, curricula, poisoned training data in medical uses, financial and environmental cost, jobs. For philosophers there are “LEGO-like” modular AI with LLMs as an “AI2AI” bridge (fn. 10), form versus content, personalisation, interpretability, human replaceability, and “power as the control of questions”.
• Thesis (pp. 5–6). ChatGPT is “a form of agency never seen before, because it is successful and can ‘learn’ and improve its behaviour without having to be intelligent to do so”. Humans used to meet agency only in animals and in what they took for spiritual agents. Now they must “learn to interact with, artificial agents created by us, as new demiurges”. People in AI-supervising jobs (“green collars”) become “shepherds of AI systems, in charge of this new form of artificial agency (Floridi & Sanders, 2004)” (p. 6).
• Appendix (p. 6). ChatGPT rated the opening paragraph 4 stars out of 5 (Fig. 3). This is a joke, not evidence.

Claims

id	claim	strength	support
C1	LLMs do not think, reason or understand. They process the formal structure of texts, not their meaning	weak	assertion with one citation (Bishop 2021), a list of failure modes from blogs and preprints, and one screenshot (Fig. 2, p. 4). Floridi himself notes that most such errors are patched (“like zero-day exploits”, p. 3), which weakens failures as evidence for a principled lack
C2	”Stochastic parrot” is only a partial analogy, because LLMs synthesise rather than repeat	moderate	informal observation (p. 3), consistent with Fig. 1. It is not tested
C3	LLMs are agents: they act successfully and “learn” to improve their behaviour	weak	asserted (p. 5). No criteria are stated, and none of Floridi & Sanders’ criteria (LIT-898), which the editorial cites, are applied
C4	This agency is new: no culture before met successful, learning agency except in animals or as natural or supernatural forces	weak	assertion (p. 5). It skips earlier artefacts that act and learn, such as adaptive controllers and pre-LLM machine-learning systems
C5	Agency has been “liberated … from intelligence”: acting successfully no longer requires understanding	weak	follows from C1 and C3 if both are granted (pp. 5–6). It is only as strong as C1
C6	Learning to make a different or better difference is what separates an agent from a mere cause such as a sea wave	assertion	one sentence (p. 5). It is the only demarcation the editorial offers

Concepts

• Agency (as used here): successful action that can be improved by learning. It is not defined beyond the sea-wave contrast (p. 5).
• Intelligence (as used here): understanding, reflecting, considering and grasping (p. 6). It is not defined.
• Agere sine Intelligere: the slogan for “intelligence-free (as in fat-free) AI” (p. 6).
• Confederated AI: modular AI systems made interoperable by LLMs acting as an “AI2AI kind of bridge” (p. 5, fn. 10, credited to Wang, with ChatGPT and Wolfram Alpha and Socratic Models as examples).
• Green collars / shepherds of AI systems: the human roles that supervise artificial agency (p. 6, citing Floridi 2017).

Connections

The agency concept is attributed to Floridi & Sanders 2004 (LIT-898) and to nothing else. That paper’s method, setting agency at a level of abstraction without mental states, is the natural way to make the thesis precise, but the editorial does not use it. Levels of abstraction (also central to LIT-152, NOTE-099) are not mentioned. Floridi & Chiriatti 2020 on GPT-3 is cited for failure modes only. The editorial’s title and thesis are taken up again in Floridi’s 2025 Philosophy & Technology paper (LIT-897, DOI-10.1007/s13347-025-00858-9), read later the same day in NOTE-702.

Within the record:

• Shanahan et al. (LIT-460, NOTE-370) also deny that the bare model has beliefs or goals, but they deny it agency too: no goals, “not even simulated versions”. Floridi grants agency on success and learning alone. The two disagree on whether agency needs anything inside the system beyond adaptive behaviour. Neither engages the other.
• Dennett (LIT-398, NOTE-353) would read the same behaviour through the intentional stance and attribute belief where prediction works. Floridi’s denial of understanding is made from outside any such stance and is not defended against it.

Bearing on the record

• Agency line, QUESTION-009. The editorial is a clear permissive data point: success plus learning is enough for agency. That is the kind of criterion the record’s permissiveness objection targets, and the editorial does not answer it. It gives no reason why a thermostat with adaptive gain, or a gradient-trained classifier, is not an agent in the same sense. It offers no criterion for when something is an additional agent.
• Organizational minimal agency (LIT-566, NOTE-458, THEORY-071). On THEORY-071, LLMs would not be agents. They have no self-individuating organization and no norms they generate for their own maintenance. Floridi’s thesis is the opposite verdict reached on a thinner notion. The record should cite this editorial as the statement of the view THEORY-071 excludes, not as an argument against it. It gives none.
• Agent, individual, self, person (ADR-024). Floridi keeps agency apart from intelligence and understanding, and says nothing about individuality, selfhood or personhood. That is compatible with the record’s separation and does not test it.
• No THEORY filed. A seven-page editorial whose central claim is asserted cannot ground one. If Floridi’s 2025 paper (LIT-897), once read, gives the multiple-realisability argument its title names, a THEORY would be drawn from that reading, not this one.
• No instruction for machine-learning practice. The teaching advice (p. 2) is about using ChatGPT in education, not about building models.

Limitations

• It is an editorial, and it reads as one. The thesis takes three paragraphs at the end. Everything before it is context.
• The no-understanding premise (C1) is carried by anecdotes from early 2023 and by assertion, and the author grants that the failures get patched. If the premise fails, the thesis changes from “agency without intelligence” to something weaker.
• No criteria for agency are given, so the thesis cannot be checked against cases. The sea-wave contrast rules out only systems that cannot learn.
• The claim of novelty (C4) ignores artefacts older than LLMs that already act successfully and improve by learning. The 2004 paper it cites is itself about artificial agents.
• Placing the source of agency in RL fine-tuning is a passing remark credited to a colleague. If it were taken seriously, a base model would not be an agent and a fine-tuned one would be. The editorial does not draw that consequence.
• It does not say whose agency this is: the model, the deployed system, or the model together with its users. The “confederated AI” remark makes the question urgent and leaves it open.

Open questions

• What criteria for agency would make “agency without intelligence” checkable? Floridi & Sanders’ criteria are the obvious candidate. Their criteria are interactivity (the agent and its environment “(can) act upon each other”), autonomy (it can “change state without direct response to interaction”) and adaptability (its interactions “(can) change the transition rules by which it changes state”) (LIT-898, p. 9 of the preprint). They offer these as “guidelines”, an “effective characterisation” and not a definition, and they hold only at a chosen level of abstraction (pp. 3, 9). The verdict can change with the level: their MENACE example is an agent when only its games are observed and not one when its mechanism is observed (pp. 11–12). Applying the criteria to an LLM would therefore mean first saying which observables count. Whether the 2025 paper (LIT-897) supplies them is for its reading to settle.
• Is a base LLM, before RL fine-tuning, an agent on this view? The RL remark suggests not, and the “learn” criterion is met by pretraining as well.
• What would show that an LLM understands, so that C1 could fail? The editorial names no test. Its “Forget passing the Turing Test” (p. 3) rejects one without proposing another.