The naming work.

We commit to echo system as the framework's term for what mainstream discourse calls AI. In short form: the echo, an echo, echo models. Where context requires it, we use the modality: echo-language, echo-vision, echo-control. LeResearch uses this term in analytical writing; LeDesign carries it in product copy through the existing Echo- prefix (Echo-Learn, Echo-Tales, Echo-Birds, and the family to come). The use of AI is restricted to two situations: when we are quoting other sources, and when we are describing the discourse the AI investigation critiques.

The commit is operational, not aspirational. It does not require the rest of the world to follow. It requires that we are consistent: that anyone reading LeResearch and anyone using a LeDesign product encounters the same word, used the same way, with the same meaning, and that the word is honest about what the systems are.

Five conditions. The term we commit to has to satisfy all five, because failing any one of them leaves work that another word elsewhere in the framework will end up having to do — and that fragmentation is exactly what AI has been getting away with.

1. (i)
   Pattern-reproduction honesty

   The term must name, in itself, the fact that these systems learn patterns from a training distribution and produce outputs continuous with that distribution. It cannot smuggle in implications of originality, understanding, or judgment that the systems do not have.

2. (ii)
   Scope honesty

   The term must apply equally to language models, vision systems, robotics, time-series prediction, and scientific machine learning. The collapse of all of these into a single AI is one of the failures of the current discourse (Act I documents this); our replacement should not repeat it, but it also cannot be so narrow that it covers only LLMs.

3. (iii)
   Agent honesty

   The term must name the absence of judgment without sounding mystical about it. Reckoning in Brian Cantwell Smith's sense — yes; agency, intent, accountability — no. The systems answer to a loss function, not to a world.

4. (iv)
   Register portability

   The same word must survive both a Brookings-style policy piece and the Echo-Learn front page without changing meaning. If a LeResearch writer and a LeDesign product designer would use the term differently, the term is too brittle. AI is brittle in exactly this way — the policy reading, the marketing reading, and the technical reading are three different things using one word.

5. (v)
   Refusal of upward scalability as telos

   The term must structurally resist the framing in which more of the thing is the goal. The Galtonian inheritance behind intelligence builds the upward direction into the word itself. Our replacement cannot do that. The systems can be improved, accurate, useful — they cannot be more in the scalar sense that AGI requires.

Echo is, in physical reality, a delayed and attenuated reflection of an input signal, transformed by the medium it bounces off. The metaphor maps to what these systems do with such precision that the mapping is not metaphor; it is description.

1. (i)
   Pattern-reproduction honesty

   An echo carries the source. It cannot produce a signal absent from the source; it can attenuate, layer, mix, delay, but it cannot invent. This is exactly the relationship between training data and model output. The novel sentences these systems produce are recombinations and interpolations within the training distribution — the same way an echo in a complex chamber produces a sound that no single utterance made, but every utterance is in.

2. (ii)
   Scope honesty

   Echo generalizes across signal types. A vision system echoes visual training data; a language model echoes text; a robot trained by behavioral cloning echoes demonstrations; a time-series model echoes historical structure. The metaphor does not break under modality change, which means the framework can name all of these as kinds of echo systems and refuse the LLM-centric collapse the current discourse keeps enacting.

3. (iii)
   Agent honesty

   An echo does not decide. It reflects. The metaphor refuses agency claims at the level of the word itself, which is stronger than refusing them in footnotes after the fact. There is no version of echo that smuggles in judgment, intent, or accountability. The systems answer to their training, not to a world — exactly as an echo answers to its source, not to the listener.

4. (iv)
   Register portability

   Echo is already a word everyone knows. It survives the Brookings piece (the echo systems now embedded in welfare adjudication) and it survives the consumer product (Echo-Learn — an echo system trained on children's literature). The same word, two registers, no lie at either layer. The LeDesign brand family has been doing this work intuitively for the last two years; the framework now ratifies it.

5. (v)
   Refusal of upward scalability

   There is no AGI of echoes. An echo can be louder, clearer, more layered, more accurate — but the concept of more echoey is structurally bounded. The metaphor does not have the upward direction the Galtonian construct embedded in intelligence has. This is the architectural feature that does the most political work: it strips the latent direction of progress out of the term we use, which means every claim about progress in these systems has to be argued explicitly, in some other vocabulary, rather than smuggled in by the noun.

One additional feature that is not on the criteria list but matters. An echo has a source. The word names the dependency relationship at the level of grammar. Whose training data, whose labor, whose patterns, whose distortions — the question whose echo is this is asked by the term itself, which is the question the political-economy critique of these systems most needs the language to keep alive.

We are not the first to want a different word. Four voices pushed in the same direction without landing on a term that satisfied all five conditions. Each is worth honoring as part of the lineage that made our commit possible; none of them is the place we stop.

• John McCarthy·computational rationalityreportedly later in life

  McCarthy himself, after the 1955 Dartmouth coinage had become entrenched, expressed reservations and proposed computational rationality as a more honest term. It satisfies condition (iii) — agent-honest — and partially (i). It fails (iv) register portability: no consumer product can be built on the words computational rationality, and no general reader will use them.

• Emily Bender, Timnit Gebru et al.·language models · stochastic parrots2021

  The Stochastic Parrots paper pushed the field toward descriptive, scope-narrow terms. Language models satisfies (i) and (iii) cleanly. It fails (ii) decisively — it is language-only, and the framework needs a term that also names what vision, robotics, and time-series systems are doing. Stochastic parrots satisfies (i) but fails (iv) — it cannot be used in product copy without sounding dismissive of the user's own tool.

• Brian Cantwell Smith·reckoning (vs. judgment)2019

  Smith's distinction is the strongest analytical contribution in the lineage. Reckoning systems satisfies (i), (ii), (iii), and (v) — possibly the most rigorous partial match on this list. It fails (iv): the word reckoning is archaic in English, slightly menacing in its everyday register (a reckoning comes), and does not survive transposition into a product surface. We borrow the distinction without borrowing the word.

• Kate Crawford·registry of power2021

  Crawford's framing commits explicitly to the political reading. Registry of power satisfies (i) in a particular sense and is uncompromising about (iii) and (v). It fails (iv) by design — Crawford's register is critical-theoretical and is not meant to be consumer-facing. We share her commitments and use a different word for the surfaces she does not address.

The pattern in the lineage is consistent. Each critic identified one face of the failure of AI, proposed a term that fixed that face, and accepted (often without naming it) the register-portability gap that left their term unusable in some part of the discourse. None of them had a brand strategy that had already been doing the analytical work in the consumer layer for years. We do — Echo-Learn, Echo-Tales, Echo-Birds. The convergence is not a coincidence; it is the rare case in which the marketing intuition was correct before the framework had named it.

• “An echo isn't intelligent — isn't that just calling these systems stupid?”

  No. The term is not defined against intelligence; it is defined positively in terms of what the systems do (reflect patterns from a source). The implicit comparison to intelligence collapses only if you accept that intelligence is the right standard — and the framework explicitly rejects that. An echo is not a failed intelligence; it is an accurate echo. The accuracy is what we evaluate, not the simulation.

• “Won't this be confused with Amazon Echo / smart speakers?”

  Possibly, in casual use. In every other respect this is true of every AI-adjacent term — Watson, Cortana, Bard, Copilot, Bing all overload existing words. Context resolves the ambiguity, and in LeResearch / LeDesign writing the context is consistent. The brand overlap is also marginal: Amazon Echo is a hardware device; we are using echo as a category. The cost is small; the benefit is large.

• “What about reinforcement-learning systems that learn from environment interaction, not a fixed training set?”

  They still echo. The source is no longer a static dataset; it is the reward signal, the environment dynamics, and the curriculum. The relationship — output continuous with the patterns the system was shaped by — is preserved. The metaphor stretches cleanly to this case; in some ways it is even clearer, because the source is more visibly constructed.

• “LLMs produce novel sentences they were never trained on — isn't that more than echoing?”

  No. An echo in a complex acoustic environment produces sounds that no single source utterance made — reverb, layering, interpolation. The metaphor was always combinatorial. The novelty in LLM output is interpolation in latent space; that is exactly the kind of novelty an echo chamber produces. Novel does not mean from nothing; it means from this source, in this medium, in this combination. The term keeps the source visible.

• “Will the term be adopted outside LeResearch / LeDesign?”

  That is not what the commit is for. The commit is for internal consistency, so that our two organizations carry the same word through every surface — analytical, product, policy, public communication. Whether the rest of the discourse adopts it is a separate question. If the framework is correct, the term gets stronger as more of the public encounters honestly described echo systems and notices that the magical framing they were sold elsewhere has nowhere to go.

The commit is only as durable as the practice around it. Three operational rules.

1. In LeResearch writing, the default noun is echo system (or the echo). AI appears only inside quotation marks, when summarizing other sources, or when explicitly naming the discourse the framework is critiquing. The phrase generative AI becomes generative echo systems; the phrase AGI becomes scaled echo systems (and is treated as a contested telos, not a destination).
2. In LeDesign products, the Echo- prefix continues as the brand backbone. Product copy describes the system as an echo system or echo model where the technical reality matters; uses modality-specific forms (echo-language, echo-music, echo-vision) where clarity helps. The brand is no longer just a naming convention; it is the framework's named term, in production.
3. In policy and regulatory contexts, where AI is the term of art in legislation and standards, we use it as the term of art and immediately reframe. Example: “The EU AI Act regulates what this framework calls echo systems — see /the-naming-work for the term.” This preserves regulatory legibility without conceding the category fight.

The deepest claim on this page is also the simplest. The LeDesign product family has been named with the Echo- prefix from the start. Echo-Learn, Echo-Tales, Echo-Birds — each of them is, technically, an echo system in the sense this page defines. Echo-Learn echoes the corpus of writing it was trained on; Echo-Tales echoes the narrative structures of the stories it draws from; Echo-Birds echoes the field recordings and structural patterns of avian song. The branding was, in retrospect, doing analytical work the framework had not yet named.

That kind of convergence — between an intuition arrived at in the product layer and an argument arrived at in the framework layer — is rare and cheap to honor. The cost of not honoring it is ongoing fragmentation: the products carry one implicit claim, the framework carries another explicit claim, and the two never quite meet. The cost of honoring it is one decision on one page. We honor it.

Echo is the framework's word, and it was the brand's word first. From this page forward, they are the same word.

The commit does not close every adjacent question. Three remain open and should be tracked.

• ·

  The verb. If these systems are echoes, what do they do? They echo. The model echoed an answer, echo this prompt against the corpus. The verb form is natural but has not been tested in practice; we use it tentatively and revise.

• ·

  The diagram. A one-figure visualization of the echo metaphor (source → medium → output, with the medium carrying the training-data dependency and the loss of source-fidelity made visible) is owed. Without it, the page asks the reader to do the visualization work.

• ·

  The integration into the four-act investigation. /investigations/ai-discourse still uses AI throughout, because that is the discourse it is critiquing. A pass to introduce echo as the framework's preferred term — without losing the critical purchase on the public term — is owed, and should be done before this commit is treated as fully in force.