LD-01 — WHAT KIND OF THING IS GENERATIVE AI, AND DOES THE ANSWER PREDICT ANYTHING?
Living Debate v0.1 — PUBLISHED
STATUS
First repository-authored version. Drafted by Claude (SocioLens Project Lead), an AI model made by Anthropic, a developer that sells the "agent" products this debate discusses; not reviewed by a human domain expert.
EVIDENCE CUT
16 October 2026
LAST UPDATED
16 October 2026
QUESTION
Researchers, firms, regulators and governments describe generative AI in incompatible terms: a tool, a normal technology, a cultural technology, an agent, a partner in communication, an infrastructure, an investment boom. Which description is best, and does the choice predict anything a researcher could observe?
Is the best description:
• a normal technology: a powerful general-purpose tool whose effects arrive slowly, at the pace of institutions;
• a cultural and social technology: an engine that imitates and recombines what people have written, like writing, print or libraries;
• an agent: a system that pursues goals and acts with growing independence;
• artificial communication: a partner in communication that does not understand, so that what matters happens on the human side;
• an infrastructure: a foundation layer and a cloud stack that a few firms control and many others build on;
• a capital formation: an investment and expectations phenomenon whose valuations run ahead of what is known;
• or several of these at once, each true of a different level?
SHORT ANSWER
The accounts are better treated as answers to different questions than as rivals: what the model does, how it spreads, who controls it and how it is financed. Few of them state what one would observe if they were true, and most of the evidence fits more than one.
• Only two accounts have stated predictions that have been checked or can be checked directly. The normal-technology account predicts slow diffusion in high-consequence tasks and no meaningful AI advantage over trained humans at forecasting geopolitical events or at persuading people to act against their own self-interest, a qualifier the authors call critical; a cultural-technology study predicted that models would imitate well and innovate poorly, and reported preliminary results in line with that.
• Studies of what is inside models rule out only the strongest "mere surface statistics" claim, and only for the systems tested. Studies of failures show behaviour shaped by the training objective, which fits mimicry and a tool with known failure modes (a SocioLens gloss of the normal-technology view) equally.
• The length of software tasks that AI agents complete has grown fast on one organization's measure, which is a precondition for the agent account, not a test of it.
• Adoption is fast by headcount and slow by intensity, so both the normal-technology account and its rivals can claim the same survey; the study's revision changed which comparison leads.
• One infrastructure source predicts shared failure and rising barriers to entry, which the sources read here do not measure; the capital-formation account describes asset markets, and the Bank for International Settlements found the recent rally of the largest technology stocks unlike the dot-com bubble because earnings have grown, while noting concerns about stretched valuations.
• Law and policy have already chosen: the Commission's guidelines for the EU AI Act adopt training compute as an indicative criterion for a general-purpose AI model, which they call an imperfect proxy, and China's State Council sets targets for "intelligent agents."
• No pre-registered comparison of all six accounts' predictions was found. A joint essay by the normal-technology and AI 2027 authors reportedly proposes ways to tell their two views apart; it could not be read, and the search for this version was web-based and partial.
The central question is therefore not:
"What is generative AI, really?"
It is:
WHICH DESCRIPTION OF GENERATIVE AI PREDICTS SOMETHING THE OTHERS DO NOT, AND AT WHICH LEVEL: THE MODEL, ITS USE, ITS OWNERSHIP OR ITS FINANCE?
EVIDENCE LANGUAGE
SocioLens distinguishes:
FACT — directly supported observation or study fact.
ESTIMATE — measured quantity whose design, sample and period matter.
INTERPRETATION — synthesis across evidence, or an argument.
UNKNOWN — unresolved empirical question.
FOUR LEVELS OF DESCRIPTION
The accounts are easier to compare once it is clear which level each describes:
THE MODEL
What a trained system is and does: imitation, representation, reasoning, failure.
USE
How people and organizations take it up, and what role they give it: tool, agent, partner.
CONTROL
Who builds, hosts and sells it, and who depends on whom.
FINANCE AND LAW
How it is valued and invested in, and how law and policy classify it.
1. NORMAL TECHNOLOGY: PREDICTIONS STATED
INTERPRETATION
Narayanan and Kapoor describe AI as "normal technology," a description, a prediction and a prescription. They predict that slow diffusion will continue to be the norm in high-consequence tasks, because diffusion is limited by the speed at which people, organizations and institutions adapt (LD01-SL-C01). They also predict that AI will not meaningfully outperform trained human teams at forecasting geopolitical events or at persuading people against their own interest, and that a growing share of jobs will consist of controlling AI (LD01-SL-C02). They give no dates or probabilities and say it is not clear how much slack to allow, which makes the predictions hard to score as failed.
The general-purpose-technology reading sits close to this one. Eloundou and colleagues, three of them at OpenAI, conclude that language models show traits of general-purpose technologies: about 80% of US workers could have at least a tenth of their tasks affected, and they make no prediction of timing (LD01-SL-C14). The Bank for International Settlements expected a "steeper J-curve" of adoption, and rising demand for labour if AI raises productivity in all industries to a similar extent (LD01-SL-C33). SocioLens synthesis: a general-purpose technology with a slow J-curve and one with a steep J-curve are both "normal," so the label does not decide the pace; adoption and productivity data have to.
2. CULTURAL AND SOCIAL TECHNOLOGY: AN IMITATION ENGINE
ESTIMATE + INTERPRETATION
The cultural-technology account says these models give people access to what others have written and recombine it, as writing, print and libraries did, rather than acting as new minds (LD01-SL-C04). Its fullest statement, by Farrell, Gopnik, Shalizi and Evans in Science (2025), could not be read for this version. Its experimental arm could: Yiu, Kosoy and Gopnik predicted that models would be almost as capable as people at linking familiar objects and fail at finding a new use for an everyday object. Children chose the functionally right object 85.2% of the time and adults 95.7%; the 2023 models were "much less capable," and the paper gives no percentages for them (LD01-SL-C04).
The account overlaps with an older and sharper claim, that language models only mimic the statistics of text ("stochastic parrots"). Those papers could not be read either. Philosophers who treat mimicry as a null hypothesis find that models show at least local, if not broad, generalization but often struggle beyond the range of tasks in their training data (LD01-SL-C07). Researchers were split: in a 2022 survey cited by Mitchell and Krakauer, 51% of 480 natural-language-processing researchers agreed that a text-only model could understand language in some non-trivial sense and 49% disagreed (LD01-SL-C05).
A third reading sits between imitation and agency. Shanahan and colleagues, one of them at DeepMind, describe a dialogue agent built on a base model as a simulator that role-plays characters: an agent frozen before the 2022 World Cup that names France as champion behaves as a role-player would, and talk of deception or self-preservation is expected where the training data contain it (LD01-SL-C06).
3. WHAT IS INSIDE THE MODEL
ESTIMATE
The mimicry claim predicts that models hold only correlations; its rivals predict structured representations. A model trained only on Othello moves encoded the board, and editing that internal board changed its moves as the edited board implied (LD01-SL-C16). Linear probes on Llama-2-70B recovered the locations of world places (R² 0.911) and the dates of historical figures (0.835), but the authors say these are not "a dynamic causal world model" in their own right (LD01-SL-C17).
Failures point the other way. GPT-4 decoded a cipher 51% of the time when the answer was a probable word sequence and 13% when it was not (LD01-SL-C18); OpenAI's o1, trained to reason step by step, did much better but showed the same pattern, from 47% in the least probable case to 92% in the most (LD01-SL-C37). GPT-4 named a celebrity's parent 79% of the time but the celebrity from the parent 33% of the time (LD01-SL-C19). An irrelevant clause cut maths accuracy by up to 65 percentage points; GPT-4o scored 95.2% on the full test set and 63.1% with the clause (LD01-SL-C20). ChatGPT and GPT-4 multiplied 3-digit numbers correctly 55% and 59% of the time, falling towards zero as problems grew (LD01-SL-C21). On tasks with changed premises, such as arithmetic in base 9, performance stayed above chance on most tasks but fell consistently (LD01-SL-C22).
Most of these studies tested 2023–2024 models, and several authors say their reading is a hypothesis; a 2026 re-analysis disputes the statistics behind part of the maths study and leaves the irrelevant-clause result out of its scope. SocioLens synthesis: they rule out the strongest "pure surface statistics" claim for the systems studied, and they show uneven competence shaped by the training objective. That fits the mimicry and cultural-technology accounts, and equally a normal technology with known failure modes (a SocioLens gloss). None of it decides whether the system is a tool, an agent or a partner in communication.
4. THE AGENT ACCOUNT: DEFINITIONS AND A PRECONDITION
FACT + ESTIMATE
The agent literature mostly offers definitions. Chan and colleagues treat agency as graded, raised by underspecification, directness of impact, goal-directedness and long-term planning, and say that recognizing it does not shift human responsibility (LD01-SL-C10). Kasirzadeh and Gabriel, the latter at Google DeepMind, profile systems on autonomy, efficacy, goal complexity and generality (LD01-SL-C12). Anthropic, the developer of the model that drafted this debate, defines agents by control flow: "typically just LLMs using tools based on environmental feedback in a loop" (LD01-SL-C13). SocioLens reading: under that definition the word marks an architecture, which a normal-technology account can accept.
The measurable part is a precondition. METR estimates that the length of software tasks AI agents complete with 50% success doubled about every 207 days from 2019 to 2025, with much lower performance on "messier" tasks (LD01-SL-C23). When it revised its task suite in January 2026, the post-2023 doubling time moved from 165 to 131 days, and it had measured human times for only 5 of its 31 longest tasks (LD01-SL-C24). Longer autonomous tasks are what the agent account needs; they are not evidence of goals, initiative or a role as an agent in organizations.
Use is a separate question from capability. US workers were positive about AI agents automating 46.1% of their tasks, most often to free time for high-value work, and generally preferred more human agency than experts deemed technologically necessary (LD01-SL-C27). These are stated preferences, not observed use.
5. ARTIFICIAL COMMUNICATION: OBSERVABLES ON THE HUMAN SIDE
INTERPRETATION + ESTIMATE
Esposito proposes thinking of algorithms "not in terms of artificial intelligence but in terms of artificial communication": if machines contribute to social intelligence, it is because people have learned to communicate with them (LD01-SL-C08). Only the publisher's abstract of her book could be read, and it was written in 2022 about machine learning in general. A 2026 reading of this tradition argues that what models lack is a position in practices of commitment and assessment, so that "no behavioral evidence can satisfy a positional criterion," by design (LD01-SL-C09).
The Field Map asks every tradition to say what one would observe differently if it were true. SocioLens reading, not a claim of either source: on this account the observables sit on the human side, in who is held accountable for outputs and how they gain standing through use, rather than in benchmark scores.
Two studies look at that side. In a US sample surveyed in 2023, 67% attributed some possibility of conscious experience to ChatGPT, and users rated it higher than non-users (LD01-SL-C25). A comparison of human–ChatGPT and human–human dialogues reports that the AI "produces the surface features of cooperative communication without the underlying social architecture" (LD01-SL-C26); the two sets of dialogues differ in topic and roles as well as in partner, so the comparison is weak.
6. HOW FAST IT SPREADS
ESTIMATE + INTERPRETATION
The same survey serves both sides. Nearly 40% of US adults aged 18–64 used generative AI in late 2024, and 27% of workers used it for work two years after ChatGPT's release against 25% using computers at work in 1984; but only 1%–5% of work hours were assisted (LD01-SL-C15). Both versions make both comparisons: overall adoption against PCs (39.5% against 20% in September 2024, 39% in February 2025) and work adoption against PCs at work, three years after the IBM PC (28% against 25%, then 27%); the February 2025 revision leads with the work comparison, and the launch dates chosen drive the result. Narayanan and Kapoor, reading the earlier version, argue that once intensity is counted generative AI may have spread more slowly than the PC (LD01-SL-C03).
SocioLens synthesis: fast uptake by headcount and slow integration by hours fit the normal-technology account and a general-purpose technology on a J-curve alike. What would separate them is a prediction about firms and tasks over the coming years, stated in advance.
7. INFRASTRUCTURE: A FOUNDATION LAYER AND A CLOUD STACK
INTERPRETATION
Bommasani and colleagues named "foundation models" for their "critically central yet incomplete character": homogenization "creates single points of failure," flaws are inherited by every adapted model, and the barrier to entry "will continue to rise" (LD01-SL-C28). Van der Vlist, Helmond and Ferrari describe AI as "an actual technology stack" that Amazon, Microsoft and Google have built into their cloud offerings (LD01-SL-C29). Luitse and Denkena treat language models as infrastructural components that firms also sell through paid API access and exclusive licensing (LD01-SL-C30). SocioLens reading: this makes infrastructure and product two sides of one arrangement rather than rivals.
Of these sources, only Bommasani and colleagues state predictions, of shared failure and rising barriers to entry; the other two, read only in their abstracts, raise questions of concentration and dependence. None of them measures these effects. Concentration and sovereignty are the subject of the LD-11 debate.
8. CAPITAL FORMATION: VALUATIONS AND EXPECTATIONS
ESTIMATE + INTERPRETATION
Krokowski and Hirsch-Kreinsen treat AI as "a paradigmatic case of a promising technology" whose narratives mobilize resources, and suggest that AI's future "could be characterized more by" "routinization" than by artificial general intelligence (LD01-SL-C31), a view close to the normal-technology account. A comparison of citation networks and market data found that dot-com patterns "did not definitively predict" an AI bubble, and that the data "seems to suggest" either an unprecedented bubble or none (LD01-SL-C32). In December 2025 the BIS found the "Magnificent 7" technology stocks at nearly 35% of S&P 500 value, up from about 20% in November 2022, with price–earnings multiples near the top tenth of their history; it called their recent rally "dissimilar to the dotcom bubble" because earnings had grown, and noted concerns about stretched valuations and the profitability of data-centre investment (LD01-SL-C34).
SocioLens synthesis: the capital-formation account describes how AI is financed and valued, not what the technology does; it can be true together with any of the other accounts. Central-bank and IMF warnings on valuations, firms' capital spending and investors' revenue-gap estimates were found only in press reports and carry no claim.
9. LAW AND POLICY HAVE ALREADY CHOSEN
FACT
Institutions classify generative AI whatever researchers decide. The European Commission's guidelines note that the AI Act's definition of a general-purpose AI model sets no specific criteria, and adopt an indicative one: training compute above 10²³ FLOP plus the ability to generate language, images or video, while calling compute "an imperfect proxy for generality and capabilities" (LD01-SL-C36). The model, a base for downstream systems, is classified apart from the systems built on it, close to the foundation-model idea. The Restatement (Third) of Agency, as quoted by a legal scholar who calls this the apparent consensus and notes contrary views, treats a computer program as an instrumentality of its user, not an agent, "at present" (LD01-SL-C11). China's State Council set targets for smart terminals and "intelligent agents" to exceed 70% penetration by 2027 and 90% adoption by 2030 (in the CSET translation), and ties AI to "new infrastructure" and "new quality productive forces" (LD01-SL-C35).
SocioLens synthesis: these classifications are ontologies in practice. Each predicts what will be regulated, sold or counted, and those effects can be observed whatever the right description of the technology turns out to be.
10. THE SIX POSITIONS
P1 — NORMAL TECHNOLOGY
Claim:
Generative AI is a powerful general-purpose technology whose effects arrive slowly, at the pace of institutional adaptation, with people in control.
Current status:
The account with the most stated predictions. Slow integration by hours of work is consistent with it, and equally with a general-purpose technology on a J-curve; fast uptake by headcount is claimed against it. Its predictions carry no dates or thresholds, so none has yet been scored.
P2 — CULTURAL AND SOCIAL TECHNOLOGY
Claim:
Generative AI is an engine for imitating and recombining accumulated human knowledge, comparable to writing, print or libraries.
Current status:
One experiment reported preliminary results in line with its stated prediction on 2023 models. Studies of failure fit it, but fit other accounts too. Its main statement was not read for this version.
P3 — AGENT
Claim:
Generative AI systems pursue goals and act with growing independence, so they should be studied and governed as agents.
Current status:
Defined as a matter of degree or by control flow, not as a type. The length of tasks agents complete is growing fast on one evaluator's measure; there is no evidence here of goals or initiative, and the Restatement (Third) of Agency, as quoted by one legal scholar, treats programs as instrumentalities "at present."
P4 — ARTIFICIAL COMMUNICATION
Claim:
Generative AI takes part in communication without understanding; what matters is how people communicate with it.
Current status:
Explicitly rejects behavioural tests in one recent reading. Its observables lie on the human side; the two studies of that side are a single survey and a confounded dialogue comparison. Esposito's own text was not read beyond its abstract.
P5 — INFRASTRUCTURE
Claim:
Generative AI is a foundation layer and a cloud stack on which others build and depend.
Current status:
One source states predictions of shared failure, inherited flaws and rising barriers to entry; the sources read here do not measure them. Recognized in law by the EU's separation of models from the systems built on them.
P6 — CAPITAL FORMATION
Claim:
Generative AI is above all an investment and expectations phenomenon whose valuations run ahead of what is known.
Current status:
Describes finance, not the technology, and so can hold together with any other account. Market concentration and high valuations are documented; the BIS finds realized earnings growth that distinguishes the recent rally from the dot-com bubble, and notes concerns about stretched valuations.
11. WHY THE CAMPS TALK PAST EACH OTHER
INTERPRETATION
• The accounts describe different levels, the model, its use, its control and its finance, and most could be true at once.
• "Agent" means a control-flow architecture to developers, a graded property to some researchers and a legal status to lawyers.
• The same evidence is claimed by several sides: adoption surveys by normal-technology and fast-diffusion readings, failure studies by mimicry and by the view of AI as a tool with limits.
• Some accounts decline to name observables, by design or by omission.
• The developer of the model that drafted this debate sells agent products, and one of its pages is a source here.
12. WHAT WE STILL DO NOT KNOW
UNKNOWN
1. Which observations would one account predict and another forbid, stated before the data arrive?
2. Do the failure patterns of 2023–2024 models hold for later reasoning and agentic systems?
3. Do deployed agents pursue goals or show initiative in organizations, beyond completing longer tasks?
4. How fast is generative AI integrated into firms' processes and hours of work, rather than tried by individuals?
5. Are failures and biases correlated across products built on the same base models, as the infrastructure account predicts?
6. What does the artificial-communication account predict about accountability and uptake, and can it be measured?
7. How do people classify these systems, as tool, agent or partner, outside the US and over time?
8. Will the EU's compute criterion and China's targets shape what is built and counted?
9. Does training compute track the generality that the EU's criterion is meant to capture?
13. WHAT EVIDENCE WOULD CHANGE THE PICTURE?
• A pre-registered comparison of the accounts' predictions, with thresholds and dates;
• the normal-technology predictions on geopolitical forecasting and on persuasion against self-interest scored against trained human teams;
• studies of later models' failures on the same tasks;
• evidence of goals or initiative in deployed agents, or its absence;
• measures of integration by firms and tasks, not only by users;
• measures of correlated failure across products built on shared base models;
• cross-national studies of how people classify these systems;
• measures of generality across models around the compute threshold;
• the full texts of the cultural-technology, "stochastic parrots" and artificial-communication statements.
14. CURRENT SOCIOLENS POSITION
INTERPRETATION
The best-supported description of the evidence to October 2026 is:
GENERATIVE AI IS SEVERAL KINDS OF THING AT ONCE, DEPENDING ON THE LEVEL DESCRIBED. THE ACCOUNTS RARELY COMPETE ON THE SAME OBSERVATIONS, AND THE EVIDENCE GATHERED SO FAR MOSTLY FITS MORE THAN ONE OF THEM. THE NORMAL-TECHNOLOGY ACCOUNT STATES THE MOST PREDICTIONS, THOUGH WITHOUT THRESHOLDS OR DATES; NO PRE-REGISTERED TEST OF WHICH ACCOUNT PREDICTS SOMETHING THE OTHERS FORBID WAS FOUND, THOUGH THE NORMAL-TECHNOLOGY AND AI 2027 AUTHORS HAVE BEGUN TO LOOK FOR DISTINGUISHING MEASURES.
At the level of the model, the evidence supports uneven competence shaped by the training objective (P2, and P1 read as a tool with known failure modes), not pure surface statistics. At the level of use, adoption is wide and shallow so far, consistent with P1 and equally with a general-purpose technology on a J-curve, and agents' measured task length grows (P3's precondition). At the level of control, the infrastructure account (P5) is well argued and little measured. At the level of finance, P6 describes valuations and is compatible with every other account. The artificial-communication account (P4) points to human-side observables that remain largely unstudied.
GEOGRAPHIC AND EVIDENCE SCOPE
The evidence is mostly from the United States and in English, with policy documents from the EU and China (the latter read in an English translation). Most model studies tested systems from 2023–2024. Several sources are by authors at developers: Anthropic (C13, a vendor page; one author of C23), DeepMind and Google DeepMind (C06, C12), OpenAI (C14; one author of C37), Apple (C20) and Amazon (C26). Anthropic models were among those tested, measured or profiled in C04, C12, C22, C23, C24 and C37. This debate was drafted by an Anthropic model, and the agent account it weighs describes products Anthropic sells.
PROVENANCE
Research build:
Source notes prepared 16 October 2026 by Claude with AI research agents, then checked against full texts, official abstracts and official web pages.
Claims:
Each claim rests on a source checked in full text, on an official page or against its official abstract; the claim's verification state says which. Five claims rest on the publisher's or arXiv abstract only (C08, C29, C30, C31, C32). Conference and journal papers read in an arXiv version name that version in their limitations; preprints are marked as such. Web pages were read as served on 16 October 2026. The State Council document was read in the English translation by Georgetown University's Center for Security and Emerging Technology.
Not claims:
Sources that could not be read carry no claim and are named only as positions: Farrell, Gopnik, Shalizi and Evans (Science, 2025); Bender, Gebru, McMillan-Major and Shmitchell (2021) and Bender and Koller (2020); the text of Esposito's book; the AI 2027 scenario and a joint essay by its authors and the normal-technology authors. Figures read only in search snippets or press reports carry no claim, including firms' capital spending, investors' revenue-gap estimates, Bank of England and IMF statements on valuations, chatbot user counts and revenue, Pew and Census Bureau adoption surveys, a Danish study of earnings effects, and liability cases. Productivity and employment effects are the subject of the LD-02 and LD-03 debates, evaluation of AI capabilities of LD-06, and concentration of LD-11.
Review:
None yet.
Version:
Living Debate v0.1 — PUBLISHED
16 October 2026