LaclauGPT: Ideological contestation over AI

Status: Phase 1 current executable research-plan paper. Publication status: This is a draft research plan and working manuscript. It has not been formally published or peer reviewed, and its arguments, methods, scope, and wording may change as the research develops. Phase progression: Phase 1 · Phase 2 draft plan · Phase 3 draft plan · Phase 4 draft plan

Abstract

Artificial intelligence is not only a heterogeneous set of technologies and sociotechnical arrangements but also a contested political signifier around which competing visions of society are increasingly articulated. Different actors connect AI to economic growth, abundance, unemployment, surveillance, existential risk, democratic control, technological emancipation, or the concentration of political and economic power. This paper studies this ideological contestation through Ernesto Laclau and Chantal Mouffe’s discourse theory, Emilia Palonen’s Formula of Populism, and the concept of sociotechnical imaginaries. Accelerationism, existential-risk discourse, Critical AI, opposition to AI, and left-wing techno-optimism are treated as starting points for analysis rather than fixed ideological categories. The paper also develops LaclauGPT, an LLM-assisted methodology for computational discourse analysis designed to identify candidate signifiers, articulations, demands, collective identities, political frontiers, affects, and sociotechnical imaginaries in large textual datasets. The model does not perform the final discourse analysis but produces theory-guided pre-analysis that remains linked to source material and can be accepted, rejected, or revised by human researchers. The proposed research programme compares three arenas of AI politics: AI elites, grassroots mobilisation, and parliamentary and electoral politics. The paper therefore makes two connected contributions by developing a theoretical framework for analysing the ideological struggle over AI and a computational method for scaling Laclaudian discourse analysis while keeping interpretation traceable to evidence.

Keywords: artificial intelligence; ideology; discourse theory; LaclauGPT; hegemony; sociotechnical imaginaries; large language models; critical AI studies; computational social science

Author: Tomi Toivio, Helsinki Hub on Emotions, Populism and Polarisation (HEPPsinki), University of Helsinki. ORCID: 0000-0002-1335-0478.

Funding acknowledgement: LaclauGPT has been developed in connection with CO3, ENDURE, and PLEDGE, supported by the European Union and the Research Council of Finland.

1. Introduction

Debates about artificial intelligence are not only debates about what machines can do but increasingly about what kind of society technological development should produce. Questions concerning who should control AI, who should benefit from it, what kinds of labour should be automated, what risks are acceptable, whether development should accelerate or slow down, and who has the authority to make these decisions connect technological change directly to political conflict. AI can consequently be articulated as economic progress, a path towards abundance, a threat to employment, an infrastructure of surveillance, an existential danger, or a technology of emancipation. These articulations do not simply describe a technology from different perspectives but connect AI to competing political projects, collective identities, demands, threats, and imagined futures. The social consequences of AI therefore cannot be derived from technological capabilities alone, because they also depend on ownership, governance, labour, infrastructure, institutions, and struggles over the distribution of benefits and harms.

Critical AI Studies provides the broader intellectual context for analysing these struggles by treating AI as a sociotechnical and political object rather than as an autonomous technological force (Lindgren, 2023). This paper focuses on one dimension of this wider problem: ideological contestation over AI. Ernesto Laclau and Chantal Mouffe’s discourse theory is particularly useful for this purpose because it does not begin from fixed ideological categories but asks how political meanings and identities are produced relationally through articulation (Laclau & Mouffe, 2001; Laclau, 2005). From this perspective, the central analytical question is not simply whether a text should be classified as accelerationist, critical, techno-optimist, or doomer. The more important question is how AI becomes connected to other political demands and signifiers such as growth, freedom, humanity, capitalism, safety, labour, democracy, extinction, abundance, or control, and how these relations give AI a particular political meaning within a discourse.

The same signifier may perform several different functions in these struggles. AI can function as a nodal point around which a discourse is organised, become a floating signifier contested between rival political projects, or under particular conditions become a tendentially empty signifier representing a wider equivalential chain and a more general political project. These roles cannot be inferred from semantic ambiguity alone but have to be demonstrated through political articulation. This creates an immediate methodological problem, because Laclaudian discourse analysis depends on contextual and relational interpretation while contemporary political discourse is produced at a scale that cannot be analysed through exhaustive close reading. LaclauGPT addresses this problem by using large language models for theory-guided pre-analysis. The model identifies candidate relations, signifiers, subjects, demands, affects, frontiers, and sociotechnical imaginaries, while human researchers evaluate these interpretations against the underlying source material. The aim is therefore not to automate political interpretation away but to make interpretive discourse analysis computationally scalable while keeping its theoretical assumptions explicit and its claims connected to evidence.

The research questions are:

What ideological formations and sociotechnical imaginaries compete to define AI and its social future?
How are these formations articulated across AI elite discourse, grassroots mobilisation, and parliamentary and electoral politics?
Under what conditions does AI function as a nodal point, floating signifier, or tendentially empty signifier?
How can LLM-assisted analysis operationalise Laclaudian concepts, and how should its reliability, validity, and characteristic errors be evaluated?

Questions 1–3 define the empirical research programme, while Question 4 concerns the methodological contribution of the paper. An earlier version of LaclauGPT was used in research on the 2024 European Parliament elections, but that work provides development experience rather than validation of the revised methodology presented here. The present paper consequently treats validation as an integral part of the research design rather than assuming that previous use demonstrates reliability. Its central argument is that LLMs can help extend the reach of interpretive discourse analysis if their output is treated as provisional, theory-guided interpretation whose relation to the source remains visible and contestable.

2. AI as a sociotechnical and ideological object

2.1 Critical AI Studies, assemblage, and ideology

Critical AI Studies begins from the recognition that AI systems are socially constituted rather than autonomous technical objects. They are developed inside institutions and depend on labour, data, computational infrastructure, economic incentives, political priorities, and assumptions about what intelligence, efficiency, optimisation, and progress should mean. Lindgren (2023, pp. 17–18) emphasises that apparently technical objectives therefore already contain social choices concerning what is valued and whose interests count. Critical analysis cannot consequently stop at evaluating whether a system performs a technical task successfully but must examine the social arrangements through which the task itself has been defined, the institutions that benefit from its automation, and the forms of authority that make particular definitions of intelligence or efficiency appear natural.

An assemblage perspective helps define this heterogeneous object. DeLanda (2016, pp. 10–11) describes assemblages as wholes composed of heterogeneous elements whose identities are not exhausted by the whole and whose relations remain partly exterior to one another. Applied to AI, this directs attention to models, datasets, GPUs, data centres, human labour, organisations, interfaces, users, platforms, regulation, law, and institutional power. AI is produced through these relations and cannot be adequately understood as an isolated algorithm or model. Assemblage theory and Laclaudian discourse theory are nevertheless used for different analytical purposes in this paper. Assemblage thinking helps describe the heterogeneous sociotechnical object under investigation, while Laclaudian discourse theory provides the primary conceptual framework for analysing its political articulation. This distinction avoids reducing AI either to discourse alone or to a technology whose social meaning is already determined by its technical properties.

Following Lindgren, the ideological dimension of AI can be divided into several connected questions. Ideologies shaping AI concern assumptions, interests, and priorities informing technological development and deployment, while ideologies reproduced through AI concern classifications, perspectives, and social assumptions carried by AI systems and their outputs. This paper adds ideological contestation over AI as a third analytical dimension: the struggle among political projects to define what AI is, what it should become, whom it should serve, how it should be governed, and what kind of future it represents. These dimensions overlap, because the ideological assumptions shaping technical development can become objects of political contestation while political narratives about AI may in turn shape investment, regulation, organisational priorities, and technological design.

Ideological formations are therefore understood here as provisional configurations of meanings, demands, identities, affects, and political projects rather than as internally coherent doctrines. Ideology also does not simply mean factual error or conscious manipulation. A technological prediction may be empirically plausible and still perform ideological work if it presents a particular organisation of ownership, authority, or distribution as the inevitable consequence of technological development. Critical AI scholarship itself must consequently remain open to analysis. Critical AI provides part of the intellectual orientation of this study, but Critical AI actors can simultaneously become objects of discourse analysis alongside technology companies, accelerationists, existential-risk organisations, labour movements, anti-AI campaigns, and institutional policymakers. This does not require treating all of these actors as politically equivalent or ignoring their very different access to institutional power. It requires applying the same evidentiary standards when analysing how they articulate AI politically.

2.2 Sociotechnical imaginaries

The concept of sociotechnical imaginaries provides an additional way to understand the politics of AI because contemporary AI discourse is saturated with competing accounts of the future. Jasanoff (2015) defines sociotechnical imaginaries as collectively held and institutionally stabilised visions of desirable social orders made possible through science and technology. Richter et al. (2023) argue that this perspective is particularly useful for analysing AI because the category encompasses heterogeneous technologies while its broader public and political meaning remains unsettled. AI may be imagined as producing abundance, eliminating undesirable work, generating unprecedented economic growth, reproducing structural inequality, enabling authoritarian surveillance, or creating an existential threat to humanity. These futures matter politically before they occur because expectations about technological development shape investment, regulation, research agendas, institutional priorities, and political mobilisation in the present.

Sociotechnical imaginaries complement Laclaudian discourse theory by making projected social order explicit within political articulation. An imaginary of AI-driven abundance immediately raises questions about property, distribution, employment, and ownership, while an imaginary of catastrophic AI risk raises questions about expertise, authority, legitimate intervention, acceptable risk, and who has the right to slow or halt technological development. The two conceptual traditions should nevertheless not be collapsed. A prediction, fear, personal aspiration, or speculative scenario does not by itself constitute a sociotechnical imaginary. Document-level analysis can identify candidate future visions, desirable outcomes, feared alternatives, authorised agents of change, and diagnoses of the present, but demonstrating a collectively held and institutionally stabilised imaginary requires comparison across actors, organisations, and institutions. The concept therefore adds a temporal and future-oriented dimension to the analysis without replacing the Laclaudian focus on articulation and hegemonic struggle.

2.3 An open field of competing formations

The contemporary politics of AI already contains several recognisable ideological formations, although these should function as sensitising concepts rather than a closed taxonomy imposed on the corpus. Oldenburg and Papyshev (2025), for example, analyse accelerationist, existential-risk, and critical imaginaries. Accelerationist and techno-optimist discourse connects technological development to desirable social transformation, although techno-optimism, singularitarianism, transhumanism, and effective accelerationism have distinct intellectual histories and should not be treated as synonyms. Andreessen’s (2023) Techno-Optimist Manifesto explicitly articulates technology together with markets, growth, abundance, and human flourishing, while Kurzweil’s singularitarian work (2005, 2024) connects AI to a much broader transformation of human capabilities and the human condition. These positions overlap in their positive orientation towards technological development but differ in their assumptions about markets, human enhancement, political authority, and the future relationship between human and machine intelligence.

Existential-risk discourse centres on the possibility that sufficiently capable AI could become an existential threat to humanity. Yudkowsky and Soares (2025) provide an unusually explicit formulation of this argument. From a discourse-theoretical perspective, however, the important questions extend beyond whether catastrophic risk is technically plausible. The analysis concerns how threat, responsibility, expertise, humanity, and legitimate intervention are constructed: what threatens humanity, who can speak for humanity, who possesses the authority to define acceptable risk, and what political measures become justified by the scale of the threat. The category “doomer” should therefore be analysed as a situated political label rather than adopted uncritically as a neutral description, because not all AI-safety discourse is existential-risk discourse and not everyone positioned within that formation accepts the designation.

Critical AI perspectives foreground power, inequality, labour, discrimination, surveillance, political economy, and the organisation of technological development. DAIR’s research philosophy offers one institutional example of an alternative account of how AI research should be organised and whom it should benefit (Distributed AI Research Institute, 2022). Gebru and Torres’s (2024) TESCREAL critique identifies intellectual connections among transhumanism, extropianism, singularitarianism, cosmism, rationalism, effective altruism, and longtermism. TESCREAL is used here as their critical analytical concept rather than as evidence that everyone associated with these traditions belongs to a single homogeneous ideology. The broader analytical question is how critical actors articulate AI through relations of power, extraction, discrimination, labour, expertise, and social justice, and how these articulations compete with accounts centred on progress, safety, or technological inevitability.

The research design also includes opposition to AI, including mobilisation around employment, creative work, surveillance, data extraction, environmental impacts, and data-centre development. Whether these struggles are becoming linked into a broader anti-AI identity is itself an empirical question. Opposition to one form of AI does not necessarily imply opposition to AI in general, and a trade union resisting workplace automation, an artist opposing generative-AI training, and a community resisting a local data centre may articulate very different grievances and political identities. The interesting Laclaudian question is precisely whether such heterogeneous demands remain separate or become connected through equivalential relations and a shared frontier. Left-wing techno-optimism further complicates any simple opposition between technological enthusiasm and social critique. Srnicek and Williams (2015) and Bastani (2019) connect automation to post-work or postcapitalist futures, while Haraway’s (1991) cyborg politics challenges established human-machine boundaries from a socialist-feminist perspective. These traditions demonstrate that support for technological transformation does not necessarily imply support for capitalist ownership, existing technology corporations, or unrestricted development.

The ideological field should consequently remain analytically open. Accelerationists may criticise corporate concentration, Critical AI actors may support particular forms of automation, safety advocates may favour technological development under stronger political control, and labour movements may oppose specific deployments without opposing AI as such. Actors can move between formations, combine elements from several traditions, or reject the labels used to describe them. LaclauGPT therefore begins from claims, signifiers, relations, demands, and political subjects rather than assigning documents immediately to ideological boxes. Broader formations should emerge through comparison of recurring articulations across the corpus.

3. From discourse theory to computational analysis

3.1 Articulation and the political production of meaning

For Laclau and Mouffe (2001), articulation is a practice that establishes relations among elements in ways that modify their identities. Meaning is therefore relational rather than intrinsic. “AI” does not carry one fixed political meaning into every discourse: connecting AI to national competitiveness produces a different political logic from connecting it to labour exploitation, while connecting it to human extinction produces another again. The analytical object is not therefore the isolated term but the network of relations through which it acquires a situated political meaning. This has direct methodological consequences for computational analysis. The co-occurrence of “AI” and “employment” tells us relatively little by itself, because a speaker may be predicting unemployment, promising employment growth, dismissing fears of displacement, quoting an opponent, or discussing someone else’s argument. Computational discourse analysis must therefore identify relations, speaker positions, and argumentative context rather than merely count words or topics.

Several Laclaudian concepts are central to this analysis. A nodal point partially organises a discourse by fixing relations among other signifiers, while a floating signifier becomes the object of competing attempts at fixation across political projects. An empty signifier comes to represent a wider equivalential chain and a broader political project whose total meaning exceeds any particular demand (Laclau, 1996, 2005). These concepts describe functions within articulation rather than permanent properties of individual words. AI may therefore operate as a nodal point inside one discourse while also functioning as a floating signifier between rival discourses. Whether it becomes tendentially empty requires stronger evidence that it has come to represent a wider chain of demands or an absent social fullness. Ambiguity, semantic breadth, and multiple definitions are not sufficient evidence by themselves.

A second important distinction is between equivalence, difference, and antagonism. Equivalence connects heterogeneous demands through a shared relation to an obstacle or political order, while difference maintains distinctions among demands or positions. Antagonism concerns the frontier through which the fulfilment of an identity is represented as being blocked by an opposing force. These distinctions matter computationally because negative sentiment, disagreement, and criticism can easily be overinterpreted as political antagonism. A policy dispute does not automatically constitute a frontier, and dislike of an actor does not necessarily mean that the actor functions as the constitutive outside of a political identity. The model must therefore identify the relation being proposed and provide evidence for it rather than infer antagonism from polarity alone.

The same relational principle applies to political subjects and affect. Terms such as “workers,” “humanity,” “artists,” “citizens,” or “innovators” do not automatically constitute collective political identities simply because they appear in a text. The analytical question is how membership, common demands, representation, and opposition are constructed. Affective investment introduces a further dimension that cannot be reduced to sentiment classification. Fear of AI, pride in technological progress, anger at technology corporations, hope for abundance, or fascination with machine intelligence become politically significant through the identities, demands, and objects to which those affects are attached. Palonen’s (2025) Formula of Populism offers a compact heuristic for representing these relations:

\[\text{Populism} = \text{Us}^{\text{Affects}_{1}} + \text{Frontier}^{\text{Affects}_{2}}.\]

The formula is an interpretive device rather than a mathematical model. “Us” refers to the collective subject being constituted, “Frontier” to the political boundary through which that subject is defined, and the affective components to investments in both sides of the relation. This makes the formula particularly suitable for computational operationalisation because an LLM can propose a candidate collective subject, frontier, associated affects, and supporting textual evidence. Identifying an in-group and opponent nevertheless does not mechanically establish populism, because the result remains a theoretical interpretation requiring contextual evaluation. Palonen’s formulation is therefore used as a heuristic within a broader Laclaudian analysis rather than as an algorithmic test that turns every conflict into populism.

Finally, hegemony concerns the partial stabilisation of a particular articulation as a wider organising principle. Frequency, visibility, repetition, engagement, or virality may provide useful empirical signals, but none of them is equivalent to hegemony. Claims about hegemonic influence require evidence that an articulation becomes normalised, institutionalised, reproduced across arenas, capable of excluding alternatives, or connected to political consequences. Power therefore remains an essential part of interpretation. A technology CEO, government minister, activist organisation, academic researcher, and anonymous social-media user may all articulate political meanings around AI, but they do not possess equal capacity to organise institutions or make particular meanings authoritative.

3.2 LLM Structuralism as a methodological bridge

LaclauGPT develops from the Anarcho-Computational / Discourse-Theoretical approach, or AC/DT, which combines discourse theory with computational experimentation and methodological pluralism (Koljonen et al., 2025). AC/DT draws partly from Feyerabend’s (1975) argument against rigid methodological monism and from computational approaches to interpretation developed in works such as Text as Data (Grimmer et al., 2022) and Lindgren’s Data Theory (2020). LaclauGPT extends this approach through the use of large language models, and the theoretical connection between LLMs and relational theories of language provides one rationale for doing so. This connection can be described as LLM Structuralism: modern language models operate through relational representations in which linguistic elements acquire computational significance through patterns of relation with other elements rather than through dictionary-like intrinsic meanings. This produces an obvious methodological affinity with structuralist accounts of linguistic value.

For Saussure, linguistic value is differential: a sign acquires significance through its position within a system of differences rather than through an intrinsic correspondence with an object. Recent work by Kozlowski on computational structuralism and Weatherby’s (2025) account of language machines makes the relevance of this tradition to contemporary language modelling increasingly explicit. Laclau and Mouffe extend the relational insight of structuralism into a poststructuralist theory of political meaning in which relations are never permanently fixed but are continuously articulated, contested, and temporarily stabilised. Political discourse can therefore be understood as a struggle over relational structures of meaning. LLMs are potentially useful for such analysis precisely because they are unusually capable of processing contextual relations in language at scale.

This affinity should not be mistaken for an argument that LLMs somehow validate structuralism or Laclaudian discourse theory. Statistical associations do not by themselves establish political equivalence, antagonism, affective investment, empty signifiers, or hegemony. The connection is methodological rather than ontological: LLMs can be treated as instruments for proposing relational interpretations that researchers then evaluate theoretically and empirically. The division of labour can therefore be summarised quite simply: the model proposes candidate relations, the source provides evidence, and the researcher interprets. Even this division should not be exaggerated, because extraction itself already involves interpretation. Identifying a speaker, deciding what counts as a demand, selecting a relevant passage, or recognising irony all require judgement. The important distinction is therefore not between objective machine extraction and subjective human interpretation but between different kinds of interpretive claim and the evidence required to support them.

Nelimarkka’s (2026) MarxistLLM provides an interesting parallel because it makes a theoretical orientation explicit through fine-tuning. LaclauGPT takes a different route by specifying its theoretical protocol through prompts, schemas, codebooks, and contextual material rather than attempting to train a model into a Laclaudian worldview. This makes the analytical framework easier to inspect, modify, and compare without retraining the underlying model, but it does not make the system theoretically neutral. LaclauGPT sees the material through concepts supplied by the researcher, and those concepts influence what the system notices and how it organises the corpus. The methodological question is consequently not whether theory can be removed from the pipeline but whether its influence can be made explicit, traceable, and empirically evaluated.

3.3 Development context

LaclauGPT was originally developed in research at HEPPsinki connected to CO3, ENDURE, and PLEDGE. An earlier version was used to support analysis of TikTok and Instagram material from the 2024 European Parliament elections, with collection running from 1 May to 9 June 2024 and covering material from Bulgaria, Croatia, Finland, France, Germany, Hungary, Portugal, Spain, and Sweden. The earlier workflow combined multimodal analysis with LLM-assisted discourse analysis: audiovisual material first had to be transformed into textual representations, after which LLMs were used to extract topics, entities, sentiment targets, and discourse-theoretical information. This produced useful analytical material but also exposed several practical problems. Similar entities proliferated under different names, overlapping topics were generated independently, context was lost between analytical runs, and interpretive categories could become inconsistent. These problems became particularly visible when results had to be consolidated across large collections.

The revised LaclauGPT architecture treats these weaknesses as methodological problems rather than merely software bugs. Persistent context and codebooks are used to help consolidate recurring entities and expressions, while provenance and review status distinguish established information from provisional model interpretations. The earlier system therefore matters as development history and as a source of concrete failure modes, but it does not demonstrate that the current system is reliable. The revised workflow must still be validated against human interpretation, and improvements in model quality do not remove the need to test whether better language generation actually produces better discourse analysis.

3.4 An evidence-linked workflow

The basic unit of analysis is a source document or a context-preserving segment, which may be a social-media post, manifesto section, parliamentary intervention, interview turn, podcast segment, or another meaningful textual unit. Each unit retains metadata such as source identifier, date, language, speaker information, collection method, and its relationship to a larger document. Context is particularly important because political statements become difficult to interpret when quotation, negation, irony, speaker position, or surrounding argument disappear during segmentation. Where speech recognition, translation, OCR, or visual description is used, these derived representations remain identifiable as transformations of the original source rather than being silently substituted for it.

The workflow begins by preserving and describing the source before moving to theoretical coding. LaclauGPT identifies speakers, claims, topics, attributed positions, and relevant context, distinguishing the author’s position from quotation, rejection, irony, speculation, and hypothetical argument. It then proposes candidate articulations, demands, subject positions, signifier roles, equivalential relations, differences, political frontiers, affective investments, and sociotechnical imaginaries. Each interpretation should include supporting evidence, uncertainty, and where relevant counter-evidence, while the model must also be allowed to abstain when the theoretical category is not supported. This is followed by identity resolution and canonicalisation, where recurring actors, organisations, expressions, and concepts are linked through persistent codebooks without assuming that semantic similarity means political equivalence. A signifier such as “freedom,” for example, may play radically different roles in different ideological projects and should not be collapsed merely because the lexical form is identical.

Corpus-level comparison then organises candidate articulations across actors, time periods, languages, ideological arenas, and source types. This makes it possible to investigate how signifiers move between contexts, how demands become connected or separated, how political frontiers are constructed, and how competing actors attempt to fix the meaning of AI. The Formula of Populism forms one conditional layer within this broader process: where the material supports such an interpretation, LaclauGPT can propose a collective subject, frontier, and associated affects, while documents that do not display a populist structure may still contain relevant articulations or imaginaries. Human researchers then accept, modify, reject, or reinterpret the model’s proposals and return to the source material before making corpus-level claims.

Every theoretical proposition should retain its provenance, including the document identifier, exact evidence span, source field, proposed relation, model version, prompt version, uncertainty, and review status. This makes model interpretation inspectable rather than allowing fluent summaries to become detached from their evidence. A quotation can be mechanically checked against the retained source representation, but its presence does not prove that the theoretical interpretation is valid because the quotation can still be misattributed, taken out of context, or irrelevant to the claim being made. Mechanical verification and human interpretation therefore solve different problems. Persistent context is similarly useful but potentially dangerous: LaclauGPT should remember established entities, recurring expressions, aliases, rejected merges, and previous human decisions, but model-generated interpretations must not silently become established facts in later runs. The context system therefore needs to preserve the epistemic status and history of its information.

The overall design should remain independent of one particular LLM family or hardware configuration. Each research run should record the model, prompt, contextual material, generation parameters, serving environment, and preprocessing steps so that changes in the analytical pipeline remain visible. Local inference can be useful for privacy, control, and reproducibility, but running a model locally does not automatically make the analysis valid, just as setting temperature to zero does not convert interpretation into deterministic measurement. The current pipeline should therefore be understood as a methodological design under development whose reliability depends on empirical validation rather than technological sophistication alone.

4. Research design and validation

4.1 Three arenas of AI contestation

The empirical research programme compares AI discourse across three arenas. The first is AI elite discourse, including entrepreneurs, researchers, public intellectuals, research organisations, technology companies, and ideological movements that have significant influence over the public meaning and development of AI. Possible sources include manifestos, blogs, interviews, podcasts, forums, speeches, and public social-media material. Actor selection should not depend exclusively on prior ideological classification but should also be justified through institutional position, visibility, demonstrated influence, or relevance to the discourse. This makes it possible to analyse ideological formations without building the sample solely from actors already assumed to represent them.

The second arena is grassroots mobilisation around and against AI, including movements and campaigns concerned with labour, cultural production, surveillance, safety, democratic control, environmental effects, infrastructure, and resistance to particular uses of AI. The important question is how specific grievances become politically connected. Resistance to workplace automation may or may not become part of a broader anti-AI identity, just as opposition to data-centre development may connect to environmental politics without producing opposition to AI in general. Creative-worker campaigns may articulate AI through ownership, exploitation, consent, or cultural autonomy, while safety movements may articulate regulation through existential risk and the representation of humanity as a political subject. These relations should be discovered empirically rather than assumed from issue labels.

The third arena is parliamentary and electoral politics, where relevant material includes parliamentary debates, party programmes, policy documents, campaign communication, and election-related social-media datasets. This arena makes it possible to examine how AI enters established political institutions and cleavages: which actors connect it to growth, sovereignty, labour, regulation, national security, education, democracy, or social policy, and which ideological formations originating in elite or grassroots discourse acquire institutional expression. The three arenas should not be treated as completely separate social worlds, because think tanks, advocacy organisations, academics, entrepreneurs, political parties, and movements can operate across several of them. Comparison across arenas therefore allows the study to examine not only different discourses but also movement and translation between them.

Collection combines automated retrieval with digital ethnography. Automated methods provide scale, while ethnographic observation provides contextual understanding and can identify emerging vocabulary, new communities, changing ideological labels, irony, local meanings, and material missed by keyword-based collection. Researcher fieldnotes nevertheless remain researcher interpretations rather than direct participant speech and should be stored as a distinct source type. The initial geographical emphasis is on the United States, the European Union, and Finland, which should be understood as a sampling decision rather than a claim that these regions represent global AI politics. AI is global in its infrastructures, labour relations, environmental effects, political economy, and imagined futures, and a corpus dominated by Western discourse cannot be used to infer perspectives from populations whose own discourse has not been collected.

Each corpus therefore needs an explicit observation period, inclusion and exclusion rules, language scope, collection provenance, and documentation of missing material. Source lists and queries should be versioned, and exploratory collection should remain distinguishable from the final analytical corpus. Engagement metrics can describe platform visibility but should not be interpreted as direct measures of ideological support or public opinion. The analytical movement proceeds from passages to documents, documents to actor trajectories, and actors to broader patterns of articulation. Movement between arenas must likewise be demonstrated rather than inferred from shared vocabulary: the same word appearing in an industry manifesto and a parliamentary speech does not prove ideological diffusion, while stronger evidence might include explicit citation, temporal sequence, repeated relational structures, institutional uptake, or shared networks of actors and concepts.

4.2 Validation as part of the method

The main methodological question is not whether LaclauGPT can produce plausible discourse analysis, because contemporary LLMs are extremely good at generating plausible interpretations. The more important question is whether those interpretations are defensible, useful, traceable, and systematically distinguishable from unsupported theoretical speculation. Validation must therefore be built into the method. A development sample can be used to refine prompts, schemas, definitions, and codebooks, but a separate evaluation sample should then be held apart from those revisions. The evaluation corpus should include the major arenas, source types, languages, expected formations, difficult documents, uncertain cases, and documents where the model identifies little or nothing. Negative cases are particularly important because evaluation restricted to successful model outputs would make false negatives and theoretical forcing difficult to detect.

At least two trained researchers should independently code the evaluation material before seeing LaclauGPT output, reducing the risk that model interpretations anchor human judgement. Their agreement and disagreement provide a benchmark against which model output can be compared, while human annotation itself should not be treated as absolute ground truth because discourse analysis contains legitimate interpretive disagreement. The benchmark should therefore preserve disagreements rather than collapsing them automatically into one authoritative answer. Different analytical tasks also require different evaluation strategies. Speaker attribution and evidence-span identification can be assessed relatively directly, bounded categories can use agreement statistics or precision and recall, while more complex judgements such as antagonism, empty signifiers, collective subject construction, or affective investment require closer interpretive assessment. Evaluation should consequently distinguish category presence, evidence accuracy, relation identification, speaker attribution, signifier role, collective-subject construction, frontier construction, and affective investment rather than reducing performance to one global score.

Abstention also requires explicit evaluation. A model that proposes almost no theoretical interpretations may achieve high apparent precision while contributing little to the research, whereas a model that confidently finds Laclaudian structures in almost every document may simply be forcing the theory onto the corpus. Coverage and error therefore need to be reported together, and model confidence should be treated as another output to evaluate rather than as a calibrated probability of theoretical correctness. Robustness testing should compare the primary configuration with at least one alternative model and with prompts that remove predefined ideological seed labels. This helps determine whether ideological formations are emerging from patterns in the material or being reproduced because the researcher supplied them to the model in advance.

Error analysis is particularly important because different mistakes have different methodological consequences. Relevant categories include invented evidence, unsupported theoretical inference, missed context, irony, mistranslation, incorrect speaker attribution, quotations mistaken for author positions, excessive entity merging, theoretical overcoding, and missed articulations. Different theoretical constructs also require different negative controls. A political text unrelated to AI may still contain antagonism, while a contentious AI policy disagreement may contain no constitutive frontier. A text quoting Yudkowsky does not automatically share Yudkowsky’s position, and a text discussing accelerationism does not automatically belong to an accelerationist formation. Validation should therefore test these specific failure modes rather than relying on a generic expectation that the model should either classify or abstain correctly.

4.3 Limitations and reflexivity

The most obvious danger in theory-guided LLM analysis is that if a language model is instructed to find Laclaudian discourse structures, it will often find them. The resulting interpretation may be sophisticated, fluent, theoretically convincing, and nevertheless empirically overstated. Wachinger et al. (2025) identify similar problems in LLM-assisted qualitative analysis, including apparently convincing interpretations supported by unreliable evidence. LaclauGPT therefore needs safeguards against theoretical forcing: prompts should explicitly permit the absence of a category, require evidence, allow uncertainty, and where useful request counter-evidence or alternative interpretations. These safeguards themselves cannot simply be assumed to work but have to be tested in the validation process.

LLMs also inherit ideological assumptions and linguistic asymmetries from their training data, can hallucinate, are sensitive to prompts, and perform unevenly across languages. Transcription, translation, OCR, and visual description introduce additional layers of transformation. These should not be collapsed into one generic problem of “LLM bias,” because they represent distinct failure modes requiring different forms of evaluation. Multimodal analysis makes this particularly visible: a transcript is not a video, OCR is not an image, and an image description is not the visual object itself. Tone, editing, humour, gesture, visual symbolism, music, and contradictions between speech and imagery may disappear when audiovisual material is transformed into text. Derived representations should therefore remain labelled as derived representations, and validation should compare at least a sample of them against the original media. Multilingual evaluation should likewise compare source-language and translated interpretation where feasible.

Human review is necessary but does not automatically solve these problems. Researchers can share the assumptions of the model, become persuaded by fluent outputs, overlook systematic absences, or reproduce the same theoretical expectations encoded in the prompt. Independent coding, explicit disagreement, examination of negative cases, and repeated return to the original source help reduce these risks. The researcher’s dual role as developer and analyst should also be made visible when interpreting favourable results. LaclauGPT cannot validate Laclaudian discourse theory by finding Laclaudian concepts supplied through its own prompts, because that would be circular. Its methodological contribution is more modest but also more useful: it can make theoretically informed interpretations traceable, comparable, and contestable across datasets that would otherwise be difficult to analyse at scale.

4.4 Research ethics and data stewardship

Public data should not be treated as ethically frictionless data. The fact that material has been posted publicly does not remove ethical obligations concerning its collection, aggregation, analysis, publication, or long-term storage. The research should therefore document processing purposes, relevant permissions and legal bases, access arrangements, retention, sensitive-data risks, and quotation practices. Searchable quotations require particular care because exact wording can make ordinary participants identifiable through search engines even when usernames have been removed. Publication strategies should therefore consider whether quotations are necessary, whether paraphrase is more appropriate, and whether particular sources involve vulnerable or non-public-facing participants.

Data should be processed through infrastructure appropriate to the project’s data-management requirements, and automatic model routing should never override those arrangements. Reproducibility does not necessarily require releasing sensitive source material. Code, prompts, schemas, run manifests, synthetic examples, and methodological documentation can often be published even where raw data cannot. Synthetic data are useful for demonstrating software behaviour and reproducibility of computational procedures, but they cannot replace naturally occurring political discourse when validating substantive interpretations. Ethical and technical reproducibility therefore need to be treated as related but distinct goals.

5. Conclusion

AI is becoming one of the major political signifiers through which competing futures are articulated. For some actors it represents abundance, growth, autonomy, human enhancement, and technological liberation; for others it represents extinction, surveillance, unemployment, environmental harm, extraction, or the concentration of economic and political power. These meanings are not properties of the technology itself but are produced through political articulation. Laclau and Mouffe’s discourse theory provides a framework for studying this struggle relationally by asking how actors connect AI to demands, identities, affects, opponents, futures, and wider political projects rather than sorting texts immediately into predefined ideological categories. Sociotechnical imaginaries add another dimension by showing how technological futures also contain visions of social order, while an assemblage perspective keeps the material side of AI visible through models, infrastructures, corporations, labour, users, data, and institutions.

LaclauGPT translates this theoretical framework into a computational research method. Its central proposition is that LLMs can help researchers identify and compare candidate political relations in datasets too large for exhaustive close reading, but that the model should not replace interpretation. Its outputs remain provisional and linked to source evidence, uncertainty, provenance, and human review. This requires keeping several distinctions clear throughout the analysis: co-occurrence is not articulation, difference is not antagonism, ambiguity is not emptiness, sentiment is not affective investment, and frequency is not hegemony. These distinctions are not only theoretical niceties but practical safeguards against reducing Laclaudian analysis to generic text classification.

The next step is empirical validation and application across AI elite discourse, grassroots mobilisation, and parliamentary and electoral politics. The value of LaclauGPT will not depend on whether a language model can produce convincing Laclaudian prose, because that is already technically easy. It will depend on whether the method helps researchers discover, compare, challenge, and explain political patterns that would otherwise remain difficult to see, while also making its own mistakes visible enough to be corrected. The aim is therefore not automated discourse analysis but computationally assisted discourse analysis in which theory remains explicit, evidence remains traceable, and interpretation remains open to disagreement.

References

Andreessen, M. (2023, October 16). The techno-optimist manifesto. Andreessen Horowitz.

Bastani, A. (2019). Fully automated luxury communism. Verso.

DeLanda, M. (2016). Assemblage theory. Edinburgh University Press.

Distributed AI Research Institute. (2022). DAIR research philosophy, version 1.0.

Feyerabend, P. (1975). Against method: Outline of an anarchistic theory of knowledge. New Left Books.

Gebru, T., & Torres, É. P. (2024). The TESCREAL bundle: Eugenics and the promise of utopia through artificial general intelligence. First Monday, 29(4).

Grimmer, J., Roberts, M. E., & Stewart, B. M. (2022). Text as data: A new framework for machine learning and the social sciences. Princeton University Press.

Haraway, D. J. (1991). A cyborg manifesto: Science, technology, and socialist-feminism in the late twentieth century. In Simians, cyborgs, and women: The reinvention of nature (pp. 149–181). Routledge.

Jasanoff, S. (2015). Future imperfect: Science, technology, and the imaginations of modernity. In S. Jasanoff & S.-H. Kim (Eds.), Dreamscapes of modernity: Sociotechnical imaginaries and the fabrication of power (pp. 1–33). University of Chicago Press.

Koljonen, J., Carrilho, K., & Palonen, E. (2025). The struggle over masks on Twitter: An AC/DT approach to Finnish pandemic governance. In E. Kerr, E. Bužinkić, & J. Foley (Eds.), The organisation of irresponsibility? Reassessing COVID-19 in Europe (pp. 132–163). Brill.

Kozlowski, A. C. (2026). Computational structuralism: Toward a formal theory of meaning in the age of digital intelligence. Theory and Society.

Kurzweil, R. (2005). The singularity is near: When humans transcend biology. Viking.

Kurzweil, R. (2024). The singularity is nearer: When we merge with AI. Viking.

Laclau, E. (1996). Emancipation(s). Verso.

Laclau, E. (2005). On populist reason. Verso.

Laclau, E., & Mouffe, C. (2001). Hegemony and socialist strategy: Towards a radical democratic politics (2nd ed.). Verso.

Lindgren, S. (2020). Data theory: Interpretive sociology and computational methods. Polity.

Lindgren, S. (2023). Introducing critical studies of artificial intelligence. In S. Lindgren (Ed.), Handbook of critical studies of artificial intelligence (pp. 1–19). Edward Elgar Publishing.

Nelimarkka, M. (2026). MarxistLLM: Fine-tuning a language model with a Marxist worldview. Big Data & Society, 13(2).

Oldenburg, N., & Papyshev, G. (2025). The stories we govern by: AI, risk, and the power of imaginaries. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, 8(2), 1939–1950.

Palonen, E. (2025). The birth and death of liberal democracy in Hungary: The populist logic of polarisation as hegemony. Helsinki University Press.

Richter, V., Katzenbach, C., & Schäfer, M. S. (2023). Imaginaries of artificial intelligence. In S. Lindgren (Ed.), Handbook of critical studies of artificial intelligence (pp. 209–223). Edward Elgar Publishing.

Srnicek, N., & Williams, A. (2015). Inventing the future: Postcapitalism and a world without work. Verso.

Toivio, T. (2025a). LaclauGPT multimodal analysis [Computer software]. GitHub.

Toivio, T. (2025b). LaclauGPT TikTok scraper [Computer software]. GitHub.

Wachinger, J., Bärnighausen, K., Schäfer, L. N., Scott, K., & McMahon, S. A. (2025). Prompts, pearls, imperfections: Comparing ChatGPT and a human researcher in qualitative data analysis. Qualitative Health Research, 35(9), 951–966.

Weatherby, L. (2025). Language machines: Cultural AI and the end of remainder humanism. University of Minnesota Press.

Yudkowsky, E., & Soares, N. (2025). If anyone builds it, everyone dies: Why superhuman AI would kill us all. Little, Brown and Company.