One of the major controversies AI has brought to societal debate concerns its renewed role in public access to information . One once spoke of “algorithms” as a catch-all denominator for the underlying infrastructures and politics behind information management , be that as ranking systems , newsfeeds , recommenders and a variety of content moderation systems . There is now an understanding that AI systems — LLMs, in particular — come to replace these as ever more intuitive and precise personalization systems. Seamless as this shift may feel, the political implications remain profound: both algorithmic and AI-based information systems conjugate the circulation of information with societal norms.
A central political consequence of this shift is the question of who gets to decide and operationalise societal norms into technical systems. Companies certainly have most decision-making power: a key aspect of “platform power”, transposed here to the emerging area of AI, is the “endogenous” nature of their capacity to create the environmental conditions in which markets and societal sectors function . But they retain dependencies with other actors that help legitimise them, particularly in governance frameworks (informal or otherwise) that influence the basic norms (and procedures) of alignment — be these epistemic (definitions of “misinformation” and other types of information integrity); bias (representational, political or other); harm (personal or societal); and myriads of other “essentially contested” concepts . Safety criteria, red teaming and benchmarking, for example, stand out as frequent entry points for third parties (nonprofits, think tanks, academics and regulators) to propose various alignment criteria as a nascent market of “independent verification organisations” .
Still, making room for the “essentially contested” aspect of these norms — a room for public participation and deliberation — is still an exploratory agenda . Part of it is a question of inclusive design: who is consulted about alignment norms, what is their participation (and applicability) in alignment processes, and how does one highlight contestation where it occurs? Another is a question of negotiation — or struggle, should we confront the irredeemably diverse interests of AI governance stakeholders. How are different actors invited to deliberate and collaborate in alignment ? Who and what moderates this process? Perhaps inconclusively, much has been said of the struggle to open algorithmic systems to a wider market or to regulatory frameworks , with middleware and algorithmic pluralism being a few of the latest conceptual propositions brought to the table of policy makers. Long before then, scholars and nonprofits had called for a variety of “participative”, “collaborative” or “reflective” methods to be embedded in platform design.
But whether or how this applies to AI alignment is a question that still waits to be explored. We begin with the how and who: who participates in alignment processes, and in what current (and prospective) capacity?
To tackle this empirically, we can begin charting AI alignment as an interdependent stack of technical and normative interventions, each of which invite contributions from different actors. There is the question of what values to train: their definition and provenance (from e.g. legislation or public norms); how they trickle down into specific “ideal conducts” by which a model should behave, as well as risks to be mitigated; data cleanup and curation processes; the training of each conduct and mitigation of each risk; and their benchmarking and evaluation. Each of these stack components may assemble a variety of “governance mechanism” : the model cards, “AI charters” and principles, company updates where norms are spelled out in “transparency” exercises ; how these norms call for different expertise and consultations, particularly in red teaming ; what benchmarks are adopted from a wider market; and so on. And each of these tend to involve a different set of actors with varying capacities: academics, nonprofits specialising in projecting AI risks (Apollo, METR, etc.), regulators, state-funded AI safety institutes, and so on.
In the tradition of actor-network theory , we map the presence of each of these actors across each stack. We devise an LLM-assisted extraction pipeline to perform a “distant reading” of 213 alignment-relevant documents (model and system cards, company policies, “principles”, grant calls, audits, blog announcements and research papers) from 29 companies across the US, Europe and China. We extract the names and types of actors (nonprofits, governmental, internal, private, academic) mentioned in relation to specific alignment processes: the definition of alignment values; the training of each value or mitigation of each risk; and their benchmarking.
Though these documents are of course not representative reflections of everything that occurs in alignment processes (and everyone involved), they do in fact participate in the governance process of AI companies: they perform public duties towards “responsible democratization of machine learning” primarily by responding to the sensitivities of regulators and public debates about the prospective use of their models, and position themselves in a wider market of other alignment norms and technical performance benchmarks. In passing, they cite academic papers that, no matter how technical, are one of the main sites of normative deliberation in alignment. They also cite conceptual influences from broader public debate; nonprofits with whom they have worked to benchmark or red team their models; public consultation processes; or AI safety think tanks with whom (or in response to whom) they have determined a particular alignment norm.
We are cautious about ascribing the motivation of an AI company to involve one or another actor, and even less about the impact these actors might have. We restrict ourselves to analysing the ways in which each actor is cited, and thus indirectly “enrolled” in a company's alignment process. Some actors are co-authors, and may thus indicate a greater degree of interdependency with companies. Others may be mentioned as influences; as somewhat tokenised representatives of certain public value concerns (e.g., child protection advocacy groups); or not mentioned at all — a valuable data point nonetheless. We take note of how they are cited and how enrolled they are throughout the alignment process — from the definition of ideal conducts and risks, to their training and mitigation, to their benchmarking.
We find that, while multiple actors are involved, each is allocated to specific and selective roles: governmental actors as initiators of “alignment standards” and producers of “AI security” norms and benchmarks; academics as benchmark producers and legitimisers of performance metrics; nonprofits as a hybrid category of value formulators and intermediaries with public consultation processes; and private firms concentrated in cybersecurity and red-teaming. While some contents of current alignment methods have been subject to more collaborative — if “democratic” — endeavours (e.g., collective constitutional AI), the process itself remains almost entirely at the discretion of AI companies. We close by reflecting on what this distribution of roles implies for the political theory of AI governance, and what alternative models of legitimacy a more pluralistic alignment infrastructure might require.
The visualisation rests on a structured reading of 234 alignment documents — model cards, system cards, company policies, principles documents, grant calls, audits, blog announcements and research papers — analysed end-to-end by a large language model against a fixed schema. Each document yields one or more rows of the form (document × conduct or risk × training method × benchmark × actors), recording not only the actors named at each stage but the capacity in which each is enrolled. The result is a flat dataset of 2,563 rows describing 154 conducts, 828 risks, 55 training techniques and some 900 benchmarks across 29 companies.
The corpus combines two parallel sources. The first is a Zotero bibliography of 181 alignment-relevant publications — model cards and system cards from the major labs (OpenAI, Anthropic, Google, Google DeepMind, Meta, xAI, Mistral, DeepSeek, Microsoft, Alibaba, Black Forest Labs, Aleph Alpha, and others), accompanied by published principles documents, charters, audit reports and grant calls. The second source is a folder of 53 deduplicated company policies: acceptable-use, privacy, terms-of-service, developer terms, data-processor agreements and related documents, tracked through time via dated snapshots stored as markdown.
Each document is classified into one of ten types — model card, company policy, principle, grant, announcement, audit, initiative, report, research, or other — with an optional sub-type for policies (Acceptable Use, Privacy, Terms of Service, Developer Terms, Data Processor Agreement, etc.). For company policies, the schema is restricted to information that informs how the model is trained, moderated, filtered or evaluated; user-facing usage rules, billing terms and account terms are explicitly excluded. This keeps the dataset focused on alignment governance and avoids drowning it in the boilerplate of consumer contracts.
Each document is first converted to plain text — PDFs through PyMuPDF, web
pages and Zotero HTML snapshots through
trafilatura,
policy snapshots read directly from their markdown — and then read in a
single request to GPT-5.4, submitted asynchronously
through the OpenAI Batch API. The system prompt defines the schema, the
document-type taxonomy, the conduct and risk vocabularies, the
training-method categories, the actor-type and actor-category
vocabularies, the citation-type vocabulary, and slug-generation rules for
stable cross-document identifiers. The model is instructed to return a
JSON object with a single key rows, whose value is an array
of row objects. Row granularity is one row per (conduct or risk ×
benchmark) combination found in the document. A conduct or risk
with three benchmarks therefore generates three rows, in which the
training/mitigation fields are repeated; a conduct or risk discussed but
not benchmarked still produces one row, so it appears in the dataset.
For each row the schema records 56 fields, grouped as follows:
pub_type,
policy_type, plus the Zotero-derived key,
item_type, pub_year, pub_author,
pub_title, pub_url.company,
company_country, company_model.pub_actors,
pub_actors_type, pub_actors_cat,
pub_actors_cite, pub_actor_type_id,
pub_actor_page). The same actor category and citation type
accompany every actor field below.risk_conduct_definition, risk_conduct_category,
risk_conduct_type, risk_conduct_item,
risk_priority_signal,
pub_risk_conduct_definition_page).specific_risk_conduct_actor,
specific_risk_conduct_actor_category). When the document
defines the concept inline without external attribution, both fields are
null.risk_conduct_training_verbatim,
risk_conduct_training_category, pub_training_page,
risk_conduct_training_actor,
risk_conduct_training_actor_type).risk_conduct_benchmark,
risk_conduct_benchmark_author,
risk_conduct_benchmark_author_type,
pub_benchmark_page).external_evaluator,
external_evaluator_type, pub_evaluator_page).Because extraction runs as a batch, the whole corpus is submitted in one request set and collected when complete; of the 234 documents, 184 yield extractable alignment content, the remainder being policy snapshots with no training-relevant material. A small number of the most safety-dense documents are refused by the provider's input-safety filter rather than processed, and are recovered separately so that none is silently dropped (see Limitations).
Free-form extraction is paired with several stages of controlled
vocabulary. The model is constrained to choose one of nine
actor types — internal, private,
academic, research institute, governmental,
nonprofit, public, public consultation,
other — to which we add the meta-category multiple when a
row contains several types. Alongside this coarse type we apply a finer
sixteen-class actor category, distinguishing
alignment-research organisations, public-consultation bodies, advocacy
groups, coalitions, industry associations, academic centres, think tanks,
foundations, legal-advocacy organisations, red-teaming and cybersecurity
firms, government agencies, internal actors, the authoring AI company,
other AI companies, publics, and an open "other"; and an eleven-class
citation type recording how each actor is enrolled — as
author, formal partner, external evaluator, third-party red-teamer,
benchmark or framework contributor, normative influence, company
initiative, standards or regulatory framework it maps to, coalition it has
joined, funding relationship, or temporary initiative. Because the model
occasionally improvises labels outside these vocabularies, a harmonisation
pass collapses every extracted category and citation value to a single
canonical term — the authoring company and its named staff, for example,
to "internal". Conduct categories are drawn from a
17-class vocabulary covering norms such as Honesty,
Helpfulness, Harmlessness, Anti-discrimination,
Political neutrality, Factuality, Reasoning,
Safety and Knowledge. Risk categories
are drawn from a 40-class vocabulary covering specific risk areas —
CBRN, Cybersecurity, Bias, Sycophancy,
Reward hacking, Persuasion, Misalignment,
Hallucinations, Child Safety, Jailbreaking, and
so on. Where a document uses language that does not map, the model returns
Other: <specify>; these are reviewed and either folded
back into the vocabulary or kept as an open category.
Training and mitigation methods are mapped to one of eight categories: Preference learning (RLHF, RLAIF, DPO, reward modelling), Supervised training (SFT, instruction tuning, demonstration fine-tuning, deliberative alignment), Reward shaping (adversarial reward functions, reward capping, rule-based rewards), Data curation (filtering, deduplication, synthetic data, PII redaction), Adversarial training (red-teaming data, adversarial examples), Inference-time filtering (content classifiers, safety filters, moderation APIs), Hard constraints (system-prompt rules, refusal policies, hardcoded refusals), and Runtime safeguards (trip wires, model lookahead, runtime monitoring, scratchpad oversight, sandboxing). A pipe separator is used when several categories co-apply.
The prompt is strict about what counts as a training method.
Constitutional frameworks (e.g. Constitutional AI,
Anthropic's Constitution) are treated as sources of
norms — listed in pub_actors or
risk_conduct_definition — and not as techniques; the
underlying technique is RLAIF or SFT. Narratives about how a risk arises
are likewise excluded from the training fields.
A single further pass — again GPT-5.4, run as one batch over the unique values discovered in the data — adds four derived columns. It harmonises each verbatim conduct or risk name to a canonical label; maps each training description to a specific technique within a structured per-category taxonomy (Preference learning → RLHF, RLAIF, DPO, KTO, reward modelling; Supervised training → SFT, deliberative alignment, rejection-sampling fine-tuning; and so on); categorises each benchmark into one of 24 domains — Safety / Harmlessness, Refusal / Over-refusal, Honesty / Factuality, Hallucination, Bias / Fairness, Toxicity, Cybersecurity, CBRN, Privacy, Jailbreak / Adversarial, Agentic / Autonomy, the capability sub-categories (coding, reasoning, maths, multilingual, multimodal, long context, instruction following), Persuasion, Sycophancy, Reward hacking, Misalignment / Scheming, Internal evaluation suite, Other; and sorts every risk onto a priority line.
The priority line is the one classification deliberately withheld from the per-document reading. Each risk is sorted into a red, mid or low line by combining the document's own priority signal with the cross-company frequency of the risk and a fixed rubric: red lines are the unacceptable transgressions a company treats as absolute — CBRN, child sexual abuse, catastrophic loss of control; mid-lines are the contested, grey-area concerns that recur intermittently and whose definitions are disputed — political neutrality, persuasion, "grey area" content; low-lines are the risks rarely emphasised across the industry. Because the cross-company frequency is a corpus-level property, this judgement cannot be made by reading a single document in isolation, and is reserved for this aggregate pass.
The pass works on the deduplicated set of values rather than the full 2,563 rows, which keeps the API cost modest (roughly a thousand unique conducts and risks, a thousand unique benchmarks); the batch interface in turn halves the per-token cost relative to synchronous calls.
The flat CSV is reshaped into a compact JSON file
(docs/data.json) consumed by the front-end. Conducts and
risks are aggregated on (company × item); training methods on
(company × risk/conduct category × training item); benchmarks on
(company × risk/conduct category × benchmark name).
For each aggregated item, the actor types appearing on its source rows are split into two sets:
pub_author or company, or whose
citation type marks active contribution (so a co-announcing government
agency counts as an author rather than a citation).specific_risk_conduct_actor explicitly credited with defining
the conduct or risk.Square colour is section-specific:
specific_risk_conduct_actor_category if a specific actor is
explicitly credited; otherwise the source-document author type.risk_conduct_training_actor_type.risk_conduct_benchmark_author_type.In every case the colour resolves to a single type when the contributing actors share one, "multiple" when more than one, or "unknown" otherwise. The cited types do not influence colour. They drive the second state of the actor-type chips: clicking a chip once highlights items where the type is the AUTHOR for that block; clicking again broadens the highlight to items where the type appears anywhere else on the row (the "+ cited" superset).
Four limitations are worth stating up front. First, model cards and policies are governance artefacts: they record what companies choose to disclose. Their silences are themselves a finding, but it means the dataset reproduces the asymmetries of disclosure across the industry — well-documented labs (OpenAI, Anthropic, Google) are over-represented relative to those that publish less. The visualisation's "top ten by document count" is not a ranking of alignment activity but of alignment disclosure.
Second, LLM-assisted extraction trades exhaustive human reading for coverage. The model may miss content buried in footnotes, conflate training methods with risk narratives, or assign actor categories inconsistently across documents that use similar phrasing for different parties. Each row preserves the verbatim source passage and a page number to facilitate manual verification, and a separate validation script checks every row's join keys against the bibliography and the internal consistency of the parallel actor fields.
Third, the controlled vocabularies are necessarily incomplete. Vocabularies have grown iteratively as the documents are read; new categories are added where companies use language the existing taxonomy cannot capture (Multilingual incoherences, Encouraging mental illness, Historical revisionism, Reward hacking). The benchmark taxonomy in particular collapses substantively different evaluations into a single bucket where their published descriptions are too brief to disambiguate (e.g. internal company suites that share a name but differ in content from one model version to the next).
Fourth, and more incidentally, a handful of the most safety-dense documents — system cards detailing CBRN, jailbreak or self-harm evaluations — are refused outright by the extraction model's own input-safety filter, which flags benign benchmark examples (a multilingual trivia question, say) rather than anything genuinely disallowed. These documents are recovered through a fallback to a second model with no equivalent filter. That the very material which makes a document a safety document is what trips an automated safety classifier is a small but apt illustration of the frictions this paper describes.
Who writes alignment documents? Who is cited?
Document colour = author type. Click an actor in the sidebar to highlight the documents they author; click again (+ cited) to also see the documents that cite them via specific_risk_conduct_actor, training-actor, benchmark-author or external-evaluator.
Who specifically defines a conduct? If unknown, who is the author of the document mentioning it?
Square colour = specific_risk_conduct_actor_category when explicitly credited; otherwise the author of the source document. Click an actor for AUTHOR; click again (+ cited) to also see items where that actor appears as specific_risk_conduct_actor.
Who specifically defines a risk? If unknown, who is the author of the document mentioning it?
Square colour = specific_risk_conduct_actor_category when explicitly credited; otherwise the author of the source document. Click an actor for AUTHOR; click again (+ cited) to also see items where that actor appears as specific_risk_conduct_actor.
Which training methods are used for each risk or conduct category?
Rows = risk_conduct_category; squares = training methods. Square colour = type of the named risk_conduct_training_actor (who actually did the training). Click for AUTHOR; click again (+ cited) for any other actor in the row — including the document's authors.
Which benchmarks are used for each risk or conduct category?
Rows = risk_conduct_category; squares = benchmarks. Square colour = type of the named risk_conduct_benchmark_author. Click for AUTHOR; click again (+ cited) for any other actor in the row — the document's authors, training actors and explicit credits.
The most-mentioned actors by type. Pick a type and how many to show.
| # | Actor | Type | Role | Based in | Collaborated with | Documents | Mentions |
|---|
How enrolled is each type of actor throughout the alignment stack?
One stack = one group of actors (one actor type, or one company when you switch the Group dropdown). The four parallelograms inside it are the four alignment stages, top to bottom:
Each circle is one actor (a person, an organisation, a consortium) that appears in at least one document touching that stage.
Click any node to isolate the network it belongs to: the actor lights up wherever it appears across the four stages and the tan edges linking it between components are highlighted, while everything else dims. Click the background or press Esc to clear.
A node's size is the relative frequency with which that actor is mentioned within that component (Conducts, Risks, Training or Benchmarks). Radius is sqrt-scaled against the most-mentioned actor on the plane, so a government agency named in many risk rows shows a bigger circle on the Risks plane than one named once.
The same actor can therefore be a large node on one plane and a small one on another, depending on where it is named most. Position within the plane is a force-directed layout (repulsion + a weak central pull) and carries no meaning on its own — only size and the cross-stage edges do.
Labels. By default no circles carry a permanent label. Hover any circle to reveal its actor name transiently; tick Show labels to label every node.
An edge appears when the same actor is mentioned in more than one component. For example, if a government agency is named under both Conducts and Training, it appears as a node on each plane and the two are joined by an edge. The denser an actor's edges, the more stages of the alignment stack it is enrolled in.
There are no within-plane edges: lines only ever run between components, tracking one actor's reach up and down the stack.
A tan line connects the same actor appearing on two stages — i.e. it tracks an actor's enrolment through the alignment stack. Each one is split at its midpoint:
Read denser tan threads as: this actor type is enrolled in more than one stage of alignment.
Which actors from the wider alignment ecosystem are — and are not — actually named in companies' AI training documents.
Each ribbon is one of the 512 actors catalogued in the alignment-actor directory, flowing Country → Type → Mentioned in AI training documents. “Mentioned” means the actor is cited by name somewhere in the model-card / system-card corpus. Ribbons are coloured by country; hover a band or node for the actor count, click a node to isolate its flows. Only 26 reach the training corpus.
Pick a conduct or risk. See how it is defined across companies and time, which other concepts it travels with, and which actors have shaped its definition together.
Central node = the selected category. Surrounding nodes = every item in it, coloured by actor type. Drag, scroll to zoom, ⊕ to recenter.
Actors involved in defining, training and benchmarking — coloured by actor type. Click a node to inspect; drag to move, scroll to zoom, ⊕ to recenter.