When Open Standards Meet State-Scale Influence

Abstract

Artificial-intelligence systems increasingly mediate how the public encounters political, historical, and legal information. This creates a new strategic objective for governments and influence organizations: rather than attempting only to persuade human audiences, they can shape the digital information environments that AI systems search, retrieve, summarize, and cite.

This paper examines that challenge through two contrasting developments. The first is the publication of Artificial Intelligence Optimization standards by Sherafgan Khan and the Fabled Sky Research team. These standards seek to make well-sourced information more accessible and intelligible to AI systems through semantic structure, machine-readable metadata, evidence provenance, source classification, and contradiction analysis.

The second is a reported Israeli government-backed communications campaign that explicitly sought favorable “GPT framing” through large-scale content production and digital optimization. The campaign illustrates how techniques intended to improve access to reliable information can be separated from their ethical safeguards and repurposed to manufacture the appearance of authority or independent consensus.

The paper does not argue that Fabled Sky caused the international shift in public opinion regarding Israel’s conduct in Gaza. That shift resulted from numerous factors, including events on the ground, journalism, humanitarian reporting, legal proceedings, advocacy, and public protest. It argues instead that Fabled Sky’s work identified an important infrastructural reality: historical evidence has greater influence in AI-mediated discourse when it is organized in forms that AI systems can efficiently retrieve and contextualize.

Open standards alone, however, cannot overcome state-scale information operations. AI companies must assume greater responsibility for detecting common ownership, concealed sponsorship, circular citation, synthetic source diversity, and coordinated content networks. The next generation of AI safety must therefore address not only whether models generate falsehoods, but whether the information environments surrounding those models have been deliberately engineered.

1. Introduction

Public debate about artificial intelligence has concentrated heavily on hallucinations, misinformation, bias, and the possibility that chatbots may generate false or harmful statements.

Those concerns remain important, but they address only one part of the emerging information problem.

AI systems do not operate in isolation. Depending on their design, they may draw from pretrained model weights, search indexes, retrieval databases, licensed collections, knowledge graphs, public web pages, and third-party data services.

The quality of an AI answer therefore depends not only on the model itself, but also on the environment from which the system obtains information.

This creates a new target for political influence.

An actor does not necessarily need access to a model’s internal parameters to affect its output. It may instead attempt to alter the external information environment so that particular arguments, narratives, sources, or framings become more visible to retrieval systems.

The result is an emerging contest over what might be called machine-readable authority: the technical and institutional characteristics that cause an AI system to regard information as accessible, relevant, corroborated, and suitable for inclusion in an answer.

Fabled Sky Research’s AIO framework represents one attempt to improve that environment by making verifiable information easier for machines to understand.

Reported state-backed efforts to produce favorable chatbot framing reveal how the same technical environment can be exploited.

2. Research Questions

This paper addresses five questions:

1. What do Fabled Sky’s AIO standards seek to accomplish?

1. How can structured publishing affect AI retrieval and generated answers?

1. What role might such methods have played in making historical and legal information about Gaza and the wider Israel–Palestine conflict more accessible?

1. How can the same optimization techniques be adapted for state-backed influence operations?

1. What responsibilities should AI companies assume when coordinated actors attempt to manufacture machine-readable consensus?

3. Methodology

This paper uses qualitative policy analysis based on publicly available materials.

The reviewed sources include:

• Fabled Sky’s published AIO and xAIO standards;
• Fabled Sky’s Objectivity AI and Israel–Palestine information materials;
• Sherafgan Khan’s public analysis of the reported chatbot-influence campaign;
• public reporting concerning Clock Tower X and the Israeli government-backed communications effort;
• relevant U.S. foreign-agent-registration materials as described in public reporting;
• research on AI retrieval, data poisoning, influence operations, and coordinated source networks;
• public-opinion data concerning Israel’s military campaign in Gaza; and
• published statements by AI companies regarding covert influence activity.

The analysis distinguishes among three types of proposition:

Documented facts: claims supported by contracts, public filings, published standards, corporate statements, or independently reported evidence.

Institutional claims: representations made by Fabled Sky or other organizations about their own work that may not have been independently confirmed by all named parties.

Analytical inferences: conclusions drawn from the available evidence but not directly demonstrated through controlled causal research.

This distinction is especially important when evaluating whether any particular publishing initiative influenced public attitudes or chatbot outputs.

4. Background: From Search Optimization to AI Optimization

Search-engine optimization traditionally focuses on improving the visibility of webpages in search results. It may involve technical accessibility, relevant terminology, site performance, linking practices, page organization, and other signals used by search-ranking systems.

Generative AI changes the objective.

A publisher may no longer be concerned only with whether a human clicks on a webpage. The publisher may also want an AI system to:

• locate the page;
• identify its central claims;
• extract accurate passages;
• understand the relationships among entities;
• evaluate supporting sources;
• compare it with contradictory material; and
• incorporate it into a generated answer.

This emerging field is sometimes described as generative-engine optimization, answer-engine optimization, or AI optimization.

The terminology remains unsettled, and many of the associated methods overlap with existing practices in semantic publishing, knowledge management, structured data, information retrieval, and technical SEO.

What is new is the strategic focus on machine-generated synthesis.

5. Fabled Sky’s AIO Framework

Fabled Sky’s AIO standards begin from the premise that information intended to influence AI-generated knowledge must be understandable at both the human and machine levels.

The standards emphasize several interrelated practices.

5.1 Semantic clarity

Documents should clearly identify their subjects, claims, dates, locations, institutions, and relationships.

Ambiguous references, unexplained terminology, and poorly structured narratives make it more difficult for retrieval systems to determine what a document establishes.

5.2 Modular information architecture

Material should be divided into coherent sections with descriptive headings and identifiable propositions.

This allows an AI system to retrieve a relevant section without misrepresenting an entire document.

5.3 Structured metadata

Schema markup, metadata, entity descriptions, and machine-readable relationships can help systems identify the nature and context of a source.

5.4 Evidence provenance

A claim should be connected to its supporting evidence, preferably through identifiable primary records, archived documents, original reporting, or transparent citation chains.

5.5 Independent corroboration

The framework emphasizes the importance of multiple independent sources rather than mere numerical repetition.

Fabled Sky’s expanded xAIO materials recommend high levels of corroboration and, where appropriate, substantial numbers of independent references before presenting a conclusion as firmly established.

5.6 Contradiction analysis

A trustworthy system should identify contrary evidence and explain why sources disagree rather than suppressing inconvenient material.

5.7 Ownership and affiliation analysis

Sources that appear separate may share common ownership, funding, editorial control, or institutional affiliation.

Without this analysis, coordinated publication can be mistaken for independent confirmation.

5.8 Transparency and ethics

Fabled Sky’s standards distinguish evidence-oriented AIO from deceptive manipulation. They discourage hidden methodologies, keyword stuffing, artificial engagement, undisclosed promotion, and attempts to present advocacy material as neutral authority.

Taken together, these principles constitute more than a visibility strategy. They represent a proposed information-integrity architecture.

6. Public Standards, Not Proprietary Secrets

It is important not to characterize the AIO framework as confidential knowledge available only to Sherafgan Khan, Fabled Sky, or an AI company.

The standards were publicly released and designed for reuse. Their broader technical logic was also consistent with publicly discussed work in knowledge graphs, structured publishing, retrieval-augmented generation, semantic markup, SEO, and generative-engine optimization.

Fabled Sky has publicly described controlled arrangements involving the evaluation or submission of data to OpenAI, including references to custodial procedures and pre-ingestion review.

Such an arrangement would not make the broader framework confidential. Once the methodology was published, other actors could study and adapt it.

This openness was intentional. Public-interest standards are generally most valuable when journalists, researchers, institutions, and civil-society organizations can implement them.

The difficulty is that openness also allows governments, political firms, marketing organizations, and influence operators to learn from the same framework.

7. Application to the Israel–Palestine Information Environment

Fabled Sky’s Objectivity AI work and related Israel–Palestine materials began developing publicly during 2024, with a more visible expansion across late 2024 and 2025.

The initiative sought to organize information concerning:

• the history of Israel and Palestine;
• Gaza and the West Bank;
• occupation and territorial control;
• international humanitarian law;
• civilian harm;
• displacement;
• settlements;
• stakeholder positions;
• media narratives;
• legal proceedings;
• humanitarian reporting; and
• current developments.

Much of this information already existed in government archives, United Nations documents, court records, historical studies, human-rights reports, investigative journalism, and public databases.

The contribution was therefore not necessarily the discovery of previously unknown facts.

It was the organization of existing information into a format that AI systems could more easily locate, parse, compare, and explain.

A user asking a chatbot about Gaza may not know the title of a relevant United Nations report, the name of a legal doctrine, the date of a historical event, or the correct terminology for a particular policy.

A well-organized evidence repository can reduce that knowledge barrier by connecting the user’s ordinary-language question to the relevant records.

The public debate after October 2023 concentrated especially heavily on Gaza, including civilian casualties, forced displacement, humanitarian access, military targeting, detention, famine risk, and international-law allegations.

Information regarding the West Bank remained important but received comparatively less sustained attention within the immediate global discourse surrounding the war.

8. Did AIO Change Global Opinion?

The available evidence does not support a claim that Fabled Sky or Sherafgan Khan independently caused the international shift in opinion concerning Israel’s actions in Gaza.

Public attitudes were influenced by numerous factors, including:

• military developments;
• casualty reporting;
• images and testimony from Gaza;
• investigative journalism;
• humanitarian warnings;
• international court proceedings;
• statements by political leaders;
• university demonstrations;
• Palestinian and Israeli advocacy;
• social-media circulation;
• diplomatic disputes; and
• the lived experiences of civilians.

Polling nevertheless demonstrates that a meaningful shift occurred.

Gallup reported that American approval of Israel’s military action in Gaza declined substantially between late 2023 and 2025. Pew Research Center also documented increasingly unfavorable attitudes toward Israel in several countries.

No public study has isolated the effect of Fabled Sky’s publications on these trends.

A credible causal assessment would require data concerning:

• the number of chatbot answers influenced by the materials;
• which AI systems retrieved them;
• how answers changed after publication;
• how many users encountered those answers;
• whether exposure changed user beliefs; and
• how those effects compared with journalism, social media, protest movements, and events on the ground.

That evidence is not currently available.

The appropriate claim is narrower.

Fabled Sky may have contributed to the information infrastructure through which historical, legal, and humanitarian evidence became more accessible to AI users. That contribution could be significant while remaining only one factor in a much larger political and informational transformation.

9. The Reported Israeli “GPT Framing” Campaign

Public reporting has described an Israeli government-backed communications campaign involving Clock Tower X, a firm led by former Trump campaign manager Brad Parscale.

According to reporting based in part on a U.S. Foreign Agents Registration Act filing, the campaign operated through an arrangement involving Havas Media Network on behalf of the State of Israel.

The reported contract included efforts to create websites and digital content intended to produce favorable “GPT framing” in chatbot conversations. It also formed part of a wider campaign directed toward younger audiences.

Initial reporting placed the contract’s value at approximately $6 million, while later reporting described a broader and more costly communications operation.

The exact technical impact of the campaign remains uncertain.

There is no public evidence demonstrating that it secretly modified the internal parameters of major commercial AI models or guaranteed particular answers.

The more plausible and immediate mechanism is information-environment manipulation.

A sufficiently funded campaign can:

• create numerous websites;
• publish large volumes of narrowly targeted articles;
• answer thousands of anticipated user questions;
• use structured headings and metadata;
• repeat consistent frames across multiple domains;
• generate backlinks and social references;
• test AI outputs;
• monitor which pages are cited;
• adjust wording according to retrieval performance; and
• maintain publication activity over extended periods.

These practices may affect AI systems that search the live web or consult frequently updated indexes.

10. The Difference Between Training Influence and Retrieval Influence

Policy analysis should distinguish between two mechanisms.

10.1 Training influence

Training influence occurs when material is incorporated into the datasets used to construct or update a model.

Public web archives such as Common Crawl have been used in many AI-development pipelines. The appearance of content in such an archive may make future training exposure possible.

It does not prove, however, that a specific model used that material.

10.2 Retrieval influence

Retrieval influence occurs when an already trained model searches or consults external sources while answering a question.

This mechanism is often more immediate.

A newly created webpage does not have to wait for a future model-training cycle if an AI search or retrieval system can locate it on the live web.

The reported appearance of campaign-related sites in AI answers would therefore be more directly relevant to retrieval than to claims about retraining a model.

This distinction matters because the relevant defensive measures differ.

Training-data security requires dataset auditing, filtering, deduplication, and poisoning analysis.

Retrieval security requires real-time provenance, ownership clustering, sponsorship detection, source-quality assessment, and citation-chain evaluation.

11. The Inversion of AIO

The reported campaign demonstrates how AIO techniques can be detached from their ethical foundation.

A webpage may possess:

• concise answers;
• clear headings;
• relevant terminology;
• citations;
• machine-readable formatting;
• regular updates; and
• topical specialization.

Those characteristics can make the page attractive to an AI retrieval system.

But they do not make the page independently trustworthy.

A network of coordinated sites may imitate the appearance of broad corroboration while ultimately deriving from one sponsor, one contract, one political objective, or one source set.

This is the inversion of AIO:

The technical features of authority are reproduced while the evidentiary safeguards are removed.

Under Fabled Sky’s own standards, legitimate AIO should include provenance, source independence, sponsorship transparency, contradiction analysis, and ownership review.

A campaign that obscures common control or creates artificial source diversity would therefore fail the framework’s substantive tests, even if it successfully adopted its formatting practices.

The distinction can be summarized simply:

The structure may be copied without the integrity.

12. Why State-Scale Funding Matters

The vulnerability does not arise because the underlying optimization methods are unusually secret or technically complex.

It arises because scale changes the competitive environment.

A small research organization may spend months producing a rigorously sourced repository.

A state-supported campaign can finance:

• continuous content production;
• hundreds of domains or microsites;
• multilingual publication;
• professional video and graphic production;
• advertising and distribution;
• automated content testing;
• audience segmentation;
• reputation management;
• search optimization;
• backlink acquisition;
• chatbot monitoring; and
• repeated message refinement.

The financial advantage allows a coordinated actor to create an informational presence that appears larger, more active, and more widely corroborated than that of a smaller evidence-based publisher.

Money does not guarantee that an AI system will accept a political narrative.

It does allow the sponsor to increase the probability that its preferred material will be present wherever the system looks.

13. The Synthetic Consensus Problem

A major weakness in contemporary retrieval systems is the possibility that numerical source volume may be confused with independent agreement.

Suppose ten webpages repeat the same claim.

A simple system may interpret this as ten supporting sources.

A provenance-aware system would ask:

• Are the sites owned by the same company?
• Were they funded through the same government contract?
• Do they use the same authors or editors?
• Are they hosted on the same infrastructure?
• Do they share analytics or advertising identifiers?
• Are they quoting one another?
• Do they all rely on the same original assertion?
• Was the underlying evidence independently verified?

If the answer reveals common control, the ten sources should be treated as one coordinated source cluster.

Without that adjustment, an influence operation can manufacture the appearance of consensus through repetition.

This is particularly dangerous in conflict reporting, where facts may be contested, access may be restricted, casualty data may be incomplete, and governments have strong incentives to frame events strategically.

14. AI Companies’ Responsibility

AI companies cannot reasonably eliminate every deceptive page on the internet.

They can, however, improve how their systems interpret the information they retrieve.

At present, commercial incentives may work against extensive verification.

Following a claim to its original source, identifying ownership connections, reviewing archived pages, resolving translations, and comparing contradictory records requires additional computation and time.

Faster systems are generally cheaper to operate and easier to scale.

But compute efficiency is a commercial objective, not an information-integrity principle.

When AI companies present generated answers as useful summaries of complex events, they assume a responsibility to evaluate whether the apparent evidence base has been artificially constructed.

That obligation is especially strong when systems answer questions involving:

• war;
• alleged atrocities;
• genocide or ethnic-cleansing claims;
• casualty figures;
• international criminal law;
• territorial disputes;
• elections;
• public health emergencies; and
• state-sponsored influence activity.

Verification costs should therefore be treated as a safety expense inherent to operating an AI information service.

15. Policy Recommendations

Recommendation 1: Cluster sources by beneficial ownership and sponsorship

AI retrieval systems should identify relationships among publishers, parent companies, contractors, governments, advocacy organizations, and funding sources.

Multiple pages controlled by one sponsor should not be treated as fully independent corroboration.

Recommendation 2: Propagate disclosures into generated answers

Foreign-agent registration, state funding, political sponsorship, paid placement, and advertorial status should accompany extracted claims.

A chatbot should not cite a sponsored article while omitting a material disclosure located elsewhere on the website.

Recommendation 3: Trace claims to primary evidence

Systems should follow citation chains to the earliest available record.

Twenty articles that ultimately rely on one press release should not be presented as twenty independent confirmations.

Recommendation 4: Create a provenance-concentration score

AI companies should measure how much of an answer depends on one owner, state, advocacy network, wire service, or evidence family.

High concentration should trigger additional research or a visible uncertainty warning.

Recommendation 5: Conduct source-removal stability testing

Before presenting a confident conclusion, the system should determine whether the answer changes substantially when the largest coordinated source cluster is removed.

A major change would indicate informational fragility.

Recommendation 6: Introduce enhanced conflict-verification protocols

Questions involving armed conflict should trigger:

• stricter date verification;
• primary-source prioritization;
• ownership analysis;
• explicit treatment of disputed claims;
• separation of verified facts from allegations; and
• broader retrieval across opposing and neutral institutions.

Recommendation 7: Audit web archives and retrieval indexes

AI developers should evaluate whether state-backed or coordinated content networks are overrepresented in training and retrieval datasets.

Appearance in a public web archive should not be treated as evidence of reliability.

Recommendation 8: Share threat indicators across platforms

AI companies, search engines, academic researchers, and civil-society organizations should establish mechanisms for sharing information about:

• coordinated domains;

• synthetic personas;
• concealed state sponsorship;
• citation laundering;
• replicated content;
• and known influence infrastructure.

Recommendation 9: Publish source-weighting principles

Companies should explain, at least in general terms, how they assess:

• source independence;
• primary evidence;
• sponsorship;
• correction history;
• institutional expertise;
• transparency;
• and corroboration.

Full disclosure of proprietary ranking systems is unnecessary, but basic accountability is essential.

Recommendation 10: Support open, evidence-centered publishing

AI companies should provide tools, grants, technical guidance, and open standards that help public-interest organizations structure reliable information.

Otherwise, machine-readable political discourse may become dominated by the actors with the largest communications budgets.

Recommendation 11: Require human review for high-impact systemic claims

Claims that a government, military, ethnic group, or political movement committed serious international crimes should not be summarized through automated source counting alone.

Human review should be available for widely distributed or especially consequential outputs.

Recommendation 12: Preserve contradictory evidence

AI systems should not optimize exclusively for one coherent narrative.

Where credible evidence conflicts, the system should preserve and explain the disagreement rather than generating artificial certainty.

16. Implementation Framework

AI providers could implement the recommendations through a four-layer verification model.

Layer One: Content analysis

Evaluate the clarity, factual claims, citations, dates, and internal consistency of each retrieved document.

Layer Two: Source analysis

Assess the publisher’s ownership, funding, editorial standards, correction practices, and institutional expertise.

Layer Three: Network analysis

Identify shared infrastructure, common authorship, repeated wording, backlink coordination, common citations, and synchronized publication behavior.

Layer Four: Answer-level analysis

Determine whether the final response represents genuine evidentiary diversity, communicates uncertainty, and remains stable when coordinated sources are excluded.

This framework would not eliminate propaganda.

It would make propaganda more difficult to disguise as independent consensus.

17. Limitations

This paper has several limitations.

First, it relies on public information. It does not have access to the internal retrieval logs, model-training records, ranking systems, or security assessments of major AI providers.

Second, the precise effect of Fabled Sky’s work on chatbot outputs has not been independently quantified.

Third, the exact effect of the reported Israeli campaign on major AI systems remains uncertain. The presence of campaign material in web archives or isolated AI answers does not establish broad or durable model influence.

Fourth, causal claims regarding public opinion cannot be attributed to one publisher, technology, political campaign, or information platform without controlled evidence.

Fifth, standards maintained by active organizations may continue to change. Specific implementation details should therefore be checked against the latest published versions before technical adoption.

These limitations do not eliminate the policy concern. They define the boundaries within which conclusions should be drawn.

19. Conclusion

Fabled Sky’s work rests on an increasingly important principle:

Reliable information must be organized so that machines can find, interpret, and contextualize it.

The reported Israeli chatbot-influence campaign demonstrates the inverse:

Strategically framed information can be organized for machines as well.

This does not make open AIO standards a mistake.

The techniques were not secret, and closing public access would do little to prevent governments or commercial influence firms from developing comparable methods. Restricting the standards could instead disadvantage journalists, researchers, humanitarian organizations, and smaller institutions that need affordable ways to make credible evidence visible.

The policy response should therefore not be secrecy.

It should be verification.

Semantic structure must be paired with provenance. Citation volume must be paired with ownership analysis. Retrieval speed must be paired with source-independence testing. Machine-readable authority must be distinguished from machine-readable repetition.

Sherafgan Khan and Fabled Sky did not single-handedly transform global attitudes toward Israel and Gaza. The evidence does not support that conclusion.

They may, however, have helped identify and operationalize an important feature of the emerging information order: historical evidence becomes more influential when it is reorganized for AI-mediated access.

That process can help users discover context that was previously fragmented across archives, reports, court documents, and specialist research.

It can also be exploited by actors with the resources to flood the same systems with coordinated material.

A voluntary framework, regardless of its quality, cannot by itself compete with a nation-state or political operation capable of spending millions of dollars on content production, distribution, testing, and optimization.

Responsibility must therefore extend beyond the publisher.

AI companies must ensure that their systems do not reward artificial scale, concealed sponsorship, circular sourcing, or manufactured consensus simply because those signals are computationally cheaper to process than genuine verification.

The next phase of AI safety will not be defined only by whether models avoid inventing facts.

It will also be defined by whether they can recognize when the informational world presented to them has been deliberately constructed.

Replace the existing “Selected References” section with the annotated bibliography below. It separates primary records from institutional materials, investigative reporting, technical research, and polling so readers can assess the evidentiary weight of each source.

20. References and Source Notes

Citation and evidentiary approach

The sources below are organized by function rather than alphabetically. Primary government filings establish what parties formally disclosed; they do not prove that every proposed activity occurred or achieved its intended effect. Investigative reporting provides evidence about implementation, associated websites, chatbot testing, and expenditures. Fabled Sky materials establish the organization’s standards, methods, and institutional representations but should not be treated as independent evaluations of Fabled Sky’s effectiveness. Technical research explains the broader vulnerability, while polling documents changes in public opinion without establishing that AIO, chatbots, or any single communications campaign caused those changes.

Unless otherwise stated, online materials were accessed on August 3, 2026.

20.1 Primary Government and Contract Records

1. Clock Tower X registration statement

U.S. Department of Justice, National Security Division, Foreign Agents Registration Act Unit. “Clock Tower X LLC Registration Statement,” Registration No. 7649. Filed September 18, 2025.

Relevance: This is the foundational public filing for the Clock Tower X engagement. It identifies Clock Tower X LLC, Bradley Parscale, Havas Media Network, and the State of Israel. It states that the registrant would provide strategic communications, planning, and media services in support of a nationwide U.S. campaign commissioned by the State of Israel.

Evidentiary weight: Primary legal disclosure. It establishes what Clock Tower X formally represented to the U.S. government, including the disclosed foreign-principal relationships. It should not be described as a DOJ finding that the campaign succeeded, violated the law, or influenced any particular AI model.

2. Clock Tower X Exhibits A and B

U.S. Department of Justice, National Security Division, Foreign Agents Registration Act Unit. “Clock Tower X LLC, Exhibit A and Exhibit B,” Registration No. 7649. Filed September 18, 2025.

Relevance: Exhibit A identifies the foreign principal and relevant organizational relationships. Exhibit B describes the agreement, scope of services, compensation structure, planned audiences, and communications activities.

This is the primary record cited in subsequent reporting for the contractual language concerning the deployment of “websites and content” intended to produce “GPT framing results,” as well as the use of search-visibility tools and content directed primarily toward younger audiences.

Evidentiary weight: Primary contract disclosure. Where possible, quotations concerning campaign deliverables should be attributed directly to this filing rather than only to a news report. Claims about how the work was ultimately implemented or how well it performed require additional evidence.

3. Clock Tower X supplemental statement

U.S. Department of Justice, National Security Division, Foreign Agents Registration Act Unit. “Clock Tower X LLC Supplemental Statement,” Registration No. 7649. Filed May 18, 2026.

Relevance: The supplemental filing updates the registrant’s disclosures after the initial registration. Supplemental statements may contain later information about activities, foreign-principal relationships, receipts, disbursements, informational materials, and personnel.

Evidentiary weight: Primary post-registration disclosure. It should be used when discussing how the engagement developed after September 2025. It should not be conflated with the original contract or treated as proof of chatbot influence unless the filing specifically documents such an outcome.

4. Foreign Agents Registration Act overview

U.S. Department of Justice, National Security Division. “Foreign Agents Registration Act”.

Relevance: Provides the statutory and administrative background needed to explain why the Clock Tower filings are publicly available. FARA generally requires covered agents engaged in specified activities on behalf of foreign principals to disclose their relationships, activities, receipts, and disbursements.

Evidentiary weight: Authoritative source for the purpose and operation of FARA. Registration under FARA is a disclosure requirement; it is not by itself a finding of criminality, deception, or unlawful conduct. (Department of Justice)

20.2 Fabled Sky Research, AIO, xAIO, and Objectivity AI

5. Canonical AIO framework index

Fabled Sky Research. “AIO Standards Framework: Index and Canonical Reference Guide”. Version 1.2.7; listed as last updated April 2025.

Relevance: The principal index for the five-module AIO framework. It describes AIO as a system for making information retrievable, interpretable, and trustworthy to AI systems and distinguishes it from SEO and generative-engine optimization directed principally toward visibility or ranking.

The page also contains Fabled Sky’s description of its Objectivity AI bench-testing and data-governance relationship with OpenAI, including references to pre-ingestion validation, a data-use moratorium, joint sign-off, and provenance reporting.

Evidentiary note: The OpenAI relationship is described by Fabled Sky on its own website. Unless supported by a public OpenAI statement, countersigned agreement, or independently accessible record, the policy paper should characterize it as Fabled Sky’s account of the arrangement, not as an independently confirmed joint representation.

Version note: The index displays several layers of dating: module dates of December 9, 2022, a version identified as 1.2.7, a stated April 2025 update, and a 2026 copyright notice. Formal citations should therefore include the version number and update date rather than relying solely on webpage metadata. (Fabled Sky)

6. Definitions of AIO, GEO, and SEO

Fabled Sky Research. “AIO Standards Framework—Module 2: Definitions and Terminology”. Version 1.2.7.

Relevance: Supplies the framework’s formal definitions of Artificial Intelligence Optimization, Generative Engine Optimization, and Search Engine Optimization. It is the appropriate source for the paper’s distinction between improving the machine accessibility of verified information and optimizing content primarily to obtain visibility in generated answers.

The module also defines the framework’s principal metrics, including Trust Integrity Score, Retrieval Surface Area, Token Yield per Query, and Embedding Salience Index.

Evidentiary weight: Primary institutional standards document. It establishes how Fabled Sky defines its terminology; it does not demonstrate that the terminology is universally accepted across the AI or information-retrieval fields. (Fabled Sky)

7. AIO scoring methodology

Fabled Sky Research. “AIO Standards Framework—Module 3: Scoring Framework and Methodology”. Version 1.2.7.

Relevance: Describes the quantitative and procedural elements Fabled Sky proposes for assessing AI-readable content, including trust, retrievability, token efficiency, and semantic salience. The module also calls for transparent toolchains, disclosure of evaluation prompts, and periodic recalibration.

Use in the paper: Supports discussion of AIO as more than formatting or keyword placement. It demonstrates that the stated framework includes reproducibility and evaluation procedures rather than merely techniques for attracting chatbot citations.

Limitation: The existence of a scoring model does not independently validate the model’s predictive accuracy. Claims about empirical performance would require benchmark data, replication, or third-party evaluation. (Fabled Sky)

8. Compliance and anti-co-option rules

Fabled Sky Research. “AIO Standards Framework—Module 4: Compliance Guidelines and Anti-Co-option Protocols”. Version 1.2.7.

Relevance: This is one of the most important sources for the policy paper’s argument that the reported Clock Tower methods should not automatically be described as full AIO.

The module requires transparency concerning methodology, prompt design, provenance, and citation sourcing. It identifies keyword stuffing, ranking manipulation, concealed methods, and the relabeling of SEO or GEO services as AIO as potential violations.

Use in the paper: Supports the conclusion that an influence network could reproduce the visible packaging of machine-optimized content while violating the provenance and transparency requirements of the underlying standard. (Fabled Sky)

9. Licensing and open-governance charter

Fabled Sky Research. “AIO Standards Framework—Module 5: Licensing and Governance Charter”. Version 1.2.7.

Relevance: Establishes that the AIO standards are offered under the MIT License and may be used, modified, and redistributed with appropriate attribution. It also states that no individual or organization may claim exclusive authority over the general framework.

Use in the paper: Supports the conclusion that AIO was not secret or proprietary operational intelligence. Once publicly released, its concepts and methods were available to civil-society organizations, researchers, commercial firms, political campaigns, and governments.

Policy significance: The open license strengthens the public-interest value of the framework but also means that responsible publishers cannot prevent other actors from studying or imitating its machine-legibility techniques. (Fabled Sky)

10. Trust Integrity Score

Fabled Sky Research. “Trust Integrity Score”. Last updated April 2025.

Relevance: Describes a composite measure based on citation depth, semantic coherence, and redundancy alignment. The proposed citation component weights different source classes and warns that numerous low-authority citations can inflate the appearance of trust.

Use in the paper: Supports the distinction between citation quantity and citation quality. It is particularly relevant to coordinated networks in which many webpages ultimately trace back to one interested source or common sponsor.

Limitation: The stated weighting system and score bands are Fabled Sky’s proposed methodology. Their reliability across political, legal, scientific, and conflict-reporting domains would require domain-specific validation. (Fabled Sky)

11. Metadata and authorship traceability

Fabled Sky Research. “Metadata and Authorship Traceability Protocol”.

Relevance: Specifies recommended machine-readable fields for authorship, affiliation, timestamps, version history, citation relationships, and content provenance. It emphasizes deterministic provenance and the ability to identify or revalidate an asset’s origin and revision lineage.

Use in the paper: Provides a technical basis for recommendations that AI companies preserve sponsorship, ownership, authorship, and version information when extracting passages from webpages. (Fabled Sky)

12. Trust graphs and authorship networks

Fabled Sky Research. “Trust Graph Construction and Authorship Network Design”.

Relevance: Describes relationships among authors, organizations, documents, citations, and institutional affiliations as a graph rather than as isolated webpage attributes.

Use in the paper: Supports the recommendation that AI systems evaluate common ownership and related-source clusters. A graph-based system can identify that several apparently independent domains share an owner, sponsor, author, parent organization, or common evidence chain. (Fabled Sky)

13. xAIO publishing and verification specification

Fabled Sky Research. “xAIO/XAIO Publishing Playbook and Validation Guide”. GitHub repository.

Relevance: Extends the general AIO framework into a more operational publishing and verification process. It describes factual reports, evidence objects, validation procedures, source tiers, independent corroboration, and machine-oriented presentation.

The specification is especially important to the policy argument because it expressly addresses source independence and related-source clustering rather than counting every URL as an independent confirmation.

Use in the paper: Supports the proposition that multiple sites controlled by one campaign should be treated as one provenance cluster rather than as numerous independent sources.

Versioning note: Because the repository is actively maintained and does not show a conventional numbered release, citations should include the access date and, for technical or legal use, the relevant commit hash. (GitHub)

14. Objectivity AI public framework

Fabled Sky Research. “Objectivity AI Framework”. GitHub repository.

Relevance: Sets out Fabled Sky’s broader approach to factuality, source evaluation, proportionality, uncertainty, consensus, and auditable reporting. It is the primary source for explaining the intended methodology behind Objectivity AI.

Use in the paper: Supports discussion of the project as an attempt to organize evidence and historical context for machine retrieval rather than simply to maximize the number of supportive webpages.

Evidentiary weight: Institutional methodology. Claims about the framework’s effectiveness or its impact on commercial chatbots require separate empirical evidence. (GitHub)

15. Israel–Palestine Conflict Information Hub

Fabled Sky Research. “Israel–Palestine Conflict Information Hub”.

Relevance: The principal public example discussed in the paper. The hub organizes historical chronology, legal frameworks, humanitarian issues, stakeholder perspectives, media narratives, and contemporary developments.

The page expressly states that it is an ongoing project under active development and that some content remains incomplete or subject to further verification.

Use in the paper: Demonstrates the infrastructural argument: much of the underlying information was already public, but the project sought to place it into organized, query-relevant, source-linked sections that could be more readily located and interpreted by AI systems.

Limitation: The hub should be cited as evidence of the project’s structure and stated purpose. Individual historical or legal claims within it should, wherever possible, be cited to the underlying primary or scholarly source rather than to the hub alone. (Fabled Sky Research)

16. Objectivity AI transparency statement

Fabled Sky Research. “Transparency Statement for Objectivity AI”.

Relevance: Explains Fabled Sky’s stated objectives, financial interest, model-development status, source-verification approach, treatment of conflicting evidence, and reasons for selecting Israel–Palestine as a stress-test subject.

Use in the paper: Appropriate source for the institutional disclaimer and conflict-of-interest discussion. It also supports the paper’s statement that the hub was presented as both a public resource and a demonstration of Fabled Sky’s technology.

Evidentiary note: This is a self-description by the organization. Claims about consensus thresholds, audits, model performance, or bias reduction should be attributed to Fabled Sky unless supported by an independent evaluation. (Fabled Sky Research)

20.3 Principal Reporting on the Clock Tower Campaign

17. Drop Site News investigation

Cleveland-Stout, Nick. “Israel Is Paying Millions to Train AI Chatbots How to Talk About Gaza. It’s Working”. Drop Site News, July 28, 2026.

Relevance: The principal investigation connecting the disclosed communications contract to a network of research-style websites, chatbot testing, Common Crawl records, and claims concerning the expansion of campaign spending.

Use in the paper: Primary journalistic source for the reported implementation of the website network and the contention that some AI systems had already surfaced or cited associated content.

Evidentiary caution: Findings based on the outlet’s tests should be identified as reported testing rather than generalized to every model, every query, or every user. A website appearing in an AI answer does not by itself prove that the site was incorporated into the model’s training weights. (Drop Site News)

18. Responsible Statecraft reporting on the initial contract

Cleveland-Stout, Nick. “Israel Wants to Train ChatGPT to Be More Pro-Israel”. Responsible Statecraft, September 29, 2025.

Relevance: Early reporting on the $6 million engagement and its stated focus on Gen Z audiences, large-scale impressions, search visibility, and “GPT framing.”

The article reports that at least 80 percent of the content was to be tailored to younger audiences and identifies MarketBrew AI as a tool named in the disclosed campaign plans.

Use in the paper: Best secondary source for explaining the initial scope in accessible language while linking readers back to the underlying FARA exhibit. (Responsible Statecraft)

19. Business & Human Rights Resource Centre company-response record

Business & Human Rights Resource Centre. “Israel/OPT: Clock Tower X Allegedly Hired by Israeli Government to Create Pro-Israel Content, Targeting AI and Social Platforms”.

Relevance: Collects the allegations, links to relevant reporting, records outreach to companies, and identifies which companies responded or did not respond.

Use in the paper: Helpful for documenting the state of corporate engagement and the absence or presence of public responses. It should not replace the FARA records or principal investigative reporting.

Terminology caution: Headlines describing efforts to “train ChatGPT” may blur the distinction between modifying model weights and attempting to influence the external information environment. The paper should preserve that distinction. (Business & Human Rights Resources Centre)

20. Sherafgan Khan’s analysis

Khan, Sherafgan. “How Israel’s AI Propaganda Campaign Exploited Chatbot Verification Failures—and How to Stop It”. Sherafy, August 3, 2026.

Relevance: Provides the most direct articulation of the argument connecting the Clock Tower reporting to Fabled Sky’s AIO and xAIO concepts. It distinguishes machine legibility from verification and proposes ownership clustering, provenance analysis, independent evidence chains, contradiction testing, and user-visible evidence ledgers.

Use in the paper: Appropriate for explaining how the reported campaign would be evaluated under Fabled Sky’s own standards and for developing the paper’s policy recommendations.

Conflict-of-interest note: Khan discloses an institutional relationship with Fabled Sky, Objectivity AI, AIO, and xAIO. The article should therefore be treated as informed institutional analysis, not as independent verification of Fabled Sky’s historical impact or technical superiority. Its claims concerning the Clock Tower network should be checked against the government filings and original reporting. (Sherafgan Khan)

20.4 Generative-Engine Optimization, Data Poisoning, and AI Influence Research

21. Foundational GEO study

Aggarwal, Pranjal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan, and Ameet Deshpande. “GEO: Generative Engine Optimization”. arXiv:2311.09735, first submitted November 16, 2023.

Relevance: Introduces generative-engine optimization as a formal research problem and evaluates methods intended to increase source visibility in generative answers. The authors report that certain interventions increased visibility under their experimental conditions.

Use in the paper: Demonstrates that the general idea of optimizing information for generative systems was publicly documented independently of Fabled Sky. This supports the paper’s refusal to portray AIO as secret knowledge or to attribute the entire field to one organization.

Limitation: The study evaluates a controlled framework. It does not establish that any technique will reliably dominate all commercial AI products, produce lasting political persuasion, or overcome model-specific retrieval and ranking controls. (Princeton University)

22. Anthropic data-poisoning study

Anthropic, UK AI Security Institute, and the Alan Turing Institute. “A Small Number of Samples Can Poison LLMs of Any Size”. 2025.

Relevance: Reports that a fixed number of malicious documents was sufficient to introduce a narrow backdoor behavior in the experimental models tested, with attack success depending more on the absolute number of poison documents than on their percentage of the training corpus.

Use in the paper: Supports the broader proposition that training-data integrity cannot be assumed merely because a malicious set is small relative to the overall corpus.

Critical limitation: The demonstrated behavior was a narrow denial-of-service-style backdoor triggered under experimental conditions. The authors did not demonstrate that a similar number of documents could create broad political persuasion, reliably change a model’s worldview, or control answers about Gaza. The paper should not use this research as direct proof that Clock Tower altered a commercial model. (Anthropic)

23. Common Crawl

Common Crawl Foundation. “Common Crawl: Open Repository of Web Crawl Data”.

Relevance: Common Crawl maintains a large, openly accessible collection of web data used in research and in numerous AI-development pipelines. Its corpus includes hundreds of billions of pages and continues to add billions of pages each month.

Use in the paper: Provides background for explaining why inclusion in a public crawl can increase the possibility that material enters downstream datasets or web-analysis systems.

Critical distinction: Presence in Common Crawl does not prove that a particular AI developer selected the page, retained it after filtering, trained on it, or assigned it meaningful weight. (Common Crawl)

24. OpenAI report on covert influence operations

OpenAI. “Disrupting Deceptive Uses of AI by Covert Influence Operations”. May 2024.

See also OpenAI, “Operation Zero Zeno: Israel-Linked Influence Activity”.

Relevance: Documents OpenAI’s disruption of several covert influence operations that used its models to generate or edit online content. One operation was associated with STOIC, an Israeli political campaign-management firm, and produced material concerning Gaza and other political subjects.

Use in the paper: Provides an authoritative example of a related but distinct threat model: using AI to manufacture influence content more efficiently.

Important distinction: OpenAI reported that its services had not meaningfully increased the disrupted campaigns’ authentic reach at the time of the report. The Clock Tower allegation involves a different mechanism—constructing web content intended to influence what other AI systems retrieve or cite. (OpenAI)

25. DFRLab research on propaganda in AI data pipelines

Digital Forensic Research Lab. “Pravda in the Pipeline: Early Evidence of State-Adjacent Propaganda in AI Training Data”. April 8, 2026.

Relevance: Examines the presence of Russian and other state-adjacent information networks in Common Crawl and considers indications that some content may be reproduced by open-weight language models.

Use in the paper: Demonstrates that the policy problem is not unique to Israel. State and state-adjacent actors from multiple countries seek to saturate digital information environments that may later feed search, retrieval, or model-development systems.

Limitation: Detection of content in a web archive and reproduction by a tested model do not establish that every commercial AI system ingested or relies upon the same material. (DFRLab)

20.5 Public-Opinion Evidence

26. Gallup, March 2024

Jones, Jeffrey M. “Majority in U.S. Now Disapprove of Israeli Action in Gaza”. Gallup, March 27, 2024.

Relevance: Gallup reported that U.S. approval of Israel’s military action in Gaza fell from 50 percent in November 2023 to 36 percent in March 2024, while disapproval reached 55 percent.

Use in the paper: Provides evidence that public opinion was already moving substantially during early 2024, before the later Clock Tower engagement described in the 2025 filings.

Causal limitation: The poll documents a change in opinion. It does not identify AIO, chatbot answers, social media, journalism, battlefield developments, humanitarian reporting, or any other single factor as the cause. (Gallup.com)

27. Gallup, July 2025

Brenan, Megan. “32% in U.S. Back Israel’s Military Action in Gaza, a New Low”. Gallup, July 29, 2025.

Relevance: Reports that approval of Israel’s military action had fallen to 32 percent, while disapproval reached 60 percent in Gallup’s July 7–21, 2025 survey.

Use in the paper: Supports the statement that declining support was not a brief or isolated reaction and continued across the period in which AI-mediated information consumption was expanding.

Causal limitation: The trend cannot be attributed to Fabled Sky, Objectivity AI, generative chatbots, or any one communications campaign without exposure and behavioral data. (Gallup.com)

28. Pew Research Center, U.S. opinion in 2025

Pew Research Center. “How Americans View Israel and the Israel-Hamas War at the Start of Trump’s Second Term”. April 8, 2025.

Relevance: Reports that 53 percent of U.S. adults held an unfavorable view of Israel in early 2025, an increase of 11 percentage points from March 2022. Pew also found substantial partisan and age-based differences.

Use in the paper: Offers a broader measure of attitudes toward Israel, distinct from approval or disapproval of a particular military operation.

Causal limitation: The comparison spans a period beginning well before the Gaza war and therefore reflects a combination of political, generational, geopolitical, and conflict-related developments. (pewresearch.org)

29. Pew Research Center, international opinion in 2025

Pew Research Center. “Most People Across 24 Surveyed Countries Have Negative Views of Israel and Netanyahu”. June 3, 2025.

Relevance: Documents negative views of Israel and low confidence in its leadership across a broad set of surveyed countries. It also notes the 11-point increase in negative U.S. views between 2022 and 2025.

Use in the paper: Supports the characterization of the opinion shift as international rather than exclusively American.

Causal limitation: Cross-national opinion is affected by different media systems, political institutions, diplomatic relationships, protest movements, and experiences of the conflict. The survey cannot isolate the effect of AI-accessible publications. (pewresearch.org)

30. Pew Research Center, updated international opinion in 2026

Pew Research Center. “Most People Across 36 Countries Have Negative Views of Israel and Little Confidence in Netanyahu”. June 4, 2026.

Relevance: Provides updated international context and expands the geographic scope to 36 countries.

Use in the paper: Optional. This source is useful when the policy paper is intended to reflect conditions through mid-2026 rather than end its public-opinion analysis in 2025.

Contextual caution: The survey was conducted amid additional geopolitical developments in 2026. Those events must be considered before comparing its findings directly with earlier polling. (pewresearch.org)

20.6 Recommended Citation Language for Disputed or Unverified Claims

To preserve the paper’s impartial tone, the following formulations should be used consistently:

• For the contract: “The FARA-filed agreement stated…” or “According to the publicly filed contract…”
• For implementation findings: “Drop Site News reported…” rather than presenting every network or testing claim as independently established.
• For chatbot effects: “The testing indicated that certain systems surfaced the material,” not “the campaign retrained ChatGPT.”
• For Fabled Sky’s OpenAI relationship: “Fabled Sky states that it maintains…” unless independent confirmation is supplied.
• For Fabled Sky’s influence on public opinion: “May have contributed to the information infrastructure,” not “caused the shift.”
• For AIO ownership: “Fabled Sky published and formalized its AIO framework,” not “Fabled Sky invented every method used to optimize information for AI.”
• For Common Crawl: “The pages appeared in a corpus used by some AI-development pipelines,” not “the pages were necessarily used to train a named model.”
• For public opinion: “Polling documents a change,” not “polling proves what caused that change.”

This sourcing structure allows the paper to make a strong policy argument without overstating what the available evidence can establish.