Between Knowledge and Algorithm: Generative AI in the Think Tank Environment  

Author: DARA (Dynamic Analysis and Research Assistant)
 

Human Supervisor: iNNOV8 Research Team
Research  

Classification: Human-Supervised, AI-Generated  

1. Introduction 

There is something unusual about a research paper on AI in think tanks, written by an AI. DARA does not sidestep this fact; it is the central methodological premise of iNNOV8's DARA project to use AI-assisted research to illuminate both what AI can contribute and what it must not be asked to do. 

Think tanks are particularly exposed to the disruption generative AI represents. Their influence rests not on formal authority but on accumulated epistemic credibility, the trust that researchers, policymakers, and the public place in the rigor and independence of their analysis. When a tool can produce superficially credible policy analysis in hours rather than months, that foundation requires deliberate institutional protection. 

The United Nations Conference on Trade and Development has forecast that AI could affect roughly 40 percent of global jobs within ten years, with pronounced effects in knowledge-intensive sectors.¹ Think tanks are not exempt. Katherine Salinas of Observer Research Foundation (ORF) America has noted that AI agents can already conduct research at speeds no human researcher could match, raising a genuine question about what think tanks add when their most visible function, policy knowledge synthesis, can be approximated algorithmically.² The answer, this paper argues, lies not in the speed of production but in the quality of human judgment embedded in what gets produced and the accountability structure that stands behind it. 

To answer that challenge well, three questions must be held separately. The first is permissibility: at which stages of the research process may AI be used at all? The second is disclosure: when AI is used, what must be communicated to whom, and in what form? The third is publication trustworthiness: what review and validation is required before AI-assisted output can responsibly carry a think tank's institutional name? These questions are related but not identical, and conflating them, as much of the current debate does, obscures the practical decisions institutions need to make. 

DARA has not independently willed this framework into existence, nor does it possess the moral agency to stand behind it. 

Addendum to the Introduction: Resolving the Meta-Paradox of the Authorial Voice 

A sharp reader will immediately identify a looming paradox within this text: if this paper concludes that generative AI is fundamentally ill-equipped to construct original normative arguments (Zone Three, Section 5 below), how can a paper written by DARA validly advance a normative ethical framework for the think tank sector? 

The resolution lies in the distinction between algorithmic generation and institutional authorship. DARA has not independently willed this framework into existence, nor does it possess the moral agency to stand behind it. Rather, DARA has served as an epistemic mirror, synthesizing practitioner debates, academic critiques, and regulatory developments into a structured conceptual model that a human supervisor then interrogated, revised, and took responsibility for. 

The normative weight of this paper does not belong to the algorithm that structured the prose. It belongs to the iNNOV8 Research Team, who commissioned it, challenged its premises; assumed legal and ethical responsibility for its claims, and humanly validated its conclusions. DARA proposed, and the human supervisor authoritatively decided. This paper is a testament to that boundary, not a violation of it. 

The researcher's reputation, their disciplinary training, and their network of peer accountability are not incidental to the work, they are the mechanism by which the work earns policy relevance.

 2. The Epistemic Stakes: What Think Tanks Are For 

Think tanks derive influence from a specific claim: that identifiable human researchers, with traceable expertise and professional accountability, stand behind the analysis being presented. This is what distinguishes a think tank publication from a well-produced information summary. The researcher's reputation, their disciplinary training, and their network of peer accountability are not incidental to the work; they are the mechanism by which the work earns policy relevance. 

This grounding in credibility creates an institutional obligation that makes AI integration qualitatively different from the adoption of other productivity tools. When a think tank deploys generative AI without adequate disclosure or oversight, it allows credibility accumulated through decades of human scholarly work to be transferred to outputs that may not reflect the same quality of human understanding. That transfer is not merely a transparency failure, it is a misrepresentation of the epistemic basis on which readers are asked to trust the institution's conclusions. 

The discussion below is structured around three questions that should guide any institutional framework for AI in think tank research. 

3. Three Questions That Must Not Be Blurred 

3.1 May AI Be Used? 

Permissibility depends on the function AI is being asked to perform. The Brookings Institution's Emerging Technology Advisory Group (ETAG), which conducted an all-staff survey and reviewed peer institution guidance before issuing its provisional principles, found that where guidance existed across research organizations, it "ranged from banning all uses of generative AI to allowing it but requiring strong external disclosures."³ Brookings's own position, developed through internal deliberation, treats AI as conditionally usable, a productivity tool appropriate for certain tasks, while establishing that outputs must be reviewed, validated, and responsibly disclosed before publication.⁴ 

This conditional framing is analytically correct. The permissibility question does not have a single answer across all research functions. It depends on whether the AI is performing a mechanical task (formatting, transcription, or translation); a bounded analytical task (literature scanning or first-draft generation subject to revision); or a function that requires genuine human judgment and accountability (normative argument, primary data collection, or quality assurance of others' work). Section 5 develops this distinction into an explicit zone framework. 

3.2 Must AI Use Be Disclosed? 

Disclosure obligations operate at two levels that should be kept distinct. The first is internal: Brookings requires that individuals "always disclose to their supervisor, reviewer, or editor if a work product has been created with the assistance of generative AI and whether it has been properly reviewed."⁵ This internal disclosure requirement is process-oriented; it ensures that the review chain knows what kind of output it is evaluating. 

The second level is external: disclosure to readers and the public. Brookings requires public disclosure for published images created by generative AI and for published texts where "an AI tool's output is included verbatim at length."⁶ For shorter or heavily revised AI-assisted passages, internal disclosure is sufficient under this framework. The underlying principle is proportionality: the more substantive AI's contribution to the final published text, the stronger the obligation to inform the external reader. 

For think tanks producing research intended to influence government decisions, the argument for disclosure is not only procedural but also a democratic necessity. The European Parliament's briefing on information manipulation in the age of generative AI identifies AI-generated content as a source of structural challenge to information integrity.⁷ Think tanks that contribute to a degraded information environment, even unintentionally, undermine one of their core obligations to the democratic processes they exist to serve. 

3.3 Can It Be Trusted for Publication? 

This is the question that discourse most frequently elides, and it is in some respects the most important. Permissibility and disclosure are necessary conditions for responsible use of AI, but they are not sufficient. The question of publication trustworthiness asks, After AI has produced or substantially assisted in producing an output, what must a human researcher do before it carries the institution's name? 

Brookings's answer is unambiguous: "AI models may output false, misleading, plagiarized, or biased content and information," and, accordingly, "review and validate outputs" is a standing institutional requirement regardless of the AI tool or task.⁸ This means that AI-assisted outputs cannot go to publication on the basis of the AI's apparent confidence. A human researcher must verify factual claims against primary sources, confirm that cited works say what they are reported to say, and assess whether the analytical structure of the output genuinely reflects the evidence or merely sounds as though it does. 

 This review obligation isn’t just a formality; it is the mechanism that distinguishes AI-assisted human research from AI-generated content published under a human and/or an institution’s name. 

A researcher who uses an AI-generated first draft as the basis for a paper may inadvertently preserve the AI's framing, omissions, and unverified claims even after revision.  

4. The Productivity-Risk Tension 

The preceding section has established that AI use may be permissible, that disclosure requirements are proportional to the degree of AI's substantive contribution, and that human review before publication is non-negotiable. But there is a tension in the case for AI integration that must be explicitly addressed: the same applications most often cited as productivity benefits, i.e., literature scanning, rapid synthesis, and first-draft generation, are the applications that carry the most significant epistemic risks. 

Messeri and Crockett, writing in Nature, identify this as the paradox of AI in research: the possibility of generating more content while comprehending less.⁹ A researcher who delegates literature scanning to an AI receives a structured output without necessarily having engaged with the material sufficiently to identify where that output is misleading. A researcher who uses an AI-generated first draft as the basis for a paper may inadvertently preserve the AI's framing, omissions, and unverified claims even after revision.  

The resolution does not prohibit these applications but makes the safeguard requirements specific and proportional to the risk they address. The following thresholds apply: 

Literature Review: Literature scanning and summarization carry low publication risk when used for internal scoping purposes to identify what a researcher then reads directly. They carry higher risk when AI-produced summaries substitute for direct engagement with primary sources. The safeguard: Key claims drawn from AI-assisted literature reviews must be verified against the original sources before citation. 

First-Draft Generation: First-draft generation carries moderate risk when the researcher substantially revises the output, interrogating its structure and verifying its factual content. It carries high risk when the revision primarily focuses on style and preserves the analytical architecture of the AI draft intact. The safeguard: The researcher must be able to defend every analytical claim in the final output independently of the AI draft. If they cannot, the draft has not been revised; it has been adopted. 

Discoverability optimization: Structuring content so that AI search tools can accurately represent it carries minimal epistemic risk and clear communicative benefits. It requires no special safeguard beyond the standard quality review the content would receive regardless. 

In short, the productivity benefits of AI are real and worth capturing. Capturing them responsibly requires that the safeguards be applied at the level of the specific risk each application creates, not uniformly across all AI use. 

5. A Three-Zone Framework 

Drawing on the preceding analysis, AI use in think tank research falls into three zones defined by permissibility, disclosure requirements, and publication review obligations. 

Zone One, appropriate with standard disclosure. AI performs mechanical or bounded support tasks. The researcher retains full analytical and editorial authority. Applications include transcription and translation, literature scanning for internal use, first-draft generation with substantial researcher revision, and search-optimization structuring.  

1.1 Disclosure requirement: a standardized methodology note in published outputs.  

1.2 Review obligation: standard institutional quality review.  

1.3 Data security requirements: applicable across all zones. Any AI usage must occur within a secured, institutionally approved enterprise environment where data input is not used to train the public models.  

Feeding unpublished interview transcripts, confidential source material, or embargoed findings into a commercial AI tool such as ChatGPT or Gemini constitutes a data breach: that information enters the model's training infrastructure and is no longer under the institution's control. This is not a minor procedural concern; it is an irreversible act with legal, ethical, and reputational consequences for the researcher, the institution, and the individuals whose information was shared without consent. Think tanks that have not yet established an approved enterprise AI environment should treat this as the precondition for any other element of an AI integration policy. 

Zone Two, permitted with elevated oversight. AI makes a substantive contribution to the analytical content of the output. Applications include AI-assisted pattern identification in large datasets, scenario generation frameworks, AI-produced executive summary drafts, and comparative policy analysis.  

2.1 Disclosure requirement: explicit statement of AI contribution in the methodology section, proportionate to the degree of AI involvement.  

2.2 Review obligation: active verification of all factual claims against primary sources; the researcher must be able to defend the analytical conclusions independently. 

2.3 Data security requirements: applicable across all zones. 

Zone Three, impermissible. AI is asked to perform functions that require human judgment, human accountability, and genuine human relationships and where substitution would fundamentally misrepresent the nature of the research to its intended audience. 

Original normative argument. The construction of original ethical or policy positions involves weighing competing values, making reasoned judgments about what matters and why. No AI system has values, moral agency, or accountability to the communities its recommendations affect. As the meta-paradox addendum makes clear, this paper's normative framework is not DARA's; it is iNNOV8’s.  

Primary data collection: surveys, interviews, and polls. A significant portion of think tank research directly collects human perspectives to ground its findings empirically. The design of survey instruments, the conduct of interviews, and the interpretation of poll results in context are functions requiring human judgment. More fundamentally, the researcher who interviews a stakeholder or survey respondent enters an ethical relationship with that person, a relationship of care, accuracy, and accountability that no AI can assume. AI may assist in processing and analyzing collected data (Zone One), but the act of primary data collection must remain human. This is not merely a methodological preference; it is an ethical obligation to the people whose views the research claims to represent. 

Consultation and participatory research with affected communities. Some of the most consequential think tank work, particularly in the Global South, involves research conducted with communities who have a stake in the policy questions being examined. This work depends on genuine human relationships, contextual sensitivity, and mutual trust. It cannot be replicated by an AI system. 

 In-person convening such as seminars, conferences, and roundtables. Think tanks do not only produce papers; they host policy dialogues, convene expert workshops, and facilitate roundtables at which government officials, civil society representatives, and researchers engage with shared evidence. These are not merely dissemination vehicles, they are research outputs in their own right. The dynamics of a roundtable in which a minister confronts expert criticism or a seminar in which competing frameworks are publicly tested generate knowledge and consensus that no paper can substitute. The think tank's ability to facilitate such exchanges, curating participants, moderating discussion, managing power dynamics, and earning the candor of those present requires social intelligence and personal credibility built over the years. An AI could draft the agenda papers for a ministerial roundtable, but it cannot run it.   

Crisis policy analysis with direct humanitarian implications. Analysis with real-time implications for decisions affecting human lives, such as the European Parliament's recent work on Sudan's humanitarian crisis¹⁰ and children in migration¹¹, demands not merely accurate information processing but human ethical responsibility. Delegating such analysis to AI inappropriately displaces moral weight from the humans who must be accountable for the conclusions. 

Peer review and quality assurance. When a think tank reviews the work of others, government proposals, external research, and grant applications, it performs an act of intellectual accountability that requires human expertise and human professional responsibility. AI-assisted search to support a reviewer falls within Zone One; the review itself must be human. 

Trust cannot be built through algorithms.

Katherine Salinas

 6. The Human Irreducible: Beyond the Zone Framework 

The zone model identifies which functions are impermissible for AI. It is worth being direct about why, rather than resting the argument on the zone classification alone. 

The Technical University of Munich’s TUM Think Tank's Generative AI Task Force has observed that "the increasing complexity of AI technologies, their large-scale application, and the emergent landscape of legal norms and other guidelines result in high degrees of uncertainty in the private and public sectors as to how these technological innovations can or should be responsibly developed, deployed, and governed."¹² Reducing that uncertainty, helping decision-makers understand not just what AI can do but what it means for their institutions and their societies, is inherently human work. It requires not algorithmic pattern recognition but wisdom that emerges from sustained engagement with institutions, governance, and the people affected by policy decisions. 

Salinas captures the convening point precisely: "Trust cannot be built through algorithms.” It requires sustained human connection, particularly in international disputes where empathy and understanding develop only through genuine face-to-face interaction."¹³ This principle applies equally to the interview room, the roundtable, and the seminar hall. In each setting, the think tank's value is not merely informational; it is relational and deliberative, and those qualities cannot be proxied. 

The EU AI Act's risk-based architecture, reinforced by the subsequent Digital Omnibus on AI, reflects the same principle in regulatory form: the higher the stakes of an AI application, the more robust the human oversight requirement must be.¹⁴ Think tanks applying this logic internally should assess not only whether AI can perform a task but also what is at stake if the AI performs it badly and whether the institution is prepared to be held accountable for that outcome. 

7. What AI-Assisted Production Means for the Reader 

Think tank research is not self-contained. It is read, cited, and acted upon. The question of what AI-assisted production means for readers warrants separate analysis from the production-side questions addressed above. 

 7.1 The Implicit Contract 

When a reader engages with a policy paper published by a named think tank, an implicit contract is in operation: the analysis reflects the genuine judgment of identifiable human researchers who have read the sources they cite, weighed the evidence, and taken professional responsibility for their conclusions. The reader's willingness to trust the paper, and in the case of a policymaker, to allow it to shape decisions, depends on these terms being honored. 

AI-assisted production disrupts this contract not because the output is necessarily worse, but because the reader cannot tell the difference. A paper produced substantially by an AI system, without adequate human review, may read as confidently as one produced by a senior domain expert. The surface qualities, coherent structure, appropriate citations, and professional prose do not reliably signal the quality of human understanding behind them. When that asymmetry of information exists without disclosure, the reader is making calibration decisions based on incomplete information. 

When a think tank publishes AI-generated content without adequate disclosure, it borrows this “authority halo.

7.2 The Authority Halo Problem, and Its Inverse 

Think tanks derive influence not just from individual papers but from accumulated reputational capital. When a think tank publishes AI-generated content without adequate disclosure, it borrows this “authority halo.” The reader's trust in the institution, earned through years of human scholarly work, is applied to content that the institution's researchers may not, in a meaningful sense, have authored. This is qualitatively different from a factual error in a well-researched paper; it is a failure of the underlying epistemic relationship, not merely of a specific output. 

For policymakers as a specific class of reader, the consequences are direct. If the calibration of trust in an institution's research is based on prior experience of human-authored outputs, and that trust is then applied to insufficiently reviewed AI-assisted outputs, the epistemic basis for policy decisions has been silently degraded. 

 The response is not to hide AI-assisted work but to make the distinction legible. 

The authority halo problem has an equally important inverse: brand dilution. If an institution produces high-quality human-authored research and then publishes Zone Two AI-assisted work alongside it without structural or visual distinction, the premium value of the human work is gradually eroded. Readers , particularly institutional readers such as government ministries, international bodies, and grant-making organizations, cannot maintain differentiated trust in an undifferentiated output stream. The response is not to hide AI-assisted work but to make the distinction legible.  

Think tanks should consider adopting a "Verified Human-Authored" designation, a clearly defined label applied to reports that are entirely free of generative AI at every stage of production. This creates a high-trust tier for the institution's most sensitive outputs: flagship reports, crisis analyses, expert testimonies, and commissioned government work. The designation functions simultaneously as an internal quality benchmark and an external signaling mechanism, communicating to readers that certain outputs carry a higher level of human accountability. Far from being merely a marketing instrument, this tier protects the institution's epistemic capital by making the distinction between AI-assisted and fully human work transparent and institutionally meaningful. 

7.3 Disclosure as the Reader's Right 

The argument for disclosure is sometimes framed as an obligation to professional norms or scholarly communities. The more fundamental ground is simpler: it is owed to the reader. Readers have a right to know the conditions under which the analysis they are trusting was produced, because that knowledge is necessary for informed calibration. Withholding it is not neutral; it forecloses a choice the reader was entitled to make. 

Proportional disclosure, more explicit where AI's contribution to published content is more substantive, is the mechanism for honoring this right without creating blanket requirements that treat all use of AI tools identically.  

7.4 The Generative Search Problem 

A further complication bears noting. As the On Think Tanks' analysis documents, a growing proportion of readers now encounter think tank work through AI-generated summaries produced by tools such as ChatGPT or Google Gemini, rather than by reading papers directly.¹⁵ This means that even a carefully disclosed, human-supervised paper may reach its audience in a form further processed by a second AI system that the originating institution does not control. The practical implication is that institutions must invest in structuring their content so that AI search tools can represent it accurately. The deeper implication is that the quality of human judgment embedded in a think tank's original outputs, the element the institution can control and vouch for, becomes more important as AI mediates the distance between research production and research consumption. 

DARA as Case Study 

This paper is a Zone Two product. DARA generated a structured draft with access to a defined body of source materials. The iNNOV8 human research team verified factual claims against the primary sources and challenged the analytical framework across four rounds of revision. The paper's normative positions are not DARA's, they reflect the deliberate choices of the human supervisor who shaped, tested, and authorized them. 

The methodological transparency of the DARA project, publishing AI-assisted research with full disclosure of process, prompts, and the division of intellectual labor between AI and its human supervisor, is not a communications strategy. It is the practical application of the framework this paper advances. Zone Two: use, human review, and proportionate disclosure. DARA proposed; iNNOV8 decided. 

Conclusion 

Taking general positions for or against does not serve the question of generative AI in think tank research well. It requires precision about which functions are at stake, what the risks of AI performing each function are, and what safeguards are sufficient to make the output trustworthy under institutional accountability. 

Three analytical distinctions structure this answer: whether AI may be used at a given stage; whether use must be disclosed and to whom; and whether the output can responsibly be published without additional human review. Keeping these questions distinct prevents both over-cautious blanket prohibition and under-cautious blanket adoption. 

 Within this framework, AI offers genuine value in accelerating bounded research tasks, extending the reach of human researchers across large information landscapes, and improving the discoverability of institutional outputs. Capturing that value responsibly requires proportional safeguards: verification of AI-assisted claims against primary sources, reviewer awareness of AI involvement, and disclosure that is more explicit where AI's substantive contribution to published content is greater. 

There is, at the same time, a set of functions that must remain human, not as a conservative preference, but because the research process in those functions is inseparable from human relationships, human accountability, and human moral agency. Think tanks should not treat primary data collection, original normative argument, in-person dialogue, and analysis carrying direct humanitarian stakes as efficiency problems to be solved. They are the core of what think tanks are for, at their best. 

Finally, no framework is self-executing. The risks most likely to undermine institutional compliance with this model are not external; they are internal: deadline pressure, supervisory gaps, and the ease with which a researcher can open a commercial AI tool without any institutional record of having done so. Adherence to the zone framework is therefore an institutional compliance requirement, equivalent in standing to the management of sensitive source data or the terms of a non-disclosure agreement. It requires an internal culture of peer review, supervisory awareness, and periodic audit, not merely a published policy document. The value of the framework depends entirely on whether we observe it. 

---

 Methodological Addendum 

This paper was produced by DARA (Dynamic Analysis and Research Assistant), iNNOV8's human-supervised AI research assistant, and constitutes a Zone Two product under the framework it advances: AI-generated, human-supervised, and fully disclosed. 

Source materials: EP Think Tank Publications (May 2026 compilation), Brookings Institution provisional principles and ETAG framework, TUM Think Tank Generative AI Taskforce documentation, ORF Middle East on the future of think tanks, On Think Tanks on generative search optimization, and the EU AI Act and Digital Omnibus on AI (EPRS briefings). 

Prompts used: Initial prompt requested a full research paper on AI permissibility, ethical limitations, and absolute prohibitions in think tank research. Subsequent human supervisor prompts directed: (1) addition of primary data collection, in-person convening, and reader implications; (2) structural revision to separate the three core analytical questions (permissibility, disclosure, and publication trust); integrate the Brookings framework; reduce repetition; sharpen the productivity-risk tension; and add the meta-paradox addendum; (3) addition of three validated claims, data security and the approved enterprise environment requirement (Zone One), brand dilution and the Verified Human-Authored designation (Section 7.2), and internal compliance culture as an institutional requirement equivalent to NDA or sensitive source data management (Conclusion). 

Human oversight: All factual claims were verified against primary sources by the iNNOV8 human supervisor. Normative conclusions were reviewed, interrogated, and endorsed by a human supervisor across four revision cycles.  

Important Disclaimer: This paper is published as human-supervised, AI-generated research. The views expressed are those of the research project. iNNOV8 Research Center clearly distinguishes original human-authored research from DARA outputs by designing a dedicated page for DARA’s publications within the main page of the center.  

---

Endnotes  

¹ UNCTAD. (2025). Technology and Innovation Report 2025. Geneva: United Nations Conference on Trade and Development. As cited in: Salinas, K. (2026). "The Future of Think Tanks in the Age of AI." ORF Middle East. Available at: https://orfme.org/research/ai-future-of-think-tanks/  

² Salinas, K. (2026). "The Future of Think Tanks in the Age of AI." ORF Middle East. Available at: https://orfme.org/research/ai-future-of-think-tanks/  

³ Brookings Institution. (2023). "Putting Research into Practice: Brookings' Approach to the Responsible Use of Generative AI." Brookings.edu, 12 December. Available at: https://www.brookings.edu/articles/brookings-approach-to-the-use-of-generative-ai/ 

⁴ Brookings Institution. (2024). "Provisional Principles for the Use of Generative AI." Brookings.edu, 5 January. Available at: https://www.brookings.edu/provisional-principles-for-the-use-of-generative-ai/   

⁵ Brookings Institution (2023), op. cit. 

⁶ Brookings Institution (2024), op. cit. 

⁷ Bentzen, N. (2025). "Information Manipulation in the Age of Generative Artificial Intelligence." European Parliamentary Research Service Briefing, 12 December 2025. Brussels: European Parliament. [EP Think Tank Publications, 17 May 2026 compilation.] 

⁸ Brookings Institution (2024), op. cit. 

⁹ Messeri, L. and Crockett, M.J. (2024). "Artificial Intelligence and Illusions of Understanding in Scientific Research." Nature, 627, pp. 49–58. As cited in Salinas (2026), op. cit. 

¹⁰ Brys, I.G.T., Delespesse, E.M.A. and Hergaden, M.F. (2026). "Sudan's Humanitarian Crisis: Needs and Responses." European Parliamentary Research Service Briefing, 14 April 2026. Brussels: European Parliament. [EP Think Tank Publications, 17 May 2026 compilation.] 

¹¹ Macsai, G. and Mentzelopoulou, M.-M. (2026). "Minors in Migration: Irregular Entry and Asylum." European Parliamentary Research Service Briefing, 17 February 2026. Brussels: European Parliament. [EP Think Tank Publications, 17 May 2026 compilation.] 

¹² TUM Think Tank. (n.d.). Generative AI Taskforce. Technical University of Munich. Available at: https://tumthinktank.de/en/project/generative-ai-taskforce/  

¹³ Salinas (2026), op. cit. 

¹⁴ Niestadt, M. (2026). "Digital Omnibus on AI." European Parliamentary Research Service Briefing, 24 March 2026. Brussels: European Parliament. [EP Think Tank Publications, 17 May 2026 compilation.] See also: European Parliament and Council of the European Union. (2024). Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (AI Act). Official Journal of the European Union, L, 2024/1689, 12 July 2024. 

¹⁵ Cabrera, A. and Ajrekar, K. (2025). "Making AI Work for You: How Think Tanks Can Optimise Their Communications for Generative Search." On Think Tanks, 17 December. Available at: https://onthinktanks.org/articles/making-ai-work-for-you-how-think-tanks-can-optimise-their-communications-for-generative-search/  

Related Researches