Response to the UK Media Green Paper

The Complex AI Institute is a research institute being established in London as a non-profit company limited by guarantee. It works on the deployment of advanced AI inside complex real-world systems: what it takes to make such systems work, and keep working, where objectives conflict, consequences are real, and every decision must remain inspectable.

We respond because the green paper’s proposals on prominence are not, at their core, media policy. They decide who may shape the ranking functions of the algorithmic and AI systems that now mediate most public attention, and the design choices made here will set precedents well beyond news.

They also present a problem of precisely the shape our founding thesis describes. Deciding what a reader should see is multi-objective: a claim may be accurate but stale, or well-sourced but misattributed, and those objectives pull against one another. It is high-stakes, because the cost of getting it wrong is borne by public understanding rather than by the system operator. And it demands traceability, because a ranking decision nobody can inspect is a decision nobody can be held accountable for.

That framing drives our central recommendation. A register of approved publishers reduces a multi-objective question to a single variable, the identity of the speaker, and attaches the audit trail to the masthead rather than to the claim. It is a proxy that will demonstrate well and degrade in deployment, which is the failure pattern this Institute exists to study. We therefore recommend verification of content ahead of designation of publishers.

Summary of recommendations

We support the government’s objective and question its chosen instrument. An information environment ranked purely for engagement has documented civic costs: Ofcom identified the role of algorithmic recommendation in amplifying misleading and divisive content during the disorder of summer 2024. But the remedy proposed, a state-recognised register of trustworthy publishers, answers a narrower question than the one readers have, and a prominence regime built badly would be worse than none. A state-preferred press list embedded invisibly in ranking systems is a serious and, in our view, avoidable risk. Our recommendations therefore begin with verification of content rather than designation of publishers, and treat designation as a fallback that must clear a high bar.

The evidence base, in three numbers

People who reported encountering misinformation or deepfakes in 202443%

AI assistants’ news answers judged to have significant issues51%

Adults confident judging whether a source is truthful45%

Sources: DCMS green paper, June 2026 · Ofcom, 2025 · BBC research, as cited therein

Our recommendations consolidate, for the UK’s specific proposal, a body of existing practice: the Council of Europe’s 2021 guidance on prioritising public interest content and its 2022 recommendation on quality journalism; Germany’s Interstate Media Treaty findability regime; the European Media Freedom Act; the Digital Services Act; the Journalism Trust Initiative’s machine-readable certification; and the UK’s own Media Act 2024, Data (Use and Access) Act 2025 and Ofcom crisis-protocol work.

To that body of practice we add a specific UK package that we have not identified in any existing statutory prominence regime or published proposal: designation bound to content-level provenance with a duty to preserve it, an express statutory bar on ministerial direction over designation decisions, and statutory triggers and accountability for the incident tier. Convergence, in a consultation, is evidence; we cite our fellow travellers throughout.

First, verify claims rather than licensing publishers. Our primary recommendation is that the regime should attach obligations to content that can be verified, not to a register of approved names. A publisher list optimises for a single variable, the identity of the speaker, in order to answer a multi-objective question: whether a specific claim is accurate, current and fairly attributed. It also places the audit trail on the masthead rather than on the claim, so nothing about an individual article becomes inspectable.

We therefore recommend the government first exhaust verification-led instruments: provenance duties, claim-level traceability, open audit and support for bridging-based public correction mechanisms of the kind X’s Community Notes demonstrates, where a correction surfaces only when raters who have historically disagreed both find it helpful, and where the algorithm and rating data are public.

If, having done so, the government still concludes designation is necessary, the safeguards below are the minimum conditions. Designation must sit with an independent authority, with Ofcom the natural candidate. Criteria should be made by the regulator after statutory consultation and approved by Parliament, not amendable by ministerial regulation, with a statutory bar on ministerial direction over designation decisions, whether individual or by class. Because no statute can entrench itself, the architecture must also address the residual channels of influence: appointment safeguards, published reasons and evidential standards, a merits appeal with interim relief, and periodic re-designation.

We say two things honestly. Designation by an independent regulator remains state-created preferential treatment; the question is whether its process makes it defensible. And concentrating designation, audit and existing Online Safety Act enforcement in one body carries its own risk, so the functions should be internally separated.

The government should also price the alternative of plural accredited routes, co-regulatory recognition of multiple certification schemes on the model of the Journalism Trust Initiative, rather than a single official list.

Second, bind verification to content provenance, with a duty to preserve it. This is the centrepiece of the verification-led route in our first recommendation, and it works whether or not any register of publishers exists. Machine-readable designation and content credentials (C2PA-class provenance, which the green paper already recognises) can carry verified status with the journalism itself, into embeds, short-form reuse and AI retrieval, so any obligation attaches to credentialed content rather than to accounts or approved names.

We are explicit about what this is not: a credential proves designation and provenance, never accuracy, and a credential cannot itself create prominence; it is the trustworthy input to platform ranking and interface mechanisms, which remain necessary and must be regulated as outcomes. We are equally explicit about the known failure mode: provenance manifests are routinely stripped by upload pipelines and re-encoding, and do not survive screenshots.

That failure mode is why the flagship duty we propose is credential preservation: platforms should be required to preserve and verify content credentials through their ingest and transcode pipelines, support fingerprint or watermark recovery for stripped copies, and fall back to verified account-level designation where content-level status is unrecoverable. The credential layer also needs lifecycle engineering: key revocation, emergency suspension distinct from due-process de-designation, a registry with signed entries and freshness guarantees, and interim status during ownership change.

The components exist: Project Origin and the Content Authenticity Initiative built the provenance rails, the Journalism Trust Initiative has offered machine-readable certification since 2019, and the EU’s 2022 Code of Practice on Disinformation carried voluntary commitments on prioritising authoritative information.

Those commitments did not produce consistent, interoperable implementation across major platforms, and Europe drew the obvious conclusion: the code was converted into a Code of Conduct under the Digital Services Act, integrated in February 2025 and applicable from July 2025. The shortfall of the voluntary version is the case for statutory footing, and the EU has already acted on it. The regulatory binding of designation to content-level provenance, with a preservation duty, is the Institute’s proposal.

Third, make the regime AI-native from day one. Define the regulated surface functionally: any service that mediates news discovery at scale, including AI assistants and AI-generated news summaries. BBC research cited in the green paper found 51 per cent of AI assistants’ answers about news had significant issues; the EBU and BBC’s international study (October 2025, 22 public service media organisations in 18 countries) found the same pattern at scale, and broadcasters have already demanded that assistants cite, attribute and link.

Our recommendation differs in scope rather than direction: duties should attach to operators of high-reach consumer retrieval and answer services meeting a UK-links test, above defined thresholds, and they should be claim-level source traceability, freshness and faithful attribution rather than a mandatory preference for designated publishers, because the 19 per cent finding shows citation alone is not a remedy. Private, enterprise and user-supplied contexts should be exempt, and every safeguard in this response applies to this channel as to feeds. The work should be coordinated with the AI Labelling Taskforce and the digital replicas consultation.

Fourth, adopt a two-tier architecture built on UK machinery. A modest, clearly labelled, always-on discoverability baseline, plus an enhanced information-incident mode. The EU’s Article 36 crisis mechanism is a statutory precedent but not a template: it has never yet been activated, it is triggered by the European Commission, an executive body, and it carries no mandatory published post-incident review.

The UK should borrow its existence and improve its design, drawing on Full Fact’s Framework for Information Incidents and Ofcom’s crisis-response protocol, whose measures Ofcom finalised in June 2026. Our additions: statutory triggers; activation by the regulator, not ministers; default duration in weeks, renewable by the regulator only with published reasons; a mandatory published post-incident review; and preservation of the user opt-out during incidents.

Fifth, outcomes require audit, and floors are not quotas. We recommend discoverability floors: availability of designated news surfaces, one-interaction reachability, incident-time discoverability, local-provider reach.

We expressly oppose impression-share targets, which would function as quotas for approved sources and invite gaming through the denominator. Because “news intent” is not a stable category across platforms, we recommend a regulator-led measurement pilot before any statutory target, with published classification error, and measurement that separates availability, exposure and engagement. Displacement and plurality effects must be measured, not assumed; the Council of Europe’s 2021 Guidance Note already contemplates publishing the compared viewing of prioritised and non-prioritised content.

For research access, the UK should use the regulation-making power in section 125 of the Data (Use and Access) Act 2025 to establish its researcher-access framework, guided by Ofcom’s report on researcher access under section 162 of the Online Safety Act, with the DSA’s Article 40 architecture as comparator.

Sixth, preserve user agency. Default on, labelled at the point of display, switchable off in one step, in both tiers, with “no degradation” specified as no punitive throttling, no loss of functionality and no dark patterns. Opt-out rates should be published, as the Council of Europe recommended in 2021; government should treat them as one evaluation signal among several, read with caution for selection effects.

Seventh, if designation proceeds, guard its boundary with mechanisms, not labels. This recommendation applies only in the fallback case described above, where the government legislates for designation despite the verification-led alternative. Beneficial-ownership disclosure with material-control thresholds triggering event-driven review and interim status; emergency suspension distinct from due-process de-designation, with credential revocation propagating through the registry on a defined timescale; consistency with the foreign-state-influence regime for newspaper enterprises introduced in 2024. Criteria and tooling must be achievable for small, local and minority-language publishers: if designation requires signing infrastructure, the registry should provide hosted tooling so the credential layer does not become a barrier to exactly the publishers the green paper wants included.

Eighth, tie benefits to responsibilities narrowly, and learn from section 40. Conditionality is where a prominence regime and press freedom collide, and the UK has been here before: section 40 of the Crime and Courts Act 2013 was designed to use costs rules to incentivise membership of a recognised regulator; it was never brought into force, was opposed by most major newspaper groups for a decade, and was repealed by the Media Act 2024.

Any prominence conditionality must be distinguishable from that architecture, and the distinction is scope. We support only regime-facing responsibilities: credential integrity, transparency about designation-relevant facts, and complaints handling about designated status. We do not support conditioning prominence on internal editorial practice, and we specifically do not support the green paper’s suggestion of “more stringent standards in AI content use” as a designation condition: how a newsroom uses its own tools is editorial practice, and making reach conditional on it would hand the designating authority a lever over newsroom operations. Obligations should scale with size.

Question 1: overall views

The questions this green paper raises about news are a subset of a larger question: how should democracies govern the algorithmic and AI systems that allocate public attention? We welcome the government’s decision to confront it, and we welcome the six principles in chapter 2.2: audience-led, future-proofed, outcomes-based, comprehensive and relevant, plurality of views, and criteria “decided in an open and transparent manner with regard to protecting media freedom”. Our recommendations map onto them directly: user agency serves the audience-led principle, AI-native scope serves future-proofing, floors-and-audit serves the outcomes-based principle, the small-publisher provisions serve comprehensiveness and plurality, and independent designation implements the transparent-criteria principle.

We apply three tests throughout: independence (can a minister influence what is promoted?), auditability (can an outsider verify what the system actually did?) and adaptability (does the regime still function when news discovery moves from feeds to AI assistants?).

Scope: we answer where our focus, AI and algorithmic governance, lets us add evidence or design detail. We respectfully decline Q2 (production sector) and Q13 to Q21 and Q25 to Q29 (television distribution and public service media structure), which fall outside that focus; Q3 and Q22, which ask about individual experience; and Q24, Q24a and Q24b, which are directed at content creators, which we are not. Where we answer attitude questions we do so in an organisational capacity on the basis of the published evidence cited.

Question 4: should “trustworthy” news content be more visible on social media?

Yes to the objective; no to the instrument as drafted. Trustworthy news should be easier to find. It does not follow that the state should maintain a register of approved publishers to achieve it, and we recommend the verification-led route in our first recommendation be exhausted before designation is legislated for. Where designation proceeds regardless, our answer is qualified by the process safeguards in the first, second and seventh recommendations.

Both action and inaction carry risk. Inaction leaves news discovery to engagement-optimised ranking, whose failure mode Ofcom documented in the summer 2024 disorder. Poorly designed action creates a government-adjacent list of preferred publishers embedded invisibly in ranking systems. We have also considered the alternative of acting without designation at all, through transparency duties, user-side controls and provenance without a state register; those tools are valuable and we recommend them regardless, but on their own they leave incident-time discoverability of reliable information to platform discretion, which is the gap 2024 exposed. If designation is to exist, the determining variable is who decides, against what criteria, with what appeal, and with what audit.

On the suggested starting point, section 56 of the Online Safety Act 2023: it is a workable eligibility gateway, described accurately rather than generously. The definition automatically includes the BBC, S4C and licensed broadcasters; other entities must satisfy cumulative conditions including a principal purpose of publishing news, content created by different persons, editorial control, a standards code, which may be the publisher’s own self-authored code, a complaints process and UK establishment requirements, with exclusions for proscribed and sanctioned entities.

Three consequences follow. The self-authored-code condition is the definition’s known weakness and the reason it games easily; a prominence regime needs a stronger code condition, red-teamed before enactment. The automatic inclusion of broadcasters means not every provider earns status through the same process; designation should regularise this. And section 56 was drafted as a shield, where over-inclusion is tolerable; as a prominence gateway the error-cost asymmetry flips, because over-inclusion becomes amplification. Its multiple-author, business and UK-establishment conditions may also exclude some solo publishers, informal or volunteer-run outlets, and overseas or diaspora publishers without a qualifying UK establishment, a plurality question the government should answer deliberately, potentially through plural accreditation routes.

Question 4a: how should this look in practice?

Seven mechanisms, consistent with the recommendations above. A public designation registry with a machine-readable interface, signed entries and freshness guarantees, so every platform, researcher and AI developer works from the same list and changes propagate on a defined timescale.

Designation bound to content credentials with a preservation duty, as set out in our second recommendation, including soft-binding recovery and account-level fallback. Interface treatments defined as outcomes rather than pixel prescriptions: dedicated news surfaces, surfacing in search for news queries, clearly labelled placement in feeds, implemented natively per product. AI assistants and answer engines within scope on the terms of our third recommendation. Discoverability floors with a measurement pilot first, expressly excluding impression-share quotas. Researcher access through the UK framework authorised by section 125 of the Data (Use and Access) Act 2025, guided by Ofcom’s section 162 report and extended to prominence and designation data.

And point-of-display labelling: every promoted placement carries a short “why am I seeing this” explanation naming the regime, because silent ranking adjustment, even benevolent, breeds the distrust it is meant to counter. Platforms should also provide standardised system cards and notice-and-explanation to publishers whose distribution the regime materially changes; we make that recommendation here, rather than under the creator-directed Question 24, because it belongs to the regime’s transparency architecture.

Question 5: always on, or only in crises?

In the survey’s fixed options we would select ongoing and “always on”, with a qualification: design two tiers of different strengths.

A crisis-only regime underperforms because audiences turn in a crisis to sources they already know, as the government’s own research on local media during the 2024 riots shows; habitual discoverability is what builds that familiarity. The always-on tier should therefore be modest: discoverability and labelled surfaces, not aggressive re-ranking of every feed.

The enhanced tier should activate only in defined information incidents: statutory triggers; activation by the regulator, not ministers; default duration in weeks, renewable by the regulator with published reasons; mandatory published post-incident review; opt-out preserved throughout. We would anchor this tier in existing UK machinery, Full Fact’s information-incidents framework and Ofcom’s crisis-response protocol finalised in June 2026, rather than in the DSA’s Article 36, which has never been activated and whose executive trigger the UK should not copy; Article 48’s crisis protocols are the closer European comparator. The safeguard against normalisation of emergency powers must be structural, not rhetorical.

Question 6: should users be able to switch prominence off?

Yes. Default on, one-step off, fully reversible, in both tiers, with “no degradation” specified as no punitive throttling, no loss of functionality and no dark patterns in the opt-out flow.

User agency is load-bearing for legitimacy, but it does not resolve the regime’s rights questions on its own.

Three interests are engaged under Article 10 of the European Convention: users’ freedom to receive information, non-designated publishers’ freedom to impart it without state-mandated relative demotion, and platforms’ own editorial and curation interests. An opt-out mitigates the first; it does not answer the second or third, and the state’s positive obligations on media pluralism cut in both directions. Legislation should therefore be accompanied by a published ECHR and plurality assessment covering the regime as a whole, with the opt-out treated as one proportionality safeguard among several. Published opt-out rates, as the Council of Europe recommended in 2021, give government and Parliament a continuing evaluation signal, read with caution for selection effects.

Questions 7 and 8: concern about misleading information

Very concerned, on both counts, answering in an organisational capacity on the basis of published evidence: 43 per cent of people reported encountering misinformation or deepfakes in 2024, only 45 per cent of adults are confident judging whether a source is truthful, and research cited in the green paper indicates that confidence in spotting AI-generated material tends to exceed actual ability. Two implications follow. Interventions should target the mismatch between confidence and competence, which exposure-reduction alone does not fix; and prominence should be understood as a supply-side complement to demand-side media literacy, justified on availability grounds, not offered as a cure for misinformation.

Questions 9 to 12a: public service media and media literacy

Public service media has a role in supporting critical engagement and should do more, provided the obligation is matched by sustainable funding and is not treated as a substitute for action by platforms, schools and civil society. The BBC’s reach makes it an effective delivery channel, and its Other Side of the Story media literacy programme is an existing route into schools. But media literacy is a property of the whole information system, not a deliverable of any single broadcaster, and judging the accuracy of information is becoming a baseline civic competence.

On the model, our answer to Question 12 is a structured forum with agreed objectives and reporting. The House of Lords Communications and Digital Committee found the media literacy sector highly fragmented, and informal coordination has not yet produced consolidation. A structured forum should set a small number of shared outcome measures, commission independent evaluation that measures effectiveness rather than activity volume, publish results annually, and include schools, libraries and civil society alongside public service media and platforms. Joint delivery is valuable where it follows agreed objectives rather than preceding them.

Questions 23 to 23c: prominence of public service media content on video platforms

We agree, with the same process disciplines, and we note the limits of the comparison honestly. Designating public service media content is cleaner at provider level, because PSM providers are already subject to independent Ofcom regulation. But the Media Act 2024 established a service-level framework, designated internet programme services receiving appropriate prominence on regulated television selection services, with Ofcom’s designation and code work still in progress; it is not a general guarantee for all PSM content everywhere. And individual PSM clips circulating on video platforms sit largely outside the Broadcasting Code’s application, which is precisely why the content-level credential layer matters: provider-level designation answers who, content-level provenance answers whether this clip is really theirs.

On benefits: a counterweight to purely engagement-driven discovery for the demographic that no longer arrives via broadcast; Ofcom research cited in the green paper shows social media is the main route to news for most 16 to 24 year olds.

On content types: news and current affairs first; children’s content and education second, as the genres where reliable provision has the clearest civic value and audiences the least ability to judge reliability themselves. On platforms, answering the question as asked, about video platforms: principally YouTube and TikTok, together with the video surfaces of Instagram and Facebook. Separately, and beyond this question’s scope, we recommend the statute define its surface functionally, services mediating news discovery at scale, so the regime does not need reopening for each new class of service, including AI assistants.

Question 30: equalities impacts

Three points, mapped to protected characteristics. First, age and disability: older and disabled audiences rely disproportionately on linear television for news, as Ofcom’s audience research consistently shows, so trusted-news prominence must not be designed only for app interfaces, and decisions on digital terrestrial television, on which we take no position, should include an assessment of their interaction with this regime.

Second, race and national or linguistic minority status: designation criteria and credential tooling must be achievable for minority-language and nations-and-regions providers, including Welsh and Gaelic language services, or prominence will structurally amplify London-centric national voices; hosted registry tooling is a concrete mitigation. Third, disability at the interface: prominence surfaces must themselves meet accessibility standards, subtitles, audio description, screen-reader compatibility, or the audiences most reliant on trusted information will be least able to reach it.

Low income is not a protected characteristic, but it correlates with several that are and with reliance on free-to-air news; the equality assessment should treat affordability of access as an indirect-discrimination question.

Closing

The government has asked the right question at the right time, and the green paper’s six principles are the correct ones. This response argues those principles hold only if the regime is built on independent designation, provenance-bound credentials with a preservation duty, tiered and time-limited crisis powers, published audit and genuine user agency, and it locates each recommendation in the existing body of UK and European practice, adding a package we have not identified elsewhere.

The Institute’s founding research programme includes comparative work on prominence, information integrity and AI-labelling regimes across the UK, the EU and the US. We would welcome the opportunity to discuss any part of this submission with the Department, including participation in the planned research on prominence methods and engagement with the AI Labelling Taskforce. Enquiries: engage@complexai.org.