When something you read turns out to be false, you usually find out because somebody corrected it. A reply, a note, someone who knew better. Rarely a government.
The UK is now considering a different arrangement: a list of news organisations the state considers trustworthy, which platforms would be required to rank above everything else.
That is what sits in chapter two of “Watch this Space: a new strategic direction for UK media”, a green paper published by the Department for Culture, Media and Sport on 23 June. Most of the document is television admin: switch-off dates, broadcasting licences. Chapter two is something else entirely. The government wants to explore legislation requiring social media platforms, and possibly video platforms too, to make “trustworthy” news prominent and easy to find.
It is a decision about who gets to shape the ranking systems that decide what millions of people see. That makes it an AI governance question, and it should be judged as one. It is a consultation rather than a plan, and the paper is honest about that. But options papers are where the architecture gets set.
Why the government thinks it must act
Start with why they think they have to do something, because that part is reasonable.
After the Southport attack in the summer of 2024, Ofcom found that recommendation algorithms had amplified misleading and divisive content during the disorder that followed. Not created it. Amplified it. The machines chose what people saw, and what they chose made things worse.
The numbers around that are uncomfortable. Most people now meet misleading or synthetic material regularly, fewer than half feel confident judging whether a source is truthful, and the research cited in the paper found something worse: people’s confidence in spotting AI-generated content runs ahead of their actual ability to do it. We are getting worse at this at exactly the moment the fakes are getting better.
The paper says so directly. Generative AI, it argues, “makes the case for intervention more urgent than ever”. The BBC has put numbers on the same worry, and the EBU and BBC’s study of AI assistants, covering 22 public service broadcasters across 18 countries, found the pattern holds internationally.
None of this is happening in isolation either. The government is standing up an AI Labelling Taskforce, running a consultation on deepfakes and digital replicas, and backing a detection framework built with Microsoft and others. So this is not nostalgia for the six o’clock news. It is Britain’s most serious attempt yet to write integrity rules into the layer where people actually find things out.
What is actually proposed
The green paper is genuinely open on the biggest questions. Ministers commit to six principles: measures should be audience-led, future-proofed, outcomes-based, comprehensive across national and local news, protective of plurality, and governed by criteria for a “trustworthy” provider “decided in an open and transparent manner with regard to protecting media freedom”. Three design choices are put directly to the public. Who counts as trustworthy at all? Should prominence run continuously, or only in periods of heightened vulnerability such as civil unrest? And should individual users be able to turn it off?
The paper also floats making prominence conditional. Get the benefit, take on extra duties, possibly including “more stringent standards in AI content use”. I would be careful here. We support conditions only where they are about the regime itself: keeping your credentials honest, being transparent about facts relevant to your designation, handling complaints properly. We oppose tying reach to how a newsroom runs itself, including how it uses AI. That is editorial interference with a compliance label on it.
Britain has been here before and it did not end well. Section 40 of the Crime and Courts Act 2013 tried to use costs rules to push publishers into a recognised regulator. It was never brought into force, it was fought over for a decade, and Parliament repealed it in 2024. Any prominence regime that forgets that history will refight it.
Then there is the question of who makes the list in the first place. The paper suggests starting from a definition that already exists in law: the Online Safety Act’s “recognised news publisher” test.
It is a reasonable place to start and a bad place to finish, and the common shorthand about it gets things backwards. That definition was written as a shield. Its job is to stop platforms over-removing legitimate news. Being generous about who qualifies was fine for that, because the worst case was that a marginal publisher avoided a takedown.
Turn the same definition into a gateway for promotion and the worst case inverts. Now over-inclusion means amplification, paid for by everyone else’s reach. It also includes all licensed broadcasters automatically, and lets a publisher satisfy the standards condition with a code it wrote itself. Before anything is attached to that test, somebody should try hard to game it and see what gets through.
Nobody else has gone this far
Britain is not acting alone, but the regimes it is joining are less similar than they look. Germany requires “easy findability”, but on television-style interfaces, not social feeds. The EU’s Media Freedom Act protects media from being taken down; it does not promote anyone. The UK’s own Media Act 2024 covers television selection services.
None of them re-ranks an ordinary social feed. This proposal would.
That is worth sitting with, because the reaction has already started. On 14 July Jim Jordan, the Republican chair of the US House Judiciary Committee, wrote to the culture secretary, alleging the proposals would give prominence to state-approved outlets and infringe the rights of American companies. YouTube’s European policy lead has warned that prominence rules would force platforms to promote government-picked channels.
You can dismiss the politics of those interventions. The lesson survives either way: any state-backed trust list will be called state media by its opponents, and that charge sticks wherever the process behind the list is hard to see.
The wrong object
I named this problem Complex AI in 2019. A trust list has exactly the shape I was describing: several objectives pulling against each other, real consequences, and no way to inspect how the answer was reached.
The question a reader actually has is narrow. Is this claim true? Is it current? Is it fairly attributed? Three things at once, pulling in different directions. That is a multi-objective problem.
A publisher list answers something else: who the speaker is. One variable. It optimises for one thing because one thing is tractable, and my argument since 2019 has been that real problems never optimise for one thing.
Think of it like hiring. You would not employ someone purely because they went to a good university. The university is a proxy. It correlates with something you care about, loosely, and it tells you nothing about whether this particular person can do this particular job. A masthead is the same proxy, applied to a sentence.
The audit trail lands in the wrong place too. Under this regime you could inspect why a newspaper was designated. You could never inspect why one sentence inside one article is reliable. We would be certifying the letterhead and calling it verification.
There is also a failure mode I have watched play out in every sector we work in. A list demonstrates beautifully. It is legible, it fits inside a policy document, a minister can hold it up. Then it is deployed, and it is gamed, captured, and slowly stops tracking the thing it was a proxy for. The models became general; the deployments did not. That gap is the reason this Institute exists, and a trust list walks straight into it.
What I would build instead
Let me be plain about where I stand, because it is not where most of this debate sits. Governments should not be deciding whose journalism gets amplified. That is a propaganda machine with a review process attached, however good the intentions of the people building it. What a government can legitimately do is make it easier for a reader to verify what is in front of them.
The most interesting working example is not in any statute. It is X’s Community Notes, and it deserves far more attention from policymakers than it currently gets, precisely because it is engineered against the failure I have just described.
The mechanism is a bridging algorithm. A note is only shown once it has been rated helpful by people who historically disagreed with each other. Consensus across opposing priors is the qualifying condition, not agreement with an authority. That is a multi-objective test built into the ranking itself, which is rarer in deployed systems than it sounds. The code and the ratings data are both public, so the apparatus is auditable by anyone who cares to check. And it attaches to the claim, not to the speaker.
The effects are measurable. X’s own A/B testing found notes cut decisions to like and repost false material by 25 to 34 per cent, and independent work published in PNAS found roughly half as many reshares once a note is attached.
I would not hold it up as finished, and here my own discipline applies to a mechanism I happen to like. Notes surface too slowly to catch the first hours of a falsehood, which is when most of the damage is done. Around 30 per cent of displayed notes later lose their helpful status and disappear. Coverage is thin outside the largest language markets. Those are deployment failures rather than design failures, and they are exactly the sort of problem an institute like ours should be working on rather than admiring from a distance.
So my answer to the green paper’s central question is not one of the options the survey offers. Do not build a register of approved publishers. Build the verification layer instead: provenance that survives the journey into clips, embeds and AI answers; claim-level traceability duties on anything that mediates news at scale; open audit so outsiders can check what the ranking actually did; and public, bridging-style correction mechanisms that no minister sits above. If prominence is used at all, it should attach to verified content, never to an approved name.
What we are actually asking for
The Institute has filed a full response to the consultation. It runs to eight recommendations and answers the questions one by one, and if you work on this for a living that is the document to read rather than this one.
The principle underneath all eight fits in a sentence.
In short: designation, if it happens at all, sits with an independent authority and ministers are barred by statute from touching it. Trust status is bound to the content itself and platforms must preserve it rather than strip it. The rules cover AI assistants from day one, not just feeds. Outcomes are audited by outsiders. And any reader can switch the whole thing off in one step.
Three questions to ask of any version of this
Prominence rules are coming in more than one country, and most readers will meet them as a headline rather than a statute. So here is the test I apply, and it takes about a minute.
1. Can a minister influence what gets promoted? Not in theory, in the drafting. Is there a statutory bar on ministerial direction over who is designated, or does the power sit with a department? If a politician can pick up a phone and change the list, nothing else in the design matters much.
2. Can an outsider check what the system actually did? Not what the platform reported. What an independent auditor and a vetted researcher could reconstruct from the data. If the honest answer is no, you are being asked to take the outcome on trust, which is the opposite of what a trust regime should require.
3. Does it still work when the news arrives through an AI assistant? Most of these proposals are written for feeds. A regime that regulates feeds and ignores answer engines will be obsolete before it commences.
A proposal that passes all three is worth arguing about on the merits. One that fails any of them is a governance problem wearing media-policy clothes, and it should be sent back for redrafting rather than debated as though the details were minor.
What happens next
The consultation closes at 11:59pm on 31 August 2026. You can respond through the survey on GOV.UK, by email to watchthisspace@dcms.gov.uk, or by post. If you have a view on this, that is the address for it.
After that, several things run in parallel: research into how prominence would actually be measured, the AI Labelling Taskforce’s interim report in the autumn, and the deepfakes consultation. We will be following all of them together, here and in other jurisdictions, because they are converging on one question that will shape the next decade of information policy.
Not whether algorithms should carry integrity obligations. They should. The question is who defines them, and who checks the checker.
The UK has asked the right question. My answer is that the machinery should verify what is said, not license who gets to say it.
Our full response to the consultation, eight recommendations and question-by-question answers, is published here: Response to the UK Media Green Paper. The filed version will replace it on submission. Organisations and individuals can respond to DCMS until 11:59pm on 31 August 2026 via GOV.UK.
Continue reading: Response to the UK Media Green Paper