A Beyond the Headlines thought experiment
A note before you read
This isn’t a proposal to replace elected politicians with artificial intelligence. Nor am I suggesting that AI is unbiased, infallible or capable of deciding what is best for a country.
Quite the opposite.
This is a thought experiment about whether technology could give citizens another way of scrutinising decisions made in their name.
You may think it’s a brilliant idea. You may think it’s dangerous.
Either response is reasonable.
It started with something much smaller
When I’m developing an article, I sometimes use more than one artificial intelligence system to examine what I’m writing.
The original idea is mine, but I can ask AI to challenge an argument, look for weaknesses, question assumptions, suggest another interpretation or identify something I’ve overlooked. Sometimes the systems agree. Sometimes they don’t. Sometimes they tell me something I don’t particularly want to hear.
That’s useful.
I still decide what I believe. I still decide what gets written. And I remain responsible for what I publish.
Recently, though, this made me wonder:
What would happen if we applied the same principle to government?
Imagine this
Britain creates a national AI system operating at arm’s length from government.
Not an AI Prime Minister. Not an electronic Parliament. Not a machine with the power to make laws.
An adviser.
Its purpose would be to examine major policy proposals and expose the arguments, evidence, uncertainties, competing options and potential consequences surrounding them.
And, wherever legally possible, the analysis and evidence behind it would be available for the public to examine too.
Calling such a system “independent” is easy. Making it genuinely independent would be considerably harder.
Perhaps it could operate through a statutory body whose independence was protected by Parliament, with leadership, funding and auditing arrangements that couldn’t easily be changed by whichever party happened to be in government.
That wouldn’t guarantee independence.
But it gives us somewhere to start.
Give it a difficult question
Take irregular migration.
I’m choosing it deliberately because it is difficult. The debate involves law, economics, border enforcement, housing, public services, administrative capacity, labour markets, humanitarian obligations, integration, social cohesion and deeply contested political judgements.
Instead of asking:
“Is immigration good or bad?”
we might ask:
“What realistic options does Britain have for reducing irregular migration, and what would be the likely consequences of each?”
The system wouldn’t be instructed to produce the politically acceptable answer. It would be instructed to investigate the question as widely as possible.
That might mean examining British and international law, enforcement capabilities, fiscal costs, economic effects, housing and public-service pressures, labour-market consequences, administrative backlogs, integration and social cohesion, humanitarian consequences, crime and security evidence where reliable evidence exists, international comparisons and the underlying causes of irregular migration.
But I’d give it another instruction:
What important factors have we failed to include in the question?
That matters because our assumptions can enter an analysis before we’ve even begun looking for answers.
If reliable evidence supports a concern, report it. If reliable evidence contradicts one, report that too. If credible evidence conflicts, show us the disagreement.
And sometimes the most truthful answer may simply be:
We don’t have enough evidence to know.
Don’t give us one answer
I wouldn’t necessarily want one AI producing one recommendation.
I’d want competing analyses.
One part of the system could construct the strongest evidence-based argument for a proposal while another attempts to demolish it. Others could examine economics, law, individual rights, implementation difficulties and unintended consequences.
Then show us where those analyses agree and, perhaps more importantly, where they disagree.
Instead of:
“The computer says do this.”
we might receive several realistic options, each showing likely benefits, costs, legal issues, supporting and contradictory evidence, uncertainties and possible unintended consequences.
AI provides the analysis.
Then elected politicians make the decision.
Humans still decide
This is the line I wouldn’t cross.
AI should advise. Humans should decide.
Interestingly, that principle isn’t entirely removed from the direction government AI policy is already taking.
The UK government’s AI Playbook for the UK Government, published in February 2025, contains ten principles for the use of AI in government. One is explicitly about having meaningful human control at the right stages.
The guidance says humans should validate high-risk decisions influenced by AI. It also warns that AI models can produce unwanted or inaccurate results and calls for testing, assurance, monitoring and mechanisms for human intervention.
The government’s Data and AI Ethics Framework provides further guidance on the responsible development, procurement and use of data and AI across the public sector.
My proposal goes considerably further than either document.
But the underlying principle already exists:
Using AI doesn’t remove human responsibility.
In this thought experiment, government remains responsible. Parliament remains responsible. Ultimately, voters decide whether they approve of the choices politicians make.
AI doesn’t get a vote.
Suppose the analysis recommends Option A and the government chooses Option B.
That’s perfectly legitimate.
But imagine if the government then published its reasoning:
We considered the analysis and chose a different course because...
Now we’d have something valuable: a contemporary record of the evidence, alternatives and reasoning available when the decision was made.
Where practical, the analysis could happen before major policy decisions. Later reviews could compare what the system anticipated with what actually happened.
That matters for another reason.
AI needs accountability too.
If its predictions repeatedly prove wrong, we should be able to see that.
Beyond the Headlines explores difficult questions without asking you to adopt a predetermined answer. If you’d like to continue the conversation, subscribe to Third Act Life.
Don’t let the original disappear
The reports should be timestamped, cryptographically signed and permanently archived so subsequent alterations could be detected.
The particular technology matters less than the principle:
Once published, the original analysis shouldn’t be capable of being quietly rewritten.
Years later, we could look back.
What did the evidence suggest? What risks were identified? What did government decide? Why did it make that decision? What actually happened?
And importantly:
How accurate was the AI?
Politicians aren’t the only people — or machines — that should be accountable for their previous claims.
But democracy should work in both directions
So far, information has travelled one way:
Government proposes something → AI scrutinises it → the public sees the analysis.
That’s about accountability.
But the same architecture could work in the opposite direction:
A citizen proposes something → AI scrutinises it → a worthwhile proposal can potentially reach government.
That’s about participation.
Imagine someone sitting at their kitchen table who has spotted a problem nobody in Westminster seems to be discussing.
Or a nurse, builder, teacher, scientist, pensioner, small-business owner or university professor.
They submit an idea.
Their status shouldn’t determine its value.
The proposal gets tested against evidence. What problem does it solve? Is it legal and practical? What would it cost? Who benefits? Who could be disadvantaged? Has something similar been attempted elsewhere? What are the strongest arguments against it? What could go wrong?
Most ideas probably wouldn’t survive serious scrutiny.
That’s fine.
But occasionally one might.
Perhaps a genuinely useful proposal could rise because of its merits rather than because its author happened to have political influence.
Now let’s attack my own idea
There are enormous problems with this.
If government funds the system, can it genuinely challenge government? Who appoints the people overseeing it? Who trains the models, chooses the evidence standards and determines what constitutes a reliable source?
Then there’s something even more difficult.
Values.
Imagine two policies.
One produces greater economic growth but increases inequality. Another produces less growth but distributes its benefits more evenly.
Which is “better”?
Or consider liberty against security. Individual rights against collective interests. Short-term costs against benefits that might not appear for twenty years.
AI cannot magically turn those into objective questions.
They are political and moral judgements.
So perhaps the system shouldn’t pretend to be value-neutral at all.
If an analysis prioritises economic efficiency, individual liberty, equality, national security or some other value, that assumption should be visible and contestable rather than buried inside the system.
Moderation creates another problem.
A public system couldn’t simply accept everything submitted to it. Threats, illegal material and incitement would require moderation.
But the moment an authority starts determining which political ideas are unacceptable, another danger appears.
Who watches the gatekeeper?
Organised manipulation would also be a serious risk. Corporations, lobbyists, campaign groups, political movements and foreign governments could attempt to influence the system or overwhelm it with submissions.
And perhaps the most important problem of all:
AI can be wrong.
It can misunderstand evidence, reproduce biases and produce extremely convincing arguments that turn out to be false.
The government’s own AI Playbook explicitly warns that AI models can produce unwanted or inaccurate results.
Giving such a system political authority would therefore be dangerous.
Which is precisely why, in this thought experiment, it doesn’t get political authority.
This isn’t entirely science fiction
Britain already has something called the Algorithmic Transparency Recording Standard, or ATRS.
It provides a standardised way for public-sector organisations to publish information about how and why they use certain algorithmic tools.
Its mandatory scope currently covers central government departments and specified arm’s-length bodies. Within those organisations, it applies to certain algorithmic tools that significantly influence operational decisions with public effect or directly interact with the public.
But there’s an important distinction.
The government’s own guidance says this mandatory scope is aimed at tools influencing specific operational decisions about individuals, organisations or groups.
Analytical models supporting broad government policymaking will typically fall outside that mandatory scope.
So the system I’m imagining would go substantially further than today’s transparency requirements.
That’s important because it separates what exists from what I’m proposing.
The ATRS demonstrates that transparency around government algorithms is already an established principle.
My thought experiment asks:
What happens if we extend that principle to the analysis behind major government policy?
Would it improve democracy?
I don’t know.
It might expose weak evidence, make political decisions easier to scrutinise and allow worthwhile ideas from outside traditional political institutions to receive serious consideration.
It might make it harder for governments to say afterwards that foreseeable risks hadn’t been identified.
But it could also create an enormously powerful new institution.
Whoever controlled its architecture, evidence standards, funding or governance could acquire considerable influence over political debate.
And there’s another danger.
People might start trusting an AI recommendation simply because a machine produced it.
That could be every bit as dangerous as blindly trusting a politician.
So I’m not suggesting that we hand government to artificial intelligence.
I’m asking whether we could use artificial intelligence to make human government more accountable and public participation more meaningful.
The practical questions eventually come down to three:
Who controls it?
Who gets access to it?
Can everyone see how it reached its conclusions?
But perhaps there’s an even bigger question.
Would a system like this strengthen democracy — or would we simply create another powerful institution that we eventually learn not to trust?
I genuinely don’t know.
That’s why I’m putting the idea out there.
Don’t let me tell you what to think.
Over to you
Would you trust an independently governed AI system to advise government if elected politicians retained the final decision?
Would you submit your own idea to it?
What safeguard have I missed?
And if you think the entire concept is dangerous, I’d particularly like to hear why.
Sources and further reading
UK Government — AI Playbook for the UK Government
Government guidance covering safe, responsible and effective use of AI, including meaningful human control.
UK Government — Data and AI Ethics Framework
Guidance on responsible development, procurement and use of data and AI across the public sector.
UK Government — Algorithmic Transparency Recording Standard
The government’s framework for publishing information about certain algorithmic tools used in the public sector.
UK Government — ATRS Mandatory Scope and Exemptions Policy
The detailed rules defining which central-government organisations and algorithmic tools fall within mandatory transparency requirements.