YouCongress MCP: Turning expert disagreement into structured decision reports
Connecting YouCongress to Claude to ask about AI safety, governance, impact on jobs, and more. For example, we analyse here 'What should Spain do to prepare for a post-AGI world?'
I’ve been building YouCongress - a non-profit, open-source platform for mapping expert disagreement on policy questions (previous post on the coordination problem it addresses). It has two layers. The data layer is an open dataset where anyone can track who said what about specific policy proposals. We use AI to automatically discover candidate quotes from public sources, then humans verify attribution, context, and accuracy. Then, experts’ positions (for/against/abstain) are inferred from their verified quotes. The participation layer lets users vote directly or delegate to experts whose judgment they trust, a form of liquid democracy applied to policy questions. The core idea: on hard questions, what matters isn’t what one expert thinks, but how and why they disagree.
Recently I connected YouCongress to Claude via MCP. The MCP server currently exposes the data layer - the curated quotes and inferred positions. The result surprised me: Claude can pull quotes from YouCongress statements (policy proposals and claims), cluster them into coherent perspectives, map the axes of disagreement, identify the real tension points, and produce actionable recommendations. All in a single conversation, with full source traceability.
This post shows what that looks like in practice, and why I think the combination of a curated quote database + LLM reasoning produces something qualitatively different from asking ChatGPT a question.
The experiment: What should Spain do to prepare for a post-AGI world?
I asked Claude to pull quotes from YouCongress and answer the question. It retrieved 78 expert quotes and produced the analysis below.
A note on methodology and limitations. Claude selected the quotes from the YouCongress database using its own judgment - I did not constrain the selection or the clustering. The output below is lightly condensed for readability but not editorially altered. This means it inherits both Claude’s strengths (pattern-finding across many sources) and its biases (tendency toward balanced framing, possible anchoring on prominent voices). The YouCongress database itself skews toward English-language, public-facing figures - tech CEOs and Western economists are overrepresented; Global South voices, labor organizers, and non-English-speaking experts are underrepresented. A different LLM, or a human analyst, might cluster the same quotes differently. Treat this as one structured reading of the evidence, not the definitive one.
Four clusters of perspective
Not “pro vs con.” Each cluster represents a coherent worldview with distinct assumptions, and each implies different policy.
Cluster A: “AI will augment jobs and create net new employment”
Core thesis: Technology has always created more jobs than it destroyed. AI is powerful but not fundamentally different from prior revolutions.
Key assumptions: Labor markets adapt. Workers can reskill. Companies reinvest productivity gains into expansion. Historical precedent is the best predictor.
Representative voices:
David Solomon (CEO, Goldman Sachs, 2026): I’m not in the job apocalypse camp. Technology has been disrupting jobs, changing the way people work, destroying jobs, and forcing us as a vibrant economy to create new jobs for decades. It’s no different this time.
Sundar Pichai (CEO, Google & Alphabet, 2025): AI is making engineers more productive, driving new product development and creating demand for more employees.
Lisa Cook (Federal Reserve Governor, 2026): AI will usher in new tasks and occupations, many impossible to conceive today. [...] AI is poised to profoundly change the economy and our lives—I believe ultimately for the better.
Torsten Slok (Chief Economist, Apollo, 2026): AI is everywhere except in the incoming macroeconomic data. Today, you don’t see AI in the employment data, productivity data, or inflation data.
Policy implication: Light regulation. Invest in reskilling. Accelerate AI adoption in SMEs.
Cluster B: “AI will displace more jobs than it creates, and fast”
Core thesis: This time is different. AI automates cognition, not just physical labor - and it’s advancing faster than institutions can adapt.
Key assumptions: AI is qualitatively different: it replaces cognitive work across all skill levels simultaneously. Speed of change outpaces education systems. Companies optimize for cost, not workforce preservation.
Representative voices:
Dario Amodei (CEO, Anthropic, 2026): New technologies often bring labor market shocks, and in the past humans have always recovered from them, but I am concerned that this is because these previous shocks affected only a small fraction of the full possible range of human abilities, leaving room for humans to expand to new tasks. AI will have effects that are much broader and occur much faster, and therefore I worry it will be much more challenging to make things work out well.
Kristalina Georgieva (Managing Director, IMF, 2026): AI is like a tsunami hitting the labor market. Tasks that are eliminated are usually what entry-level jobs present, so young people searching for jobs find it harder to get to a good placement. [...] The middle class, inevitably, is going to be affected.
Geoffrey Hinton (Godfather of Deep Learning, 2025): I don’t believe there’ll be nearly as many new jobs created as it destroys, as it replaces.
Sam Altman (CEO, OpenAI, 2026): If it’s hard in many of our current jobs to outwork a GPU, then that [power balance between labor and capital] changes. [...] The next few years are going to be a painful adjustment.
Policy implication: Proactive safety nets. UBI pilots. Regulate pace of deployment. Urgently reform youth employment pipeline.
Cluster C: “The job count is the wrong question - distribution is”
Core thesis: AI may create abundance, but under current capitalism, gains flow to capital owners. The crisis isn’t unemployment - it’s inequality.
Representative voices:
Gita Gopinath (Former IMF Deputy Director, 2026): If AI leads to a transformation, where the labor share goes down by a lot more and the capital share goes up by a lot more, you can’t run the kinds of programs you’re running, in terms of entitlements, without having a higher capital income tax. It’s just not viable.
Daron Acemoglu (Nobel laureate, MIT, 2024): The US tax system favors introducing machinery instead of people. If you hire people, you pay 30 percent tax. If you hire machinery, you pay zero tax. These are the kinds of things that we can easily change.
Larry Fink (CEO, BlackRock, 2026): Early gains are flowing to the owners of models, owners of data and owners of infrastructure. What happens to everyone else if AI does to white-collar workers what globalization did to blue-collar workers? We need to confront that today directly. It is not about the future. The future is now. [...] Capitalism can evolve to turn more people into owners of growth, instead of spectators.
Dario Amodei (CEO, Anthropic, 2025): If AI creates huge total wealth, a lot of that will, by default, go to the AI companies and less to ordinary people. So, you know, it’s definitely not in my economic interest to say that, but I think this [a tax on AI companies] is something we should consider and I think it shouldn’t be a partisan thing.
Policy implication: Shift taxation from labor to capital. Create broad-based ownership of AI-productive capital (sovereign wealth funds, equity stakes). Pilot UBI or universal basic compute.
Cluster D: “Without sovereign AI capability, Europe becomes a digital colony”
Core thesis: The country that owns AGI controls the economic surplus. Europe (and Spain) cannot outsource intelligence and expect to share the gains.
Representative voices:
Jensen Huang (CEO, NVIDIA, 2024): Every country needs to own the production of their own intelligence. [...] It codifies your culture, your society’s intelligence, your common sense, your history — you own your own data.
Ursula von der Leyen (President of the European Commission, 2025): A European AI is essential for our future independence. It will help power our industries and our societies. From healthcare to defence. So we will focus on the first key building blocks – that’s from the Cloud and AI Development Act to the Quantum Sandbox.
Christoph Strnadl (CTO, Gaia-X, 2026): No US company can guarantee the US government will never access your data. If a US administration decided to stop service delivery to Europe, everyone would be out of business.
Policy implication: Invest in EU sovereign AI infrastructure. Support European AI labs. Create national compute reserves.
The disagreement map
The clusters don’t disagree randomly. They disagree on specific axes:
Axis A: Augment B: Displace C: Redistribute D: Sovereign Speed of AI capability growth Gradual Explosive Fast Explosive Worker reskilling feasibility High Low Irrelevant alone Depends on infra Historical precedent applies? Strong guide This time different Partial Partial Role of government Light regulation Heavy safety nets Structural reform Industrial policy Who captures value Broad economy Capital owners Capital owners (fixable) US/China unless you act Urgency Low - watch and learn High - now High - now Critical - now
The key insight: Clusters B, C, and D all agree that urgency is high, even though they disagree on the mechanism of harm. Only Cluster A counsels patience - and notably, Cluster A’s most recent voices (2025–2026) are increasingly hedged.
Three tension points
If you resolve these, the “right” policy becomes obvious. If you can’t, you need hedged strategies.
T1: Is cognitive automation qualitatively different from prior automation? Cluster A says no - steam engines, electricity, and the internet triggered the same fears. Cluster B says AI automates the thing that made humans adaptable: cognition itself. Spain’s economy is ~75% services, making it more exposed than manufacturing-heavy economies.
T2: Can Spain reskill fast enough? Spain’s FP (vocational training) system is growing, but university reform is slow and youth unemployment already exceeds 27%. If AI disruption hits in 2–5 years, the education system cannot retrain fast enough without radical reform.
T3: Can Europe capture enough AI value to redistribute? Clusters C and D converge: redistribution requires something to redistribute. If Spain’s AI value chain runs on US models, US compute, and US platforms, there is no domestic surplus to tax. Sovereignty and redistribution are coupled problems.
What should Spain actually do?
I used Spain as a test case because I’m based here and know the policy context, but the pipeline works for any country or policy question — swap the prompt and the same clusters → disagreement map → recommendations structure applies. The smart move for Spain: take actions that are valuable under multiple scenarios. Here are eight that work regardless of which cluster turns out to be right.
1. Shift tax base from labor to capital + consumption - Even if Cluster A is right, equalizing the tax treatment of labor and capital removes the perverse incentive to automate for tax reasons. Spain should push this at EU level and begin domestic reform immediately. Act now.
2. Build a national AI compute reserve - Even augmentation requires compute. Spain should join the EU Cloud & AI Development Act investments and build national inference capacity, giving SMEs affordable access. Act now.
3. Radical FP + university reform for AI literacy - Not “learn to code” but reform curricula so every profession integrates AI tools. Partner with employers for AI apprenticeships. Spain’s current cycle for curriculum reform (3–5 years) is too slow. Act now.
4. Negotiate bilateral AI talent visas with Latin America - Spain’s unique advantage in the global AI race is the 600M+ Spanish-speaking world. Create fast-track visas for AI engineers and researchers from Latin America. This is something France and Germany cannot replicate. Spain could become the EU’s bridge to LatAm AI talent - a structural advantage no other member state can claim. Act now.
5. Launch a UBI pilot in one Comunidad Autónoma - Critical under Clusters B and C. Even if not needed yet, running a controlled experiment now provides data and institutional readiness (expert positions on UBI). The Basque Country or Navarra (with existing Renta Garantizada programs) are natural candidates. Start 2027.
6. Create a Spanish Sovereign AI Fund - If C and D are right, Spain needs ownership of AI-productive capital. A sovereign fund that invests in European AI companies and infrastructure gives the state a mechanism to redistribute future AI dividends. Norway’s oil fund is the model; AI is Spain’s equivalent moment. Design now, fund 2028.
7. Mandate AI-augmented (not AI-replaced) public services - Government is the one employer that doesn’t optimize for profit. Mandate that public sector AI deployment augments civil servants rather than replaces them, creating a floor of human-in-the-loop employment. Ongoing policy.
8. Lead the EU push for an “AGI Preparedness” framework and an EU+ AI alliance - Spain should push for an EU-wide AGI Preparedness Index benchmarking member states on fiscal resilience, reskilling capacity, compute sovereignty, and social safety nets. Beyond the EU, push for a broader EU+ AI alliance - with the UK, Switzerland, Canada, Australia, and Norway - focused on shared compute infrastructure, open-weight frontier models, and talent mobility. These countries share a growing recognition that dependence on US AI infrastructure is a strategic vulnerability; Spain can be the convener. An institutional model for this exists - a “CERN for AI” focused on shared compute and safety research, which already has broad support from figures like Tim Berners-Lee, Demis Hassabis, and Yoshua Bengio (YouCongress statement with 56 contributors). Start now.
Why this matters
Three things I think this demonstrates:
1. Structured disagreement > consensus-seeking. Most AI governance discussion tries to find “the right answer.” But on genuinely uncertain questions (will AI create or destroy jobs?), mapping why smart people disagree is more useful than pretending we can resolve the disagreement. The 54%–42% split among 69 experts with sourced positions on whether AI will create more jobs than it destroys isn’t noise - it reflects genuinely different models of how economies adapt to technology.
2. LLMs + curated data > LLMs alone. John Burn-Murdoch’s analysis for the FT found that all major LLMs nudge users away from extreme positions and toward moderate, expert-aligned stances — the opposite of social media’s polarising effect. YouCongress leans into this strength: instead of relying on the LLM’s training data alone, it feeds it a curated database of real expert positions with the disagreements preserved. When you ask ChatGPT “what should Spain do about AI and jobs?”, you get a generic answer with few or no sources. When Claude pulls 78 verified expert quotes from YouCongress and clusters them, you get traceable reasoning grounded in real positions real people hold. The quotes aren’t decoration - they’re the epistemic foundation that makes the analysis trustworthy.
3. MCP makes this composable. Anyone can connect YouCongress’s MCP server to Claude (or any MCP-compatible tool) and generate their own analysis. Ask about AI safety policy proposals, AI governance in Brazil, autonomous weapons policy, or whether to open-source frontier models. If YouCongress has quotes on the topic, the analysis pipeline works. No new product needed. The data is the product.
Try it yourself
The YouCongress MCP server is open — connect it to Claude or any MCP-compatible assistant and run your own analysis. Current coverage: 60+ statements, 2,000+ sourced expert positions, 1,300+ authors. It’s free and it doesn’t require sign up.
What I’m looking for
Feedback on the analysis framework. Does the clusters → disagreement map → tension points → actions pipeline produce useful output? What’s missing? What other use cases do you find?
Statement suggestions. What policy questions should YouCongress cover next? I’m particularly interested in topics where expert disagreement is high and the stakes matter.
Stress-test the dataset. We are starting online quotathons every two weeks — open calls where anyone can join, pick unverified quotes, and check them against sources together. If you want to poke at the data or help improve it, join the next one.
YouCongress is non-profit and the code, the MCP server, and the analysis method is open. If structured disagreement mapping is useful, connect and build on it.