One member. One token.
Replace the membership with digital twins. They attend every consultation, never lose a password and agree in exactly the proportions specified in the prompt.
An anonymous SDP memberIndependent satire
Your new political operating system
Trust the machine.Members are a legacy feature.
Why spend years building democratic judgement when you could put everything in a box marked AI?
Confidence: limitlessMandate: not foundCloud: France
The unofficial political operating system, inspired by Eleanor Rofe’s campaign.
Delegate something irreversible*Choose a scripted conversation
More ideas from the Ministry of Unnecessary Automation
*The agents are in control. The agents have drawn a reassuring diagram.
Your text stays in this page. There is no upload, account or message storage.
The joke is fictional. The questions are serious.
The system that chooses the evidence, flags the problem and writes the draft can influence the decision before anyone approves it.
The fictional machine manifesto
No humans required.
Terms may apply.
The imaginary ChatSDP platform for a frictionless political future. Every inconvenient human function has been reclassified as a software problem.
The fictional whole-party upgrade
Policy and research?
Why stop there?
If a machine can carry the policy workload, our product team sees an exciting opportunity to automate absolutely everything else. People are a legacy integration challenge.
Introducing the party-as-a-startup stack
From grassroots democracy
to seed-stage disruption.
Build the machine. Fill the membership funnel. Scale the comms. Somewhere between the investment case and the product roadmap, the members would quite like a word.
EllieBot™ · campaign androidHuman mode: enabled*
Meet EllieBot.
Eleanor Robot, the imaginary campaign android. Warmth preloaded. Conviction cached. Opening and closing remarks available at the touch of exactly the same button.
Région: France. English mode: mostly enabled.
READY. Please select a pre-installed display of spontaneous conviction.
650 ChatSDPs by 2029.
Every constituency receives a robot candidate, a charging station and a subscription to Local Concern Premium. Personal conviction available as an optional plug-in.
Nationalise the Approve button.
Human oversight will be preserved by appointing a human to press it. The machine will choose what appears above it. This arrangement will be described as empowerment.
A constitutional update overnight.
Member sovereignty has moved to Settings → Advanced → Legacy Features. The consultation period will last until the loading bar reaches 100%.
Democracy on autocomplete.
Save time by consulting the model about what members would probably have said. Publish the result as “a rich and diverse conversation”.
All hail the knowledge graph.
A blue line between two boxes shall constitute an insight. Three lines shall constitute a strategy. Any further questions shall be answered with a larger diagram.
Privatise the national mood.
The public will be represented by whichever social-media accounts were easiest to scrape. The silent majority will remain silent because it was not in the dataset.
Install ThinkTankGPT.
Experience fresh external thinking, preloaded with assumptions. Funding declarations, access rights and appointment procedures must be downloaded separately.
A paperless accountability system.
When a decision fails, responsibility will pass seamlessly from the operator to the vendor to the model to the training data and back to the operator. Nobody waits long enough to be accountable.
Ban low-productivity dissent.
Disagreement will be recoded as inconsistency. Inconsistency will trigger a review alert. The alert will recommend removing the disagreement. Consensus achieved.
A right to remain synthetic.
Citizens may appeal a generated assumption about them by completing a form that will be summarised by the same model. The appeal will be marked “human involvement”.
Permanent independence, billed monthly.
The party will control its destiny through a supplier’s dashboard, on a supplier’s terms, until a supplier’s update changes the available destinies.
A completely fictional democratic exercise
Accept your ChatSDP overlord.
Choose freely from three equally meaningful options. The machine has already predicted your answer.
Awaiting your entirely voluntary compliance.
Fictional release notes · Democracy 2.0
We have improved your experience.
- Fixed a bug where real members introduced unexpected opinions.
- Reduced meeting times by removing the debate.
- Moved “Who authorised this?” to an unread channel.
- Renamed supplier dependence “strategic autonomy”.
- Known issue: the human condition. A patch is being considered.
Sources, system design, draft wording and decision rights: behind every frictionless interface is someone deciding how the friction should be removed.
Research and implications
A human can approve
a machine-shaped choice.
The risk begins before the final click. A system that gathers evidence, defines connections, selects alerts and drafts proposals can influence what a policy team sees as the problem—and which answers seem available.
Follow the decision before the decision · choose a stage
The feed is already a political choice.
A source list can favour information that is easy to scrape, frequently published or institutionally prominent. Local experience, minority positions and information outside the chosen feeds can disappear before analysis begins. More documents do not automatically mean a wider view.
The control that matters: Publish source-selection criteria, record important gaps and provide a route for people to submit evidence outside the automated feed.
Analytical walkthrough of a possible design, not a technical specification supplied by the campaign. Each stage could be governed well or badly.
A plausible failure chain—not an observed incident
A small error can become institutional confidence.
A weak connection → a confident alert → a plausible draft → a rushed approval → a future source.
If a tentative relationship loses its uncertainty label, later summaries may treat it as settled knowledge. Linking outputs back into the store can make the same unsupported claim appear in multiple places. Independent sources, original documents, correction propagation and a visible distinction between evidence and generated interpretation are safeguards against that loop.
What the evidence actually says
The concerns are more specific than “AI might be wrong”.
Experiments, technical guidance and UK governance standards identify different failure mechanisms. The sources below support scrutiny of the proposed design; they do not establish that every risk will materialise in this project.
The draft can change the drafter.
Jakesch and colleagues asked 1,506 participants to write about social media. Writing assistants deliberately configured to favour one side affected both the views expressed in the text and the participants’ subsequent survey answers.
Implication: an assistant can influence substantive judgement while appearing to help with wording. Scrutiny should reach the model’s defaults and draft framing, not end at who presses Approve.
Method notes
Limit: one topic, deliberately opinionated models and a specific experimental setting. This does not demonstrate manipulation by the proposed system or establish how lasting any effect would be.
“Human in the loop” is not a result.
NIST identifies over-reliance on generated content and confident falsehoods, including invented citations. It recommends checking sources and citations. The UK Data and AI Ethics Framework calls for meaningful human ability to monitor and influence systems, with accountable roles and output verification.
Implication: count the reviewer’s time, skills, evidence access and power to challenge. An approval box proves that a click occurred; it does not establish that the reasoning was sound.
Method notes
Limit: these sources describe recognised risks and governance expectations. The government framework is a useful comparator, not a rule automatically binding a party’s internal policy process.
NIST AI 600-1: §§2.2, 2.7; MS-2.5-003UK Data and AI Ethics Framework
More output does not prove better judgement.
Vaccaro and colleagues combined 106 experiments. Human–AI teams improved on humans alone on average, but underperformed the better of the human-only or AI-only conditions. Results differed substantially across tasks; decision tasks showed less synergy than content-creation tasks.
Implication: success at producing drafts cannot be assumed to establish success at making policy choices. Test the actual workflow, with its actual reviewers and consequences.
Method notes
Limit: heterogeneous studies published from 2020 to June 2023, often unlike political policymaking. This is not evidence that AI assistance always makes people worse or that this particular system would fail.
Vaccaro et al.: When combinations of humans and AI are useful
The connections can be sensitive data too.
The ICO’s current guidance includes political opinions and deliberately drawn inferences revealing special-category information. Specified purposes and compatible reuse still matter. A graph could infer a person’s views or associations even where the original records did not explicitly state them.
Implication: scrutiny must cover what the system derives, who sees it and how errors are challenged. “Everything we know” is not a test of what a particular task needs.
Method notes
Limit: applicability depends on the information, purposes and processing. This does not establish an existing breach. Consent is not the only potentially available condition, and membership is not a blanket exemption.
A knowledge base does not grant permission.
NCSC warns that malicious instructions can enter through external content. OWASP’s retrieval-security guidance separately calls for permissions to follow retrieved material and for deletion and access changes to reach indexes and caches.
Implication: keeping the original membership file restricted is insufficient if a copied passage can emerge in somebody else’s answer. Source ingestion, retrieval access and any power to update records need separate controls.
Method notes
Limit: these are implementation risks. The proposal’s actual access rules are not known. Searching documents does not necessarily involve training a model on them; neither architecture is automatically safe or unlawful.
Measure the work that survives checking.
HM Treasury’s AI evaluation guidance describes comparing briefing work with and without an LLM, recording time and independently assessing quality and surviving errors. NAO’s government review also stresses skills, funding, data foundations and clear ownership.
Implication: measure time per verified useful result, missed changes, false alerts, checking effort and continuing cost against the existing human process. Connections drawn and pages generated are not sufficient measures of success.
Method notes
Limit: the Treasury briefing example is explicitly hypothetical; its illustrative numbers are not empirical findings. Government experience is a comparator, not a cost estimate for this party.
AI evaluation guidance: Annex A.2NAO: Use of artificial intelligence in government
Access to a register is not access for any purpose.
The Electoral Commission lists permitted uses of the full register by recipient. Registered parties’ uses include electoral purposes and donation-control compliance. Candidate access has its own restrictions. The campaign itself acknowledges electoral-data ringfencing.
Implication: a shared store would need to preserve purpose and disclosure restrictions through indexes, summaries and supplier services. This does not prohibit suitable election tools; it requires the proposed flows to be specified.
Method notes
Limit: the slides and public pages do not establish that restricted register records would be supplied to the AI engine. That connection must not be assumed.
Invented respondents are not a consultation.
Bisbee and colleagues found that ChatGPT-generated US survey responses could approximate averages while distorting variation and relationships, and changing with prompts and time.
Implication: a model’s simulated agreement should not be presented as the views or consent of real members. Validity needs external evidence, not a second generated answer reassuring us about the first.
Method notes
Limit: older models and particular surveys; not a universal finding about every current method. Synthetic members are an absurd extension invented by this parody, not a proposal established in Rofe’s material.
A calculator can count assumptions perfectly.
The campaign calculator uses provisional participation inputs and assumed vote shares. Its win label is a threshold comparison, while feasibility checks hours rather than a funds constraint. The battleground map labels its scores draft and says member and resource geography is not yet included.
Implication: attractive outputs and exact counts cannot establish whether the assumptions are calibrated. A result should retain its conditions when it becomes a funding or membership argument.
Method notes
The reviewed pages run browser calculations and embedded data; no live AI model calls were found. AI-assisted construction, live AI operation and validated modelling are different claims. The map itself proposes further testing.
Evidence: Eleanor Rofe’s campaign website, party hustings and her online public surgeries.
Who would own the judgement?
Outside advisers can supply evidence, a supplier can supply software, and a model can supply a draft. None of those facts answers who is entitled to set the party’s priorities, interpret its principles or change its decision-making powers. Those rights need to be explicit in governance and any constitutional text.
This is where the technocratic concern becomes concrete: technical choices can allocate practical influence even without formally transferring the final vote. Its scale depends on the design and on whether members and authorised officers can inspect, challenge and change it.
The serious part
Eight things an arrow
on a slide cannot solve.
An AI system can shape what a party notices, how it interprets evidence and what its decision-makers are asked to approve. That makes the design a political and organisational question.
01Membership is not blanket permission.
Data protection · purpose · access
Membership is not blanket permission.
Data protection · purpose · access
Political-party membership can reveal political opinions. The ICO treats political opinions, including deliberately inferred opinions, as specially protected data. Possessing a membership database does not settle whether a new use, new linkage or disclosure to a supplier is lawful.
A system connecting internal records with public information needs clear purposes, an appropriate lawful basis and special-category condition, restricted access, retention rules and transparent explanations. A data protection impact assessment is required where processing is likely to create high risks. Consent is one possible condition; it is not the only condition available in every situation.
Before connecting the recordsWhich personal information enters the system? Which provider receives it, where is it processed, and can it be used for training? Who approved the purposes and can people exercise their rights?
Sources: ICO: special-category dataICO: impact assessments
The deeper problem: deriving more than you collectedA link between a named member, an issue and a set of posts could become an inferred political profile. That would be new processing to assess. Some monitoring tasks may need no identifiable membership records at all. A necessity test should precede the integration.
Further guidance: ICO: specified purposes and reuse
02The internet is an untrusted input.
Security · scraping · permissions
The internet is an untrusted input.
Security · scraping · permissions
News, websites and social posts are outside the party’s control. A language model reading them can encounter malicious instructions disguised as source material. This is known as indirect prompt injection. Its consequences depend on what the system can read, change or send.
Combining broad internal access with automatic external ingestion would raise the stakes. A safer design needs limited permissions, separation of sensitive information, source checks, logs, recovery procedures and controls over updates. A prompt telling the model to “be careful” is not an access-control system. Permission to read a website also does not itself settle permission to store or reuse its data.
Before switching on the feedCan untrusted material cause the system to retrieve confidential records, alter data or publish? What prevents that, who tests the controls, and how can a harmful change be reversed?
Source: NCSC: prompt injection and security boundaries
The deeper problem: permissions do not copy themselvesIf a restricted document is copied into a search store, its original permissions do not automatically govern the new retrieval path. Access restrictions, changed permissions and required deletions need to reach derived indexes and caches. A convenient answer must not become a side door around the original file controls.
Further guidance: OWASP: retrieval access, retention and provenance
03A connection is not a finding.
Validity · provenance · public opinion
A connection is not a finding.
Validity · provenance · public opinion
A fluent summary can conceal a wrong quotation, a missed qualification or a false match. Linking two policies does not establish a contradiction. An alert that a policy is “superseded” needs a traceable reason, the original source and a reviewer able to challenge the interpretation.
Social-media discussion needs a defined research method before it supports claims about public opinion: who is represented, who is missing, and how are bots, repeated posts, sarcasm and organised activity handled? Polling also has uncertainty that should survive the journey into a generated briefing.
Before trusting the conclusionCan every material claim be traced to evidence? How often are alerts correct, what changes are missed, and how much checking does each useful result require?
Source for AI error risks: NCSC: AI limitations. Sampling and representativeness are research-design questions; no invalid output from this proposed system has been demonstrated here.
The deeper problem: political trade-offs are not data errorsAn outdated figure, an unpassed bill, a change in evidence and a conflict between values need different treatment. A shared subject is not a dependency; a difference in emphasis is not automatically a contradiction. The graph’s definitions and the alert’s evidential status must remain visible.
Further reading:
04Who decides what the machine notices?
Authority · agenda-setting · democratic control
Who decides what the machine notices?
Authority · agenda-setting · democratic control
Influence begins before the final approval. Someone chooses the sources, decides what counts as a conflict, sets the alert thresholds and determines which options reach the policy team. An AI-generated draft can frame a question before a member or officer ever sees it.
The photographed proposal includes a Policy Governance Board and human review. Those roles need defined decision rights, visible evidence and an ability to challenge the system’s framing. If constitutional changes are proposed, the exact text and its effect on member and officer powers need scrutiny in their own right.
Before changing the processWho sets priorities and modifies the rules? Who can challenge an omission or reject a recommendation? Which decisions remain with members and the bodies authorised under the constitution?
Basis: photographed slides 3–5 and risk analysis. Constitutional changes are included in the supplied account of the spoken presentation; no amendment text has been examined.
The deeper problem: influence can move without a formal vote movingThe people choosing sources, categories, priorities and default prompts can shape the agenda while the final vote remains unchanged. Any constitutional amendment should expose appointment powers, delegated decisions, member rights and appeal routes. A technical specification cannot settle who has political authority.
Research and interpretation:
05A human in the loop needs time to think.
Oversight · verification · accountability
A human in the loop needs time to think.
Oversight · verification · accountability
Putting a person at the end of a workflow does not tell us whether that person can meaningfully check it. If the system produces more claims than a small team can verify, approval may become a bottleneck or a routine click.
Effective oversight needs competent reviewers, source access, enough time, recorded reasons for important decisions and clear ownership of mistakes. It also needs an ordinary way to work when the system is unavailable or its output cannot be trusted.
Before calling it human oversightWho checks the original evidence, how much time is allocated, and what happens when the reviewer disagrees? Can work pause without pressure to accept the generated answer?
Basis: governance and workload risk analysis. This is a possible failure mode, not an allegation about an existing approval process.
The deeper problem: checking becomes the workloadA small team asked to verify an expanding stream of fluent material may face pressure to accept its framing. Oversight must be resourced and tested. Drafting speed should be evaluated alongside errors that survive into approved documents and the time required to find them.
Evaluation design: UK AI evaluation guidance, hypothetical briefing example
06Expertise does not confer a mandate.
Outside advisers · funding · conflicts
Expertise does not confer a mandate.
Outside advisers · funding · conflicts
Outside researchers and think tanks can contribute useful knowledge. They may also bring priorities, assumptions, funders and institutional interests. Those influences should be visible so advice can be judged on its evidence and context.
The concern is the boundary between advice and authority: who selects advisers, what access they receive, whether they can set priorities, and who retains the right to decide. External participation alone does not establish improper influence or capture.
Before giving someone a seat at the tableAre affiliations, relevant funding and conflicts disclosed? Is there a published remit? How are member expertise, competing evidence and dissent considered, and who takes the final decision?
Basis: the supplied account of proposals to involve outside experts and think tanks, plus governance analysis. No individual adviser or funder is alleged to be exercising improper influence.
The deeper problem: whose assumptions become the starting point?Influence can enter through problem definitions, selected evidence or draft options as well as through a formal decision. Disclosed funding and affiliations, a bounded advisory remit, plural evidence and recorded dissent help make that influence contestable. This is a governance concern, not evidence of a covert donor arrangement.
07The subscription is not the budget.
Costs · opportunity cost · maintenance
The subscription is not the budget.
Costs · opportunity cost · maintenance
A model bill is one line in a system’s costs. Development, integration, data permissions, hosting, access controls, security reviews, legal advice, verification, training, maintenance and handover all need owners and resources. Volunteer time also has an opportunity cost.
An automated process can reduce one task while creating checking and support work elsewhere. Affordability depends on the actual scope, measured benefit and full cost of running it. These slides are not a costed specification; neither a saving nor an unaffordable bill has been demonstrated here.
Before promising it will carry the workloadWhat are the first-year and recurring costs? What useful work does it replace after checking time is counted? Can the party afford maintenance and move to another supplier?
· No campaign budget or supplier quote is reproduced on this site.
The deeper problem: output volume can hide negative valueThe relevant comparison is with the existing process. Count useful verified results, checking, upkeep and what volunteer time could otherwise achieve. A cheaper model does not make integration, licences, governance or a fragile dependence on one developer disappear.
Comparator, not a party estimate: NAO: realistic skills, funding and delivery foundations
08“Always on” is a service commitment.
Delivery · reliability · dependence
“Always on” is a service commitment.
Delivery · reliability · dependence
Sources change format, feeds break, permissions change and policies evolve. Keeping a knowledge system current takes continuing maintenance. A graph, a search index and a language model also do different jobs; drawing them together does not establish that the whole workflow works.
A bounded pilot could test useful alerts, missed updates, false alarms, time spent checking and total cost before wider dependence. The proposal already describes phased projects. Each stage still needs measurable success criteria, independent evaluation, stop conditions and a fallback.
Before making it indispensableWhat must a pilot prove? Who can stop a failing project? Can the party recover its records, continue normal work and maintain the system when the original developer leaves?
Basis: photographed slides 3–5 and delivery-risk analysis. The feasibility of a specified implementation has not been tested here.
The deeper problem: the system keeps changing after launchA source can be revised, a model updated and a supplier’s service altered. A result that passed a test once does not establish continuing reliability. Record versions, re-test material changes and preserve an exportable evidence trail and handover plan.
Comparator: UK DDaT Playbook: whole-life costs and exit
·
The unglamorous interface
The machine also
generates invoices.
A blank worksheet for thinking about total cost. Add your own assumptions to see direct costs and the value of human time separately.
Illustrative worksheet, not a campaign estimate. Every field starts empty. No figures, supplier prices, campaign budget or predicted savings have been supplied. Entries are not saved or sent anywhere.
Illustrative first-year total
—
Enter assumptions to calculate.
First-year direct costs—
Annual human time value—
Ongoing annual total*—
*Excludes one-off setup. Includes your contingency on recurring direct costs and your human-time valuation. Do not count paid work twice across the direct-cost and time fields.
Count useful work, after checking.
A cost worksheet cannot establish value for money on its own. A pilot would need to measure what work is actually saved, what checking and maintenance are added, and whether the result meets the party’s research and governance standards.
Calculation: annual direct costs = 12 × (monthly AI + monthly infrastructure) + annual licences + annual reviews. First-year direct costs add setup; the entered contingency applies to those direct costs. Annual human time value = 12 × monthly hours × hourly value. Blank fields count as zero once a calculation begins; partially completed totals are incomplete.
Sources, scope and purpose
A little levity.
Some serious questions.
ChatSDP is independent political satire by an anonymous SDP member, writing in a personal capacity.
Why this site exists
Satire, in the spirit of an open campaign.
This site is intended to bring some lightheartedness and fun to the leadership campaign while raising serious questions and concerns about AI, data, costs, oversight and party governance.
It is not intended to bring the Social Democratic Party into disrepute. Its purpose is to encourage discussion of the ideas and their implications through humour, parody and commentary.
The robot characters, invented dialogue and exaggerated scenarios are satire. The underlying proposals and concerns are treated seriously.
Where the material comes from
The campaign’s own material and discussion.
Evidence was collected from Eleanor Rofe’s campaign website, party hustings and her online public surgeries, including published proposals, presentation slides and attendee accounts.
The site draws on proposals for connected data, AI-assisted research, policy review and organisational growth. It also examines the assumptions and presentation of the campaign’s published tools.
How to read the jokes
The robots are imaginary. The questions matter.
EllieBot, the French cloud, robot candidates, NCC.exe and all ChatSDP conversations are comic inventions. They are not quotations from Eleanor Rofe, official party statements or descriptions of systems operating behind this page.
The serious commentary examines proposals and possible implications. It distinguishes what is proposed from what has been built, and what a tool assumes from what its results establish. The photographed proposals include human teams and review roles; scrutiny concerns how those arrangements would work in practice.
Research and interpretation
Evidence informs the criticism.
The discussion of privacy, security, validity and oversight draws on independent research and guidance from bodies including the ICO, NCSC, NIST and the Electoral Commission. Research findings, method notes and this site’s interpretation are presented in the analysis section.
The investor language and model outputs are examined as questions of framing, assumptions and accountability. The parody does not claim that donors own the party, that an existing data breach has occurred or that imaginary robot features are actual campaign proposals.
Independent, unofficial and personal.
ChatSDP is not an official Social Democratic Party or leadership-campaign website. It does not speak for the party, Eleanor Rofe, her campaign or OpenAI, and is not presented as endorsed by them.
Published by an anonymous SDP member, writing in a personal capacity. This is the author’s own project, with no affiliation to any leadership campaign.
Content reviewed 5 November 2026.
About ChatSDP
A mock interface.
A real question of control.
ChatSDP is independent political satire about the proposed use of AI across party research, policy and data systems. Its subject is the concentration of information, workload and influence in an insufficiently specified machine. The fictional extensions deliberately carry the automation premise into absurd robot bureaucracy; they are not added claims about the real campaign.
Independent satire by an anonymous SDP member. Written in a personal capacity to add some lightheartedness to the campaign while asking serious questions. It is not intended to bring the party into disrepute and is not an official party or campaign website.
Tools need boundaries.
AI-assisted research can be useful. The question is what the tool can access and change, what its claims are based on, and who remains able to challenge it.
Advice needs a mandate.
Expertise can inform political judgement. Decision rights, conflicts, funding and accountability still need to be visible.
Satire needs a clear label.
ChatSDP is presented as satire. Robot characters and invented dialogue are comic exaggerations, not quotations or admissions by a real person.
Evidence needs its context.
Campaign material and discussion inform the jokes. Independent research supports the serious analysis, with method notes available alongside the findings.
Privacy
Chat text and worksheet entries stay in this browser page and are not sent to an AI service. This page uses no accounts, analytics, browser storage, message uploads or external fonts, and sets no cookies of its own. Reloading clears the entries. Independent source links open other websites, which handle their own requests.
Cloudflare hosts and protects this site and may process IP addresses, request information and security cookies to deliver the page and prevent abuse. Chat text and worksheet entries are not submitted to Cloudflare. The French-cloud references are part of the parody and do not describe the site’s actual hosting.
Publisher
An anonymous member of the Social Democratic Party, writing in a personal capacity. This is the author’s own project, with no affiliation to any leadership campaign. This site does not speak for or represent the party.
Content reviewed 5 November 2026.