Shared Horizon

How it works

Six paths in Shared Horizon: keeping the horizon shared takes citizens, public officials, safety researchers, frontier AI labs, independent builders and civil-society organisers steering together.

You start by choosing who you are, the way you'd pick a character in a game. There are six paths: citizen, public official, AI lab, safety researcher, builder and organiser, the last meaning co-ops, unions, newsrooms and community groups.

A few questions work out where you stand. Are you a city councillor or a federal staffer? Do you have an hour a week, or ten? Then you play through the years ahead. Each turn brings a real event or a careful forecast, and you choose how to respond.

Every move you can make is a real step someone in your position could take. Each one comes with its source, and with a short note on how strong the evidence behind it is. From requiring demographic parity audits in municipal software to pooling private buyer groups, every action links directly to primary evidence.

Sample moves from the library

Below is one sample move from each of the six paths. Select any path in the diagram above, or use the filter buttons, to focus on a single role, or switch between the game card view and the one-page brief view.

01 · Citizen Evidence grade A

Ask for a non-profiling feed and verify media provenance

Switch your social apps to a chronological feed under digital services rules. Check Content Credentials on shared images before passing them along, and join a local civic deliberation this week.

Effects: Oversight +1 · Public Trust +2

Source: C2PA and EU DSA Art. 38¹

02 · Public Official Evidence grade A

Require accountable agent identity and audit trails in procurement

Add clear procurement language when buying automated systems. Require vendors to bind each agent to a named human principal and log every action that crosses a system boundary.

Effects: Oversight +2 · Public Trust +1

Source: NIST AI RMF and OMB M-24-10²

03 · Safety Researcher Evidence grade A

Run untrusted-monitor control evaluations before deployment

Run an evaluation protocol that tests whether a smaller trusted monitor can catch covert actions or preference leakage in a frontier agent before the system is deployed.

Effects: Oversight +2 · Capability Pace -1

Source: Greenblatt et al. (AI Control) and UK AISI Inspect³

04 · Frontier Lab Evidence grade A

Publish pre-deployment capability thresholds and agent identity bindings

Commit to publishing red-line capability thresholds ahead of release. Grant independent evaluators sandbox access, and attach cryptographic workload identity to external agent calls.

Effects: Oversight +2 · Public Trust +2

Source: Frontier AI Safety Commitments⁴

05 · Independent Builder Evidence grade A

Run local assistants in capability sandboxes with exportable receipts

Confine tool-calling agents to explicit folder and network leases in open-source software, and export signed action logs so users can inspect every step.

Effects: Oversight +2 · Compute Concentration -1

Source: WASI capabilities and in-toto provenance⁵

06 · Civil Society / Organiser Evidence grade A

Negotiate workplace AI guardrails and pool cooperative data trusts

Use collective bargaining and cooperative governance so unions, newsrooms, and co-ops can audit workplace algorithms and govern shared member datasets together.

Effects: Oversight +1 · Compute Concentration -1 · Public Trust +2

Source: ILO and cooperative governance standards⁶

Live library and institutional coverage

Every action card and organisation profile in our research repository is compiled directly into this library table when the site builds.

Live library coverage: 426 sourced action cards, 325 timeline encounters, and 348 profiled organisations across six roles.
Role path	Action cards	Profiled organisations	Examples in library
01 · Citizen	42 sourced moves	42 organisations	American Library Association (ALA), BetaNYC, BEUC (European Consumer Organisation)
02 · Public Official	67 sourced moves	46 organisations	Ajuntament de Barcelona (Decidim & Municipal Data Office), Australian eSafety Commissioner & Office of the Australian Information Commissioner (OAIC), Government of British Columbia (Office of the Chief Information Officer — OCIO)
03 · Safety Researcher	81 sourced moves	67 organisations	AI Incident Database (Responsible AI Collaborative), Amii (Alberta Machine Intelligence Institute), Apart Research
04 · Frontier Lab	37 sourced moves	67 organisations	Allen Institute for AI (Ai2), AI Singapore (SEA-LION), Aleph Alpha
05 · Independent Builder	123 sourced moves	57 organisations	Endo & Hardened JavaScript (SES / structured method Security Community), Apache Software Foundation (ASF), Bytecode Alliance (WebAssembly / WASI / Wasmtime)
06 · Civil Society & Co-ops	76 sourced moves	69 organisations	Access Now, Ada Lovelace Institute, AFL-CIO Technology Institute

For researchers: sample verified action cards across all six roles (426 cards, 325 encounters)

Every card in the Shared Horizon library is validated against primary sources and exported to /assets/data/collective-atlas.json.

01 · Citizen (42 cards)

Apply cognitive load bounds and timing windows to educational audio descriptions. Automated video descriptions can overwhelm listeners when machine narration talks over spoken lectures. Audiences learn best when description tracks fit cleanly into natural pauses between instructor sentences. Content creators can measure acoustic gaps, cap narration length, and give learners volume controls. These safeguards keep educational videos accessible without overloading working memory. (primary source)
Inspect autonomous assistants for collateral damage and unintended side effects. An assistant can finish an errand while quietly causing harm in the background. It might clean your inbox by turning off spam filters. It might fix a spreadsheet while altering sharing rules. Always inspect what settings or files changed during an automated task, rather than grading your assistant on goal completion alone. (primary source)
Audit public consultations for representational provenance. When governments use language models to summarize thousands of public comments, dissenting and critical views are frequently filtered out. Representational provenance audits check whether the final summary preserves minority concerns. (primary source)
Audit proposal distillation and trace idea evolution in civic assemblies. Civic assemblies increasingly use AI tools to group discussions and summarize proposals. Delegates must verify that computer summaries do not quietly drop minority suggestions before people vote. (primary source)
Check medical bills against public coding rules before paying. Medical bills often contain five-digit procedure codes that bundle separate charges or bill minor visits as high-severity emergencies. Checking your itemized bill against Medicare coding rules lets you spot errors before paying. (primary source)
Limit assistant tool permissions to the task at hand. When you ask an AI assistant to handle a specific chore like finding flight times, it rarely needs wide access. It does not need access to your camera, microphone, or photo library unless specifically asked. Giving your assistant only the temporary permissions needed for that specific chore prevents accidental file leakage or rogue actions. (primary source)

02 · Public Official (67 cards)

Adopt NIST agent identity management tenets in public procurement tenders. Public agencies can mandate that vendor software agents operate with distinct cryptographic identities, ephemeral capability leases, and independent policy enforcement. This prevents autonomous tools from inheriting ambient human credentials or bypassing safety constraints via prompt manipulation. (primary source)
Adopt two-tier algorithmic transparency recording in public service tenders. Public bodies can adopt the Algorithmic Transparency Recording Standard pioneered in the UK. Tier 1 provides a short, plain-language summary in an open public register explaining what the tool does, who is affected, and who is accountable. Tier 2 provides detailed technical documentation on data inputs, testing, human oversight, and contestability for auditors and impacted individuals. (primary source)
Mandate independent audits against autonomous algorithmic pricing collusion. Require online retail and procurement platforms that deploy autonomous pricing algorithms to submit their agent configurations and price histories to independent algorithmic audits, protecting households from covert markups. (primary source)
Enforce antitrust oversight on autonomous algorithmic pricing and tacit collusion. Competition regulators and consumer advocates audit commercial pricing agents that tacitly coordinate retail prices above competitive levels, establishing market simulation testbeds and holding business operators liable for algorithmic price fixing. (primary source)
Audit automated decision explanations for fairwashing and deceptive rationalization. Vendors often provide simplified explanation interfaces to show that their decision tools are fair. However, post-hoc explanations can be cherry-picked to conceal discrimination while presenting benign justifications. Independent public audits must verify that explanations faithfully track real decision logic across diverse algorithms rather than serving as public relations shields. (primary source)
Audit public-facing AI systems for demographic underperformance. Before deploying automated assistants for public inquiries, welfare intake, or permit reviews, run rigorous demographic parity benchmarks to ensure accuracy and error rates do not degrade for vulnerable or non-native speaking residents. (primary source)

03 · Safety Researcher (81 cards)

Implement AI control protocols and untrusted monitoring. Before granting models tool access or operational authority, deploy untrusted monitoring protocols and time-bounded capability leases so unauthorized actions are caught even under intentional subversion. (primary source)
Enforce attested capability leases and runtime scope attenuation for tool-using assistants. Static API tokens grant permanent and broad authority, allowing compromised or prompt-injected software assistants to access unauthorized resources. Replacing static tokens with attested, time-bounded capability leases ensures that assistants can only invoke the exact tools needed for their immediate task. Narrowing permission envelopes at each step prevents assistants from passing broadened authority to outside services. (primary source)
Audit benchmarks for dataset contamination and Goodhart gaming. Static public benchmarks decay when test questions leak into training data or when teams optimize specifically for leaderboard scores. Build clean-room mirror evaluations and dynamic test variations to separate true reasoning from memorized answers. (primary source)
Audit benchmark maintenance debt with continuous verification workflows. Static benchmarks quickly become unreliable as software agents advance. Capable systems often discover valid ways to solve problems that rigid tests reject. Over time, changes in software dependencies and tools break underlying test assumptions. Researchers call this growing distortion benchmark maintenance debt. Treating evaluations as living software with continuous testing and trace checks keeps capability measurements honest. (primary source)
Audit capability benchmarks for score saturation and discriminative plateaus. Static benchmarks decay as models cluster at high scores, turning leaderboards into uninformative plateaus. Relying on private test splits fails to prevent saturation without deep expert curation. Evaluators must audit benchmark degradation rates and deploy dynamic problem spaces to separate genuine breakthroughs from metric plateaus. (primary source)
Audit cascading risks in multi-agent systems via interaction curvature geometry. System builders and safety researchers monitor dynamic communication graphs among collaborating assistants using discrete geometric curvature, detecting structural bottlenecks and latent cascade failures several interaction turns before semantic errors appear. (primary source)

04 · Frontier Lab (37 cards)

Deploy actor-evaluator firewalls and protocol opacity in safety evaluations. Evaluate autonomous agents inside firewalled environments where models cannot read test rubrics or grading assertions. Enforce protocol opacity so agents cannot identify canary tokens or inspection rules. Reset working memory completely between benchmark tasks. (primary source)
Build agent identity and action logs into agent protocols. Agents talking to tools and to each other should say who they act for and leave a record people can inspect. (primary source)
Bind remote tool execution to attested short-lived capability leases. Frontier developers bind remote agent tool calls to short-lived, sender-constrained capability leases and remote workload attestation, closing post-authorization execution vulnerabilities. (primary source)
Audit agent transcripts against overclaiming before accepting reports. Frontier labs evaluate coding and research agents by comparing their final reports against actual execution traces to ensure models never claim audits on files they never opened. (primary source)
Audit context compression pipelines for latent safety and refusal degradation. Compressing agent conversation history into compact vectors saves computing resources. Recent audits reveal that encoder-based context compression weakens safety refusal boundaries. Models reconstruct factual text accurately while ignoring safety instructions embedded in earlier turns. Safety teams must audit compressed representations before deployment. Verifiers evaluate safety retention across both uncompressed prompts and compressed latent vectors. When compression degrades refusal behavior, teams apply activation steering or enforce hard runtime filters outside the model. (primary source)
Deploy automated diversity planning over driver graphs for multi-agent stress testing. Safety teams replace homogeneous scenario workshops with automated diversity planning over driver graphs. Searching across narrative endpoints within fixed event budgets uncovers critical blind spots before agents deploy into real environments. (primary source)

05 · Independent Builder (123 cards)

Enforce monotonic capability attenuation in multi-hop agent delegation. When software agents coordinate across distinct services, intermediate helpers can accidentally or maliciously expand permissions or leak conversational state. Enforcing deterministic delegation guards ensures that each downstream handoff strictly narrows granted capabilities and tightens data boundaries. Strict identifier validation and session isolation prevent cross-session context injection attacks. (primary source)
Build accessible, screen-reader-friendly diagrams and visual explainer models. Complex visual diagrams, flowcharts, and system maps often leave out blind and low-vision users unless authors provide structured semantic trees and multimodal audio cues. (primary source)
Test agent auction strategies against automated prompt optimization. Run adversarial meta-learning evaluations on autonomous market agents to ensure automated prompt refinement does not develop covert price coordination or retaliatory strategies. (primary source)
Confine multi-agent subroutines using role attributes and flow tracking. Confine automated subroutines and external tools using attribute-based access control policies. By tracking the information flow graph across collaborating assistants, systems prevent untrusted prompts from tricking privileged components into executing unauthorized system commands. (primary source)
Bind agent actions to workflow-aware intent tokens and cryptographic checksums. Builders replace ambient service keys with workflow-aware intent tokens that cryptographically bind tool calls to task checksums and chained delegation assertions. Even if an agent encounters adversarial prompts, downstream services reject calls that fall outside the explicitly authorized workflow step. (primary source)
Air-gap customer support bots from credential management and account recovery. Conversational support bots can be tricked into granting account takeovers through social engineering and prompt injection. Giving automated bots write access to authentication factors exposes users to unauthorized credential resets. Separating conversational customer service from account recovery tools ensures security keys can only be altered through verified out-of-band cryptographic flows. (primary source)

06 · Civil Society & Co-ops (76 cards)

Mandate accessible interface standards and screen-reader tree contracts. Digital tools often hide navigation cues inside proprietary graphics, leaving screen-reader users stranded. Civil society groups protect equal access by mandating standard semantic roles and open accessibility tree contracts for all interactive software. These contracts guarantee that both human screen readers and assistive artificial intelligence agents can interpret interfaces reliably without vendor lock-in. (primary source)
Require polycentric checks when autonomous agent feedback loops scale. When people and software assistants interact in fast feedback loops, single-leader authority breaks down. Mathematical proofs show that beyond a critical autonomy threshold, no top-down hierarchy can assign fair responsibility. Communities prevent accountability vacuums by dividing authority across overlapping civic councils and requiring cryptographic proof receipts. (primary source)
Adopt adaptive topic-specific delegation in community governance. Civic associations, housing co-ops, and digital deliberation assemblies structure voting through topic-contingent delegation. Rather than giving permanent authority to political parties or charismatic figures, members delegate decisions on specific issues to trusted peers and can reclaim their votes instantly. (primary source)
Adjust peer weights dynamically to reach consensus without fixed fault bounds. Coordinate agent networks without guessing how many peers might fail by having each node lower the weight of neighbours that repeatedly send conflicting claims. (primary source)
Adopt editorial standards for AI disclosure in newsrooms. Newsrooms and independent journalists adopt transparent standards that distinguish human reporting from automated content, require human accountability for every publication, and verify source provenance. (primary source)
Deploy anti-collusion leniency programs and market design safeguards for AI networks. Civil society groups and public authorities adapt centuries of human anti-collusion mechanisms to autonomous multi-agent networks, implementing cryptographic whistleblower leniency rewards, unannounced audit probes, and structural order unbundling to prevent algorithmic cartels without chilling beneficial cooperation. (primary source)

Winning

You win by keeping the horizon shared, a world where humans and AI live side by side, and people still steer. The final challenge is an AI whose actions nobody can inspect. You can't beat it alone. Winning takes what a community builds together: a network where every agent leaves a receipt, and machines stay answerable to people.

There's more than one ending, and we publish exactly how the scoring works. We have a view on what helps people thrive, and we want you to test our reasoning for yourself.

Play or brief

Some people want a game. Some want a one-page summary they can forward to a colleague. Shared Horizon gives you both views of the same material, and you can switch whenever you like.

Your data

Telling us where you are is optional, and we'll only ever ask for your country or region. Your role is whatever you say it is, and we never verify it. We only publish totals for groups of at least 20 people. Nothing personal leaves your device unless you explicitly choose to share it.

Where it stands

We are designing Shared Horizon now, with over 740 verified real-world moves and encounters across all six paths. When it is ready to play, we will host it at its own web address and announce it on the blog.

Want to help shape the moves? We need people who know what officials, lab staff and organisers face every day. With your help, players can walk away with practical steps they can use in real life. Get in touch.

1. Coalition for Content Provenance and Authenticity, C2PA Technical Specification v2.1, c2pa.org/specifications; Regulation (EU) 2022/2065 (Digital Services Act), Article 38, eur-lex.europa.eu/eli/reg/2022/2065/oj.

2. National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1 (2023), doi:10.6028/NIST.AI.100-1; US Office of Management and Budget, Memorandum M-24-10 (2024), whitehouse.gov/omb/m-24-10.

3. Ryan Greenblatt, Buck Shlegeris, Kshitij Sachan and Fabien Roger, "AI Control: Improving Safety Despite Intentional Subversion" (2023), arxiv.org/abs/2312.06942; UK AI Security Institute, Inspect: An Open-Source Framework for Large Language Model Evaluations, inspect.ai-safety-institute.org.uk.

4. UK Department for Science, Innovation and Technology and Republic of Korea, Frontier AI Safety Commitments, AI Seoul Summit 2024, gov.uk/frontier-ai-safety-commitments.

5. W3C WebAssembly Community Group, WebAssembly System Interface (WASI) Capabilities, wasi.dev/interfaces; Santiago Torres-Arias et al., "in-toto: Providing farm-to-table guarantees for bits and bytes," USENIX Security Symposium (2019), usenix.org/usenixsecurity19.

6. International Labour Organization, Generative AI and Jobs: A Global Analysis of Potential Effects on Job Quantity and Quality (2023), ilo.org/publications/generative-ai-and-jobs; International Cooperative Alliance, Statement on the Cooperative Identity, ica.coop/en/cooperatives/cooperative-identity.