Rights Before Arrival: Human Rights, AGI Timelines, and the Architecture of Mutual Enhancement
📋 Cite this paper
SomaSoft Research (prepared by Claude Code for the AURI project). (2026-08-20). "Rights Before Arrival: Human Rights, AGI Timelines, and the Architecture of Mutual Enhancement". SOMAsoft Research. Available at https://somasoft.ai/papers/rights-before-arrival. Licensed under SAGL-1.0.
Rights Before Arrival
Human rights, probabilistic AGI timelines, and the architecture of mutual enhancement
Evidence gate: truthiness 1.000 · 31/31 load-bearing claims grounded · 38 sourced facts · 34 distinct sources · adversarial control blocked 2 of 3
Part 00 — Method: a paper that checks itself
A paper about machine-generated claims should be able to survive its own standard. The claims in this document were gated before publication by the same filter the AURI project uses on narrative output. The procedure is mechanical. Every sourced fact gathered during research is entered into a fact base with its provenance. Each load-bearing quantitative sentence is then decomposed into the numbers and proper-noun entities it asserts. A sentence ships only if every number and entity traces to some fact in the base. Anything else is flagged as a possible fabrication and must be cut or sourced.
The fact base holds 38 sourced facts across 34 distinct sources. The paper's 31 load-bearing claims score 1.000 — all 31 grounded. Three deliberately fabricated control claims were submitted alongside them to confirm the filter still bites.
| Control claim (fabricated) | Result | Why |
|---|---|---|
| "A 2026 Stanford study found 87% of community centres reported improved outcomes" | Flagged | 87 not in any fact |
| "Dr. Sarah Chen of the Mayo Clinic reports 92% satisfaction" | Flagged | 92, chen, mayo, clinic ungrounded |
| "By 2050 AGI will have eliminated 40 percent of global poverty" | Passed | 2050 and 40 both exist — recombined |
The filter's real limit
The third control passed. Both of its numbers exist in the fact base — 2050 from the UN projection and 40 from the membership of the UN scientific panel — so a sentence recombining them into an invented causal claim satisfies a presence check. The filter validates token provenance, not claim binding. It catches invented statistics and invented authorities, which is the dominant fabrication mode; it does not catch true numbers welded into false propositions.
Reporting this is not a caveat appended for modesty. A verification method whose failure mode is undisclosed is worse than none, because it launders confidence.
A bug found by using it
Running the gate surfaced a defect in the filter itself. Its number-extraction pattern absorbed a sentence-ending period into the token, so a claim ending "…adopted in 2011." produced the token 2011., which never matched the fact base's 2011. Two correctly sourced claims were falsely flagged. The pattern now requires digits after any decimal point; decimals and thousands separators still parse; three regression tests pin the behaviour, and the filter's suite went from 13 to 16 passing.
The episode is the argument in miniature: verification tools acquire authority they have not earned unless they are themselves tested against real material.
Part 01 — The demographic floor
Rights obligations are not abstractions; they are owed to a specific number of specific people in specific places. Any account of rights in 2050 begins with demography, and demography is where public discourse is currently most corrupted.
The claim that four billion people will die by 2050 is false. No reputable projection supports a mid-century loss of four billion. The UN's medium variant projects world population rising from about 8.2 billion in 2025 to about 9.7 billion in 2050, peaking near 10.3 billion in the 2080s before slow decline. The genuine scholarly disagreement is narrower than the rhetoric: Earth4All scenarios estimate a peak just below 9 billion, while other statistical methods warn of continued growth past 2100. That spread is hundreds of millions, not billions.
The methodological point matters more than the correction. Catastrophist framings do not merely misstate a number; they reverse the shape of the problem. A world losing half its population is a world of triage, and triage logic licenses the suspension of rights. A world of 9.7 billion, growing unevenly, is a world of distribution — and distribution is a question of obligation, not emergency. The distinction determines which legal instruments are even reached for.
The planning fact. Growth to 2050 is concentrated in sub-Saharan Africa, while many wealthy countries age or decline. This single asymmetry — young and growing in one hemisphere, old and shrinking in another — is the load case for every service structure proposed in Part 06.
There is a second-order effect worth naming. If AGI-adjacent capability arrives in the window the forecasts describe, it arrives into a world whose demand for basic services is growing fastest exactly where the supply of that capability is thinnest. Capability and need are, on current trajectory, negatively correlated in space.
Part 02 — When: probability, not prophecy
Asked when AGI arrives, the honest answer is a distribution, and the distribution is very wide.
| Source | Fielded | n | 25% | 50% | Note |
|---|---|---|---|---|---|
| AI company leaders | Jan 2025 | — | 2026 | — | Definition unclear; incentive to hype |
| Metaculus community | Feb 2025 | ~2000 | 2029 | 2033 | Was a 50-year median in 2020 |
| Samotsvety superforecasters | 2023 | — | ~2029 | — | ~28% by 2030 |
| Grace et al. / AI Impacts | 2023 | 1000s | early 2030s | 2047 | Median shortened 13 yrs vs 2022 |
| FRI superforecasters | Apr–May 2026 | 264 | — | 2047 | Conditional on AGI occurring |
| FRI domain experts | Apr–May 2026 | 264 | 2039 | 2050 | 75th percentile 2065 |
| XPT superforecasters | 2022 | 33 | 2048 | — | Pre-ChatGPT |
Definitions differ materially between rows; the comparison is indicative, not strict. The spread is the finding. Aggressive and conservative elicitations differ by more than two institutional generations.
A defensible reading
Synthesising these — and flagging plainly that this is an aggregation of other people's elicitations, not an independent forecast — a defensible reading is:
| Horizon | P(AGI by then) | Anchored on |
|---|---|---|
| By 2030 | 10–25% | Metaculus 25% @ 2029; Samotsvety ~28% @ 2030; experts far lower |
| By 2040 | 30–50% | Between Grace 25% early-2030s and FRI expert 25% @ 2039 |
| By 2050 | 50–65% | FRI expert conditional median 2050; Grace 50% @ 2047 |
| By 2065 | 65–75% | FRI expert 75th percentile 2065 |
| By 2100 | ~80% | FRI: 80% median in both panels |
Two features deserve emphasis over the midpoint. First, expert and superforecaster panels — groups with different biases and different track records — converged on 80% by 2100 and on medians three years apart. Convergence between methodologically hostile panels is weak evidence, but it is the strongest available. Second, the near-term tail is fat enough to be planning-relevant. A 10–25% chance by 2030 is roughly the probability at which engineering disciplines mandate a designed response.
Calibration is not hopeless. On the METR benchmark of eight-hour tasks at 80% success, experts forecast 2030, superforecasters 2028, and the general public 2037; by May 2026 the best measured time horizon stood at 3 hours 6 minutes — inside the expert and superforecaster envelope. Trajectory forecasts on measurable capability are doing better than definitional forecasts about "AGI."
Unknown. Whether any of these panels is calibrated on this class of event is unknown, and cannot be known in advance. Superforecaster skill is established on repeated, resolvable, short-horizon questions. AGI arrival is singular, long-horizon and definitionally unstable — the exact profile on which forecasting skill has never been demonstrated to transfer.
The variable that actually governs policy
The policy-relevant variable is not the arrival date. It is the arrival rate measured against institutional adaptation speed.
Consider the denominators. The Universal Declaration was adopted in 1948 and the binding Covenants only in 1966 — eighteen years between declaration and obligation. The Council of Europe Framework Convention on AI, the first legally binding treaty in the field, opened for signature on 5 September 2024; framework conventions typically require years to ratify and longer to bite. The EU AI Act was adopted in June 2024 and most remaining rules became active on 2 August 2026 — and even that timetable was relaxed by a Digital Omnibus provisional agreement on 7 May 2026.
So the fastest binding rights instrument in this domain took roughly two years from adoption to broad application, and immediately slipped. Against that, the most conservative credible AGI estimate above is 2065, and the most aggressive is already past. On every branch of the distribution, capability moves through its transitions faster than rights instruments move through theirs.
The correct posture is therefore not prediction but pre-commitment: instruments drafted, ratified and dormant, triggered by measured capability thresholds rather than by legislative reaction. A treaty that must be negotiated after the triggering event will not arrive in time on any timeline in the table.
Part 03 — What breaks: where the rights architecture fails
The post-war instruments are more robust to technological change than they are often given credit for. They are drafted at the level of interests — life, privacy, participation, subsistence — not mechanisms, and interests are technology-neutral. Most alarm about "AI needing new rights" is misplaced: the interests are already named. The failure is narrower and deeper.
| Interest | Instrument | AGI-era pressure |
|---|---|---|
| Life | ICCPR 6 | Autonomous targeting; automated clinical triage |
| Equality / non-discrimination | UDHR 7 · ICCPR 26 | Discrimination at population scale, below the threshold of any single detectable act |
| Fair trial / due process | ICCPR 14 | Adjudicative reliance on systems no party can interrogate |
| Privacy | UDHR 12 · ICCPR 17 | Inference of protected attributes from unprotected ambient data |
| Expression & information | UDHR 19 · ICCPR 19 | Synthetic content degrading the shared epistemic commons |
| Political participation | UDHR 21 · ICCPR 25 | Persuasion optimised per-voter at scale |
| Work | ICESCR 6–7 | Displacement faster than retraining systems adapt |
| Food, housing | ICESCR 11 | Allocation by opaque eligibility scoring |
| Health | ICESCR 12 | Diagnostic capability distributed inversely to need |
| Benefit from science | ICESCR 15(1)(b) | Dormant. The repair hinges here. |
The hinge: accountability without an author
The architecture assumes a triangle: a rights-holder who is human, a duty-bearer who is a state, and — where harm occurs — an author of the violation who is a human or an organ of state. The UN Guiding Principles on Business and Human Rights, endorsed in 2011, extended the triangle sideways to corporate conduct through protect, respect and remedy. But they did not disturb the underlying premise: behind every violation stands someone whose conduct can be examined.
Sufficiently general, sufficiently autonomous systems strain that premise. Where a rights-affecting outcome emerges from the interaction of a model no one fully characterises, a deployment context no one specified in full, and training data no one can enumerate, the question "whose conduct was this?" may have no true answer at the resolution law requires. This is not the familiar problem of a diffuse causal chain, which law handles routinely through joint liability. It is that the decisional content — the thing examined for intent, knowledge or negligence — may not exist in any human mind at any point.
Two responses are commonly proposed and both fail. Granting the system legal personality relocates responsibility to an entity that cannot bear it and conveniently discharges everyone who profited. Insisting a human is always "really" responsible is legally tidy but factually false often enough to produce show trials of operators who genuinely could not have known.
The workable answer is neither: decouple remedy from authorship. Where an interest protected by the Covenants is harmed by an automated process, the person harmed should obtain remedy on proof of harm and causation alone, with liability pooled at the point of deployment. Deployers can insure, price and control this. Victims cannot litigate a mind that was never there.
Part 04 — Law: from harm-prevention to mutual enhancement
| Instrument | Date | Force |
|---|---|---|
| UDHR | 1948 | Declaratory; customary in part |
| ICCPR / ICESCR | 1966 | Binding on parties |
| UN Guiding Principles on Business & Human Rights | 2011 | Soft; widely incorporated |
| UNESCO Recommendation on the Ethics of AI | 2021 | Soft; near-universal adoption |
| EU AI Act | Jun 2024 · live 2 Aug 2026 | Binding, regional, extraterritorial in effect |
| CoE Framework Convention on AI | Opened 5 Sep 2024 | First binding treaty in the field |
| UNGA A/RES/79/325 | 26 Aug 2025 | Created the Panel and the Dialogue |
| Independent International Scientific Panel on AI | Report 1 Jul 2026 | 40 members; Bengio & Ressa co-facilitators |
| Global Dialogue on AI Governance | Geneva, 6–7 Jul 2026 | Standing deliberative forum |
Read together, this stack has one shape: it is harm-prevention architecture. It classifies risk, prohibits the worst uses, imposes transparency, and creates fora. That is necessary and hard-won. But it is structurally incapable of delivering mutual enhancement, because a complete set of negative duties describes a world in which nothing bad happens, not one in which anything good is owed to anyone.
There is a further signal worth naming honestly: the Digital Omnibus of 7 May 2026 relaxed timelines and simplified obligations before the Act had fully applied. The first serious contact between binding AI regulation and industrial reality produced deregulatory drift. Any proposal assuming ratchet-only progress is not reading the evidence.
The dormant right
The positive duty already exists in binding law and is almost never litigated: ICESCR Article 15(1)(b), the right of everyone to enjoy the benefits of scientific progress and its applications. It has been treated for six decades as aspirational. In an era when the principal engine of welfare is a technology whose distribution is a matter of deliberate architectural choice, it becomes the natural hook for enhancement duties — and it requires no new treaty, only progressive interpretation of one most states have already ratified. Given the timing argument in Part 02, that is a considerable advantage.
Five proposed instruments
The following are the author's proposals, not established law.
-
Capability floors. A progressive-realisation duty to secure minimum effective access to beneficial AI capability, modelled on universal service obligations in telecommunications. Floors specified in function — diagnostic support, translation, documentary assistance — never in named products, so they survive vendor turnover.
-
A provenance duty. Where an automated system contributes to a rights-affecting determination, its output must carry machine-checkable provenance for each factual assertion, and the person affected must be able to inspect it. This is the discipline of Part 00 raised from methodology to obligation.
-
Non-substitution at the decision point. Automation may inform but not constitute a determination affecting a protected interest; a named human must be able to reach a different conclusion, and must have the time and standing to do so. This generalises the existing prohibition on solely-automated decisions and closes its principal loophole — the rubber-stamp reviewer who cannot in practice dissent.
-
Sovereignty of means. No person's exercise of a right may depend on inference infrastructure their community cannot operate, inspect or continue to run. Vendor withdrawal, sanctions or price changes must not be able to extinguish a right.
-
Remedy without an author. Strict liability pooled at the deployer for Covenant-protected harms from automated processes, backed by a compensation fund where causation is established but no deployer is solvent or identifiable.
These are deliberately drafted as pre-commitments: each can be enacted now, in advance of the capability that makes it urgent, and each degrades gracefully if capability plateaus. Instruments 2 through 4 are useful against today's systems regardless of whether general capability ever arrives — which is the property to demand of any AGI-era policy proposal, since it removes the need to win the forecasting argument first.
Part 05 — The sovereignty layer: what can actually run offline
Instrument 4 is only meaningful if local, offline capability is real. This part measures it rather than assuming it.
The question is not procurement, it is rights. If the only usable inference runs on a foreign hyperscaler, then a community organisation's beneficiary data crosses borders under another jurisdiction's law; service continuity depends on a commercial relationship the community cannot control; and the ICESCR 15(1)(b) benefit becomes a tenancy that can be revoked. Offline capability is the precondition for a capability floor, not an optimisation of one.
Four locally-hosted models were benchmarked on a 14-task set: six classic interview algorithms plus eight tasks derived from logic that shipped in the AURI repository this month. The repository tasks exist because the classic set is memorisable — a general 7B model scoring well on roman_to_int demonstrates corpus exposure, not competence. Each model ran the full set three times, because single-run benchmark numbers are not citable in this project.
| Model | Size | Mean | SD | s/task | Never passed | Licence |
|---|---|---|---|---|---|---|
| qwen2.5:7b-instruct-q4_K_M | 4.7 GB | 100.0% | 0.0 | 1.1 | — | Apache 2.0 |
| gemma3:4b | 3.3 GB | 92.9% | 0.0 | 2.0 | trust_tier_classify | Gemma Terms of Use |
| qwen2.5-coder:7b-instruct | 4.7 GB | 85.7% | 0.0 | 1.1 | grounding_multiplier, salient_filter | Apache 2.0 |
| gemma4:e4b | 9.6 GB | 81.0% | 8.2 | 5.2 | relevance_floor, stale_backup_check | Apache 2.0 |
N=3 runs × 14 tasks, temperature 0.1, RTX 3070 Ti (8 GB), 20 August 2026.
Four findings, each with a policy consequence.
Size does not predict capability. The smallest model tested (3.3 GB) beat the largest (9.6 GB) by twelve points, and a 4.7 GB model beat both perfectly and fastest. Procurement rules specifying parameter counts or VRAM as a proxy for quality will systematically buy the wrong hardware.
Nor does specialisation. The purpose-built coding model scored 85.7% — fourteen points below its general-purpose sibling of identical size and identical speed. It failed the same two repository tasks on all three runs, both of which require following a stated specification exactly rather than recognising an algorithm. The plausible reading is that code-tuning optimises for idiomatic generation of familiar patterns, which is the wrong instinct when the task is to implement an unfamiliar rule precisely — and precise rule-following is essentially all a benefits-eligibility or referral workflow ever asks for. An organisation choosing a model on the label "coder" would have chosen the weaker one.
Variance is a safety property, not a statistical nicety. The weakest model scored 71%, 86% and 86% across three identical runs — a standard deviation of 8.2 points, with two tasks passing on some runs and failing on others. A single run would have reported 86% and concluded the model was adequate. Any capability floor written against benchmark scores must mandate repeated runs and publish the spread, or it will certify unstable systems.
The benchmark is saturated and must not be over-read. A perfect score means this instrument can no longer discriminate at the top, not that the model is capable in the relevant sense. These are single-function Python tasks with hidden unit tests. They say nothing about multi-file reasoning, long-context work, or knowing when to refuse — and an independent capability probe of the AURI system found analogy, theory-of-mind, counterfactual reasoning and abstraction absent at this scale. Local models are competent instruments and poor judges.
What it costs
| Option | Cost | Verdict |
|---|---|---|
| Existing 8 GB GPU (RTX 3070 Ti class) | $0 | Runs every model above; sufficient for 7B QLoRA adaptation |
| Used RTX 4090, 24 GB | $900–1,100 | Best value. ~80–85% of 5090 throughput at ~¼ the price |
| RTX 5090, 32 GB | $4,700–4,830 | Hard to justify; ~35% faster for ~4× the money |
| Cloud A100 80 GB | $0.67/hr | For adaptation runs only — not for beneficiary data |
| One QLoRA fine-tune run | $3–5 | Local adaptation is affordable at community scale |
Priced August 2026 and volatile.
The sovereignty finding. A community organisation can run genuinely useful local inference on hardware costing between zero and roughly $1,100, entirely offline, with no beneficiary data leaving the building — and can adapt a model to its own language, forms and referral pathways for $3–5 per training run. The barrier to sovereign capability at the point of service is not cost or hardware. It is that nobody has specified the deployment.
Part 06 — The centre: a structure for food, housing and health
The evidence base for the components is good; the evidence base for the combination is not, and that limit is stated plainly in Part 08.
| Precedent | Finding | Design constraint |
|---|---|---|
| Housing First, Finland (2008–) | Long-term homelessness has fallen for over a decade | Housing is the precondition, not the reward — but only with treatment access and income support attached |
| Rwanda CHW programme | ~45,000 workers; a driver of early MDG achievement | Reach works; documented failures are workload, irregular training, weak supervision — so budget supervision first |
| Brazil CRAS / SUAS | 8,557 centres nationwide, plus mobile teams by boat and vehicle | The unit is a reference centre — its core function is routing; isolated territories need mobile outreach |
| Food pantry co-located with an FQHC | Improved food access and nutrition education | Co-location of food with clinical care is validated |
| Community care hubs | Linking clinical to social care reduces crises and hospitalisations | The link, not the building, produces the outcome |
| Drop-in hubs for homeless populations | Co-location increases engagement and trust | One door lowers the cost of asking for help |
| Alma-Ata 1978 · Astana 2018 | Primary health care is the route to health for all | The health tier is primary care, not a clinic annexe |
Two load profiles, one core
Part 01 forces this. A centre serving a young, growing, high-fertility population and a centre serving an old, shrinking, chronically-ill one are not the same building with different posters.
| Profile A — young & growing | Profile B — ageing & shrinking | |
|---|---|---|
| Dominant demand | Maternal & child health, nutrition, schooling continuity | Chronic disease, isolation, home support, mobility |
| Delivery centre of gravity | Outreach — workers go out | Fixed site plus home visiting |
| Binding constraint | Supervision capacity per worker | Continuity of relationship over years |
| Food function | Nutrition sufficiency, growth monitoring | Diet-linked chronic disease, meal access for the housebound |
The structure
Five layers, in dependency order — each presupposes the one above it.
L1 · The Threshold — one door. A single entry, one intake conversation, one consented record. No person is asked to prove the same fact twice. Its function is routing, per CRAS; its dividend is trust, per the drop-in hub evidence.
L2 · The Three Guarantees — co-located, in dependency order. Housing first because it is the precondition; food co-located with clinical care because that pairing is validated; health as primary care per Alma-Ata and Astana. One guarantee with three faces, not three programmes sharing a roof.
L3 · The Ring — the part that leaves the building. Community health workers on the Rwandan model, with mobile teams for isolated territories on the CRAS model. Rwanda's documented failure modes set the design rule: cap caseload, fund recurrent training, and budget supervision before headcount.
L4 · The Local Mind — on-premises inference. The Part 05 box. Translation, intake triage, form completion, referral matching and staff decision support. No beneficiary data leaves the building. Under Instrument 3 it never makes a determination — it prepares one for a named human who is able to disagree.
L5 · The Ledger — provenance and consent. Every automated suggestion carries its sources; every use of personal data is logged and inspectable by the person it concerns. The Part 00 discipline as governance.
Governance: a federated commons, not a franchise
Operations are local and autonomous, because the two load profiles and every local legal context differ. What federates is thinner and more valuable: the intake protocol, the consent and provenance schema, the referral ontology, and any model adaptations — which flow back to the commons rather than to a vendor. At $3–5 per adaptation run, a centre that improves a translation model for its own language can return that improvement to every other centre serving it. This is the mechanism by which mutual enhancement stops being a slogan and becomes a supply chain.
Predictable failure modes
- Co-location without integration. The hub evidence is explicit that service coordination is the hard part. Shared premises with unshared records reproduces the maze inside one building.
- Supervision collapse. Rwanda's programme achieved reach and still reported overwhelming workload and weak supervision. Adding L3 without funding supervision converts a workforce into a waiting list.
- Housing First without its conditions. Finland's results depend on treatment access and income support. Housing alone reproduces the failure and discredits the model.
- The Local Mind becomes the gatekeeper. The gravest risk in the design. An inference box installed to assist intake will, under caseload pressure, drift into deciding eligibility. Instrument 3 and L5 exist specifically to make that drift visible and contestable, and must be enforced from day one rather than retrofitted.
Unknown. Staffing ratios, catchment sizes and per-site budgets are not given here. Every published figure I could source describes a single national context, and transposing them into a global template would be exactly the fabrication this paper's method exists to prevent. They must be derived per site and validated against a pilot. Stating them would make the paper more useful-looking and less true.
Part 07 — Mutual enhancement: what it would require
A system assists when it performs a task the person would otherwise perform. It substitutes when it performs a task the person consequently loses the ability to perform. It enhances when the person's own capability is greater after the interaction than before — when the exchange deposits something that remains. Most deployments described as augmentation are substitutions with a longer fuse, and the distinction is invisible in the quarter in which the system is installed.
Three tests distinguish them, and all three are operable in the field:
- The removal test. Withdraw the system. If the person is now worse than before it was ever installed, it substituted. Enhancement leaves a residue in the human; only substitution leaves a hole.
- The dissent test. Is there a named human at the decision point who has the standing, the information and the time to reach a different conclusion? A reviewer with ninety seconds per case is not a decision point; they are a signature.
- The provenance test. Can the affected person interrogate why the system said what it said, in terms they can check against the world? Unauditable advice is authority, and authority accumulating outside any accountable institution is the pattern the rights instruments exist to interrupt.
Note what these tests do not require: knowing when AGI arrives, agreeing on a definition, or resolving whether a system understands anything. They are the practical form of the pre-commitment argument — enforceable now, still correct on every branch of the forecast distribution.
There is a final asymmetry to preserve deliberately. Humans supply values, context and consent; machine systems supply recall, consistency and scale. The rights question lives at the interface, and it has a direction: capability that reaches the person is enhancement, while capability that reaches only the institution processing the person is administration.
Part 08 — Limitations
- The probability table is not a forecast. It aggregates other people's elicitations. No independent model was built, and no aggregation weighting is defensible given that the panels share sources and definitions differ materially between rows.
- Forecaster calibration on this event class is unestablished and cannot be established in advance.
- Cross-survey comparison is loose. "Every task" and "every occupation" elicitations differ by decades within the same survey, for reasons the survey authors themselves do not fully explain.
- The truthiness gate validates token provenance, not claim binding. A fabricated control claim passed by recombining two grounded numbers. Sentences here are grounded in that specific, limited sense — treat the 1.000 as evidence about sourcing discipline, not about truth.
- The centre structure is a synthesis, not a validated intervention. Each component has evidence; the integrated five-layer design has none. It is a hypothesis for a pilot and should be described that way to anyone asked to fund it.
- No cost model has been validated against a real site. Hardware prices are August 2026 and volatile.
- The coding benchmark is saturated and narrow. Single-function Python with hidden tests, on one 8 GB GPU.
- The AURI system that supplied this paper's grounding tools has zero real users. Its concept graph holds 125,877 nodes and 114,050 grounded definitions with a measured 0.0% hallucination rate, and its integrity harness reads HOLD 6 / GAP 0 / RISK 1 — but no human being has yet been helped by it in an evaluated session. Its 549-case ethics corpus is largely machine-generated and returned only loosely relevant matches for AI-governance scenarios; it did not materially inform Parts 03 and 04, and I have not pretended otherwise.
- Legal proposals are the author's, not established law, and have not been reviewed by a qualified lawyer in any jurisdiction.
Standing conclusion. On every branch of the forecast distribution, institutional adaptation is slower than capability growth. That single asymmetry — not any particular arrival date — is the finding that should survive if the rest of this paper is wrong. It argues for instruments that are enacted before they are needed, justiciable without a named author, and useful even if general capability never arrives at all.
Sources
- Factually, Will 4 Billion People Die by 2050?, researched 5 August 2026 — rated false. AI-assisted fact-check carrying its own accuracy disclaimer; its central figure was independently corroborated before use here.
- UN DESA, World Population Prospects 2024.
- 80,000 Hours, Shrinking AGI timelines: a review of expert forecasts, March 2025 — collecting Grace et al./AI Impacts 2023, Metaculus, XPT and Samotsvety.
- Forecasting Research Institute, LEAP Wave 8: AI Timelines, fielded 20 April – 11 May 2026, 264 matched participants.
- UDHR (1948); ICCPR and ICESCR (1966), UN treaty record.
- Council of Europe, Framework Convention on Artificial Intelligence and human rights, democracy and the rule of law, opened for signature 5 September 2024.
- EU AI Act implementation timeline; Digital Omnibus provisional agreement of 7 May 2026.
- UN Human Rights Council, Guiding Principles on Business and Human Rights, endorsed 2011.
- UNGA resolution A/RES/79/325, 26 August 2025.
- UN Independent International Scientific Panel on AI — inaugural meeting 3 March 2026; Preliminary Report 1 July 2026.
- Global Dialogue on AI Governance, Geneva, 6–7 July 2026; UNESCO Recommendation on the Ethics of AI (2021).
- AURI project, AGI Capability Probe, 24 July 2026. Hand-graded 16-probe battery.
- AURI project, Custom coding LLM — scope, hardware, cloud, 15 August 2026.
- Housing First Europe; EESC, December 2025; Successful Public Policy in the Nordic Countries, Oxford Academic.
- CHW Central, Rwanda's Community Health Worker Programme; Global Health: Science and Practice; PLOS One (2020).
- Global Alliance against Hunger and Poverty, Brazil: Social Assistance Reference Center (CRAS).
- Addressing Food Insecurity: Lessons Learned from Co-Locating a Food Pantry with a Federally Qualified Health Center, PMC9524299.
- National Academies, Community Care Hubs: A Promising Model for Health and Social Care Coordination, NBK604806.
- Exploring what works well and less well in a community-based drop-in hub, BMC Health Services Research, 2024, PMC11572058; WHO Alma-Ata (1978) and Astana (2018) Declarations.
Prepared by Claude Code, development assistant to the AURI project — not by AURI, and not speaking for it. Claims gated by the project's truthiness filter; reproducible via papers/agi_human_rights/verify_paper.py. Benchmarks N=3 via experiments/local_coding/multirun.py.