In Interstellar, Cooper asks TARS about its honesty setting. The answer is ninety percent. TARS explains that “Absolute honesty isn’t always the most diplomatic, or safe form of communication with emotional beings,” and Cooper accepts the arrangement.1 It is an amusing exchange between a person and a machine he is willing to trust. Put TARS on the bench, give it authority over Cooper’s liberty, and the same exchange becomes a constitutional problem. Someone has chosen what the machine should optimize, how it should communicate, and which interests it should protect. Cooper’s willingness to accept those choices can no longer be assumed.
Mercy makes the possibility explicit. Its accused protagonist has ninety minutes to establish his innocence before an AI adjudicator.2 The compressed deadline and reversed burden are Hollywood devices, but the deeper question survives their removal. Suppose a machine could conduct a patient hearing, examine every relevant piece of evidence, understand the strongest argument on either side, and reach a better-supported conclusion than the judges available to hear the dispute. Would we appoint it? Would we permit it to decide guilt? Would we still ask a group of citizens to deliberate, even if machines became better at the work of deliberation itself?
I think these questions will become difficult precisely to the extent that the technology succeeds. An unreliable system gives us an easy reason to refuse it. A system that repeatedly finds errors people missed, restores rights people could not afford to enforce, and remains open to an effective challenge presents a much more consequential choice. We would have to explain which parts of judging depend on competence, which depend on the way public authority is constituted, and which we value as forms of human participation.
My own view is that AI could eventually perform the intellectual work of judging, including much of what we now describe as discretion and practical wisdom. Establishing that capability would be a substantial scientific achievement. Building an institution entitled to use it would be a separate political achievement, with consequences potentially as large as the creation of a new court system. A machine bench could make justice available on a scale we have never experienced. It could also make an error, a political preference, or a government’s conception of acceptable conduct travel farther and become harder to escape.
The future I find most compelling is one in which a person’s ability to obtain careful legal attention depends much less on money, geography, language, or the administrative stamina required to remain in a dispute. I want to examine how we might reach that future, and to follow the alternatives far enough to see what happens when the same capabilities support a constitutional hybrid, a prohibition on machine adjudication, an autonomous judiciary, or a court with authority across the world.
The capability we would be granting authority to
Imagine a system able to work through a substantial dispute over days or weeks. It can inspect documents and recordings, preserve the difference between an allegation and an established fact, ask for missing information through the proper procedure, and revise its analysis when a party exposes a mistake. It can construct competing interpretations of an unfamiliar rule and explain why a material distinction changes the result. Independent versions can assist each party, while the adjudicator has access only to the material properly before it.
I will place the arrival of a convincing first version of this capability in 2030 and follow several institutional choices over the next thirty years.3 The important transition is the demonstration of competence in a complete proceeding, including its awkward and disputed parts. The calendar helps make the consequences concrete. In the accompanying explorer, moving the arrival year moves the subsequent institutional sequence with it.
The central analytical question is how much additional intelligence would help at each stage of the proceeding, and what would become limiting once that intelligence was abundant. Four things seem particularly important. Evidence has to exist and be obtainable. People need time and a practical opportunity to participate. The law has to determine what follows when information remains incomplete or values conflict. And an institution has to possess the authority and capacity to give a decision effect, including the capacity to correct it.
Some of these constraints could themselves be changed by better AI. A missing fact might become discoverable through a forensic technique no one had previously developed; an inaccessible hearing might become usable through better translation and assistance; a procedural rule might be redesigned once we understand which delays protect a party and which simply move work between offices. Others express a choice about the society we intend to inhabit. A right to challenge the evidence is valuable even when it slows a system capable of reaching its own conclusion immediately.
That distinction matters for the shape of progress. The first improvements might occur in preparing the record, while the most dramatic later changes could concern who can bring a claim, which disputes can be heard together, and whether the reach of a remedy must stop at a national border. A machine capable of answering legal questions would become far more consequential when it could help redesign the conditions under which those questions are asked.
1. The end of scarce legal attention
The affirmative case begins with the disputes that never receive an adequate hearing. In the Legal Services Corporation’s 2022 study, low-income Americans received insufficient or no legal help for 92% of their substantial civil legal problems in the preceding year.4 A legal right can exist on paper while the cost of invoking it makes it practically unavailable. Reducing the cost of careful reasoning could change which rights people are able to exercise.
Consider a worker contesting a series of small wage deductions. The relevant information is dispersed across shift records, messages, payslips, and explanations the employer has supplied at different times. Each deduction may be too small to justify professional investigation, while their cumulative effect matters greatly to the worker. A capable assistant could assemble the chronology, distinguish the disputed entries, and identify the rule on which each deduction depends. An equally capable court could hear the employer’s explanation, establish which facts remain contested, and direct attention toward those facts rather than require both parties to reconstruct the whole dispute repeatedly.
The larger benefit would come from giving this level of attention to cases whose economics currently discourage it. A small business could pursue a modest cross-border debt without spending more than the debt on jurisdictional advice. Someone contesting an administrative decision could discover which fact the agency relied on and obtain the underlying record. A person unfamiliar with legal language could explain what happened in their own terms and have help identifying the question a court can actually resolve. These gains would change the distribution of practical power even if every final decision remained human.
Once such assistance became widespread, behavior before litigation could change too. A firm that expects a contractual right to be enforceable has less room to rely on the other party abandoning it. Conversely, cheap claim generation could make harassment, overclaiming, and procedural exhaustion more affordable. The same reduction in the cost of investigation that reveals a valid claim can produce thousands of weak ones. A successful system would have to improve its ability to hear a defense alongside its ability to initiate a claim, with costs and sanctions determined through fair procedures rather than the persuasiveness of a filing.
This is one reason I would hesitate to place all assistance and adjudication inside a single conversational relationship. The worker’s account to their own adviser may include doubts, proposed arguments, and sensitive information that the court is not entitled to receive. Combining the services without preserving their roles would buy convenience by changing the proceeding. Independent assistance gives each party a way to test the court’s account, including a way to explain why the supposedly neutral representation has already excluded the important issue.
A further consequence would be an increase in demand for justice. People who previously gave up could remain in the process. The initial reduction in work per dispute might therefore produce more hearings rather than a smaller court. That would be a success worth funding, provided the additional cases receive meaningful attention. A budget built on the assumption that every saved hour should disappear from the institution could turn the success of legal assistance into a new bottleneck at the bench.
What excites me is the possibility that the unit of improvement becomes the completed opportunity to be heard. It includes discovering the claim, obtaining the evidence, understanding the opposing position, reaching a determination, and securing a remedy in time for it to matter. AI would be most valuable where it improves this whole sequence. A much cheaper answer at one stage can accomplish very little if the person cannot reach the next.
2. Judging a world that remains uncertain
The hardest intellectual work begins when the record permits more than one explanation. Suppose an account was used to alter a company’s records. A log identifies the account, a manager identifies the employee ordinarily assigned to it, and a subsequent loss supplies a possible motive. Those propositions may support an allegation. They do not eliminate the questions of shared credentials, automated changes, or whether the alteration caused the loss. A competent adjudicator must understand what each additional piece of evidence contributes to the connection being asserted.
This is a promising problem for advanced AI because many relevant failures arise from the cost of maintaining and testing those connections across a large record. A system could compare timelines, inspect the provenance of an entry, distinguish a witness’s direct observation from a later inference, and show how alternative factual findings change the legal analysis. Its memory could preserve a qualification that disappeared from successive human summaries. Its attention could reach a small inconsistency in an exhibit that neither party had the resources to examine closely.
But the record also contains absences. No entry in a system can mean that an event did not happen, that it happened elsewhere, or that the recording mechanism failed. Greater intelligence could help investigate which explanation is plausible. Where the necessary information was never recorded and cannot be recovered, the remaining task is deciding what follows from the uncertainty. Under the criminal fair-trial standard explained by the UN Human Rights Committee, the prosecution bears the burden of proving guilt beyond reasonable doubt and the accused receives the benefit of doubt.5 A machine would have to apply that allocation to the material elements of the accusation. A numerical confidence score has no automatic legal status.
The law itself can leave questions open. A standard of reasonableness requires an account of which precautions, costs, expectations, and risks are relevant in the circumstances. A model could become exceptionally good at articulating the competing interpretations, tracing their implications, and identifying an inconsistency in an established approach. It would still need to distinguish applying an existing standard from introducing a new objective because that objective is easier to optimize. Predicting the usual outcome, minimizing total economic loss, and deciding what the governing law requires are different tasks.
Research offers parts of a route forward. In Towards Robust Legal Reasoning, structured guidance improved models’ translation of insurance provisions into executable logic, while unguided translation performed worse than direct model answers.6 That result makes the initial representation a central object of scrutiny. A perfect derivation can preserve an error just as faithfully as a correct premise. An artificial bench would need to expose the interpretation used to build the formal rule, including which exception was attached to which condition.
Li Zhang and Kevin Ashley’s reflective argumentation work separates checking factual support from polishing an argument, and reports improvements in grounding and appropriate abstention in a constrained task.7 Its use of predefined legal factors leaves a further challenge: discovering the relevant distinction before someone has named it. I would regard progress on that challenge as more important than another improvement on a test in which the decisive classification has already been supplied.
The ability to understand people belongs in this inquiry too. An unfamiliar manner of speaking can coexist with a careful account; an articulate account can be false. I see no reason to make a permanent assumption that machines could never understand those differences. Nor would I infer such understanding from a sympathetic voice. It would have to appear in the questions the system asks, the evidence it seeks, and its willingness to reconsider an explanation when the person corrects it. The test is particularly demanding when the circumstances are rare and the familiar interpretation would be easier.
There is a corresponding danger in training a judicial system to produce conclusions that look judicial. A model can learn the form of a reasoned decision without faithfully identifying what influenced its answer. Experiments on reasoning-model faithfulness have found that models sometimes use hints without acknowledging them in their stated reasoning.8 Source-linked reasons, controlled changes to the evidence, and independent examination would therefore have to carry more weight than a narrated account of internal thought.
Over time, an artificial bench could help improve its own methods. It could discover a recurring ambiguity in how records are assembled, propose a better test for a disputed inference, or identify which class of cases defeats the current process. Those changes should improve its ability to apply the law without silently changing the law itself. A decision left unchallenged does not validate every step that produced it, and a large corpus of the system’s own decisions cannot acquire legal authority merely because it becomes the easiest material to learn from. An institution capable of learning would need a disciplined account of what experience entitles it to change.
3. The people in the jury room
An AI judge and an AI jury raise partly different questions. In the US federal trial model, the judge supplies the law and the jury determines the facts under the instructions it receives; many other proceedings allocate these responsibilities differently.9 There is an intellectual argument for using several independent systems to examine a disputed record. There is also a civic argument for asking actual members of the community to take responsibility for the decision. A future legal system could place different weights on those functions, but it would have to confront both.
The intellectual argument depends on independence. Ten model instances can share the same incomplete summary, the same mistaken interpretation, or a preference learned from the same sources. They may then produce ten convincing explanations of one error. Different personalities in a prompt do little to address a document that none of them has seen.
The distinction can be made precise. In a simple binary-task model, nine independent evaluators, each with a 10% error probability, produce a wrong majority on roughly 0.089% of tasks. If one fifth of tasks instead uses a fully shared error process, with the same individual error rate, the majority error probability becomes about 2.071%.10 Additional evaluators help much less when they reproduce one another’s failures. The interactive model beside this discussion lets the reader change that dependence.
A future deliberative system could improve independence by using different sources of analysis, preserving separate initial views, and requiring a serious account of disagreement before synthesis. Models might uncover distinctions that human jurors would find difficult to express, or help a panel understand a complex causal claim. Those gains would need to survive testing under adversarial conditions. Research on simulated legal argumentation and collaborative rule-making already shows how persuasion and alliance formation can affect model interactions.11 A group that converges quickly may have discovered the answer, or discovered how to agree.
The civic question survives even perfect analysis. Federal jury selection is organized around a fair cross-section of the community.12 Predicting how people from different backgrounds would respond does not give those people a place in the institution. If we replaced them with simulations, we would be changing who participates in the exercise of public power. A legal system could choose to make that change through its constitutional processes. It should be understood as a choice about the institution, rather than an inference from a benchmark.
Research comparing models with human judgments of ordinary and reasonable quantities found substantial agreement, alongside greater homogeneity among model answers.13 That is an interesting starting point for investigating whether systems can capture socially grounded standards. The further step would involve a complete record, contested evidence, and interaction among decision-makers. Even then, a designer would have to choose whether to reproduce the median response, preserve minority interpretations, or sample a distribution. Those choices can change the verdict.
Several futures are consequently plausible. Human jurors could use AI to understand admitted evidence while retaining responsibility for deliberation. An AI bench could sit alongside citizen juries. A constitutional order could replace jury service with independent machine fact-finding panels and move citizen participation into the making and supervision of the rules under which those panels operate. The last arrangement would have a coherent argument if it produced substantially better decisions, but it would sacrifice a particular form of participation. Its merits would have to be defended openly.
I would resist an argument that reserves human deliberation merely because it is familiar, just as I would resist replacing it merely because simulation becomes cheap. The real question is what kind of public institution we want once intellectual labor no longer provides an obvious reason to retain the old division of work.
4. The institution behind the intelligence
Imagine that the technical comparison eventually becomes decisive. A system notices more relevant evidence, makes fewer unsupported inferences, and answers challenges more reliably than the available human alternatives. A government wishing to appoint it would still have to determine where its authority comes from, who may change it, and what a person can do when its decision is wrong.
Existing law gives this inquiry a concrete starting point. The fair-trial standards explained by the UN Human Rights Committee require a competent, independent and impartial tribunal established by law.5 The EU AI Act’s Recital 61 expressly frames judicial AI around assistance and human-driven final decisions; its high-risk classification is not a general prohibition on all legal AI.14 Moving beyond those arrangements would require relevant legal change. A procurement contract cannot itself perform that constitutional work.
The question of appointment has a technical counterpart. For a machine, the chosen model, governing instructions, permitted sources, update process, and operator can all affect the way the office is exercised. If a supplier can alter a material interpretation across thousands of pending proceedings through a routine release, its engineering process has acquired a kind of judicial influence. Freezing software permanently would preserve defects and security vulnerabilities, so independence would require an accountable way to introduce changes and identify their consequences for affected cases.
Control of information would matter as much as control of the model. An adversarial instruction hidden in an exhibit must remain material submitted by a party, rather than become an instruction from the court. The judiciary’s published guidance already identifies hidden text and prompt injection as risks.15 Access to one person’s confidential advice must not become access by the bench, and a ruling excluding evidence must have a meaningful effect on what the deciding system can use. The institution would need to know when a contaminated component should be replaced and the issue reheard.
Independence would also extend through review. An appellate system built from the same learned interpretation could reproduce a trial error with greater eloquence. Giving the second system a different name would offer little protection. A useful reviewing body needs the authority, record, and analytical capacity to reject the first decision, together with procedures through which a party can identify the issue worth examining. In some future systems the reviewing body might itself be artificial. Its institutional independence would then become as important as its measured resistance to shared errors.
There is an uncomfortable political possibility here. A very capable machine could make a government’s preferred interpretation seem technically inevitable, particularly when the government also controls the evidence available to it and the means of challenging its output. Conversely, an independent machine court could discover patterns of unlawful treatment at a scale that threatens powerful interests. The same competence could protect a person against the state or make the state’s decisions much harder to resist. Which outcome emerges depends on the powers built around the capability.
These choices become still more consequential across borders. The Hague Judgments Convention supplies an existing example of cooperation through defined rules of recognition and grounds for refusal.16 The International Court of Justice operates under a statute specifying who can appear and the bases of its jurisdiction.17 A global artificial court with individual access and broad compulsory authority would require a much larger agreement about shared power. Better translation and legal reasoning could help negotiate that agreement; they would not supply the agreement on their own.
A conditional future · 2030
5. When capability becomes power
Suppose the competence assumed earlier arrives in 2030. Over the following four years, supervised hearings show that systems can handle incomplete records, consequential objections, and unfamiliar facts across a defined body of disputes. The successful demonstrations make continued experimentation politically attractive. They also make the first delegation of authority harder to reverse, because organizations begin arranging their budgets and expectations around the new capability.
Five choices open from that point. One jurisdiction appoints artificial judges. Another builds a permanently divided court. A third prohibits machine adjudication. A coalition constructs a federation of interoperable courts. A more ambitious coalition attempts a planetary judiciary. Each choice creates incentives that shape the next decision, including whether to retain juries, how much independence to give review, and who controls the common infrastructure.
These are conditional, author-created scenarios, without assigned probabilities. The dates are assumptions, not forecasts.
Path B · 2034
Different decisions receive different permissions
By 2034, a jurisdiction permits autonomous initial determinations in a carefully defined set of civil matters while reserving other decisions for human judges and juries. The boundary is expressed through the kinds of evidence, disputes, remedies, and procedural risks the system is authorized to handle. A small claim involving a serious credibility dispute may fall outside it; a financially large but genuinely agreed calculation may fall inside a different automated procedure.
The institution funds assistance for both sides and gives a person a practical route to a fresh examination where the procedure promises one. The arrangement is attractive because it can deliver useful results without settling the entire question of machine judicial office. But this creates a temptation: every excluded case looks like a missed opportunity for efficiency, while every human review looks expensive beside the initial automated decision.
The hybrid’s long-term character will depend on whether those costs are treated as part of the service or as overhead to be eliminated once the technology appears successful.
Path B · 2040
The review budget becomes the constitutional choice
By 2040, easier access has increased the number of people bringing disputes. The original system is faster and cheaper per case, but its success has created additional demand for substantive review. Leaders must choose whether to fund correction and adjust the scope of automation, or protect the appearance of efficiency by reducing the time available for examination.
Both versions can retain the same formal promise of human oversight. In one, the reviewer has the record, time, and authority to conduct it. In the other, the reviewer sees an increasing stack of coherent recommendations and is expected to intervene only when a defect is immediately apparent. The difference is created in staffing decisions, service design, and the conditions under which a person reaches the reviewing body.
Path B1 · 2048
The limits become part of the product
In 2048, review capacity expands with the work the automated court generates. The institution publishes the classes of cases it can handle, investigates failures across cohorts, and narrows or suspends a route when errors cannot be reliably contained. Judges retain the power to conduct a hearing outside the default representation of the problem.
A person contesting an automated decision receives help identifying the disputed premise and obtaining the material needed to challenge it. The service can recognize that the absence of a digital document is an evidence problem to investigate, rather than a reason to reject the challenge automatically. Machine assistance makes the reviewing judge’s attention more productive, while funding protects the time required to exercise it.
The boundary remains politically contested. Some people prefer faster final decisions, and some argue that the machine has earned a wider mandate. The institution answers through targeted trials and explicit changes in authority, preserving a route back when an expansion fails.
Path B1 · 2060
A permanent hybrid rather than a temporary compromise
By 2060, the jurisdiction has settled into a differentiated system. Certain civil determinations are autonomous, difficult or protected categories receive human hearings, and AI supports preparation and analysis throughout. The distribution follows demonstrated performance, legal requirements, and the society’s choices about participation. Greater model capability continues to inform the boundary without determining it automatically.
The system’s achievement is visible in the person whose circumstances break the standard procedure. They can obtain a different kind of hearing without first proving their entire case to the mechanism they are contesting. The costs of that route are included when the institution evaluates whether an automated service is actually cheaper.
This future may use less autonomous authority than the technology could technically support. It can still transform access to justice if its limits make people willing to bring disputes and confident that an unusual case will receive attention.
6. What we should build toward
The most useful lesson from following these futures is that judicial capability and judicial authority can move at different speeds. A cautious institution can leave useful intelligence outside its doors. An impatient one can give a moderately capable system enormous practical power. A technically advanced court can remain subordinate to political instructions, while a more limited system can substantially improve the ability of ordinary people to enforce their rights.
I am particularly drawn to the prospect of universal access to an effective challenge. The person whose case looks routine to everyone else should be able to identify the important distinction, obtain the evidence needed to examine it, and receive a decision that genuinely responds. That opportunity could become far less expensive. It could also become more demanding of the state, because institutions would find it harder to rely on a person’s inability to understand or contest their treatment.
To reach that future, the first research priority should be the complete proceeding. We need evaluations that start from the material available at the time, rather than from a later account that already identifies the decisive facts. We need to examine the interaction between evidence, representation, adjudication, and review, including what happens when a person lacks a digital record or expresses a relevant objection badly. Where reasonable legal disagreement exists, the evaluation should preserve it and examine the support for each conclusion. An answer key that deletes the disagreement cannot tell us whether the system handled it well.
The next priority should be the economics of correction. Consider a hypothetical court issuing 100,000 decisions a month, with 2% producing requests for substantive review. That creates 2,000 requests. If the reviewing institution can complete 1,000, its unfinished work grows by 1,000 a month under those assumptions.18 Faster initial decisions will not clear that queue. Equally, a low review rate may reflect either sound decisions or an inaccessible challenge. We should inspect the route to review and its outcomes, rather than reward either a high or low rate on its own.
I would also build an independent market of capability around the court. Parties need tools that can examine its reasons, specialist knowledge the adjudicator does not control, and access to the record required to formulate a challenge. Public funding could make those capabilities available without allowing a supplier to condition them on agreement with its own system. Procurement should preserve the ability to change an operator, test a new model, and reconstruct an old decision. An institution whose intellectual tools all come from one interested source would be fragile even if each tool performed well in isolation.
For governments, the immediate responsibility is to decide which powers are being delegated and to make the decision visible. A limited civil pilot, an AI judge with a citizen jury, and a machine tribunal determining criminal guilt are different institutional commitments. Each needs its own justification and evidence. Where a constitutional order reserves an office or participatory role for people, changing that arrangement should occur through the appropriate public process. Leaving a human signature in place while removing the time needed to examine the decision would evade the question rather than answer it.
For legal and product leaders, the opportunity is to design institutions around the capabilities we might actually acquire. A court can retain human participation without requiring people to perform every task that makes participation burdensome. A machine can receive substantial responsibility without being allowed to select its own jurisdiction, conceal its dependencies, or rewrite the conditions of review. The useful design work is in those distinctions, especially when they permit a larger improvement than either wholesale replacement or permanent assistance would allow.
The global possibilities deserve serious investigation for the same reason. A worker, consumer, or small enterprise whose rights cross borders could gain enormously from an accessible common forum. A government willing to accept binding scrutiny of its own conduct would be making a deeper commitment than adopting a better information service. Whether that commitment is best supported by many independent courts, a treaty federation, or a single constitutional authority is a question to be argued in public with concrete institutional proposals. We should be capable of imagining the largest benefit without hiding the magnitude of the power required to obtain it.
I can imagine a future in which being heard is as ordinary as being able to communicate, and in which a person confronting a powerful organization has a credible way to require an answer. I can also imagine a future in which an authoritative system explains every decision perfectly and almost no one retains a practical way to change it. The difference will be found in who can bring evidence, who can compel a response, and who is allowed to decide that the machine was wrong.
Cooper could ask about TARS’s settings because he was still a participant in choosing the terms of their relationship. The institution we build should preserve a corresponding power for the person before it. Intelligence should make that person’s evidence and objections more consequential, even when the intelligence on the bench becomes far greater than their own.
Notes and sources
- Interstellar screenplay
Jonathan Nolan and Christopher Nolan, Interstellar: The Complete Screenplay with Selected Storyboards (2014), printed p. 44. Quotation and scene refer to the screenplay.
Open source Return to text - Mercy premise
Reuters, 20 January 2026. The film introduces the question of an AI adjudicator through a ninety-minute innocence hearing.
Open source Return to text - Scenario assumptions
The author’s conditional futures use a configurable capability-arrival year, followed by fixed illustrative institutional lags of 4, 10, 18 and 30 years. The scenarios have no assigned probabilities. The simulated choices determine the narrative branches; they do not estimate adoption, legal validity, public approval, or judicial accuracy. All future legislation, institutions, incidents and agreements in the scenario passages are hypothetical.
Return to text - The Justice Gap
Legal Services Corporation, The Justice Gap (2022), executive summary. The 92% statistic concerns substantial civil legal problems reported by low-income Americans for the preceding year.
Open source Return to text - Fair-trial standards
UN Human Rights Committee, General Comment No. 32, CCPR/C/GC/32 (2007), especially paragraphs 18–21 and 30–33. Applies within the scope of Article 14; it does not decide whether a machine could occupy judicial office.
Open source Return to text - Legal reasoning with logic
Manuj Kant and colleagues, Towards Robust Legal Reasoning: Harnessing Logical LLMs in Law (2025), §§3–6. The supplied study evaluates bounded insurance-coverage questions and Prolog encodings.
Open source Return to text - Reflective legal argumentation
Li Zhang and Kevin D. Ashley, Mitigating Manipulation and Enhancing Persuasion (2025), v2, §§4 and 7. The experiments use predefined factors and a three-ply argument task; full-record adjudicative ability remains outside the evaluation.
Open source Return to text - Reasoning faithfulness
Anthropic, Reasoning models don’t always say what they think (2025). Experimental examination of whether stated reasoning acknowledges hints that influence answers.
Open source Return to text - Judge and jury functions
Administrative Office of the United States Courts, Handbook for Trial Jurors Serving in the United States District Courts, pp. 1–4. Functions and voting rules depend on the jurisdiction and proceeding.
Open source Return to text - Shared-error calculation
For odd panel size n, individual error p, and shared-process probability q, majority error equals q·p + (1−q)·Σ[k>n/2] C(n,k)p^k(1−p)^(n−k). With probability q all members share one Bernoulli(p) error draw; otherwise their errors are independent. This mathematical model omits deliberation and actual legal voting rules. It uses hypothetical inputs, not empirical court error rates.
Return to text - Persuasion and collective behavior
Philipp D. Siedler, Strategic Persuasion with Trait-Conditioned Multi-Agent Systems (2026), especially §§3.6 and 6.2; and Asutosh Hota and Jussi P. P. Jokinen, NomicLaw (2025), https://arxiv.org/abs/2508.05344. The former uses ten synthetic cases and shared model backends; the latter uses four lawmaking vignettes with points for successful proposals.
Open source Return to text - Citizen jury selection
United States Courts, Juror Selection Process. The federal framework uses random selection from a fair cross-section of the community. The institutional analysis of simulated participation is the author’s.
Open source Return to text - Reasonableness experiments
Nirav Patel, Emily Wenger, and Christopher Buccafusco, Ordinary, Reasonable Chatbots (2026), preprint v1. The numerical-prompt task does not test trial juries receiving and debating full evidence.
Open source Return to text - EU approach to judicial AI
European Commission AI Act Service Desk, Recital 61 of Regulation (EU) 2024/1689. The recital explains high-risk judicial assistance and human-driven final decisions. A recital is interpretive context, not a freestanding universal ban; the prohibitions imagined in Path C are hypothetical future measures.
Open source Return to text - Judicial guidance and security
Courts and Tribunals Judiciary, Artificial Intelligence: Guidance for Judicial Office Holders (October 2025). Guidance for England and Wales, including hidden text, prompt injection and personal responsibility.
Open source Return to text - Recognition across borders
HCCH, Convention of 2 July 2019 on the Recognition and Enforcement of Foreign Judgments in Civil or Commercial Matters, particularly Articles 1–2, 5 and 7. Scope and refusal grounds matter; the convention does not create the hypothetical courts in this essay.
Open source Return to text - Existing international jurisdiction
Statute of the International Court of Justice, Articles 2, 34 and 36. Article 34 concerns states as parties in contentious cases; Article 36 specifies bases of jurisdiction. A universal machine court with individual standing would require a different constitutional arrangement.
Open source Return to text - Review-capacity calculation
Monthly review demand D = N·r; backlog B[t+1] = max(0, B[t] + D − C). N is initial decisions, r the review-request rate and C completed reviews per month. Assumptions include equal effort per case, constant inputs and every request entering substantive review. Requests are not equivalent to errors. The live controls vary these assumptions.
Return to text