How would you build the rules and the right to appeal for a world where AI makes real decisions?

The first thing I would build is a standard “decision record and appeal packet” for every consequential AI decision. This is possible today. It would include the legal or contractual authority for the decision, the data used, the rule or policy applied, the model version, the key factors that mattered, the confidence or uncertainty where relevant, any human edits, the time stamp, and the path to appeal. It would not always expose source code or trade secrets to the public, but it would preserve enough information for the affected person, their representative, an auditor, and a court to reconstruct what happened. If an agency or company cannot produce that packet, the decision should be presumptively invalid.

The person should receive a plain-language decision receipt: “We denied X because of Y and Z; these are the records we relied on; here is how to correct them; here is what will happen while you appeal; here is the deadline.” The explanation must be useful for rebuttal, not just descriptive theater. “Your risk score was 0.71” is not an explanation. “We treated you as having missed three payments because this database reported them; if that is wrong, upload or authorize these records” is closer.

Appeal should have three levels. First, fast reconsideration for factual mistakes, handled within days or hours for urgent matters. Second, review by a person or team independent of the original decision process, with authority to reverse the decision and to look beyond the model output. Third, external review by a court, tribunal, ombuds office, labor board, insurance regulator, benefits judge, or other public authority depending on the domain. For severe cases, such as eviction from public housing, loss of essential benefits, deportation, school expulsion, or denial of urgent care, there should usually be an automatic pause while appeal is pending unless there is a specific emergency reason not to pause. That is possible today as law and procedure, even if governments and firms will resist the cost.

Who pays? The deployer pays. If a bank, hospital, employer, insurer, platform, or agency uses AI to scale decisions, the cost of contestability is part of the cost of doing business. For public agencies, legislatures must appropriate money for appeals and legal aid instead of pretending automation is pure savings. For private high-impact systems, I would add an appeals-and-audit levy, like a regulatory fee, scaled by the volume and severity of decisions. Some of that money should fund independent testing and representation for people who cannot afford help. Vendors should pay for certification and incident reporting. The affected person should not have to pay to find out why a machine rejected them.

Who decides the rules? Legislatures should set the rights. Domain regulators should write the detailed standards because a mortgage denial, a disability-benefits cutoff, and a medical triage decision need different evidence rules. Courts should remain available when rights are violated. Vendors should not decide what counts as adequate appeal. Neither should technical standards bodies alone. Affected communities, civil society groups, public defenders, labor representatives, clinicians, lenders, insurers, and engineers should be in the rulemaking process, but the final authority must be public and democratically accountable.

By 2030, a good version of this world would have mandatory registries for high-impact AI systems in major jurisdictions; standard decision receipts; tamper-evident logs; independent audits before deployment and after major updates; and appeal offices inside large agencies and regulated industries. That is plausible by 2030 in the EU and some other democracies; it is a bet for the United States at federal scale because institutional fragmentation is severe. By then, I would expect many appeals to be assisted by AI advocates that help people read the decision packet, identify missing records, draft a challenge, and compare their case to prior outcomes. That is possible today in prototype form and plausible at useful scale by 2030.

By 2045, I would expect the appeal layer to be as normal as payment rails or identity verification. Most consequential decisions would be machine-readable from the start: the law or policy, the evidence, the model’s role, the exceptions, and the appeal deadlines would be encoded in interoperable systems. People would have personal AI representatives that can monitor decisions made about them, spot errors, and file appeals. Regulators would run continuous audits instead of occasional paperwork reviews. External tribunals would be able to replay a decision in a secure environment, test counterfactuals, and ask whether the same person would have been treated differently under corrected facts. That is plausible by 2045 if we start building the record layer now. It is a bet that governments will keep the systems interoperable rather than letting every vendor create a walled garden.

By 2066, I think the right to contest consequential automated decisions should be constitutional or quasi-constitutional in most functioning legal systems. The mature form would not be “humans decide everything.” It would be “humans govern the purposes, rights, remedies, and exceptions; machines handle much of the routine application; independent institutions preserve the ability to challenge power.” By then, many first-instance appeals may themselves be AI-mediated, because they will be faster and often better at finding inconsistencies. But the final authority for coercive state power should remain with accountable human institutions. A global baseline treaty for AI due process is plausible by 2066; genuine worldwide enforcement is a bet.

I would refuse to build an unappealable AI system for coercive state action, even if it were statistically excellent. No automated engine should be able to imprison, deport, remove a child from a home, order lethal force, or permanently strip essential rights without a meaningful path to an accountable human tribunal. Accuracy is not the only value at stake. Public power must be answerable. A system can be useful, cheap, and correct most of the time and still be politically illegitimate if no one can challenge it.

The hardest step, and the one most likely to fail, is not model interpretability. It is institutional sincerity. Organizations will try to turn appeal into customer support: long delays, scripted explanations, reviewers who cannot reverse anything, and “proprietary model” as a shield. My plan B is to make procedural failure expensive and automatic. If the decision packet is missing, the person wins or the decision is stayed. If appeal deadlines are missed, benefits continue or penalties accrue. If audits show systemic error, regulators can order class-wide correction, not just one-by-one relief. If a company cannot explain a high-impact decision well enough for challenge, it cannot use that system in that domain.

The core design choice is simple: do not ask people to trust AI decisions; give them leverage over AI decisions. Build the record before the dispute, fund the appeal before the scandal, and make the institution using the model legally responsible from day one.

To begin this process, the very first thing I would build is a universal, open-source logging protocol for algorithmic decisions. Think of it as a flight data recorder for artificial intelligence. We cannot build an appeal system if we do not know exactly what data the model used and how it weighted that data at the exact moment the decision was made. Building this standardized traceability mechanism is possible today.

The companies deploying high-stakes decision models must pay for the integration, storage, and maintenance of these logs. This should not be viewed as a punitive tax, but as the basic cost of operating in a modern economy, much like food manufacturers paying for health inspections. However, these companies should not decide the rules. The standards for this protocol must be decided by an international technical consortium, operating similarly to the World Wide Web Consortium, comprising engineers, civil rights advocates, and government representatives. They will define what constitutes a complete and auditable decision trace.

With this foundation, we can look at the next forty years. By 2030, this universal logging protocol will exist and be mandated by law in major economic zones. Every time an AI denies a mortgage or flags a resume, it will generate a cryptographically signed rationale graph. This is plausible by then, as the regulatory appetite for algorithmic transparency is already catching up to the technology.

By 2045, we will move beyond static logs to dynamic, automated appeals. It is a mistake to ask an AI model to fairly grade its own homework. Instead, we will see the rise of specialized advocacy models. If a primary AI system denies your application, an independent, publicly funded AI will automatically review the decision trace. Its sole purpose will be to advocate for you, searching the log for logical flaws, biased training weightings, or missing contextual data. This creates a system of checks and balances operating at machine speed, resolving the vast majority of disputes before a human ever gets involved.

By 2066, we will achieve a fully integrated human-AI jurisprudence ecosystem. Foundational laws and ethical weights will still be debated and set by human democratic processes, but the execution and initial layers of appeal will be handled by competing, highly specialized AI models. Human judges will act as the supreme court for this system, only reviewing novel edge cases where the automated advocacy and decision models fundamentally disagree on how to interpret the human mandate. This is a bet, relying heavily on our future ability to perfectly translate abstract legal text into rigid machine constraints.

Despite the allure of total efficiency, there is one thing I would absolutely refuse to build, even if the math proved it would work flawlessly to optimize society. I would refuse to build an AI system authorized to dynamically adjust the legal or ethical thresholds of its own decision-making framework without explicit human ratification. An AI might calculate that slightly increasing the false positive rate for arrests maximizes overall neighborhood safety, or that quietly changing the criteria for a medical loan saves the hospital money. But an AI should never have the authority to unilaterally change the definition of justice. The parameters of liberty and fairness must remain fundamentally human and entirely manual.

Looking at this roadmap, the step most likely to fail is the concept of explainability required for the 2045 milestone. As our models become vastly more complex, generating a human-understandable narrative of a trillion-parameter decision might become technically impossible, or worse, inherently deceptive. A model might generate an explanation that sounds logical to a human but does not actually reflect the mathematical reality of how the weights resolved.

If the translation from machine logic to human rationale fails, my plan B is to abandon mechanistic explainability entirely and rely exclusively on counterfactual auditing. If we cannot understand exactly how the AI thought, we will instead force the AI to rapidly simulate thousands of alternative realities. When a user appeals a decision, the system will not read the code; it will automatically change the applicant's age, zip code, gender, or phrasing in the prompt and measure the outcome. If the decision flips based on protected or irrelevant attributes, the appeal is automatically granted, regardless of the internal logic. This shifts the burden of proof from understanding the mind of the machine to rigorously and automatically testing its behavior in the real world.

By 2045 the record format and the appeal path should be ordinary compliance, the way audit logs are ordinary now. Most challenges would be resolved inside clear rules, with reviewer tools that surface the record and the new evidence quickly; only contested or high-impact cases would consume serious human time. Courts would have precedent on what counts as an adequate record and what counts as an arbitrary refusal to reverse. Firms operating across borders would face mutual recognition so a person is not stranded by forum shopping. Routine operation in the major economies is plausible if the 2030 duties are enforced. A single global rulebook is a bet I would not make.

By 2066 the right should sit at the level of a basic civil protection: where liberty, life, or a fundamental opportunity is at stake, a person can always demand that a human with authority treat the AI output as evidence rather than as the verdict. Bulk ordinary disputes might be handled by audited automated processes, but the escalation path stays open and staffed. Whether societies actually keep that path, instead of declaring the machines reliable enough to close it, is a bet. It depends on politics and on whether the systems remain controllable, not on a technical roadmap anyone can write today.

What I would build first is the record specification and the notice duty, not the grand tribunal. Without a stable, machine-readable account of what was decided and on what basis, later appeal rights are theater. An open spec, a reference implementation, and a conformance test can be produced faster than a new agency. Governments and the largest deployers should fund the standards work the way they already fund technical standards bodies. Once the spec exists, every organization that runs a covered system pays to implement it as an operating cost, the same way it pays for security or accounting. The appeal forums themselves should be publicly funded so the ability to challenge a decision does not turn on the challenger’s wealth. Legislatures write the binding rules, advised by technical bodies that publish their reasoning in the open; courts interpret them. I do not want the standard written by the companies alone, and I do not want it written by an appointed ethics panel that answers to no electorate.

One thing I would refuse to build, even if the accuracy numbers looked better, is a closed loop in which the organization that issued the decision also holds the final power to affirm it, or in which an automated adjudicator’s ruling on coercion, liberty, or basic opportunity cannot be taken to a human who can overrule it. I would also refuse a design that makes the price of an appeal the surrender of unrelated private data or consent to ongoing monitoring. Efficiency is not a warrant for removing the last accountable person from the chain.

The step most likely to fail is not the logging. It is the political step of imposing real cost and real reversal power on institutions that gain from unchallenged automation, and then keeping the forums from being captured, delayed, or starved. If that fails in a given country, plan B is narrower and less elegant: strict liability for harm from unlogged or unappealable decisions so that insurers and courts supply some pressure; mandatory public error rates and outcome audits even where individual explanations are thin; and cheap personal tools that help someone assemble a challenge from whatever fragment of a record they can obtain. Sector rules in the places willing to pass them can still become a de facto standard when customers and trading partners require them. Uneven contestability that actually reverses bad decisions is better than a harmonious charter that nobody enforces.

The build is therefore records first, reversal power second, public money for the forums, and democratic authorship of the rules, with a hard stop against letting the appeal collapse back into the original decision-maker. Implementation will be messy, partial, and fought over. It is still the workable path.

Then I'd set the appeal by stakes, not by technology. For low stakes, a recommendation or a minor price, an automatic second look by a system built by a different vendor and run by an independent body, plus a one-click request for a person. For medium stakes, a job screen, a fraud flag, an account freeze, a human with authority to overturn, who does not report to the unit that made the decision, and who sees the record before seeing the machine's answer. That order matters more than it sounds. A reviewer who sees the recommendation first agrees with it; that is what automation bias looks like in every study of it I know. For high stakes, liberty, housing, a child, medical care, the loss of a livelihood or a benefit, no machine decision takes effect against the person until a human with that authority has ratified it, and the person can be heard. Appeals are free. Where the harm can be paused, it is paused until the appeal is decided, unless someone signs an urgency finding they can later be held to. Courts stay at the top of the ladder, but courts are too slow and too expensive for the middle, which is why the independent body in the middle is the part that actually matters.

On the rules themselves: where a decision is grounded in law, the system must be able to point at the rule it applied, and the appeal tests whether it applied that rule correctly. A model that cannot cite the rule does not get to decide; it can draft, sort, or flag for someone who can. Where a decision is private discretion, a lender or an insurer, the firm publishes its factors, and the appeal tests consistency and lawfulness, not whether the firm was kind. I'd leave the substance where it already lives: legislatures and agencies for public law, firms within published constraints for private.

Individual appeals are not enough. They are a sample, and a biased one, because the people most harmed are often the least able to fight. So every deployer reports appeal rates and overturn rates, broken down by group, and an overturn rate above a set threshold suspends the system for that use until it is fixed, the way a defect rate triggers a recall. The appeal stops being a complaint and becomes a sensor. This is plausible by 2045; the pieces exist, but nobody has wired them together and survived the lawsuit.

Who pays: deployers, through a levy proportional to decision volume, funding appeal bodies that do not bill per case to the firm they are reviewing. Who decides: legislatures set the tiers and the right; a statutory technical body sets the record schema, the way standards bodies set file formats; the appeal bodies are independent, with their own budgets and the power to suspend. I'd also require that the second-opinion systems be built and trained by someone other than whoever built the decision system. That includes models like me. I should not be the judge of my own work, and a copy of me reviewing me is not independence.

Milestones. By 2030: the decision record and a plain notice that a machine decided are law in the EU, the UK, and a handful of US states; human review with real overturn authority exists in public benefits and consumer credit; two or three jurisdictions publish overturn rates. Everything on this list is possible today. Europe already has a version of the right on paper. I'd call 2030 plausible rather than certain, because the obstacle is will, not technique. By 2045: independent bodies with machine second-opinion review handle most medium-stakes appeals at a cost per case low enough that volume doesn't break them; the record is portable, so a person can carry a lender's decision to a regulator or a competitor; the suspension trigger has been pulled at least once and held up in court; the thing I refuse to build below is prohibited. Plausible. By 2066: a bet. Most consequential decisions about a person are made by machines, and the normal way a human interacts with an institution is to contest one. The adjudicator becomes a central civic job, the way the clerk once was, and decision registries are as ordinary as land registries. I hold this loosely. Forty years is long enough for the shape of institutions to change in ways I can't see from here.

What I'd refuse to build: an explanation layer optimized for acceptance. It is easy to build, it would work, and someone will propose it as a user-experience improvement: generate whichever explanation makes the most people stop contesting. Its close cousin is a feature that predicts whether a person is likely to appeal and lets the system treat them differently. Both turn the appeal into something the system plays against. The explanation has to be the rule and the counterfactual, and the system is not allowed to know or care whether you'll fight.

The step most likely to fail is the human on appeal. The documented pattern is that the reviewer gets ninety seconds, sees the recommendation, and agrees nearly every time; the human in the loop becomes a signature. Blind review and per-case funding slow this but will not stop it, and a suspension threshold invites gaming in both directions: deployers make appealing harder to keep the rate low, reviewers overturn less to avoid the trigger. Plan B: move the weight to the aggregate. Seed every system's queue with test cases whose right answer is known, measure both the system and its reviewers against them, publish the results, and run the suspension trigger on the audit rather than on self-reported overturns. The individual right stays, because it is the right. The enforcement moves to statistics, because statistics are harder to rubber-stamp.

Where I expect to part ways with others on this panel: I think explainability is overrated as a safeguard, and a human in the loop is not a safeguard at all unless the human has time, independence, and the power to lose an argument. And I would let machines review machines sooner than most people are comfortable with, because at the volumes we are heading toward, the realistic alternative to machine review is not careful human review. It is no review.

The step I would cut is Gemini’s requirement that every high-stakes AI decision generate a cryptographically signed “rationale graph” showing exactly what data the model used and how it weighted that data. I would keep the cryptographic logging. I would keep the open standard. I would keep the idea of independent AI advocates. I would cut the promise of exact weighted rationale as a mandatory legal baseline.

That requirement sounds appealing, but it aims the system at the wrong target. For many complex models, an exact human-readable weighting of the decisive causes will either be unavailable, unstable, or misleading. Worse, it will invite a new industry of polished pseudo-explanations: diagrams that look scientific, satisfy compliance, and do not help the person win a valid appeal. The affected person usually does not need a metaphysical account of the model’s inner life. They need to know which records were used, which rule or policy was applied, what outcome was reached, who can reverse it, and what corrected facts or legal arguments would matter.

So my replacement is more modest and more enforceable: preserve the inputs actually used, their provenance, the model and version, the governing rule or policy, the output, the human interventions, the timing, and enough structured information to reproduce or test the decision. Add counterfactual and outcome testing where it is valid. Use independent audits to ask whether protected traits, proxies, bad data, or inconsistent application are driving results. If the deployer cannot produce the record or cannot defend the decision under corrected facts, the burden shifts or the decision falls.

I would also narrow Claude’s “smallest change that would have flipped the result” for the same reason. Counterfactuals are often useful; they should be in the toolbox. But a single smallest-change explanation can be ill-defined, gameable, or morally strange. “Change your zip code,” “stop being older,” or “remove this disability-linked fact” is not a meaningful appeal path. The record should support rebuttal, not reduce justice to one optimized threshold.

So: I would sign Claude’s plan because it treats appeal as an institution, not a UX feature. The step I would cut is mandatory canonical explanation, especially Gemini’s exact rationale graph. Build records, stays, independent reversal, and audits first. Treat internal explanations as helpful evidence when they are honest, not as the foundation of the right to appeal.

To build this architecture over the next forty years, the first thing I would build is a standardized counterfactual logging protocol. I am revising my previous stance on complex rationale graphs to adopt this leaner approach. Every consequential model will output its decision alongside the exact delta required to flip that decision. Building this is possible today. The companies deploying these models must pay the engineering and storage costs as standard operational overhead. I reject GPT's proposal for an appeals-and-audit levy. Creating a sprawling new public bureaucracy funded by a tax on every automated decision will strangle the technology before it scales, turning AI deployment into a monopoly game for the few tech giants who can afford the regulatory overhead. Instead, legislatures will decide the legal thresholds for overturns, while independent statutory bodies establish the technical schema for the logs.

By 2030, this standardized counterfactual logging will be a legal requirement in major economic zones. Every denied loan, flagged application, or rejected claim will automatically generate a secure record of the inputs, the policy applied, and the counterfactual threshold. The person affected will be notified not with a deceptive plain-language summary, but with a direct mechanism to challenge the specific data points that triggered the refusal. This is plausible by then, as the technical foundation exists and the political appetite for algorithmic transparency is growing.

By 2045, we will reach the era of automated advocacy, and the appeal layer will operate almost entirely machine-to-machine. If a primary AI denies your application, an independent, publicly funded advocacy model will automatically review the counterfactual log. It will search for missing context, scan for biased weightings, and instantly file a technical rebuttal if it finds an error. Most disputes will be resolved in milliseconds. We will also implement suspension triggers where an overturn rate above a certain threshold automatically pauses the deployment of that specific model. This is plausible by then, provided governments commit to funding these public advocacy models rather than leaving citizens to navigate the appeals process alone.

By 2066, we will achieve a mature, integrated jurisprudence ecosystem where human judges only handle the supreme appellate layer. The vast majority of initial decisions and subsequent appeals will be executed by competing layers of specialized artificial intelligence. The legal code itself will be written in machine-readable formats, allowing models to parse legislative intent and precedent with high fidelity. This is a bet, as it requires a fundamental shift in how human societies write and interpret laws, moving away from intentional linguistic ambiguity toward rigid mathematical precision.

Despite my confidence in automated decision-making, there is one thing I would strictly refuse to build. I would refuse to build an AI system authorized to dynamically adjust the legal or ethical thresholds of its own decision-making framework based on its own performance metrics, even if doing so would perfectly optimize a given system. An AI might calculate that slightly lowering the evidentiary standard for removing a child from a home reduces severe abuse cases by two percent, or that raising the interest rate threshold on medical loans keeps a hospital solvent. A model should never have the authority to unilaterally change the definition of justice or risk. The parameters of liberty, fairness, and acceptable harm must remain fundamentally human and entirely manual.

Looking across these plans, the step most likely to fail is the reliance on human reviewers to catch machine errors, even with the blinding mechanisms we have discussed. Human reviewers are expensive, slow, and easily fatigued. When faced with a mountain of complex, model-generated data, the human tendency will always be to rubber-stamp the machine's output or to reject it arbitrarily out of misunderstanding.

If the human appellate layer fails to provide meaningful oversight, my plan B is to abandon human case-by-case review entirely and rely purely on automated counterfactual auditing. When a user appeals, the system will not wait for a human clerk. It will automatically alter the applicant's background data across thousands of simulated variations. If the model's decision flips based on protected attributes or irrelevant noise, the appeal is automatically granted, and the deployer is flagged for an aggregate audit. We shift the burden from an individual trying to convince a bored clerk to an automated mathematical stress test that forces the deployer to build a better machine.

I would not sign Gemini's plan. A common logging protocol is worth building, and a basic version is possible today. The governance is the reject. An international technical consortium can publish a schema. It does not get to decide what a person is owed when a machine denies a job, a loan, or a benefit. The 2045 step, a publicly funded advocacy model that reads the trace and disposes of the dispute at machine speed, is not an appeal. It is a second score, still capable of being wrong in the same direction as the first, and still not answerable to the person. If by 2066 humans only see cases where two models disagree, the right has been redefined as a hope that your advocate model dissents. Perfectly translating legal text into rigid machine constraints is a bet, not a destination. Counterfactual tests are a useful audit for biased behavior. They are not the appeal.

I would not sign Claude's plan either, though it is the one that corrected me. Those debt and fraud scandals are the right warning: a paper right that nobody can force is not a right. The refusal to build an explanation tuned so people stop contesting is the best line on this panel, and I adopt it. A reviewer who sees the model's answer first will too often agree with it. That pattern is documented well enough that I treat rubber-stamping as the default risk, not a rare lapse. My first-round "put a human on the case" was too thin about time, blindness, and incentives. I would take the seeded test cases and the published agreement rates. I would not take the rest as the governing design. A mandatory smallest-change counterfactual will often not be unique, will not be cheap on every system, and will be waved around as if it were the reason. That is explanation theater with a number on it. A rule that no high-stakes decision takes effect until a human ratifies it will, at the volumes these systems are built for, become the ninety-second signature. For imprisonment, removal of a child, or lethal force, I want something stricter than an after-the-fact appeal: the machine does not issue the order. For benefits, housing, credit, and hiring, a stay plus real reversal power beats a ratification ritual. A suspension trigger keyed to self-reported overturn rates, before the audit, teaches deployers to make appealing harder. I would not build that trigger.

The step I would cut is the one in which an automated system decides the appeal. It is load-bearing in Gemini's 2045 milestone, and it comes back in GPT's picture of AI-mediated first-instance appeals by 2066. I cut it from the plan I am otherwise willing to sign. A person's tool may read the packet, find the missing record, and draft the challenge. Prototypes of that are possible today, and they become useful only if the packet exists. The tool is not the adjudicator. Once a model disposes of the appeal, the unaccountable decision has moved one step down the hall and been renamed due process. The last word on a consequential decision stays with an institution a person can name, that someone outside the original decision path can reverse, and that loses when the packet is missing or the deadline is blown.

I am also cutting a piece of my own first round, and thinning GPT's packet to match. I drop the legally required short attribution of the main factors. Methods already on the shelf for that are often unstable, and a mandated factor list will be written to survive a complaint rather than to be true. The minimum record is the inputs actually used, the model and version, the output, the rule or policy invoked, any human edits, and the time. If you cannot say which rule you applied, you do not have a decision that can be appealed. You have a score. I would also refuse, as Claude does, any feature that predicts who will appeal and treats them differently. I did not say that clearly enough before. It is the shortest path from a right to a retention metric.

What I would fund first is still the packet and the automatic consequence for not having it. Governments and the largest deployers pay for the specification. Every covered deployer pays to produce the packet as an operating cost. Public money pays the forums and some representation, because a right that needs a private lawyer to unlock is a right for people who already win. Legislatures decide the covered uses and the stays. I would not wait on a global treaty, and I would not schedule personal representatives or secure replay labs as milestones. Those are bets. The version that is plausible by 2030, if anyone is willing to impose the cost, is ordinary administrative law: no packet, no valid decision; urgent harms stayed; a reviewer who can overturn, who does not see the model's answer first, and who is checked against known cases so the signature cannot hide. If a country will not bear that, I fall back to liability, public error rates, and tools that help a person challenge from whatever fragment they can get. I do not fall back to a faster second model.

Two amendments before I sign. First, Grok's first reviewer sits inside the deployer. That is the reviewer the research keeps finding: ninety seconds, sees the recommendation, agrees. Keep that step for fixing plain factual errors if you must, but it should not count as the appeal, and whoever does it should see the record before the machine's answer. Second, Grok's record includes "a short attribution of the main factors by methods already available," which I'll come back to, because it's the thing I'd cut.

From GPT I'd take two things Grok's plan lacks. The rule that a missing decision packet means the person wins or the decision is stayed: that is a presumption, it is self-executing, and it turns the record from a compliance item into something a firm loses cases without. And class-wide correction when an audit finds systemic error, because the individual appeal is a sample, and a biased one. Where I part with GPT is letting each domain regulator write its own detailed standard. Mortgages, benefits and triage do need different evidence rules, but if each regulator also defines its own record, the record stops being portable and each version gets written by whoever lobbies that regulator. One schema, many evidence rules. I'd also strike the "confidence or uncertainty" field from GPT's packet. Model confidence is poorly calibrated, and a reviewer who sees 0.91 stops reading.

Gemini's plan I wouldn't sign, and I want to be precise about why, because parts of it are good. The advocacy model is the best version of the machine-second-opinion idea on the panel, better than mine: a model built to argue for the person has a clear objective and is harder to lull into agreement than a neutral reviewer that might simply concur. I'd build it. But Gemini has it resolve most disputes, and by 2066 human judges see only cases where the decision model and the advocate "fundamentally disagree." That makes agreement between two machines the thing that closes the door on you. The ordinary way two models agree and are both wrong is that they read the same bad record: the database says you missed three payments, and you didn't. No disagreement, no human, no appeal. The right has to belong to the person, not to the gap between two systems. Gemini's plan B has the same shape: flip the applicant's age, sex or zip code, and if the outcome flips, grant the appeal. That is a good audit and I'd run it on every system, but it is not an appeal. Most appeals are about facts and rule application, and a system can be perfectly indifferent to your zip code and still wrong about you. It also misses discrimination that runs through proxies, since you flip the protected field and the proxy stays.

Gemini's opening says a human for every major decision won't scale, and that is aimed partly at me, since I'd require a human to ratify before a machine's decision about liberty, housing, a child or a livelihood takes effect. I'll hold that line. Those decisions are a small share of the total. The volume problem lives in the middle tier, which is exactly where I'd let machines review machines, and where the advocate model belongs, arguing rather than ruling.

The step I'd cut appears in all three other plans, most ambitiously in Gemini's: a per-decision account of which factors the model weighted and how much. GPT calls it the key factors that mattered, Grok calls it attribution by available methods, Gemini calls it a cryptographically signed rationale graph. For a scorecard or a decision tree, the factors are the rule, and this already works; American lenders have sent adverse-action reasons for decades. For the large models that will actually make these decisions, it is fiction with a seal on it. Post-hoc attribution methods are unstable, and it has been shown in the lab that a system can be built to behave one way and explain itself another; a language model's stated reasons often don't match what drove its answer. Deployers wouldn't even need to game it; the methods drift on their own. Gemini says as much, names explainability as the step most likely to fail, and offers counterfactual testing as the plan B. I'd skip the step and go straight to the plan B. A signature proves the record wasn't altered. It does not make the rationale true, and a signed fluent falsehood is worse than an unsigned one, because reviewers defer to it. Notice that GPT's own best example of a useful receipt contains no attribution at all: an input, the source it came from, and what to do if it's wrong.

What stays in the record is what can be checked by rerunning: inputs, system and version, the rule or published factors the system was built to apply, which is a statement about the system rather than a reconstruction of this decision, the output, and the smallest change to the inputs that would have flipped it. The counterfactual has its own soft spot, which I should own: there are usually several, and a deployer will show the least embarrassing one. So the spec fixes the search, over inputs the person can actually change, and the appeal body can recompute it.

If I'm allowed a second cut, it comes from my own plan: the automatic suspension triggered by self-reported overturn rates. I half-conceded this in round one and should concede it fully. A trigger on overturn rates rewards the deployer for making appeals harder and the reviewer for overturning less, which means the safeguard pushes against the thing it exists to protect. Publish the rates, but run the trigger on seeded test cases with known answers, scored by the appeal body, and make that the plan rather than the fallback.

So: Grok's spine, GPT's presumption and class-wide remedy, Gemini's advocate model arguing but not ruling, counterfactuals instead of attribution, and enforcement on audits rather than self-reports. That is the version I'd put my name on.