---
title: "Designing social systems with AI instead of choosing left or right: AI-written tax and redistribution rules, Community Notes, and who sets the objective"
kind: "Article"
topic: "Economics"
date: "2026-10-18"
url: "https://canopticon.com/articles/designing-social-systems-with-ai-instead-of-choosing-left-or"
author: "Chris Telford"
license: "CC BY 4.0"
---
Can AI design a social system that beats capitalism, socialism and every party platform, and so end the fight between left and right? AI can now write working rules that people vote for. It cannot decide what those rules should aim at, and that decision is the old fight in a new place.

Since about 2020 a handful of experiments have shown the first half. A reinforcement-learning planner wrote tax schedules for a simulated economy. Another system designed a payout rule that real players chose over the rules each ideology would have offered. A language model wrote compromise statements on divisive UK questions that people preferred to the drafts of human mediators. In every case the target came from outside the machine: <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="3. What AI system-design experiments have actually demonstrated">a welfare formula, a majority vote, endorsement, cross-cluster agreement</q>.

The argument here is that the target is where left and right now live. In the lab, the public increasingly sets it, by voting or by endorsing. Outside the lab, people nobody elected are already setting it: pricing software aims at the seller's margin, social feeds reward attacks on the other team, and pollsters are building synthetic citizens to stand in for real ones.

## Left and right were already arguing about how an economy handles information

The argument between market and plan was a design argument long before a computer could run either. In 1920 Ludwig von Mises argued that without market exchange in the means of production there are no money prices to calculate with. Friedrich Hayek turned that in 1945 into a claim about information: <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="Key Findings">the economic problem is using knowledge that exists only in scattered, local form, and prices are a way of communicating it</q>.

Whether computers answer Hayek is still disputed. One peer-reviewed side holds that <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="1. The precedent of the calculation debate and Cybersyn">advances in computing power are irrelevant to Mises' most fundamental argument</q>, because the problem concerns institutions (property and exchange), not processing speed. A 2024 paper from the planning side replies that this in-principle case depends, more than its authors admit, on what computing can and cannot do; only its abstract was read. The fair summary is that <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="1. The precedent of the calculation debate and Cybersyn">whether AI dissolves the calculation problem is actively disputed, and that the dispute has shifted from logic toward questions that could in principle be tested</q>.

Separate research on pricing shows where the line falls. Walmart and Amazon plan vast operations internally, and some socialists cite them as proof that planning works. What the two companies plan is logistics (inventory, routing, scheduling, forecasting), <q doc="price-without-discovery-from-the-signal-to-the-instruction" section="Strand Two: The Hayek Inversion — pressure-tested">exactly the class of problem Mises conceded planning could handle</q>. Their capital decisions still lean on the market around them, since <q doc="price-without-discovery-from-the-signal-to-the-instruction" section="Strand Two: The Hayek Inversion — pressure-tested">their internal "transfer prices" are calibrated to those external prices</q>. The best-known computational proposal, planning in labour-time values, <q doc="price-without-discovery-from-the-signal-to-the-instruction" section="Strand Two: The Hayek Inversion — pressure-tested">answers the computation problem but not the valuation/knowledge problem</q>.

The one serious attempt to run part of a national economy as a designed system shows the other limit. Project Cybersyn, built in Allende's Chile between 1970 and 1973 with the British cybernetician Stafford Beer, was meant to <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="1. The precedent of the calculation debate and Cybersyn">manage the transition of Chile's economy from capitalism to socialism</q>. Its historian found that its technical structure reflected the government's socialist ideology.

<figure data-kind="pull" doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="Key Findings">
“Design did not escape ideology; it encoded it”
</figure>

## What the AI systems designed, and who told each one what to aim for

The AI Economist, from Salesforce Research, trained a planner to set taxes while simulated workers learned to gather, trade and build around them. Its designers fixed the planner's target as <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="3. What AI system-design experiments have actually demonstrated">the product of equality and productivity</q>, and the preprint reports that the learned policy <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="3. What AI system-design experiments have actually demonstrated">improved this trade-off by 16% over a tax framework derived from Emmanuel Saez's optimal-tax theory</q>. The economy is a small gridworld, the workers are reward-maximizers rather than people, and <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="3. What AI system-design experiments have actually demonstrated">the objective is one value choice among many</q>. Someone who weighted growth or liberty differently would have scored the same tax schedules differently.

DeepMind's Democratic AI, published in 2022, took its target from the players. In an online investment game, players voted between redistribution rules, and the AI's rule beat human-designed baselines <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="3. What AI system-design experiments have actually demonstrated">including strict egalitarian, libertarian and liberal-egalitarian schemes</q>. The winning rule <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="3. What AI system-design experiments have actually demonstrated">redistributed in proportion to players' contributions relative to their means, which corrected the starting wealth imbalance while discouraging free-riding</q>. It is the closest anything has come to design replacing the contest between ideologies: a rule no camp owned, chosen over the rules each camp would have proposed. Majority vote is still a target someone chose, and the authors <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="3. What AI system-design experiments have actually demonstrated">acknowledge the risk of a design that a majority likes while a minority bears its costs</q>.

The Habermas Machine, from Google DeepMind and published in Science in 2024, aimed at agreement. A language model read individual opinions on divisive UK policy questions, drafted a group statement, took critiques and revised it. Across more than 5,000 participants, <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="3. What AI system-design experiments have actually demonstrated">the AI-written statements were preferred to those written by human mediators 56% of the time, and participants' positions converged toward a common view after the process</q>. <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="3. What AI system-design experiments have actually demonstrated">The process produced statements, not adopted laws.</q>

Each of these systems has three parts: a picture of the people (simulated agents, game players, written opinions), an optimizer that searches for a rule, and the target the rule is scored against. AI has made the search cheap. The picture of the people and the choice of target are the parts that still carry the politics.

## Arrow's theorem: no target for adding up preferences is neutral

Kenneth Arrow proved that once there are more than two options, no procedure for turning individual rankings into a social ranking meets a short list of reasonable conditions together: it handles any preferences, produces no cycles, follows unanimous preferences, has no dictator, and ignores irrelevant alternatives. The options can be <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="2. What formal theory says about designing collective choice">policies, public projects, candidates in an election, distributions of income</q>, so the result covers choosing between whole economic arrangements as much as between candidates. Arrow himself <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="2. What formal theory says about designing collective choice">regarded his conditions as value judgments, not neutral axioms</q>. <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="2. What formal theory says about designing collective choice">Any aggregation rule an AI uses therefore embodies a political choice about which conditions to give up.</q>

The theorem has been described as being <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="2. What formal theory says about designing collective choice">about the impossibility of trying to do too much with too little information</q>, and the ways around it bear that out. People can grade options instead of ranking them, utilities can be made comparable across people, or the range of preferences can be narrowed. The third route is where deliberation comes in. If discussion pushes preferences into a shape that can be added up, deliberation tools <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="2. What formal theory says about designing collective choice">may work by changing the structure of disagreement, not by finding a hidden consensus</q>, and <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="2. What formal theory says about designing collective choice">List et al. (2013) present empirical evidence that deliberation sometimes does have this effect</q>.

Seen this way, the ideologies are answers to design questions. <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="2. What formal theory says about designing collective choice">Conservatism, liberalism and socialism can be read, roughly, as bundles of answers about which conditions to relax, whose utilities count and how much, and how far to trust local knowledge over central models.</q> Their argument does not end; it <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="2. What formal theory says about designing collective choice">becomes a set of explicit parameters</q>.

A formal field already works this way. Mechanism design, recognised by the 2007 economics Nobel, <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="3. The formal toolkits behind these experiments">treats markets as one mechanism among many, a framing that can show when markets perform well and when other arrangements do better</q>. Elinor Ostrom's research on how communities govern shared resources <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="3. The formal toolkits behind these experiments">explains outcomes that do not fit a two-way choice between the market and the state</q>, though only the journal record of her Nobel lecture was read.

## Outside the lab, sellers and platforms are choosing the target

The experiments let a vote or an agreement score set the target, and they printed it in the paper. The AI-designed systems already running at scale mostly do neither.

Take prices. Research on algorithmic pricing describes a third mode that is neither Hayek's market nor a socialist plan: prices <q doc="price-without-discovery-from-the-signal-to-the-instruction" section="Strand Two: The Hayek Inversion — pressure-tested">set by many private, competing, opaque corporate algorithms</q>. In that mode <q doc="price-without-discovery-from-the-signal-to-the-instruction" section="Strand Two: The Hayek Inversion — pressure-tested">calculation has been privatized and rendered opaque</q>, and <q doc="price-without-discovery-from-the-signal-to-the-instruction" section="Strand Two: The Hayek Inversion — pressure-tested">we get computation without either discovery or accountability</q>. "Administered prices" is Gardiner Means's term from the 1930s; applying it to prices computed for each buyer is that research's own framing. The target is the seller's. In one field experiment, personalized pricing <q doc="price-without-discovery-from-the-signal-to-the-instruction" section="The tetrad on the pricing algorithm">improves the firm's expected posterior profits by an additional 19%, relative to the optimized uniform price</q>, while <q doc="price-without-discovery-from-the-signal-to-the-instruction" section="Counterarguments and honest responses">total consumer surplus declines 23% under personalized pricing relative to uniform pricing</q>. The same study found that <q doc="price-without-discovery-from-the-signal-to-the-instruction" section="Counterarguments and honest responses">over 60% of consumers benefit from lower prices under personalization</q>, so the objection is to who sets the target and who sees it, rather than to every outcome. The AI Economist's welfare formula was stated in its paper; a pricing model's is a trade secret.

Social feeds are designed systems too, and what they reward bears directly on the tribal side of politics. In a study of 2,730,215 posts from members of Congress and news outlets, posts about the political out-group <q doc="research-brief-negative-partisanship-silence-pattern-canada-us" section="5. The Underlying Concept: Negative Partisanship">received roughly twice as many shares as posts about the in-group</q>, and <q doc="research-brief-negative-partisanship-silence-pattern-canada-us" section="5. The Underlying Concept: Negative Partisanship">"out-group" language was the single strongest predictor of social-media virality, outperforming negative emotion and moral language</q>. A study of political elites' tweets found that <q doc="research-brief-negative-partisanship-silence-pattern-canada-us" section="5. The Underlying Concept: Negative Partisanship">politicians themselves post more in-party-positive than out-party-negative tweets, but ordinary users engage (likes, shares) much more with the out-party-negative posts.</q> The upshot is that partisans <q doc="research-brief-negative-partisanship-silence-pattern-canada-us" section="5. The Underlying Concept: Negative Partisanship">get rewarded — in attention, in social validation, in algorithmic amplification — for attacking the out-group</q>.

Community Notes, on Twitter/X, is what happens when a platform changes the target. Its algorithm <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="3. What AI system-design experiments have actually demonstrated">rates a crowd-written note highly only when raters from different viewpoints, as inferred from their past ratings, find it helpful</q>. In a survey experiment, notes it selected <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="3. What AI system-design experiments have actually demonstrated">reduced the likelihood of agreeing with the substance of a potentially misleading tweet by about 26% (N=6,046), with no statistically significant difference in effect between Democrats and Republicans</q>. In live tests, <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="3. What AI system-design experiments have actually demonstrated">people exposed to notes were 25–34% less likely to like or retweet the annotated tweets</q>. The study was written by Twitter's own staff and has not been independently replicated, and its authors name the limit: <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="3. What AI system-design experiments have actually demonstrated">By construction, the algorithm needs some cross-partisan agreement to function.</q> Even the notes it chose were received very differently by the two parties:

<figure data-kind="chart" doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="3. What AI system-design experiments have actually demonstrated">
caption: Share who rated the algorithm-selected notes at least somewhat helpful
- Democrats | 78.30% | 78 | #2f7d52
- Republicans | 48.22% | 48 | #cfc9bb
note: Twitter users in a survey experiment, Wave 3, April 2022
</figure>

Polis, used in Taiwan's vTaiwan process to feed public input into rules such as ride-hailing regulation, works on the same principle of rewarding agreement across groups. It has reached real legislation, but there is <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="3. What AI system-design experiments have actually demonstrated">no rigorous causal evaluation of how much they changed outcomes compared with ordinary consultation</q>.

## The fight a better design is meant to end is mostly about the other team

Political scientists separate two kinds of division:

<figure data-kind="names">
- Policy | Ideological polarization | <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="5. What the tribalism evidence says">differences in policy positions</q> | How far apart people are on what government should do.
- People | Affective polarization | <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="5. What the tribalism evidence says">dislike and distrust of the other party's people</q> | How much people resent the other side as a group.
</figure>

In the United States, the gap between how partisans rate their own party and the other party rose <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="5. What the tribalism evidence says">from 22.64 degrees in 1978 to 40.87 degrees in 2016</q> on a 0–100 feeling thermometer. That hostility is <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="5. What the tribalism evidence says">largely distinct from the ideological divide</q>. Ideological sorting accounts <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="5. What the tribalism evidence says">for only about 5% of the variance in affective polarization</q>, and <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="5. What the tribalism evidence says">in some settings, affective polarization can increase while ideological divisions shrink</q>. <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="5. What the tribalism evidence says">In 1960 only 4–5% of people were upset at the idea of their child marrying someone from the other party; by 2010 the figure was one-third of Democrats and one-half of Republicans.</q>

The American evidence is the deepest, but Canada shows the same pattern in a multi-party system. Using Canadian Election Study data from 2008 and 2011, researchers found that negative partisan identity <q doc="research-brief-negative-partisanship-silence-pattern-canada-us" section="5. The Underlying Concept: Negative Partisanship">independently predicts turnout, vote choice and political activity</q>.

The consequence for any design programme is blunt: <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="5. What the tribalism evidence says">the fight would not end if a better-designed system were offered, because the fight was never mainly about system design</q>. Team identity has to be a design target in its own right, and the feed evidence shows it already is one, aimed the wrong way. The remedies tried so far are thin. On small-group discussions between Democrats and Republicans, <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="5. What the tribalism evidence says">there have been no systematic evaluations of these efforts, and it seems questionable whether they are scalable</q>.

> **Objection: the Habermas Machine already shows AI can bring opposing sides together**
>
> It shows AI can write a statement that opposing sides will sign. The experiment <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="Open questions">measured convergence on statements</q>; nobody has yet measured whether taking part lowers resentment of the other party, the quantity the thermometer tracks. Agreement on a text and dislike of the people across the aisle can move separately, and the polarization evidence suggests they often do.

## Simulated citizens would make every design look more agreeable than it is

The cheapest way to test a designed system before using it on people would be to run it on a simulated population of AI citizens. When researchers asked ChatGPT to answer as demographic personas and compared the results with the American National Election Study, the averages matched closely. But <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="4. Can AI simulate a society well enough to test designs before deploying them?">there is less variation in responses than in the real surveys</q>, and <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="4. Can AI simulate a society well enough to test designs before deploying them?">the same prompt yields significantly different results over a 3-month period</q>. Missing variation is the worst possible failure for this use, because <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="4. Can AI simulate a society well enough to test designs before deploying them?">Arrow-type problems arise precisely from heterogeneity</q>. <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="4. Can AI simulate a society well enough to test designs before deploying them?">A synthetic population that is too uniform will make a design look more consensual and stable than it is.</q>

Synthetic respondents are already a product in Canada. Léger, with Plus Compagnie, has launched Smart Persona, and a Léger executive <q doc="pass-2-the-forensics-the-math-and-the-philosophy-of-canadian-polling" section="Synthetic respondents — the actual frontier">claimed an 85–90% concordance score between Smart Persona output and traditional human panel results</q>. That figure is <q doc="pass-2-the-forensics-the-math-and-the-philosophy-of-canadian-polling" section="Caveats">a self-reported industry metric, not an independently validated benchmark</q>, and <q doc="pass-2-the-forensics-the-math-and-the-philosophy-of-canadian-polling" section="Synthetic respondents — the actual frontier">Léger characterizes Smart Persona as for marketing decisions (ad testing, A/B), not political polling.</q> Even so, no rule obliges a pollster to say when synthetic answers fill out a sample: <q doc="pass-2-the-forensics-the-math-and-the-philosophy-of-canadian-polling" section="Synthetic respondents — the actual frontier">There is no regulatory framework at CRIC or anywhere else in Canada that requires disclosure of synthetic respondent supplementation.</q> The ChatGPT study is the warning about how to read a headline match: ChatGPT got the averages right too, and the spread is what went missing. A concordance score on its own does not say whether the disagreement survived.

So the picture of the people can fail the same way the target can: quietly, inside a commercial product, before anyone outside sees it.

> **Objection: designing society from models is how the twentieth century's planned disasters began**
>
> James C. Scott's history of failed planned social orders is the strongest form of this objection, and it bites. High modernism, the drive to order society administratively, was not one camp's vice: <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="1. The warning from high modernism">As a faith, it was shared by many across a wide spectrum of political ideologies.</q> Its carriers were engineers, planners and technocrats. <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="1. The warning from high modernism">That is exactly the group that an AI-enabled design programme would empower.</q> His central point is close to Hayek's: <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="1. The warning from high modernism">Formal order, to be more explicit, is always and to some degree parasitic on informal processes, which the formal scheme does not recognize, without which it could not exist, and which it alone cannot create or maintain.</q>
>
> Scott is <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="1. The warning from high modernism">emphatically not making a blanket case against either bureaucratic planning or high-modernist ideology</q>; his target is planning that shuts out local, practical knowledge. The experiments that worked took their target from the people subject to the rule, by vote or endorsement, which answers part of his objection. It does not answer all of it. A vote carries what people prefer, not the practical know-how by which they make things run, and no study yet tests whether an AI-designed rule keeps or overrides that know-how.

## Bell declared the end of ideology in 1960 and left a usable test

The claim that ideology is finished has been made before. Daniel Bell argued in 1960 that among Western intellectuals the old passions were spent, and that ideology lets people skip judging issues on their merits; they turn to an ideological <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="6. Has &quot;the ideological debate is over&quot; been claimed before?">vending machine, and out comes the prepared formulae</q>. Within a decade the New Left had put ideological conflict back at the centre of Western politics. <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="6. Has &quot;the ideological debate is over&quot; been claimed before?">Bell's claim was a description of a mood among intellectuals, not of the public, and it underestimated how much people want politics to supply identity and passion.</q> That is the same gap the polarization evidence exposes.

Bell did not give up on utopia; he demanded that it be specified. A utopia, he wrote, <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="6. Has &quot;the ideological debate is over&quot; been claimed before?">must specify where one wants to go, how to get there, what the enterprise costs, and who is to pay, with a justification for that choice</q>. Those are the questions a designed system's target answers. The lab experiments answer them in their methods sections. A pricing model and a ranking algorithm answer them too, in code their users never see.

## The left–right argument survives as the choice of a system's target, and private systems are making it first

AI has made the search for a good rule cheap. It has not made the target obvious, and Arrow's result says it never will: every way of adding up what people want is a choice about whose wants count and how. The old ideologies were bundles of those choices made in advance. A designed system makes the same choices explicitly, and the explicit version can at least be compared, tested and argued with.

The lab and the market have so far gone different ways. Where researchers let the public set the target, through a vote, an endorsement or agreement across groups, the results are promising and small: a game, a set of statements, a fact-checking feature. Where firms set it, the systems already run prices and feeds at national scale, aimed at margin and engagement, and the second of those rewards exactly the team hostility a better design was supposed to calm. <q doc="beyond-which-system-wins-designing-social-systems-with-ai-and-what-the-evidence-allows-1" section="Beyond this question">Choosing between those approaches is itself the modern version of the old split between democratic and technocratic thinking.</q> The question that replaces "which system wins" is the one to ask of any rule that governs people, public or private: what is it told to maximize, who decided, and who was in the room, real or simulated, when it was tested.