When AI simulates the future of our technological choices

A vir­tu­al cell is not simply a cell described in great­er detail. More than that, it enables treat­ments to be tested in a digit­al bio­lo­gic­al sys­tem – inhib­it­ing a gene, intro­du­cing a new molecule, or modi­fy­ing an envir­on­ment, and then com­par­ing the res­ult­ing tra­ject­or­ies. At the oth­er end of the spec­trum, digit­al twins of the Earth sys­tem no longer seek merely to rep­res­ent cli­mate. They enable us to explore scen­ari­os, link res­ults to water, energy or extreme events, and test policies before they are imple­men­ted. In both cases, the mod­el ceases to be a mere image for obser­va­tions. It becomes a space in which we can act before tak­ing action in the real world.

World mod­els are driv­ing this trans­form­a­tion fur­ther. An arti­fi­cial agent no longer merely pro­duces a response, it main­tains an oper­a­tion­al rep­res­ent­a­tion of its envir­on­ment, anti­cip­ates the effects of an action and explores mul­tiple tra­ject­or­ies. DreamerV3 learns by ‘ima­gin­ing’ futures; Genie gen­er­ates inter­act­ive envir­on­ments from videos; GAIA‑1 pro­duces sev­er­al pos­sible out­comes for a driv­ing scen­ario. The ini­ti­at­ive pub­lished in July 2026 by the Stan­ford Insti­tute for Human-Centred Arti­fi­cial Intel­li­gence gives this move­ment polit­ic­al sig­ni­fic­ance – as soon as a world mod­el con­trib­utes to a crit­ic­al decision, its assess­ment and gov­ernance become a mat­ter of pub­lic concern.

How­ever, the key shift lies neither in com­pu­ta­tion­al power nor in visu­al real­ism. A spec­tac­u­lar scene may be sci­en­tific­ally lack­ing if causes do not pro­duce the same effects with­in it, if actions leave no trace, or if uncer­tainty remains invis­ible. Con­versely, an aus­tere envir­on­ment can become a power­ful labor­at­ory when it retains states, imposes con­straints and allows for con­tra­dic­tions. A sim­u­lated world is not more than a back­drop – it is a hypo­thes­is made executable.

For a long time, we have relied on sim­u­la­tion to determ­ine wheth­er a tech­no­logy would work. This ques­tion has become obsol­ete. An organ­isa­tion nev­er adopts a sys­tem in the per­fectly con­trolled envir­on­ment used to test it. It intro­duces it into an incom­plete envir­on­ment, pop­u­lated with act­ors who learn, cir­cum­vent the rules, reor­gan­ise their prac­tices and trans­form the data. We must there­fore shift the focus of the eval­u­ation from simply test­ing the tech­no­logy, with the decision to adopt it.

We refer to the abil­ity of a choice to remain under­stand­able, revis­able and defens­ible as its envir­on­ment changes as ‘decision sta­bil­ity’. A stable decision is not stat­ic; it may change, but the reas­ons for that change are known: decision-makers know which assump­tions under­pin their choice, which vari­ables need to be mon­itored, and which thresholds neces­sit­ate a review. Sim­u­la­tion no longer seeks primar­ily the most prob­able future. It seeks out futures in which the decision ceases to achieve its objective.

Sim­u­la­tion no longer seeks primar­ily the most prob­able future. It seeks out futures in which the decision ceases to achieve its objective.

As such, the use­ful out­come is not a single score. It is a map of the decision’s vul­ner­ab­il­it­ies cap­able of revers­ing the choice, the com­bin­a­tions that cause a break­down, and the sig­nals to watch for after imple­ment­a­tion. This approach dis­tin­guishes between what needs to be made robust, what needs to remain adapt­able, and what requires a fall­back solu­tion. It also trans­forms the mod­el into a mon­it­or­ing tool – the sim­u­la­tion is not aban­doned once the decision has been made, it is updated as the envir­on­ment changes.

The tempta­tion with­in industry is to cre­ate an ever-lar­ger rep­lica, inten­ded to gradu­ally encom­pass all dimen­sions of real­ity. This over­looks the inher­ent power of mod­els.  Each one selects vari­ables, scales and caus­al rela­tion­ships. Hence, the right infra­struc­ture is not a defin­it­ive world, but a con­trolled plur­al­ity of par­tial worlds. For any giv­en decision, sev­er­al envir­on­ments must dif­fer in terms of their data, the beha­viour of their act­ors, their caus­al assump­tions and their cri­ter­ia for success.

The first world is the one that tech­no­logy demands. Every innov­a­tion incurs an ‘envir­on­ment­al debt’ – mean­ing the total­ity of resources, infra­struc­ture and skills neces­sary for its oper­a­tion – yet absent from the raw meas­ure­ment of its per­form­ance. Annot­ated data, com­put­ing power, sup­pli­ers, main­ten­ance, trained oper­at­ors, con­nectiv­ity and backup cap­ab­il­it­ies are often provided free of charge dur­ing the exper­i­ment­al phase. Sim­u­lat­ing this world involves meth­od­ic­ally degrad­ing these con­di­tions. We no longer meas­ure just the acquis­i­tion cost, but the cost of main­tain­ing the required world.

The second world is the one that tech­no­logy cre­ates. No sys­tem oper­ates in a pass­ive envir­on­ment. Users change their prac­tices, organ­isa­tions shift their resources towards what the sys­tem meas­ures, com­pet­it­ors adapt and adversar­ies seek out vul­ner­ab­il­it­ies. An accur­ate pre­dic­tion can become incor­rect when all act­ors fol­low the same sig­nal. This reflex­ive world requires multi-agent sim­u­la­tions, decision games, red team­ing and exper­i­ments com­bin­ing human oper­at­ors and auto­mated systems.

The third world is the one that remains after a fail­ure. Is the data still access­ible? Have human skills been retained? Is there a fall­back pro­ced­ure in place? Can the infra­struc­ture be repur­posed? Does the decision pro­duced by the sys­tem remain explain­able and con­test­able? Revers­ib­il­ity does not mean simply revert­ing to the pre­vi­ous situ­ation. It con­sists of pre­serving the capa­city to act when the main sys­tem fails, becomes undesir­able or must be abandoned.

Com­par­ing sev­er­al scen­ari­os reveals three key ele­ments. A decision invari­ant refers to a choice that remains prefer­able des­pite sig­ni­fic­ant vari­ations in the con­text. A decision frac­ture occurs when a minor change to an assump­tion is suf­fi­cient to reverse the rank­ing of options. It often reveals more than a cent­ral fore­cast, as it indic­ates the vari­able on which the choice actu­ally depends. Finally, the com­mon blind spot arises when sev­er­al sim­u­la­tions con­verge because they share the same data, the same beha­viours or the same omission.

This is why the audit must be sup­ple­men­ted by a counter-sim­u­la­tion. A tra­di­tion­al audit veri­fies that a sys­tem oper­ates in accord­ance with its design. A counter-sim­u­la­tion con­structs an altern­at­ive scen­ario, shifts the bound­ar­ies of the prob­lem and seeks out the assump­tions cap­able of over­turn­ing the con­clu­sion. Each envir­on­ment must define its scope and omis­sions, be sub­jec­ted to dis­rup­tions out­side its nom­in­al con­di­tions, and then be com­pared with anoth­er sim­u­la­tion, a phys­ic­al exper­i­ment or real-world observations.

A counter-sim­u­la­tion con­structs an altern­at­ive scen­ario, shifts the bound­ar­ies of the prob­lem and seeks out the assump­tions cap­able of over­turn­ing the conclusion.

Let us take the auto­ma­tion of a logist­ics func­tion as an example. A first scen­ario assumes abund­ant energy and stable con­nectiv­ity. A second intro­duces net­work out­ages and infre­quent main­ten­ance. A third depicts oper­at­ors reor­gan­ising their prac­tices around the sys­tem. A fourth adds an adversary seek­ing to deceive it. None of these is ‘the real one’. Com­par­ing them reveals under which con­di­tions auto­ma­tion retains its advant­age, at what threshold it cre­ates excess­ive depend­ency, and which human or tech­nic­al cap­ab­il­it­ies must be preserved.

A tech­no­logy con­sti­tutes a man­age­able advance when it deliv­ers genu­ine cap­ab­il­ity, retains its value across sev­er­al plaus­ible scen­ari­os, acknow­ledges the uncer­tainty of its out­comes, and leaves options open should it fail. This require­ment does not pit cau­tion against power. On the con­trary, the most power­ful tech­no­lo­gies demand more extens­ive sim­u­la­tions, pre­cisely because they trans­form the world in which they oper­ate to a great­er extent.

Before any major deploy­ment, four ques­tions should there­fore be asked: what kind of world must be main­tained for the tech­no­logy to func­tion? What beha­viours will its adop­tion alter? At what threshold must the decision be reviewed? What is pos­sible when the sys­tem no longer responds? These ques­tions do not guar­an­tee a per­fect decision. They high­light the implic­a­tions of a deploy­ment before they become infra­struc­ture, habits and dependencies.

For a long time, we have sim­u­lated what might hap­pen in a giv­en world. The chal­lenge now is to sim­u­late sev­er­al worlds to determ­ine which decision is worth imple­ment­ing in the one world where we will actu­ally have to bear the consequences.