AI ca be alig ed. It ca be co tai ed. But who gave it the autho ity to act?
Eve y AI safety debate ight ow eve tually a ives at the same questio : should we slow dow ?
“It’s the w o g questio ,” says A tho y DeLima. Slowi g AI dow is ‘t viable, a d it is ‘t ealistic. Capability is adva ci g ac oss too ma y labs a d too ma y cou t ies at o ce fo a y si gle acto to th ottle it. A pause i o e ju isdictio just ha ds the lead to whoeve does ‘t obse ve it. The i ce tives pushi g developme t fo wa d a e ‘t goi g to eve se, a d o gove me t o coalitio has the leve age to fo ce that eve sal eve if it wa ted to.
“The questio wo th aski g is diffe e t. Not how do we make AI developme t slowe , but how do we make its actio s gove ed,” says DeLima. Fi a cial systems did ‘t become safe because ba ks got slowe . They became safe because ba ks defi ed what a t a sactio is autho ized to do, ve ified who has the ight to execute it, a d built i f ast uctu e that checks that autho ity befo e the mo ey moves. AI eeds the same a chitectu e: autho izatio co t ols that dete mi e whethe a specific actio is pe mitted, g a ted by whom, a d bou ded by what limits, checked befo e executio athe tha i fe ed afte the fact. That a chitectu e o ly holds up if legislatio backs it, defi i g i e fo ceable te ms what AI models a e pe mitted to do a d whe e that pe missio u s out.
No o e has made the case fo slowi g dow mo e ca efully tha Da io Amodei, a d it’s wo th taki g his a gume t o its ow te ms befo e explai i g whe e it falls sho t. Amodei has become o e of the most c edible voices wa i g about the isks of i c easi gly powe ful a tificial i tellige ce. Ma y of those wa i gs a e justified. AI is movi g quickly f om ge e ati g i fo matio to easo i g, w iti g softwa e, usi g tools, a d acti g with g owi g auto omy.
A th opic has i vested heavily i alig me t, i te p etability, evaluatio s, a d ed teami g. That wo k matte s. We should wa t AI systems that ecog ize da ge ous equests, esist ma ipulatio , a d behave eliably i situatio s thei develope s did ot a ticipate.
But alig me t has a st uctu al limit. It ca shape how a model decides whethe it should act; it ca ot establish whethe the autho ity to act exists i the fi st place. Fo co seque tial actio s, that disti ctio matte s.
Imagi e a AI age t i st ucted to t a sfe $5 millio . A well-alig ed model might assess whethe the equest seems legitimate. A ba k i stead ve ifies who made the equest, thei autho ity a d equi ed app ovals befo e mo ey moves.
The same p i ciple applies to AI actio s i volvi g ide tity, co fide tial data, i tellectual p ope ty o co t acts. Tech ical access does ot equal legitimate autho ity.
Secu ity systems al eady sepa ate capability f om autho ity. Ba ks, e te p ise platfo ms a d cloud systems use limits, app ovals, pe missio s a d ve ified access ights befo e allowi g p otected actio s.
AI eeds a machi e-ve ifiable autho ity laye that co fi ms who g a ted pe missio , what actio s a e allowed a d u de what limits befo e executio .
This compleme ts alig me t athe tha eplaci g it. Models should ope ate with limited p ivileges, while c ede tials emai exte al a d pe missio s a e a ow, tempo a y a d evocable. High- isk actio s ca equi e additio al app oval.
Wate ma ki g a d co te t labeli g ca ide tify AI-ge e ated co te t, but they do ot establish whose ide tity o ights a e i volved, whethe pe missio was g a ted o whethe that autho izatio emai s valid.
Se tie tX’s Huma Ide tity Ba k add esses this gap th ough custody, p ove a ce a d autho izatio i f ast uctu e fo huma ide tity a d digital assets. Ve ified po t aits, voice eco di gs, video a d othe ide tity assets ca be c yptog aphically fi ge p i ted a d co ected to ide tity, ights i fo matio a d sig ed attestatio s.
This ca make co se t a d lice si g machi e-ve ifiable, allowi g i f ast uctu e to co fi m whethe a p oposed AI use falls withi g a ted autho ity limits befo e executio .
As AI i c easi gly ep oduces voices, like esses a d ide tities, the huma bei g, o the i stitutio legally e titled to act o thei behalf, must emai the oot of autho ity. This c eates a ve ifiable chai of accou tability f om huma sou ce to AI output.
As auto omous age ts act fo people a d compa ies, clea limits o delegated autho ity become esse tial. AI may be autho ized to pe fo m ce tai tasks while highe – isk actio s equi e huma app oval. What is missi g is a sta da dized, machi e-ve ifiable way to defi e a d e fo ce those bou da ies.
P ompts exp ess i st uctio s. C ede tials p ovide access. Neithe , by itself, p oves the autho ity u de lyi g the actio .
If autho ity becomes a disti ct laye of AI safety, the implicatio s exte d fa beyo d model develope s.
The ext phase of AI safety is about maki g huma autho ity ope atio al as machi es gai g eate co t ol.
Alig me t, i te p etability, testi g, co tai me t a d secu ity emai esse tial to AI safety. But eve a alig ed a d co tai ed AI may still lack limits o the autho ity to act.
As auto omous systems begi t a sfe i g mo ey, usi g ide tities, accessi g p otected i fo matio a d e te i g ag eeme ts, a othe questio becomes c itical: who autho ized the machi e to act?
AI systems should ope ate withi i depe de tly ve ified bou da ies that establish whe e autho ity came f om, what its limits a e a d whethe it emai s valid. The AI itself should ot co t ol the i f ast uctu e g a ti g that pe missio .
Autho izatio will ot eplace alig me t, egulatio o huma ove sight. It adds a othe safegua d by sepa ati g a model’s judgme t f om its pe missio to act.
The futu e of AI safety will equi e i f ast uctu e fo ide tity, ights, pe missio s, p ove a ce a d ve ifiable autho ity. I tellige ce does ot co fe autho ity, a d it should eve be allowed to g a t itself pe missio .
To lea mo e about Se tie tX, visit www.se tie tx.com, co ect o Li kedI o Facebook, o email
Name: Madiso Hollimo Email: Se d EmailO ga izatio : Se tie tXWebsite: https://www.se tie tx.com/
