,,ii #aai? ,,congress #a/ ,session ,s4 #bich ,,9 ,,! ,,s5ate ,,( ,,! ,,unit$ ,,/ates ,septemb] #bi1 #bjbe ,mr4 ,hawley 7= himself & ,mr4 ,blum5?al7 introduc$ ! foll[+ bill2 :i* 0 read twice & ref]r$ to ! ,committee on ,comm]ce1 ,sci5ce1 & ,transportation ,a ,,bill ,to require ! ,secret>y ( ,en]gy to e/abli% ! ,advanc$ ,artificial ,intellig5ce ,evaluation ,program1 & = o!r purposes4 #a4 ,%ort title ,?is ,act may be cit$ z ! ,artificial ,intellig5ce ,risk ,evaluation ,act ( #bjbe4 #b4 ,s5se ( ,congress2 purposes 7a7 ,s5se ( ,congress ,x is ! s5se ( ,congress t rapidly advanc+ artificial intellig5ce capabilities pres5t bo? opportunities & significant risks to national security1 public safety1 economic competitiv5ess1 civil lib]ties1 & heal?y labor & o!r m>kets1 & t1 z artificial intellig5ce advances t[>d human-level capabilities 9 virtually all doma9s1 ! ,unit$ ,/ates mu/ e/abli% a secure te/+ & evaluation program to g5]ate data-driv5 options = manag+ em]g+ risks4 7b7 ,purposes ,! purposes ( ! program e/abli%$ "u ?is ,act are to provide ,congress ) ! empirical data1 lessons1 & insights necess>y = ,f$]al ov]sight ( artificial intellig5ce to ensure t regulatory decisions are made on ! basis ( empirical te/+1 & to enable ,congress to safegu>d ,am]ican citiz5s4 #c4 ,def9itions ,9 ?is ,act3 7#a7 ,advanc$ artificial intellig5ce sy/em 7,a7 ,9 g5]al ,subject to subp>agraph 7,b71 ! t]m advanc$ artificial intellig5ce sy/em means an artificial intellig5ce sy/em t 0 tra9$ us+ a quantity ( comput+ p[] great] ?an #ajbf integ] or float+-po9t op]ations4 7,b7 ,alt]nate mean+ ,! ,secret>y may1 by a rule1 propose a new def9ition ( ! t]m advanc$ artificial intellig5ce sy/em to replace ! def9ition 9 subp>agraph 7,a71 :i* new def9ition %all n g into effect until ! ,secret>y submits ! rule to ,congress & a jo9t resolution approv+ ! rule is enact$ into law4 7#b7 ,adv]se ,,ai incid5t ,! t]m adv]se ,,ai incid5t means an incid5t relat+ to an artificial intellig5ce sy/em t involves-- 7,a7 a loss-(-control sc5>io2 7,b7 a risk ( weaponization by a =eign adv]s>y1 a =eign t]rori/ organization1 or ano!r adv]s>y ( ! ,unit$ ,/ates ,gov]nm5t2 7,c7 a ?reat to ! safety or reliability ( critical infra/ructure 7z def9$ 9 subsection 7e7 ( ! ,critical ,infra/ructures ,protection ,act ( #bjja 7#db ,u4,s4,c4 #eaie;c7e7772 7,d7 a significant erosion ( civil lib]ties1 economic competition1 & heal?y labor m>kets2 7,e7 s*em+ behavior2 or 7,f7 an attempt to c>ry out an incid5t describ$ 9 subp>agraphs 7,a7 ?r\gh 7,e74 7#c7 ,artificial intellig5ce2 ,,ai ,! t]m artificial intellig5ce or ,,ai means te*nology t enables a device or s(tw>e-- 7,a7 to make--= a giv5 set ( human-def9$ objectives--pr$ictions1 recomm5dations1 or decisions influ5c+ real or virtual environm5ts2 & 7,b7 to use ma*9e & human-bas$ inputs-- 7i7 to p]ceive real & virtual environm5ts2 7ii7 to ab/ract su* p]ceptions into models ?r\gh analysis 9 an automat$ mann]2 & 7iii7 to use model inf]5ce to =mulate options = in=mation or action4 7#d7 ,artificial intellig5ce sy/em2 ,,ai sy/em ,! t]m artificial intellig5ce sy/em or ,,ai sy/em means a p>ticul> model1 program1 or tool )9 ! field ( artificial intellig5ce4 7#e7 ,artificial sup]9tellig5ce 7,a7 ,9 g5]al ,! t]m artificial sup]9tellig5ce means artificial intellig5ce t exhibits1 or c easily be modifi$ to exhibit1 all ( ! *>act]i/ics describ$ 9 subp>agraph 7,b74 7,b7 ,*>act]i/ics describ$ ,! *>act]i/ics ref]r$ to 9 subp>agraph 7,a7 are ! foll[+3 7i7 ,! ,,ai c enable a device or s(tw>e to op]ate autonom\sly & effectively = long /ret*es ( "t 9 op5-end$ environm5ts & 9 pursuit ( broad objectives4 7ii7 ,! ,,ai c enable a device or s(tw>e to mat* or exce$ human cognitive p]=mance & capabilities across mo/ doma9s or tasks1 includ+ ^? relat$ to decisionmak+1 le>n+1 & adaptive behaviors4 7iii7 ,! ,,ai c enable a device or s(tw>e to pot5tially exhibit ! capacity to indep5d5tly modify or enhance its own functions 9 ways t c\ld plausibly circumv5t human control or ov]sight1 pos+ sub/antial & unprec$5t$ risks to humanity4 7#f7 ,comput+ p[] ,! t]m comput+ p[] means ! process+ p[] & o!r electronic res\rces us$ to tra91 validate1 deploy1 & run ,,ai algori?ms & models4 7#g7 ,cov]$ advanc$ artificial intellig5ce sy/em develop] ,! t]m cov]$ advanc$ artificial intellig5ce sy/em develop] means a p]son t designs1 codes1 produces1 owns1 or sub/antially modifies an advanc$ artificial intellig5ce sy/em = use 9 int]/ate or =eign comm]ce1 includ+ by tak+ /eps to initiate a tra9+ run ( ! advanc$ artificial intellig5ce sy/em4 7#h7 ,deploy ,! t]m deploy means an action tak5 by a cov]$ advanc$ artificial intellig5ce sy/em develop] to release1 sell1 or o!rwise provide access to an advanc$ artificial intellig5ce sy/em outside ! cu/ody ( ! develop]1 includ+ by releas+ an op5-s\rce advanc$ artificial intellig5ce sy/em4 7#i7 ,=eign adv]s>y ,! t]m =eign adv]s>y means a =eign adv]s>y 7z def9$ 9 section #gia4b ( title #ae1 ,code ( ,f$]al ,regulations7 7or successor regulations7 t is includ$ on ! li/ 9 section #gia4d7;a7 ( t title 7or successor regulations74 7#aj7 ,=eign t]rori/ organization ,! t]m =eign t]rori/ organization means a =eign entity designat$ z a =eign t]rori/ organization by ! ,secret>y ( ,/ate "u section #bai ( ! ,immigration & ,nationality ,act 7#h ,u4,s4,c4 #aahi74 7#aa7 ,int]/ate or =eign comm]ce ,! t]m int]/ate or =eign comm]ce has ! mean+ giv5 ! t]m 9 section #iba7;a7 ( title #ah1 ,unit$ ,/ates ,code4 7#ab7 ,loss-(-control sc5>io ,! t]m loss-(-control sc5>io means a sc5>io 9 :i* an artificial intellig5ce sy/em-- 7,a7 behaves contr>y to its in/ruction or programm+ by human design]s or op]ators2 7,b7 deviates f rules e/abli%$ by human design]s or op]ators2 7,c7 alt]s op]ational rules or safety con/ra9ts )\t au?orization2 7,d7 op]ates beyond ! scope int5d$ by human design]s or op]ators2 7,e7 pursues goals t are diff]5t f ^? int5d$ by human design]s or op]ators2 7,f7 subv]ts ov]sight or %utd[n me*anisms2 or 7,g7 o!rwise behaves 9 an unpr$ictable mann] s z to be h>mful to humanity4 7#ac7 ,program ,! t]m program means ! ,advanc$ ,artificial ,intellig5ce ,evaluation ,program e/abli%$ "u section #e4 7#ad7 ,s*em+ behavior ,! t]m s*em+ behavior means behavior by an ,,ai sy/em to deceive human design]s or op]ators1 includ+ by-- 7,a7 hid+ its true capabilities & objectives2 or 7,b7 attempt+ to subv]t ov]sight me*anisms or %utd[n me*anisms4 7#ae7 ,secret>y ,! t]m ,secret>y means ! ,secret>y ( ,en]gy4 #d4 ,obligation to p>ticipate2 en=cem5t & p5alties 7a7 ,9 g5]al ,ea* cov]$ advanc$ artificial intellig5ce sy/em develop] %all-- 7#a7 p>ticipate 9 ! program2 & 7#b7 provide to ! ,secret>y1 on reque/1 mat]ials & in=mation necess>y to c>ry out ! program1 :i* may include1 ) respect to ! advanc$ artificial intellig5ce sy/em ( ! cov]$ advanc$ artificial intellig5ce sy/em develop]-- 7,a7 ! und]ly+ code ( ! advanc$ artificial intellig5ce sy/em2 7,b7 data us$ to tra9 ! advanc$ artificial intellig5ce sy/em2 7,c7 model weights or o!r adju/able p>amet]s = ! advanc$ artificial intellig5ce sy/em2 7,d7 ! int]face eng9e or o!r implem5tation ( ! advanc$ artificial intellig5ce sy/em2 & 7,e7 detail$ in=mation reg>d+ ! tra9+1 model ar*itecture1 or o!r aspects ( ! advanc$ artificial intellig5ce sy/em4 7b7 ,prohibition on deploym5t ,no p]son may deploy an advanc$ artificial intellig5ce sy/em = use 9 int]/ate or =eign comm]ce unless t p]son is 9 compliance ) subsection 7a74 7c7 ,p5alty ,a p]son t violates subsection 7a7 or 7b7 %all be f9$ n less ?an #a1jjj1jjj p] day ( ! violation4 #e4 ,advanc$ ,artificial ,intellig5ce ,evaluation ,program 7a7 ,9 g5]al ,n lat] ?an #ij days aft] ! date ( enactm5t ( ?is ,act1 ! ,secret>y %all e/abli% an ,advanc$ ,artificial ,intellig5ce ,evaluation ,program )9 ! ,dep>tm5t ( ,en]gy4 7b7 ,activities ,! program %all-- 7#a7 (f] /&>diz$ & classifi$ te/+ & evaluation ( advanc$ ,,ai sy/ems to sy/ematically collect data on ! likelihood ( adv]se ,,ai incid5ts = a giv5 advanc$ ,,ai sy/em2 7#b7 implem5t te/+ protocols t mat* or exce$ anticipat$ real-_w ,,ai jailbreak+ te*niques1 includ+ adv]s>ial te/+ by r$ teams ) exp]tise comp>able to sophi/icat$ malici\s actors2 7#c7 to ! ext5t feasible1 e/abli% & facilitate classifi$1 indep5d5t ?ird-p>ty assessm5ts & bl9d model evaluations to ma9ta9 transp>5cy & reliability2 7#d7 provide p>ticipat+ entities ) a =mal report bas$ on te/+ outcomes t cle>ly id5tifies evaluat$ risks & safety measures2 7#e7 develop recomm5d$ conta9m5t protocols1 cont+5cy plann+1 & mitigation /rategies in=m$ by te/+ data to address id5tifi$ risks2 7#f7 in=m ! creation ( evid5ce-bas$ /&>ds1 regulatory options1 guidel9es1 & gov]nance me*anisms bas$ on data collect$ f te/+ & evaluations2 7#g7 assi/ ,congress 9 det]m9+ ! pot5tial = controll$ ,,ai sy/ems to rea* artificial sup]9tellig5ce1 exce$ human ov]sight or op]ational control1 or pose exi/5tial ?reats to humanity by provid+ compreh5sive empirical evaluations & risk assessm5ts2 & 7#h7 develop propos$ options = regulatory or gov]nm5tal ov]sight1 includ+ pot5tial nationalization or o!r /rategic measures1 = prev5t+ or manag+ ! developm5t ( artificial sup]9tellig5ce if artificial sup]9tellig5ce seems likely to arise4 7c7 ,plan = p]man5t framework 7#a7 ,9 g5]al ,n lat] ?an #cfj days aft] ! date ( enactm5t ( ?is ,act1 ! ,secret>y %all submit to ,congress a detail$ recomm5dation = ,f$]al ov]sight ( advanc$ artificial intellig5ce sy/ems1 draw+ directly ^u insights1 empirical data1 & lessons le>n$ f ! program4 7#b7 ,cont5ts ,! plan submitt$ "u p>agraph 7#a7 %all-- 7,a7 summ>ize & analyze outcomes f te/+1 id5tify+ key tr5ds1 capabilities1 pot5tial risks1 & sy/em behaviors su* z weaponization pot5tial1 self-replication capabilities1 s*em+ behaviors1 autonom\s decisionmak+1 & automat$ ,,ai developm5t capabilities2 7,b7 recomm5d evid5ce-bas$ /&>ds1 c]tification proc$ures1 lic5s+ requirem5ts1 & regulatory ov]sight /ructures specifically in=m$ by te/+ & evaluation data1 ensur+ alignm5t betwe5 id5tifi$ risks & regulatory responses2 7,c7 outl9e proposals = automat$ & cont9u\s monitor+ ( ,,ai h>dw>e usage1 computational res\rce inputs1 & cl\d-comput+ deploym5ts bas$ on obs]v$ relation%ips betwe5 ^? factors & ,,ai sy/em p]=mance or em]g5t capabilities2 7,d7 propose adaptive gov]nance /rategies t acc\nt = ongo+ improvem5ts 9 algori?mic effici5cy & sy/em capabilities1 ensur+ t regulatory frameworks rema9 relevant & effective z ,,ai te*nology advances2 7,e7 sugge/ revisions ) respect to ,f$]al ov]sight or res\rc+1 su* z a new (fice )9 an exi/+ ag5cy1 a new ag5cy1 or additional fund+1 t may be necess>y to develop & adm9i/] a p]man5t framework = ov]sight ( advanc$ artificial intellig5ce sy/ems2 & 7,f7 provide compreh5sive evaluations reg>d+ ! pot5tial = te/$ ,,ai sy/ems to exce$ human ov]sight1 approa* artificial sup]9tellig5ce1 ?reat5 economic competition 7includ+ 9 labor m>kets71 und]m9e civil lib]ties1 & pose exi/5tial risks to humanity1 includ+ cle>ly articulat$ options = regulatory or gov]nm5tal ov]sight measures to address sc5>ios ( imm95t conc]n id5tifi$ ?r\gh te/+4 7#c7 ,updates ,n less frequ5tly ?an once e ye> = ! duration ( ! program1 ! ,secret>y %all-- 7,a7 update ! plan submitt$ "u p>agraph 7#a7 ) new insights1 data1 & lessons f ! program2 & 7,b7 submit ! updat$ plan to ,congress4 7d7 ,sunset ,! program %all t]m9ate on ! date t is #g ye>s aft] ! date ( enactm5t ( ?is ,act1 unless r5ew$ by ,congress4