The head of the ar­ti­fi­cial in­tel­li­gence com­pa­ny that makes Chat­G­PT told Con­gress on Tues­day that gov­ern­ment in­ter­ven­tion will be crit­i­cal to mit­i­gat­ing the risks of in­creas­ing­ly pow­er­ful AI sys­tems.

“As this tech­nol­o­gy ad­vances, we un­der­stand that peo­ple are anx­ious about how it could change the way we live. We are too,” Ope­nAI CEO Sam Alt­man said at a Sen­ate hear­ing.

Alt­man pro­posed the for­ma­tion of a US or glob­al agency that would li­cense the most pow­er­ful AI sys­tems and have the au­thor­i­ty to “take that li­cense away and en­sure com­pli­ance with safe­ty stan­dards.”

His San Fran­cis­co-based start­up rock­et­ed to pub­lic at­ten­tion af­ter it re­leased Chat­G­PT late last year. The free chat­bot tool an­swers ques­tions with con­vinc­ing­ly hu­man-like re­spons­es.

What start­ed out as a pan­ic among ed­u­ca­tors about Chat­G­PT’s use to cheat on home­work as­sign­ments has ex­pand­ed to broad­er con­cerns about the abil­i­ty of the lat­est crop of “gen­er­a­tive AI” tools to mis­lead peo­ple, spread false­hoods, vi­o­late copy­right pro­tec­tions and up­end some jobs.

Sen. Richard Blu­men­thal, the Con­necti­cut De­mo­c­rat who chairs the Sen­ate Ju­di­cia­ry Com­mit­tee’s sub­com­mit­tee on pri­va­cy, tech­nol­o­gy and the law, opened the hear­ing with a record­ed speech that sound­ed like the sen­a­tor, but was ac­tu­al­ly a voice clone trained on Blu­men­thal’s floor speech­es and recit­ing Chat­G­PT-writ­ten open­ing re­marks.

The re­sult was im­pres­sive, said Blu­men­thal, but he added, “What if I had asked it, and what if it had pro­vid­ed, an en­dorse­ment of Ukraine sur­ren­der­ing or (Russ­ian Pres­i­dent) Vladimir Putin’s lead­er­ship?”

The over­all tone of sen­a­tors’ ques­tion­ing was po­lite Tues­day, a con­trast to past con­gres­sion­al hear­ings in which tech and so­cial me­dia ex­ec­u­tives faced tough grillings over the in­dus­try’s fail­ures to man­age da­ta pri­va­cy or counter harm­ful mis­in­for­ma­tion. In part, that was be­cause both De­moc­rats and Re­pub­li­cans said they were in­ter­est­ed in seek­ing Alt­man’s ex­per­tise on avert­ing prob­lems that haven’t yet oc­curred.

Blu­men­thal said AI com­pa­nies ought to be re­quired to test their sys­tems and dis­close known risks be­fore re­leas­ing them, and ex­pressed par­tic­u­lar con­cern about how fu­ture AI sys­tems could desta­bilise the job mar­ket. Alt­man was large­ly in agree­ment, though had a more op­ti­mistic take on the fu­ture of work.

Pressed on his own worst fear about AI, Alt­man most­ly avoid­ed specifics, ex­cept to say that the in­dus­try could cause “sig­nif­i­cant harm to the world” and that “if this tech­nol­o­gy goes wrong, it can go quite wrong.”

But he lat­er pro­posed that a new reg­u­la­to­ry agency should im­pose safe­guards that would block AI mod­els that could “self-repli­cate and self-ex­fil­trate in­to the wild” — hint­ing at fu­tur­is­tic con­cerns about ad­vanced AI sys­tems that could ma­nip­u­late hu­mans in­to ced­ing con­trol.

That fo­cus on a far-off “sci­ence fic­tion trope” of su­per-pow­er­ful AI could make it hard­er to take ac­tion against al­ready ex­ist­ing harms that re­quire reg­u­la­tors to dig deep on da­ta trans­paren­cy, dis­crim­i­na­to­ry be­hav­ior and po­ten­tial for trick­ery and dis­in­for­ma­tion, said a for­mer Biden ad­min­is­tra­tion of­fi­cial who co-au­thored its plan for an AI bill of rights.

“It’s the fear of these (su­per-pow­er­ful) sys­tems and our lack of un­der­stand­ing of them that is mak­ing every­one have a col­lec­tive freak-out,” said Suresh Venkata­sub­ra­man­ian, a Brown Uni­ver­si­ty com­put­er sci­en­tist who was as­sis­tant di­rec­tor for sci­ence and jus­tice at the White House Of­fice of Sci­ence and Tech­nol­o­gy Pol­i­cy. “This fear, which is very un­found­ed, is a dis­trac­tion from all the con­cerns we’re deal­ing with right now.”

Ope­nAI has ex­pressed those ex­is­ten­tial con­cerns since its in­cep­tion. Co-found­ed by Alt­man in 2015 with back­ing from tech bil­lion­aire Elon Musk, the start­up has evolved from a non­prof­it re­search lab with a safe­ty-fo­cused mis­sion in­to a busi­ness. Its oth­er pop­u­lar AI prod­ucts in­clude the im­age-mak­er DALL-E Mi­crosoft has in­vest­ed bil­lions of dol­lars in­to the start­up and has in­te­grat­ed its tech­nol­o­gy in­to its own prod­ucts, in­clud­ing its search en­gine Bing.

Alt­man is al­so plan­ning to em­bark on a world­wide tour this month to na­tion­al cap­i­tals and ma­jor cities across six con­ti­nents to talk about the tech­nol­o­gy with pol­i­cy­mak­ers and the pub­lic. On the eve of his Sen­ate tes­ti­mo­ny, he dined with dozens of U.S. law­mak­ers, sev­er­al of whom told CN­BC they were im­pressed by his com­ments.

Al­so tes­ti­fy­ing were IBM’s chief pri­va­cy and trust of­fi­cer, Christi­na Mont­gomery, and Gary Mar­cus, a pro­fes­sor emer­i­tus at New York Uni­ver­si­ty who was among a group of AI ex­perts who called on Ope­nAI and oth­er tech firms to pause their de­vel­op­ment of more pow­er­ful AI mod­els for six months to give so­ci­ety more time to con­sid­er the risks. The let­ter was a re­sponse to the March re­lease of Ope­nAI’s lat­est mod­el, GPT-4, de­scribed as more pow­er­ful than Chat­G­PT.

The pan­el’s rank­ing Re­pub­li­can, Sen. Josh Haw­ley of Mis­souri, said the tech­nol­o­gy has big im­pli­ca­tions for elec­tions, jobs and na­tion­al se­cu­ri­ty. He said Tues­day’s hear­ing marked “a crit­i­cal first step to­wards un­der­stand­ing what Con­gress should do.”

A num­ber of tech in­dus­try lead­ers have said they wel­come some form of AI over­sight but have cau­tioned against what they see as over­ly heavy-hand­ed rules. Alt­man and Mar­cus both called for an AI-fo­cused reg­u­la­tor, prefer­ably an in­ter­na­tion­al one, with Alt­man cit­ing the prece­dent of the U.N.’s nu­clear agency and Mar­cus com­par­ing it to the US Food and Drug Ad­min­is­tra­tion. But IBM’s Mont­gomery in­stead asked Con­gress to take a “pre­ci­sion reg­u­la­tion” ap­proach.

“We think that AI should be reg­u­lat­ed at the point of risk, es­sen­tial­ly,” Mont­gomery said, by es­tab­lish­ing rules that gov­ern the de­ploy­ment of spe­cif­ic us­es of AI rather than the tech­nol­o­gy it­self.

Eu­rope lead­ing

LON­DON (AP) — Au­thor­i­ties around the world are rac­ing to draw up rules for ar­ti­fi­cial in­tel­li­gence, in­clud­ing in the Eu­ro­pean Union.

A Eu­ro­pean Par­lia­ment com­mit­tee vot­ed last week to strength­en the flag­ship leg­isla­tive pro­pos­al as it heads to­ward pas­sage, part of a years­long ef­fort by Brus­sels to draw up guardrails for ar­ti­fi­cial in­tel­li­gence. Those ef­forts have tak­en on more ur­gency as the rapid ad­vances of chat­bots like Chat­G­PT high­light ben­e­fits the emerg­ing tech­nol­o­gy can bring — and the new per­ils it pos­es.

Here’s a look at the EU’s Ar­ti­fi­cial In­tel­li­gence Act:

How do the rules work?

The AI Act, first pro­posed in 2021, will gov­ern any prod­uct or ser­vice that us­es an ar­ti­fi­cial in­tel­li­gence sys­tem. The act will clas­si­fy AI sys­tems ac­cord­ing to four lev­els of risk, from min­i­mal to un­ac­cept­able. Riski­er ap­pli­ca­tions will face tougher re­quire­ments, in­clud­ing be­ing more trans­par­ent and us­ing ac­cu­rate da­ta. Think about it as a “risk man­age­ment sys­tem for AI,” said Jo­hann Laux, an ex­pert at the Ox­ford In­ter­net In­sti­tute.

What are the risks

One of the EU’s main goals is to guard against any AI threats to health and safe­ty and pro­tect fun­da­men­tal rights and val­ues.

That means some AI us­es are an ab­solute no-no, such as “so­cial scor­ing” sys­tems that judge peo­ple based on their be­hav­ior. AI that ex­ploits vul­ner­a­ble peo­ple in­clud­ing chil­dren or that us­es sub­lim­i­nal ma­nip­u­la­tion that can re­sult in harm, such as an in­ter­ac­tive talk­ing toy that en­cour­ages dan­ger­ous be­hav­ior, is al­so for­bid­den.

Law­mak­ers beefed up the pro­pos­al by vot­ing to ban pre­dic­tive polic­ing tools, which crunch da­ta to fore­cast where crimes will hap­pen and who will com­mit them. They al­so ap­proved a widened ban on re­mote fa­cial recog­ni­tion, save for a few law en­force­ment ex­cep­tions like pre­vent­ing a spe­cif­ic ter­ror­ist threat. The tech­nol­o­gy scans passers-by and us­es AI to match their faces to a data­base.

The aim is “to avoid a con­trolled so­ci­ety based on AI,” Bran­do Benifei, the Ital­ian law­mak­er help­ing lead the Eu­ro­pean Par­lia­ment’s AI ef­forts, told re­porters Wednes­day. “We think that these tech­nolo­gies could be used in­stead of the good al­so for the bad, and we con­sid­er the risks to be too high.”

AI sys­tems used in high risk cat­e­gories like em­ploy­ment and ed­u­ca­tion, which would af­fect the course of a per­son’s life, face tough re­quire­ments such as be­ing trans­par­ent with users and putting in place risk as­sess­ment and mit­i­ga­tion mea­sures.

The EU’s ex­ec­u­tive arm says most AI sys­tems, such as video games or spam fil­ters, fall in­to the low- or no-risk cat­e­go­ry.

What about Chat­G­PT?

The orig­i­nal 108-page pro­pos­al bare­ly men­tioned chat­bots, mere­ly re­quir­ing them to be la­beled so users know they’re in­ter­act­ing with a ma­chine. Ne­go­tia­tors lat­er added pro­vi­sions to cov­er gen­er­al pur­pose AI like Chat­G­PT, sub­ject­ing them to some of the same re­quire­ments as high-risk sys­tems.

One key ad­di­tion is a re­quire­ment to thor­ough­ly doc­u­ment any copy­right ma­te­r­i­al used to teach AI sys­tems how to gen­er­ate text, im­ages, video or mu­sic that re­sem­bles hu­man work. That would let con­tent cre­ators know if their blog posts, dig­i­tal books, sci­en­tif­ic ar­ti­cles or pop songs have been used to train al­go­rithms that pow­er sys­tems like Chat­G­PT. Then they could de­cide whether their work has been copied and seek re­dress. (AP sto­ries)