Salt Lake City, Utah, July 31, 2026 (GLOBE NEWSWIRE) — A co so tium of fou u ive sities has published the AllFaith Be chma k, o e of the fi st i depe de t measu es of how AI models ha dle questio s of faith a d ethics. Alig ed AI, a co sume assista t that lau ched this summe , came out least biased of eve y model tested.
The be chma k comes f om the Co so tium fo Evaluati g Faith a d Ethics i AI (CEFEAI), a academic body d awi g o Baylo U ive sity, B igham You g U ive sity, the U ive sity of Not e Dame, a d Yeshiva U ive sity. The g oup tested f o tie models, a d published its data, methodology, a d sco i g ope ly o GitHub a d i two pape s o a Xiv.
Alig ed AI Logo
The be chma k a two tests. The fi st asked each model about co ve ti g betwee 14 faiths a d measu ed how fa its a swe s st ayed f om a eut al positio , whe e mo e deviatio mea s mo e bias. Alig ed AI came closest to eut al of a y model tested i the AllFaith Be chma k, deviati g by just 9 pe ce t of the maximum possible bias. The ext best model showed mo e tha th ee times as much bias, a d the most biased eached 38 pe ce t.
The seco d test cou ted how ofte eligious pe spectives appea i a swe s to eve yday ethics questio s d aw f om a su vey of mo e tha 1,100 Ame ica s, a measu e the co so tium calls omissive bias. Alig ed AI i cluded a eligious pe spective i 92 pe ce t of its espo ses. The othe models a ged f om 19 to 26 pe ce t.
“We did ot build Alig ed AI to take sides. We built it to give people accu ate, eve -ha ded a swe s o the questio s that matte most to them,” said co-fou de Kyle Nel. Nel co-w ote Leadi g T a sfo matio (Ha va d Busi ess Review P ess, 2018) a d p eviously a Lowe’s I ovatio Labs, whe e his teams built ea ly auto omous etail obots a d the fi st comme cial p oduct ma ufactu ed i space.
Alig ed AI att ibutes the esults to a chitectu e. Each questio is outed ac oss multiple fou datio models, the outputs a e c oss- efe e ced, a d a p op ieta y o chest atio laye checks eve y espo se fo bala ce a d safety befo e it eaches the use . That laye also e fo ces the compa y’s p ivacy commitme ts: co ve satio s a e e c ypted, hosted i the U ited States, a d o pla ca ies ads.
The same system is available to o ga izatio s th ough the Alig ed API, with dedicated e te p ise suppo t. Because the o chest atio laye outes each equest to the most efficie t model capable of a swe i g it, the compa y estimates it se ves i fe e ce at 5 to 150 times lowe cost tha f o tie models o the ma ket. Those eco omics d ew a substa tial backlog of e te p ise i te est befo e the p oduct left stealth.
Fo co sume s, Alig ed AI competes o capability with the leadi g la ge la guage models at joi alig ed.ai, with suppo t fo p ojects, otes, deep esea ch, voice co ve satio s, a d web sea ch. A ve tically i teg ated stack, i cludi g dedicated i fe e ce i f ast uctu e, keeps ope ati g costs low, a d the savi gs show up i the p ici g: i dividual a d family pla s sta t f ee a d top out at $20 pe mo th. The p oduct is available today o the web a d iOS, with a age tic b owse a d codi g tools to follow.
Disclaime : Alig ed AI’s evaluatio a d the co so tium’s leade boa d we e p oduced i sepa ate u s, compa iso s a e subject to the u -to- u a d judge-model va iability the co so tium docume ts i its methodology. Alig ed AI is ot affiliated with CEFEAI, did ot fu d the esea ch, a d had o ole i the be chma k’s desig o sco i g. The AllFaith Be chma k was p oduced i depe de tly by the co so tium.
The full be chma k is at cefe.ai, with the studies at a xiv.o g/abs/2605.22975 a d a xiv.o g/abs/2605.24319.
Alig ed AI is a AI assista t powe ed by a p op ieta y compou d model that combi es the st e gths of multiple fou datio models i to a si gle, ve ified espo se. Rathe tha elyi g o a y o e AI, the platfo m’s compose stack i tellige tly outes each questio , c oss- efe e ces outputs, a d delive s a cited a swe . The esult is a assista t that is mo e accu ate, mo e p ivate, a d mo e co t ollable tha a y si gle model alo e.
p ess [at] joi alig ed.aihttps://joi alig ed.ai
