TEMECULA, Calif., July 15, 2026 (GLOBE NEWSWIRE) — PYX Labs, a esea ch lab spo so ed by Pe ceptyx, today a ou ced the elease of PYX-Voice, the i dust y’s fi st be chma k desig ed to evaluate how well f o tie AI models u de sta d employee feedback. The be chma k tested seve leadi g models f om Ope AI, Google, A th opic, a d xAI ac oss 84 employee liste i g tasks, fi di g that model eliability decli ed sig ifica tly whe i te p eti g the complex huma co text behi d what employees a e exp essi g.
The be chma k evaluates model espo ses agai st expe t-defi ed c ite ia developed by i dust ial-o ga izatio al (I-O) psychologists a d o ga izatio al behavio specialists, measu i g ot simply whethe AI ca complete wo kplace tasks, but whethe it demo st ates the judgme t equi ed to accu ately i te p et how employees expe ie ce wo k.
The fi di gs come as o ga izatio s i c easi gly ely o ge e ative AI to summa ize employee comme ts, ide tify wo kplace issues, a d i fo m decisio s about cultu e, leade ship, o ga izatio al cha ge, a d employee developme t. I a 2025 su vey of mo e tha 1,300 U.S. ma age s, six i te epo ted usi g AI to help make decisio s about thei di ect epo ts, i cludi g aises, p omotio s, layoffs, a d te mi atio s—despite the abse ce of a y established sta da d fo evaluati g how eliably these systems i te p et huma feedback.
O qua titative tasks with clea , ve ifiable a swe s, f o tie models pe fo med co siste tly well, cluste i g betwee 64% a d 82%. O i te p etive tasks—sy thesizi g ope -e ded employee feedback i to a cohe e t, accu ate takeaway—sco es d opped as low as 33%.
F o tie AI models pe fo med st o gest o employee expe ie ce topics whe e feedback is exp essed usi g co siste t la guage a d ca be eadily catego ized i to well-defi ed themes.
Fo example, employee feedback elated to pe fo ma ce e ableme t typically has ve y clea , co siste t te mi ology (e.g., goals, esou ces, tools, success met ics) that easily maps to this catego y. I co t ast, employee feedback elated to cha ge & i ovatio ca be b oad, complex, a d ua ced, eflecti g ot o ly a employee’s pe so al expe ie ce of a pa ticula o ga izatio al cha ge, but also thei u ique huma eactio to that cha ge.
The fi di gs suggest that today’s f o tie models a e effective at ide tifyi g clea ly defi ed wo kplace issues but emai less eliable whe huma judgme t is eeded to dete mi e what employees mea .
Gemi i-3.5-flash led the full be chma k with a ove all sco e of 76%, the highest of the seve models tested, though which model pe fo med best shifted depe di g o the task. O st uctu ed, qua titative tasks, seve al models sco ed withi a few poi ts of each othe , with Gemi i-3.5-flash, GPT-5, a d G ok-4 tied at the top. O i te p etive tasks equi i g ua ced huma judgme t, the field sp ead out fu the , a d leade ship shifted agai by specific capability, with diffe e t models leadi g o et ieval, calculatio , a d sy thesis. No si gle model was the st o gest pe fo me ac oss eve y dime sio measu ed.
Eve whe models could eliably et ieve the ight i fo matio a d ide tify eleva t themes, they st uggled specifically with sy thesis—pulli g scatte ed, ambiguous sig als togethe i to a si gle, cohe e t i te p etatio . Sy thesis was the lowest-sco i g capability ac oss all seve models, with sco es a gi g f om just 14% to 57%, a wide gap tha a y othe capability measu ed.
The b eakdow happe s specifically whe they must weigh i complete, emotio al, o co text-depe de t sig als a d esolve them i to o e clea takeaway.
Ac oss evaluated espo ses, PYX Labs ide tified a e but mea i gful i sta ces whe e models p oduced fab icated statistical outputs o failed to adhe e st ictly to u de lyi g dataset co st ai ts.
While i f eque t, these e o s highlight the impo ta ce of validatio a d ove sight whe AI-ge e ated i sights a e used i wo kplace decisio -maki g co texts, whe e outputs may i flue ce actio s affecti g employees.
The fi di gs highlight a key limitatio i cu e t AI models a d u de sco e the eed fo clea e sta da ds i evaluati g how they i te p et huma behavio i wo kplace co texts.
“O ga izatio s a e al eady usi g AI to i te p et employee feedback a d ge e ate ecomme datio s that i flue ce eal decisio s about people,” said Joseph F eed, Chief P oduct Office at Pe ceptyx a d Head of PYX Labs. “The questio is ot whethe these models ca p oduce flue t a swe s—it’s whethe they u de sta d what ‘good’ looks like i the co text of the wo kplace. I ou view, ‘good’ mea s g ou ded i behavio al scie ce, co siste t with how employees actually expe ie ce wo k, a d eliable e ough to suppo t decisio s that affect ca ee s, teams, a d o ga izatio al t ust. This be chma k is the fi st step i ide tifyi g whe e models fall sho t today, a d whe e ta geted post-t ai i g, evaluatio , a d expe t-guided efi eme t ca imp ove thei eliability i these domai s.”
PYX Labs was established to defi e the evaluatio sta da d fo how AI systems should i te p et a d easo about people i the wo kplace, helpi g o ga izatio s deploy AI mo e espo sibly while helpi g model develope s imp ove pe fo ma ce th ough expe t evaluatio a d post-t ai i g.
Melissa Vale ti e, P ofesso of Ma ageme t Scie ce at Sta fo d U ive sity a d Se io Fellow at the Sta fo d I stitute fo Huma -Ce te ed AI (HAI), who is advisi g PYX Labs, said the wo k add esses a c itical gap i how AI systems a e evaluated fo wo kplace use. “What makes PYX Labs’ app oach disti ctive is the atte tio paid to defi i g what ‘good’ looks like whe AI i te p ets the huma expe ie ce at wo k,” said Vale ti e. “Most be chma ks measu e whethe a AI ca complete a task. This wo k asks a ha de a d mo e impo ta t questio : whethe AI is applyi g the ight values a d expe tise whe evaluati g that task. The wo kplace is o e of the most co seque tial domai s fo AI to get ight, a d wo k like this is what the field eeds to move f om capability to t ustwo thi ess.”
A joi t PYX Labs a d Sta fo d HAI webi a will take place Wed esday, August 5 at 12:00 pm ET, b i gi g togethe esea che s a d p actitio e s to discuss the fi di gs a d implicatio s fo AI use i wo kplace decisio -maki g. Lea mo e a d egiste he e.
PYX Labs is a esea ch lab spo so ed by Pe ceptyx focused o defi i g evaluatio sta da ds fo how AI systems u de sta d, i te p et a d easo about people i the wo kplace. The lab leve ages p op ieta y employee expe ie ce datasets, expe tise i behavio al scie ce a d o ga izatio al psychology, a d st uctu ed evaluatio methodologies to assess AI pe fo ma ce i eal-wo ld o ga izatio al co texts. PYX-Voice is the fi st i a pla ed se ies of be chma ks desig ed to help the AI i dust y build systems that a e ot o ly capable, but t ustwo thy i thei u de sta di g of employees.
O ga izatio s a d AI labs ca lea mo e about PYX Labs a d the methodology behi d the be chma k eleased today at: www.pyxlabs.ai
Lau a Lomba diGlobal Head of Commu icatio sPe ceptyxllomba di@pe ceptyx.com
A photo accompa yi g this a ou ceme t is available at https://www.globe ewswi e.com/NewsRoom/Attachme tNg/4a41a1a4-b745-4aa3-b692-7522403925b4
