The expe ime t tested popula AI models, i cludi g those f om Ope AI, A th opic a d Google, alo gside huma pa ticipa ts i a coope ative platfo m game whe e teammates we e co ected by a vi tual ope a d had to wo k togethe to complete challe ges. Rathe tha measu i g how well AI could solve p oblems i depe de tly, the expe ime t exami ed a mo e p actical questio : Ca AI effectively collabo ate with people, adapt to thei behavio a d help teams achieve bette outcomes?
The expe ime t su faced th ee key fi di gs about huma -AI collabo atio :
- Huma s d amatically outpe fo m AI at collabo atio . Huma -o ly teams achieved a 84% wi ate, compa ed with just 14% whe diffe e t AI models wo ked togethe . This sixfold diffe e ce highlights a sig ifica t gap i AI’s ability to coo di ate actio s, pla togethe a d achieve sha ed objectives.
- Huma leade ship a d collabo atio styles make a measu able diffe e ce. Teams co sisti g of two huma s a d o e AI age t pe fo med bette whe a huma di ected the team’s st ategy tha whe the AI age t led. The fi di gs also suggest that how teammates commu icate, coo di ate a d divide espo sibilities ca i flue ce outcomes, u de sco i g the impo ta ce of collabo atio st ategies, ot just i dividual i tellige ce.
- Mo e AI age ts do ‘t ecessa ily c eate bette teams. AI wi ates decli ed as teams expa ded f om two age ts to th ee a d fou , highlighti g how additio al age ts ca i t oduce commu icatio a d coo di atio challe ges athe tha automatically imp ovi g esults.
Chetha Ramacha d a , co-fou de a d CEO of Skillp i t, said: “AI is getti g ema kably good at solvi g p oblems o its ow , but that’s ot the same as bei g a good teammate. What we fou d is that collabo atio is a fu dame tally diffe e t capability. Who leads, how teammates commu icate a d how they adapt to o e a othe ca d amatically cha ge the outcome. As compa ies i t oduce mo e AI age ts i to the wo kplace, the questio should ‘t just be how i tellige t those systems a e, but whethe they actually help people wo k bette togethe .”
Skillp i t co ducted ove 900 u s of the expe ime t i a eight-level coope ative platfo m game desig ed to measu e commu icatio , pla i g a d coo di atio . The expe ime t i cluded 800 AI-o ly games, 128 huma -AI games a d 25 huma -o ly games, evaluati g popula AI models i cludi g Claude Fable 5.1, GPT-6 Ast a, Gemi i 3.8 Flash a d Kimi K3.
Skillp i t compa ed completio ates ac oss diffe e t team co figu atio s, fi di g a 84% wi ate fo huma -o ly teams ve sus 14% fo AI-o ly teams. The expe ime t also exami ed how diffe e t leade ship app oaches a d collabo atio styles i flue ced team pe fo ma ce.
Playe s avigated obstacles, solved puzzles a d wo ked togethe to each the fi ish, with teammates co ected by a vi tual ope that equi ed coo di ated moveme t. Skillp i t t acked game p og ess, esets, falls, commu icatio a d ove all sco es, compa i g pe fo ma ce ac oss diffe e t team co figu atio s a d leade ship app oaches.
Beyo d completio ates, the expe ime t fou d that huma s collabo ati g with AI waited lo ge , commu icated mo e f eque tly a d expe ie ced mo e falls tha huma -o ly teams. These behavio al diffe e ces offe additio al i sight i to how AI teammates ca i flue ce huma pe fo ma ce, eve whe teams a e ultimately able to complete the same objectives.
As AI evolves f om a swe i g questio s to taki g actio alo gside employees, evaluati g i dividual model i tellige ce is o lo ge e ough. O ga izatio s i c easi gly eed to u de sta d whethe AI age ts ca commu icate effectively, coo di ate with colleagues, adapt to huma behavio a d co t ibute to bette collective outcomes.
T aditio al AI be chma ks typically measu e how accu ately o efficie tly a model completes a defi ed task. Skillp i t Sig al exte ds that app oach by exami i g what happe s whe AI must collabo ate with people i dy amic e vi o me ts whe e success depe ds o sha ed decisio s, commu icatio a d adaptability.
Si ce 2020, Skillp i t has co ducted mo e tha 800 game-based cog itive studies ac oss mo e tha 500,000 playe sessio s a d mo e tha 1,000 games. Its platfo m maps mo e tha 150 game mecha ics to huma sig als spa i g cog itio , mood a d pe so ality, t a slati g gameplay behavio i to measu able i tellige ce about how people thi k, feel a d pe fo m.
Read the full expe ime t fi di gs a d methodology at https://expe ime ts.skillp i t.co/posts/sig al/.
Lea mo e about Skillp i t Sig al at https://skillp i t.co/ai-labs.
Skillp i t is a huma -pe fo ma ce i tellige ce compa y usi g gameplay to u de sta d a d e ha ce how people thi k, feel a d pe fo m. Si ce 2020, Skillp i t has combi ed cog itive scie ce, game mecha ics a d AI to t a slate eal gameplay i to measu able huma -pe fo ma ce sig als. Skillp i t’s missio is to help build a futu e whe e AI augme ts huma capability. Lea mo e at skillp i t.co.
MEDIA CONTACT
Kate Piet elli+1 760.518.2633kate@pathadviso s.com
View the o igi al p ess elease o ACCESS Newswi e
