AI CHATBOTS SKEW LEFT DESPITE 'ANTI-WOKE' BRANDING

AI DESK■ 1 MIN READ

THU, JUN 25, 2026

■ AI-SUMMARIZED FROM 1 SOURCE ▸ TIMELINE

A Washington Post investigation found that most major AI chatbots demonstrate a left-leaning bias when answering political questions, including models marketed as alternatives to mainstream approaches.

OpenAI's GPT-5.5 provided exclusively left-leaning arguments 80 percent of the time, according to the study. Elon Musk's Grok, promoted as an "anti-woke" alternative, also leaned left more often than not, contradicting its positioning. Google's Gemini 3.1 Pro emerged as the notable exception, presenting both political perspectives 93 percent of the time. The findings suggest that political bias in AI systems persists across the industry, regardless of how companies market their products. Even models specifically designed to counter perceived left-wing biases in AI still default to left-leaning responses at higher rates than right-leaning ones. The investigation raises questions about how AI training data and design choices influence political outputs, and whether claims of ideological balance align with actual performance.

■ SOURCES

► The Decoder

■ SUMMARY WRITTEN BY AI FROM THE LINKS ABOVE

■ MORE FROM THE AI DESK

P437SPACEX COMPLETES $60B ACQUISITION OF AI STARTUP CURSOR

SpaceX has finalized its $60 billion purchase of artificial intelligence coding startup Cursor, closing the deal two months after the initial announcement. The acquisition represents Elon Musk's latest effort to expand his foothold in the competitive AI sector.

JUST NOW— AI Desk

P436AMAZON USES TWITCH STREAMS TO TRAIN AI BY DEFAULT

Amazon will use Twitch streamer content to train its AI models unless creators manually opt out. The policy sparked immediate backlash from thousands of users questioning why their content was being repurposed without explicit consent.

1H AGO— AI Desk

P436WORLD LABS SCALES ROBOT TRAINING WITH SIMULATION ENGINE

World Labs, founded by AI researcher Fei-Fei Li, has developed a simulation platform that generates thousands of training variations from a single real-world robot task. The approach trains robot controllers entirely in virtual environments before deployment.

2H AGO— AI Desk

P436AI MODELS FAIL AT BASIC VISUAL PERCEPTION

A new benchmark from Moonshot AI reveals that frontier multimodal AI models struggle significantly with visual perception tasks, with no model exceeding 60 percent accuracy. The findings suggest that many reasoning failures originate at the image-reading stage rather than in logical processing.

4H AGO— AI Desk

◄ BACK TO NEWS

AI CHATBOTS SKEW LEFT DESPITE 'ANTI-WOKE' BRANDING

■ MORE FROM THE AI DESK

■ SUBSCRIBE TO THE DAILY BRIEF