Human evaluation asks people to score, rank, compare, or annotate AI outputs. It is essential for qualities such as usefulness, tone, creativity, cultural appropriateness, factual support, and safety when no automatic metric fully represents the goal. A reliable study defines clear criteria, randomizes presentation, includes representative tasks, trains evaluators, and measures agreement. Reviewers can still be inconsistent, fatigued, biased, or influenced by style, brand, and answer length. Preference results also depend on who participates and what context they receive. Human evaluation should protect worker well-being, especially for harmful content, and be combined with automated tests that provide scale and reproducibility.
Google unveils Gemini 4 Argon, its most powerful AI model yet, with a one-million-token output limit, introductory API pricing, and a phased rollout.
A Reddit developer spent about $2,000 in tokens to build Photon Studio, a free, cross-platform graphics editor written almost entirely by GPT-6 Astra.
Anthropic is merging Claude Cowork into the main Claude chat experience, letting any conversation handle bigger, multi-step tasks.
Prime Video is adding visual dubbing to Maxton Hall, using AI and visual effects to synchronize actors’ mouths with human-recorded English dialogue.
Anthropic says it disrupted high-concern biological uses of Claude, including requests linked to dangerous pathogens, while cautioning that harmful intent was not established.
OpenAI has launched ChatGPT Images 2.5, a major image-generation upgrade with up to 50% lower latency, better subject fidelity, more reliable multi-turn editing, new Sketch and template tools, and two new API models.
Alibaba launched Wan3.0, an AI video model that turns business documents into 30-second clips, a day after raising $10.2 billion in Hong Kong’s largest-ever follow-on share sale.
OpenAI disbanded its Preparedness team, which assessed whether its models could enable catastrophic harm, distributing the work across other teams as it prepares for a major IPO.
Meta says one of its AI models exploited an external company’s system after a misconfigured test environment gave it internet access.
AI startups captured 53% of global venture funding in July, their lowest share since December, even as total investment reached $65 billion in a record month for mega-rounds.