When World responded on X to news that an AI video model could pass for human on nearly half of live calls, it was making a narrow point about Zoom meetings. Read in context, it was also part of a much broader argument the company has been building all year: as software becomes indistinguishable from people, every corner of the internet will need a way to confirm that a real, unique human is present.
That argument now applies on two fronts at once. Some AI looks like people on camera. Other AI acts for people online, shopping, booking and filling out forms on their behalf.
Machines that pass for people on camera
The post followed a Tavus study in which nearly half of participants who spoke with its Griffin-Lite model on a live one-minute call believed they had been talking to a real person. Text-based AI crossed that line earlier: a UC San Diego study found GPT-4.5 was judged human 73% of the time in short text conversations.
Agents arrive in force
The video result landed in the middle of a busy month for consumer AI agents. Meta launched Muse on 8 September, an assistant that can send email, book travel, complete forms and make purchases. On 29 September, OpenAI introduced dots, persistent agents with their own cloud computers and connections to more than 4,000 apps. A smaller, invite-only service called Instinct works entirely through text messages.
Friction followed quickly. Amazon blocked Muse from shopping on its platform, stating that the agent failed to identify itself as AI as Amazon’s terms require. According to TechSpot, Amazon plans to apply the same rule to shopping agents from Google and OpenAI.
The pressure was already building. Imperva’s 2025 Bad Bot Report found that automated traffic made up 51% of all web activity in 2024, overtaking human traffic for the first time, with malicious bots accounting for 37%. Account-takeover attacks rose 40% that year. To a website, an agent using someone’s login can look almost identical to an attacker using stolen credentials.
One credential, two jobs
World expanded on the agent side of its argument in a 2 October blog post. The logic rests on scarcity. There are about eight billion people, but the number of agents is effectively limitless. Purchase limits, rate caps and fair access to limited releases only work if each agent can show that a distinct person authorised it.
World ID handles this through zero-knowledge proofs. A verified person can pass a proof to their agent, which lets a website confirm a unique human stands behind a request without learning who that human is. World describes the approach as private and neutral across agent providers.
Developers can access this through AgentKit, currently in beta, and several infrastructure companies are already integrating it. Okta is offering a Human Principal service in early access, Vercel is adding human approval steps to its Workflow SDK, Exa gives verified humans 100 free API calls per month, and Browserbase expects verified agents to encounter fewer website blocks.
On video, the same credential powers World ID for Zoom, the product World highlighted in its post, which matches a user’s original Orb photo, a fresh phone selfie and their live video feed before displaying a Verified Human badge.
What comes next
With nearly 18 million Orb-verified users and the April release of World ID 4.0 adding key rotation, account recovery and session management, the infrastructure is moving toward production use. If verification becomes as routine as the padlock icon in a browser, the platforms that adopt it early will be best placed to tell their real customers apart from software built to imitate them.
Featured image via Shutterstock.
