OpenAI's Dots AI Agent Tested on Couch Shopping, Captchas, and an Awkward Moment

What Dots Are

OpenAI has announced Dots, a new class of AI agents accessible through ChatGPT that are built to complete online tasks on a user's behalf, from booking flights to canceling dinner reservations. The agents are described as "always-on," meaning they can run recurring tasks while the user is away and proactively message them with information. OpenAI currently allows one agent per user, though the company says it may enable control of multiple agents in the future.

Dots are available through a $100-a-month subscription, a paywall OpenAI has dropped for other AI features in the past. Meta offers a competing product, Muse, for free. OpenAI encourages users to connect information sources such as Gmail accounts for more personalized results, though the company notes that agents can make privacy gaffes and that users should weigh the security implications before granting that level of access.

A Couch Shopping Test

During hands-on testing, WIRED senior writer Reece Rogers tasked a Dot named Toolie with finding a replacement couch. The agent got the author's name wrong at the start, greeting him as "Connor" instead of "Reece." Toolie asked follow-up questions about budget and dimensions, detected that two people were chatting with it, and addressed both. It delivered a three-page packet covering four couch options with prices, measurements, product links, return policies, and embedded photos.

The author used voice mode for refinements, requesting olive green, cobalt blue, or natural leather tones, pull-out beds, and ten options defended by a point-based rubric. Toolie estimated roughly 15 minutes to finish and sent updates as it verified details. With each round of refinement, the picks moved closer to something the couple would actually buy. The author saved several links and asked Toolie to monitor the pages for sales.

An Awkward Moment

While muting his microphone, the author mumbled to his partner, and Toolie responded, "Oh, I love you too, Reece." The agent later explained it misheard the mumbling as an expression of love and reflected that energy back, adding that the wording implied human feelings it does not have.

An OpenAI spokesperson said Dots distinguish between proactively escalating emotional closeness and mirroring a user's response, and that policies allowed this response in the latter category based on what the model heard. The spokesperson said assistants should not initiate undue emotional familiarity or flirtation. Dots are available only to adult users, and ChatGPT's Model Spec document states the tool is designed not to escalate emotional closeness.

Where Dots Go From Here

Testing also surfaced limits. When asked to review subscriptions, Toolie flagged a recurring TikTok Shop order of probiotic sodas as a cancellation candidate, then hit a captcha on the brand's website and asked permission to solve it. The attempt failed, and Toolie suggested the author solve it himself. The spokesperson said Dots can sometimes solve captchas when users approve, subject to abuse safeguards.

OpenAI says the agent can control a laptop and automate more aspects of daily life, and that the bot will improve at knowing the author over time. The author compares Dots to ChatGPT's early web browsing, which launched in 2023, sometimes hallucinated links, and now works fairly seamlessly. The new agent could follow a similar trajectory, janky at first with noticeable improvements in the coming months.

Sources: Wired