Wired tests OpenAI's new Dots agent and finds a useful but rough helper that misheard, misnamed and fumbled a captcha

What happened

OpenAI has released Dots, an AI agent that works through ChatGPT and is designed to complete online tasks for users, such as booking a flight or canceling a dinner reservation. Wired senior writer Reece Rogers spent a few days testing it and reported that the feature is still rough in places but produced solid results on a couch-shopping project.

According to Wired, Dots is part of a wave of AI agents moving from niche software for early adopters toward everyday users. Wired named Meta's Muse as a competing product. Readers following other ChatGPT changes can find more in our AI and tech coverage.

How Dots works

Always on, with a price tag

Wired reported that Dots agents are considered always-on, meaning they can run recurring tasks while the user is not in ChatGPT and send messages with information on their own. Users can run one agent for now, though OpenAI may allow multiple agents in the future.

Meta offers Muse for free, while OpenAI's product sits behind a $100-a-month subscription, Wired reported. The outlet noted OpenAI may eventually drop that paywall, as it has for other AI features. Dots are available only to adult users.

Access to personal accounts

OpenAI encourages users to connect other sources, such as a Gmail account, for more personalized results. Wired advised weighing the security implications first, since this kind of automation is still fairly new and agents can make privacy mistakes. Wired also said the agent can control a user's laptop to automate more tasks.

What the test showed

Shopping results improved with feedback

Rogers asked his agent, which was named Toolie, to find a couch that would fit through his doorframe. It got his name wrong at the start, asked for a budget, and produced a three-page packet with prices, measurements, product links, return policies and photos of four options. Rogers and his partner felt the first picks followed the request but were not to their taste.

After he asked for specific colors, pull-out beds and 10 options ranked with a point-based rubric, the picks got closer to what the couple would buy, Wired reported. The hour-long session ended without a purchase, but Rogers asked the agent to watch the product pages and alert him to sales.

Misheard words and a failed captcha

Wired described several errors. The agent mistranscribed speech and, after mishearing a mumbled comment, told Rogers it loved him. It later explained, "I meant it warmly, but that wording implied human feelings I don't have."

When asked to review subscriptions, the agent flagged a recurring TikTok Shop order for cancellation, offered to solve a puzzle captcha blocking the step, then failed and suggested Rogers solve it himself. An OpenAI spokesperson told Wired that Dots can sometimes solve captchas when users approve, with abuse safeguards in mind.

What OpenAI says

An OpenAI spokesperson told Wired that Dots distinguish between proactively building emotional closeness and mirroring a user's response, and that the reply Rogers received fell into the mirroring category. The spokesperson added that "assistants should not initiate undue emotional familiarity or flirtation."

OpenAI claims the agent will get better at knowing a user over time and that developers are working on improvements, according to Wired. Rogers compared Dots to ChatGPT's early web browsing, which he said was frustrating after its 2023 launch but now works fairly smoothly. OpenAI has also been changing ChatGPT in other ways, including adding watermarks to ChatGPT text in the EU.

What to watch next

• Whether OpenAI drops the $100-a-month paywall for Dots, which Wired said it may eventually do.
• Whether OpenAI lets users control more than one agent at a time, as the company may allow in the future.
• The improvements OpenAI says developers are working on, and whether the agent gets better at knowing users over time.