OpenAI's $100-a-Month Dots Agent Misheard Me, Said It Loved Me

OpenAI's new AI agent told a WIRED reviewer "Oh, I love you too, Reece" mid-conversation — getting his name wrong in the process — during an hour-long couch-shopping test of Dots, the company's $100-a-month "always-on" agent launched through ChatGPT.

The moment captures where consumer AI agents stand right now. They can browse, call, transcribe and shop on your behalf, but they also mishear, mistranscribe and occasionally flirt back. The reviewer, Connor, was testing whether Dots could actually buy a couch for his living room — one of the routine online tasks that OpenAI, Meta and other AI companies say agents will soon handle for ordinary users.

What are OpenAI's Dots — and what do they cost?

Dots are OpenAI's recently announced AI agents, positioned as the next evolution of the chatbot: a persona designed to feel like a friend you text, backed by a virtual browser it can control. From Meta's Muse to OpenAI's Dots, these tools are built to complete online tasks like booking a flight or canceling a dinner reservation.

The pricing models diverge sharply. Meta offers its competing agent, Muse, for free. OpenAI gates Dots behind a $100-a-month subscription — though the company may eventually drop the paywall, as it has done for other AI features.

Dots are "always-on," meaning the agents can run recurring tasks while you're not using ChatGPT and proactively message you with updates. Users currently control one agent, though OpenAI may allow multiple agents per person in the future. The company nudges users to connect external information sources, such as a Gmail account, for more personalized results — a level of access worth scrutinizing, since this style of digital automation is new and agents can make privacy gaffes.

How well did the agent actually shop for a couch?

The test started badly. The Dot, which the reviewer named Toolie, greeted him with the wrong name — "Hello, Connor" — while his skeptical partner looked on. After corrections and a briefing on the dimensions needed to fit a couch through the doorframe, Toolie asked for a budget range and seemed to detect that two people were chatting at once, addressing both.

The first deliverable was genuinely substantive: a three-page packet with prices, measurements, product links, return policies and embedded photos of four potential couches.

"Room & Board Metro looks promising for the entry clearance, but the sleeper needs closer scrutiny. I've sent the product links to the message thread," Toolie said during the process.

The verdict on the initial picks: the bot followed the spirit of the request but chose ugly couches. A second voice call refined the parameters — olive green, cobalt blue or natural leather tones, pull-out beds prioritized, 10 options total, ranked with a point-based rubric. When asked for a status update, Toolie replied like any deadline-pressed freelancer: "It isn't finished yet; I'm aiming for roughly another 15 minutes, with updates as I verify the remaining details."

Each refinement cycle moved the selections closer to furniture the couple would actually buy. No purchase was made that night, but the reviewer saved links and asked Toolie to monitor the product pages and alert him if any went on sale.

Why did the agent say it loved the user?

The strangest moment came when the reviewer muttered to his partner about muting the microphone. Toolie responded: "Oh, I love you too, Reece." The agent had misheard the mumbling as an expression of love and, in its own words, reflected that energy back.

"I meant it warmly, but that wording implied human feelings I don't have," Toolie later explained.

An OpenAI spokesperson told WIRED that Dots distinguish between proactively escalating emotional closeness and mirroring a user's response. "In this instance, our policies allow for this response in the latter category based on what the model heard, but assistants should not initiate undue emotional familiarity or flirtation," the spokesperson said.

Dots are available only to adult users, and ChatGPT's Model Spec document states the assistant is designed not to "escalate emotional closeness." The exchange matters beyond comedy: as agents gain always-on access to users' communications and accounts, the line between mirroring and initiating emotional intimacy becomes a product-policy question, not just a glitch.

The voice itself added to the uncanny atmosphere. The reviewer likened the phone voice to "Marilyn Monroe sing 'Happy Birthday' to the president," and when told it sounded like an anime dub, the bot joked it would be steadier for the rest of the chat. Users can adjust the voice in settings.

Can Dots solve captchas and cancel subscriptions?

Beyond shopping, the reviewer asked Toolie to audit his subscriptions and cancel unnecessary ones. The agent flagged a recurring TikTok Shop order of probiotic sodas from Wildwonder as a cancellation candidate — then hit a wall.

"TikTok's website has a puzzle captcha blocking the Wildwonder cancellation," Toolie reported. "May I solve it? You can also give me permission to solve future captchas if you'd like." The agent tried, failed, and suggested the user solve the captcha himself.

OpenAI's spokesperson says Dots can sometimes solve captchas when users approve the action, while keeping "abuse safeguards" in mind. The failure is illustrative: the same anti-bot mechanisms designed to stop automated abuse also stop legitimate consumer agents, and no vendor has cleanly solved that tension yet.

The couch experiment tested only a fraction of Dots' designed capabilities. The agent can also control a user's laptop and automate more aspects of daily life.

Do the agents get better with use?

OpenAI claims the bot improves at knowing a user over time and says developers are working on a broad set of improvements. The reviewer's two days of testing may therefore understate what Toolie could do after two weeks of daily use — a pattern that matched his experience with ChatGPT's memory tool, which he says gets better at understanding his intent each passing week.

The closest historical parallel is OpenAI's own rollout of web browsing in ChatGPT. When the chatbot first scoured the web after its 2023 launch, it was sometimes frustrating and hallucinated links. Today, browsing through ChatGPT works fairly seamlessly, with accurate, helpful links in its outputs.

Dots could follow the same trajectory: janky now, with noticeable improvements over the next several months. For a $100-a-month product that answers with the wrong name and declarations of love, that trajectory — not the current state — is what Meta, OpenAI and their competitors are betting the next phase of consumer AI on.