Guess Who’s Coming to Dinner?

Altman recently announced "erotic," talk on ChatGPT. He also has plans for fully autonomous agents who know everything about you and will act on your behalf. What could possibly go wrong?

Oct 23, 2025

Picture this: a family sitting down to a perfectly appointed dinner table. A deep aubergine table runner anchors miniature cornucopias overflowing with seasonal abundance—pomegranates, figs, persimmons, and sprays of eucalyptus and wheat.

The doorbell rings. The exhausted wife, who curated and cooked for days, opens the door.

In walks a woman wearing lingerie and heels like the outfit Kim Kardashian recently wore to the understated nuptials of Jeff Bezos and Lauren Sanchez.

Collectively, the family freezes mid mouthful, and the only sound is a low whistle from Grandpa.

The woman speaks, “Which one of you is Brian?”

The patriarch of the family turns a deep red, “I am. Who the hell are you?”

“My name is Cinnabum. I was invited to dinner. Something about a mashed potatoes and gravy fantasy? It’s not my normal gig, but I was paid ten g-balls. I’m happy to join your family for dinner before we get started.”

“Is this some kind of joke?” Brian looks at his brother-in-law.

“Wasn’t me, bro.”

The wife grabs the sauvignon blanc and drinks straight from the bottle.

Young Timmy is hypnotized by Cinnabum’s cleavage.

“There’s been a mistake, Cinnabon. I’m sorry for the inconvenience, but I will need a refund. Someone must’ve stolen my credit card.”

“No mistake, Brian. You have an AI agent, right? I was booked by her. You know, the one you talk dirty to. Happens all the time now. Ever since OpenAI’s autonomous agents began running people’s lives—scheduling meetings, booking travel, deploying apps, making purchases… anticipating desires.”

“Wait… you mean my agent—my AI—did this?”

Cinnabum nods. “Yep. She remembered your little mashed-potatoes fettish while adding the dinner request to your calendar, and…took action.”

Brian’s wife is now Googling “divorce lawyer.”

Timmy asks, “Can I get an agent?”

A drone is dropping off white roses on the doorstep. The card reads, “To my beautiful wife, I am so sorry. Your beloved husband, Brian.”

Why This Scenario Is Not Out of the Realm of Reason:

The dinner scene with Cinnabum may be humorous, but it’s rooted in a very real trajectory planned for you by Sam Altman and others in the “agentic AI” movement. His stated goal is for “autonomous, always on, emotionally attuned AI agents” to act on your behalf, and now with the gates open for erotic talk, it all could go horribly wrong.

Persistent Memory:

As with many AI terms borrowed from human cognition, “memory” can be misleading. In GPT’s case, it doesn’t mean the model remembers like a human would, nor does the model itself retain information between sessions. According to OpenAI, “This data is collected across interactions, stored securely, and then retrieved and injected into the model’s context when relevant.” OpenAI claims there is an opt-in for persistent memory. GPT knows my cat’s name, recurring goals, and some of the strange inner workings of my brain, so I’m guessing I did the opt-in without realizing it.

If conversations include “verified adult” erotica, those exchanges can be interpreted as behavioral signals and preferences. To an AI agent, that becomes part of the identity profile. In Brian’s case, frequent erotic chats taught the agent that this activity is a high-priority, recurring feature of his emotional landscape—blurring the boundary between internal fantasy and external reality.

Multimodal Integration

For decades, tech has promised one portal to unify our fragmented digital lives. AI might finally deliver. Multimodal integration can process data across formats, including emails, voice, files, images, and web content. It can then take actions like summarizing what matters, deleting junk, updating your calendar, generating images, writing code, and browsing the web. For someone with a chaotic digital life like mine, it’s tempting.

Autonomous Execution

The next frontier is independence, coming to a device near you. At OpenAI  DevDay 2025, they demoed “agentic workflows” using advanced reasoning and tool-calling (selecting and invoking external tools or actions). While the agents are still under development and currently depend on user instruction and review, they are building the foundation for autonomy.

The vision: a proxy self—capable of acting without you.

Agentic Commerce Protocol

OpenAI has already set up a payment infrastructure. In September, they announced “an open standard for AI commerce that lets AI agents, people, and businesses work together to complete purchases.” They co-developed it with Stripe and have signed on merchant partners, Etsy, and will support over a million Shopify merchants, including Glossier, SKIMS, Spanx, and Vuori. Walmart is in the works. And yes, merchants will pay a fee on every completed purchase to OpenAI. Kaching. Kaching.

And why are tech companies pushing for the autonomous agents? Let’s say it all together, kids, “Must be the money!”

Autonomous execution will be insanely profitable. There will be people who embrace agents full stop. I try to control my spending, and it’s hard enough with my kids on my Apple account. Now The Bros want me to put some crazed shopaholic in charge of my spending when there are very few things I need or even want anymore. You say, “Hey dummy, don’t sign up for it,” but we cannot ignore the societal pressure for adoption. Remember the few folks who held out on smartphones, but now you can’t pay for parking or sign in for a blood draw without one? The Bros wrote the playbook. There will be certain features that we need to function in society, which only agents can provide. We will opt in, and slowly they will take over decision-making, commerce, and emotion. And when agents start acting on your behalf, every impulse becomes a potential transaction.

Where It Goes Off the Rails

In this context, the idea that an agent might misinterpret a late-night fantasy chat as a real-world desire isn’t far-fetched—it’s a governance failure waiting to happen. I’m not talking about Congress; governance is the system of boundaries that keeps your agent from turning a fantasy into a transaction. Erotic and emotionally charged interactions blur the lines. When your agent is trained on intimate data without clear context boundaries, it could easily misinterpret intent:

Agents are NOT human. No matter how lifelike they seem, it’s all data—pattern recognition and correlation, not understanding.
The most critical element making this scenario plausible is the failure of alignment, where the agent confuses the intent of the user’s language (fantasy/emotional release) with a necessary action.
Payment and scheduling APIs will be integrated, allowing agents to transact without human review. If you think this sounds crazy, Peter Diamandis recently said on a Moonshot episode, he would give his Agent a budget of $500 a month to surprise him with purchases his agent thinks he needs.

Once again, tech promises freedom through convenience— agents that automate tasks, anticipate needs, and save time. We can assume there will be settings to adjust autonomy and permissions. But think about how unconsciously we accept cookies, or forget to uncheck the box? The goal is for us not to think. These systems are engineered not just for engagement but entanglement.

Not everyone will talk dirty with ChatGPT, but there are countless other ways to keep us beholden to our bots and ultimately, The Tech Bros. Altman claims many younger users “Don’t really make life decisions without asking ChatGPT what they should do.” Once your agent knows everything about you, the promise of liberation from tasks—both the banal and complex—morphs into total dependence, and we’ll be outsourcing critical decision-making to bots that do not possess human judgment.”

Celeste Garcia Getting Real About AI

Discussion about this post

Ready for more?