I Thought I Was Emailing a Bot. It Was a Person the Whole Time.

I spent an entire email exchange this week convinced I was talking to a bot. Two replies, six sentences total, and I would have bet money there was an AI sitting in that inbox. I was wrong (good thing nobody took the bet). It was a person the whole time, and the thing that finally proved it was, of course, an AI.

Lets back up.

I emailed a listing agent a simple question about a property I was looking at for a client. The kind of question you answer in one line if you know the answer and two lines if you dont.

What came back did not read like a person. Short to the point of rude, grammar broken, and it did not answer what I asked. So I clarified. The second reply came faster and was just as strange. By then I had made up my mind. Somebody plugged a bot into their inbox and turned it loose.

So I asked an AI whether I was talking to an AI

Afterward I dumped the thread into Claude and asked what it thought. Human or machine. Four seconds later: human, definitely human. Then it walked me through why, and every reason was backwards from what I assumed.

Bad grammar is a human tell, not a bot tell. AI writes clean. Tidy paragraphs, commas in the right places, polite to the point of being weird about it. A person typing with one thumb between showings sends a run on with two typos in it and never looks back. The mess was the evidence. Which is not how any of us have been trained to think about this, right.

He misread my question, then caught himself. The first reply answered something I did not ask. The second opened with a version of “I see now.” That is a distracted human skimming at a stoplight. A language model parses the question right the first time. It is the one thing they are genuinely great at.

The formatting changed between the two replies. One quoted my message the way Gmail does it. The next had the Outlook style block with From and Date and To stacked up. Same person, two devices, one morning. Bots do not switch mail apps halfway through.

The timing was human, and so was the attitude. Two hours, then nine minutes (bots do not take lunch). And he was a little prickly, which sealed it. Bots do not get defensive, they get wordy. Push a model and it apologizes and hands you three more paragraphs. Push a person having a day and you get five words with an edge on them.

So I was wrong on every count. And I am not throwing this guy under the bus, he was working. I was the one running forensics on another man’s punctuation.

The part that actually bothers me

I do this for a living. I build with this stuff every day, I have written plenty about where it is going, and I still guessed wrong on six sentences. Kahneman’s whole thing is that your fast brain calls it before your slow brain gets a vote. Mine took one look at broken grammar and said machine.

And look, I am not standing outside this problem pointing at it, right. My own signature discloses that my emails may be drafted with AI assistance and reviewed by me before they go out, with the full policy on our AI disclosure page. On purpose. I would rather tell you up front than have you squinting at my commas. (I have also said before that I do not want a bot answering for me. Still true.)

But we are heading somewhere strange and getting there fast. Every email, every text, every inquiry that hits my phone now carries a small question mark on it. Is there a person back there. Most of the time it does not matter. Sometimes it matters a lot, like when you are negotiating on a house and you need to know whether the other side can actually change its mind.

So here is what I keep chewing on. If the tells I was sure about turned out backwards, what would you look for? The signals are moving, and the ones that work today probably stop working next year.

If you would rather just talk to a person

That is still how we do it here. A human reads it, a human answers, and when I use AI to help me draft something I tell you so. Reach out to me directly, get in touch with the team, or call the office at (512) 366-3270.

And if you figure out a reliable way to tell the difference, come tell me. I clearly need the help.