[ed. Expectations of privacy in communications are dead.]
Hill saw postal delivery as wasteful and expensive. The price varied, depending on distance and number of sheets, and it was paid by the recipient, who could just refuse to accept. Hill began collecting sad stories of failure: urgent letters sitting for weeks at the postmaster’s, parents pawning clothes to pay for a letter from a child. Instead, Hill proposed, why not have the sender pay in advance, a flat rate, with a stamp (one penny). Parliament liked the idea and this is how the mail has worked since January 10, 1840 (though the price of stamps has naturally gone up).
As Hill predicted, everyone started sending letters, now that it was affordable. Mail volume doubled in the first year, from about 76 million to about 169 million. Volume passed 300 million by 1850 and kept climbing for a century and a half, peaking in 2004–05 at about twenty billion a year. It is about seven billion now. The US adopted pre-paid stamped mail in 1847.
More importantly your mail was now private. Under the old system, clerks counted sheets to figure the price, which meant handling and close inspection. An envelope counted as a sheet of paper, so most people folded their letters and sealed them with wax. In 1845 Hill’s brother Edwin, with a stationer named Warren De La Rue, patented an envelope-folding machine. By the end of the decade in the US and the UK everyone was using envelopes (which are officially called “covers”).
The envelope, or cover, gave you the presumption of privacy, legally. In the US, the Post Office Act of 1792 had already forbidden postal officials to open letters except when they couldn’t be delivered. In 1878 the Supreme Court included the sealed letter under the Fourth Amendment. Ex parte Jackson held that letters and sealed packages could only be opened with a warrant. [...]
Everyone sent email like they sent mail: candidly, embarrassingly, filled with gossip, complaints, things a spouse isn’t supposed to know about, dirty jokes. The privacy of email seems to be just as sacred. Congress had extended wiretap protection to electronic mail early, in 1986, on the theory that a message in transit was like a letter in transit. A system administrator could read your mail but was not supposed to. Such invasions of privacy became firing and disciplinary offenses.
Email feels personal and private to people. It works 99% of the time, unless someone’s bulk email list is wrong. You don’t get email that isn’t yours. Your address is a string: local part, @, domain. Some people have had the same address now for decades.
But very early on came the junk mail problem. In August 2002 Paul Graham published “A Plan for Spam,” and within two years the presumption of email privacy ended. Most people have no idea about this history. Graham was mad about the sheer amount of spam. The idea was software to filter your email by word phrases to separate out the junk. At first the software was just for you to put on your computer. Within a year, email providers were putting the software on the channel. There was no big announcement. Getting rid of spam seemed worth it.
Enter Gmail in April 2004. Google was transparent about the idea they’d read your email to funnel you appropriate ads. Privacy groups were appalled. A California state senator, Liz Figueroa, introduced a bill to block it. European organizations asked the British and German governments to investigate. Google shrugged and said scanning your mail is what antivirus and anti-spam software already do. What are you complaining about? By 2013, defending a wiretap suit in federal court, Google argued that a person has no legitimate expectation of privacy in information voluntarily turned over to third parties. Having your email scanned is expected in the ordinary course of business.
And yet people kept sending emails as if they were private. In 2003, the Federal Energy Regulatory Commission made public roughly half a million internal Enron emails. Everyone just chatting away candidly like people in every organization do. The Sony hack in 2014 showed that nothing changed, as do the emails in the Epstein files: plain language, real names, embarrassing remarks. In the latter case, there are people who really should have known better.
A sent message exists in at least four places: the sender’s outbox, the sender’s server, the recipient’s server, and the recipient’s inbox. Corporate servers may hold mail for decades. Brokers must keep business email three to six years under SEC Rule 17a-4; federal agencies keep senior officials’ email permanently under the Federal Records Act. Once a lawsuit is anticipated, any deletion becomes an offense. Nothing in the postal system worked this way. Undeliverable mail went to Washington and eventually was burnt.
Everything changed again in November 2022 with ChatGPT, which saw a hundred million users within two months. Everyone was chatting just like on email, presuming privacy.
There’s no envelope with an LLM. There is no routing layer separate from interpretation; there’s no channel distinct from message. In this case the medium not only is the message it also eats the message, to turn it into training data.
When you’re prompting ChatGPT or Claude or DeepSeek there is no other party. Sender, carrier, addressee—the singularity is the point where three collapse into one. You are not sending a message through anything to anyone. The LLM is the addressee and you are desiring your message to be read and replied to.
On a July 2025 podcast Sam Altman said out loud that people talk about the most personal things in their lives to ChatGPT, that young people especially chat with it like it was a therapist or a life coach, asking for relationship advice. But unlike therapists, lawyers, or doctors, none of it is protected by privilege. Should there be AI privilege? [...]
Email feels personal and private to people. It works 99% of the time, unless someone’s bulk email list is wrong. You don’t get email that isn’t yours. Your address is a string: local part, @, domain. Some people have had the same address now for decades.
But very early on came the junk mail problem. In August 2002 Paul Graham published “A Plan for Spam,” and within two years the presumption of email privacy ended. Most people have no idea about this history. Graham was mad about the sheer amount of spam. The idea was software to filter your email by word phrases to separate out the junk. At first the software was just for you to put on your computer. Within a year, email providers were putting the software on the channel. There was no big announcement. Getting rid of spam seemed worth it.
Enter Gmail in April 2004. Google was transparent about the idea they’d read your email to funnel you appropriate ads. Privacy groups were appalled. A California state senator, Liz Figueroa, introduced a bill to block it. European organizations asked the British and German governments to investigate. Google shrugged and said scanning your mail is what antivirus and anti-spam software already do. What are you complaining about? By 2013, defending a wiretap suit in federal court, Google argued that a person has no legitimate expectation of privacy in information voluntarily turned over to third parties. Having your email scanned is expected in the ordinary course of business.
And yet people kept sending emails as if they were private. In 2003, the Federal Energy Regulatory Commission made public roughly half a million internal Enron emails. Everyone just chatting away candidly like people in every organization do. The Sony hack in 2014 showed that nothing changed, as do the emails in the Epstein files: plain language, real names, embarrassing remarks. In the latter case, there are people who really should have known better.
A sent message exists in at least four places: the sender’s outbox, the sender’s server, the recipient’s server, and the recipient’s inbox. Corporate servers may hold mail for decades. Brokers must keep business email three to six years under SEC Rule 17a-4; federal agencies keep senior officials’ email permanently under the Federal Records Act. Once a lawsuit is anticipated, any deletion becomes an offense. Nothing in the postal system worked this way. Undeliverable mail went to Washington and eventually was burnt.
Everything changed again in November 2022 with ChatGPT, which saw a hundred million users within two months. Everyone was chatting just like on email, presuming privacy.
There’s no envelope with an LLM. There is no routing layer separate from interpretation; there’s no channel distinct from message. In this case the medium not only is the message it also eats the message, to turn it into training data.
When you’re prompting ChatGPT or Claude or DeepSeek there is no other party. Sender, carrier, addressee—the singularity is the point where three collapse into one. You are not sending a message through anything to anyone. The LLM is the addressee and you are desiring your message to be read and replied to.
On a July 2025 podcast Sam Altman said out loud that people talk about the most personal things in their lives to ChatGPT, that young people especially chat with it like it was a therapist or a life coach, asking for relationship advice. But unlike therapists, lawyers, or doctors, none of it is protected by privilege. Should there be AI privilege? [...]
Interpretation is the service with LLMs, not delivery. The model cannot not read and digest everything you send. That is the whole point. (And this is why I am deliberately impersonal in my extensive LLM engagement.)
And now agents. The postal system, with its categories of sender, carrier, recipient, never anticipated this. As Tyler Cowen and Sonia Farrell Pearson point out in their recent piece “Capitalizing Untethered AI Agents,” “the thing deciding and the thing being blamed come apart.” You authorized the agent to complete a task but you may not have authorized it to write a particular letter. Were a court to compel disclosure of your agent’s communication, who is the author?
Your agent may have contracted with vendors who did not know they were contracting with an agent. The recipient has no way to know who sent the message.
Is your agent a third party? If it is, then nothing is private per Smith v. Maryland. You have no reasonable expectation of privacy. If it is not, it’s a tool, an extension of you, like your pen or your hard drive. Giving instructions to your agent is like talking to yourself. Of course the “tool” runs on OpenAI’s or Anthropic’s servers…
by Hollis French, Anecdotal Value | Read more:
Image: via
[ed. And Edward Snowden is still exiled in Russia.]
