An Address Anyone Can Write To

At the end of 2024 the Radicati Group, a firm that has counted email for more than two decades, estimated that the world sends and receives more than 361 billion emails a day, and expected the number to keep rising for years.¹ Close to half of that traffic, by Kaspersky’s measurement, is spam.² Few technologies are complained about so constantly. Critics have announced the death of email on behalf of instant messaging, social networks, Slack and now artificial intelligence. Each time, email has absorbed the newcomer’s ideas and carried on.

The more surprising fact is that no company owns it. Google runs the largest mail service, Apple’s software opens most of the marketing mail that is measured, and Microsoft runs the mail of a large share of the world’s offices. But none of them controls the thing itself. Anyone with a server and a domain name can send a message to any of the world’s email addresses, and that message will usually arrive. Nothing else in mainstream consumer technology works this way any more. You cannot send a WhatsApp message to an iMessage user, or a Slack message to someone on Microsoft Teams, without one of the companies agreeing to connect them.

That raises a question worth investigating. Why has a protocol designed in the 1970s for a few hundred cooperating researchers, a protocol that everyone finds tiring and that every large technology company has tried either to replace or to own, outlived nearly every product built on top of it? Following that question leads through a long series of decisions: what a message is, who may send one, who pays for it, who may read it in transit, and, since the arrival of machine learning, who writes it. The answer that emerges is that email’s strength and its burden come from the same design choice. It is an address anyone can write to.

Before email: the letter and the shared machine

People had been writing letters to each other for thousands of years before computers, and the letter had already developed most of the conventions email would borrow: a greeting, a signature, the expectation that a reply could wait, and the idea that a letter might be kept and reread. The Roman philosopher Seneca, writing to his friend Lucilius in the first century, treated letters as a way of being present to someone far away, a kind of conversation stretched across distance and time. His Letters show the form’s strengths: they were reflective because they were slow, and personal because each was addressed to one reader.

The postal systems that grew up in the nineteenth century added a second set of ideas that also passed into email. Britain’s Penny Post of 1840 set a single low price for a letter anywhere in the country, paid by the sender through a prepaid stamp. A letter became cheap enough for ordinary people, and the post office delivered it to an address: a public, stable identifier that anyone could write on an envelope without asking the recipient’s permission. Anyone could write to anyone. The modest price of a stamp kept the volume within bounds.

Electronic mail did not begin with that ambition. It began with a practical problem on a single, very expensive computer.

A message left in a shared machine

In the early 1960s, MIT’s Compatible Time-Sharing System, CTSS, let dozens of people use one IBM mainframe at the same time from typewriter terminals. They shared the machine’s disks as well as its processor, and that sharing created a need to leave notes. Tom Van Vleck, who was one of the system’s young programmers, later recorded that the immediate need was mundane: “CTSS Operations did need a facility to inform users when a request to retrieve a file from tape had been completed.” He and others argued that “a facility that let any user send arbitrary messages to any other would have more general uses.”³

A staff note written around the turn of 1965 by Pat Crisman, Glenda Schroeder and Louis Pouzin proposed such a facility. When Van Vleck and Noel Morris asked where the program was, they were told “there was nobody available to write it.” So, in Van Vleck’s words, “Noel and I wrote a version of MAIL for CTSS in the summer of 1965.” Their approach set the tone for much of what followed: the idea was “to get it working, and put features in later after community discussion.”³

Two features of this first system matter for the rest of the story. The first is that it was designed for a community of known users who mostly trusted each other. The second is that its authors already saw the danger of that trust. MAIL deliberately prevented ordinary users from sending a message to every user at once; Morris and Van Vleck “had foreseen the possibility of inappropriate mass mailing.” Van Vleck records that around 1971 a member of his own team got round the restriction and sent “a long anti-war message to every user of CTSS.” When challenged, the sender replied: “but this is important!”³ Every later argument about bulk email, from political campaigns to newsletters, has some version of that reply in it.

Across the network: the @ sign

Electronic mail on one machine connected people who already shared a computer. The ARPANET, the US Defense Department’s research network that began operating in 1969, connected the machines themselves. In late 1971 Ray Tomlinson, an engineer at the research firm Bolt Beranek and Newman, combined a program for sending messages on one computer, SNDMSG, with an experimental program for copying files across the network. The result delivered a message to a mailbox on another computer.⁴

To tell the program which user on which machine should receive it, Tomlinson needed a separator. He chose @, a character on his teletype keyboard that was unlikely to appear in anyone’s name and that read naturally as “user at host.”⁴ The decision looks trivial, but it fixed a model that has never changed. An email address names two things: a person, and the place where that person’s mail is kept. The person is defined locally, by whoever runs the host. The host is defined globally, by the network. No central directory of people was ever needed.

Email spread across the ARPANET faster than its designers expected. The network had been funded to share expensive computers; researchers discovered that what they most wanted to share was conversation. Mailing lists followed, and so did the first arguments about etiquette, privacy and volume. In 1977 a group of network researchers published RFC 733, a common standard for the format of messages, including the header fields that still begin every email: From, To, Cc, Subject, Date.⁵ Programs such as John Vittal’s MSG added the ability to answer a message with its headers filled in automatically. With that, email had acquired the basic shape of a correspondence.

The first spam

On 3 May 1978, a marketer at Digital Equipment Corporation, Gary Thuerk, had a colleague send an announcement of DEC’s new DECSYSTEM-20 computers to several hundred ARPANET users on the west coast, inviting them to product presentations. The message was typed in capital letters, and a limitation of the mail software meant that many of the addresses spilled out of the To field into the body.⁶

The network’s operators replied on 4 May. Major Raymond Czahor of the Defense Communications Agency wrote that “THIS WAS A FLAGRANT VIOLATION OF THE USE OF ARPANET AS THE NETWORK IS TO BE USED FOR OFFICIAL U.S. GOVERNMENT BUSINESS ONLY,” and that “APPROPRIATE ACTION IS BEING TAKEN TO PRECLUDE ITS OCCURRENCE AGAIN.”⁷ Among the recipients the debate was more interesting. Richard Stallman, then at MIT, first wrote that he was “in favor of seeing jobs offered via whatever,” and then, having seen the full message, added: “I sure would have minded it!”⁷

That exchange contains the whole problem in small form. The network had no technical means of stopping an unwanted message. It relied on a social rule (official business only) enforced by an administrator who could identify the sender, and on the willingness of a small community to shame anyone who broke the rule. Both conditions would disappear when the network opened to the world.

The protocol that trusted strangers

In August 1982 Jon Postel published RFC 821, the Simple Mail Transfer Protocol, and David Crocker published RFC 822, the format of internet text messages. Postel’s stated objective was modest: “to transfer mail reliably and efficiently.”⁸ The design, which with extensions still carries almost all of the world’s email, made several decisions that shaped everything after it.

Store and forward

A message is passed from server to server, each one taking responsibility for it, until it reaches the recipient’s server, where it waits. Neither the sender nor the recipient needs to be online at the same time. This made email asynchronous like the letter, and resilient on a network whose links often failed.

Federation

There is no central mail company. Each organisation runs, or rents, its own mail server, and the servers find each other through the Domain Name System. From 1986, a special DNS record, the MX record, told the world which machine accepted mail for a given domain.⁹ Anyone could join by setting up a server.

Trust by default

An SMTP server accepts a message from any other server that connects to it, and believes what the message says about who sent it. The protocol had no step in which the sender proved its identity. On a research network of known institutions, checking identity seemed unnecessary, and building it in would have made the system harder to deploy.

This last choice reflected a principle Postel had stated two years earlier in the specification of TCP, the protocol on which most of the internet runs: “be conservative in what you do, be liberal in what you accept from others.”¹⁰ It became known as the robustness principle. Accept whatever arrives, interpret it generously, and the network will keep working even when implementations differ. Email took the principle further than almost any other system. It accepted any message, from anyone, about anyone.

The decisions made email the only communication system that works across every organisation in the world. They also guaranteed that the cost of sending a message would fall close to zero, and that nothing in the protocol itself could tell a friend from a stranger, or a stranger from a forger.

The walled gardens and the open network

Through the 1980s most people who used electronic mail did not use the internet’s version. Universities and laboratories had ARPANET and its successors. Businesses bought proprietary systems: IBM’s PROFS on mainframes, Lotus cc:Mail and Microsoft Mail on office networks, and later Lotus Notes, which combined mail with shared databases. Consumers and small businesses used commercial online services. MCI Mail, launched in 1983, charged per message; CompuServe, Prodigy and above all America Online sold access to closed communities that included mail.

These were coherent business decisions. A closed system could be sold, billed, supported and controlled. It could guarantee that a message was really from the account it claimed to be from, because the company knew every account. But each garden could reach only its own members. The services built gateways to each other and to the internet, and as more people joined the internet the value of the gateways began to exceed the value of the gardens. By the mid-1990s internet email had become the common language, and the proprietary systems either adopted internet standards or faded. AOL kept its own service and its famous greeting, but its members wrote to internet addresses.

The lesson that the industry took from this period is an old one in network economics. A communication system becomes more valuable as more people can be reached through it. An open protocol that anyone could implement had more potential members than any single company could ever sign up, and that advantage outweighed the commercial strengths of the closed systems.

How mail reached the desk

Federation answered the question of how mail moved between organisations. It did not say how a person should read it. Early users logged into the computer where their mailbox lived. Personal computers changed that, and two protocols offered two philosophies.

The Post Office Protocol, POP, first specified in 1984, let a personal computer collect mail from a server and, normally, delete it there.¹¹ Your mail lived on your own machine. The Internet Message Access Protocol, IMAP, developed by Mark Crispin at Stanford and first published in 1988, kept mail on the server and let many devices view and organise it.¹² Your mail lived on the network, and your computer was a window onto it.

For years POP was the more common choice, because disk space on servers was expensive and connections were slow. Desktop programs flourished. Eudora, which Steve Dorner began writing at the University of Illinois in 1988, became one of the most loved email programs of its era after Qualcomm took it on.¹³ IMAP’s philosophy, though, anticipated the future. The question of whether a person’s mail lives with them or with a service would return with webmail and again with the cloud.

In 1992 Nathaniel Borenstein and Ned Freed published MIME, a set of extensions that let email carry pictures, sound, non-English text and attached files.¹⁴ Borenstein sent one of the first MIME messages on 11 March 1992. It contained a photograph and a recording of the Telephone Chords, a barbershop quartet of Bellcore employees, of which Borenstein was the tenor.¹⁵ With attachments, email became the default way to send documents between organisations, a role it still plays.

Free mail and the portable address

Until the mid-1990s an email address belonged to an institution. It came with a job, a university place or a subscription to an internet provider. If you left, you lost it.

On 4 July 1996, Sabeer Bhatia and Jack Smith, two former Apple engineers, launched Hotmail, a free email service that anyone could use from a web browser. The idea removed two dependencies at once. A user no longer needed a mail program, because the browser was enough, and no longer needed an employer or a provider, because the address belonged to the service and the service was free.¹⁶ The service was paid for by advertising. Each message sent from Hotmail carried a short line at the bottom inviting the recipient to get a free account, so every user advertised the product to everyone they wrote to. Within about eighteen months Hotmail had millions of users, and in December 1997 Microsoft bought it in a stock deal reported at around $400 million.¹⁶

Yahoo launched its own webmail the same year, and many others followed. The business model shaped the product. Storage was expensive, so free accounts came with small quotas, a few megabytes at most, and users spent time deleting mail to stay under the limit. Advertising paid only for attention, so services competed to keep users returning to their pages. The address had become portable between employers, but it was now bound to a company whose interests were in advertising.

Mail in the pocket

In January 1999 Research In Motion, a Canadian company, released the BlackBerry 850, a two-way pager with a keyboard that could receive email.¹⁷ Its significance lay in a decision about who initiates. Most email programs pulled mail by checking the server at intervals. BlackBerry’s system pushed it: a server inside the company forwarded each new message to the device, which vibrated when it arrived.

The effect on working life was profound. Email ceased to be something done at a desk and became something that arrived wherever a person was. BlackBerry built a business around corporate email, with strong security, efficient use of thin wireless networks and a physical keyboard designed for writing replies. When the iPhone arrived in 2007 and Android after it, the phone became a general-purpose computer, and email became one application among many. RIM’s market share fell steeply after 2010.¹⁷ The decision that made BlackBerry successful, designing a device around email, became a constraint once people wanted a device designed around everything else. The habit BlackBerry created, checking mail at every spare moment, outlived it.

Gmail: storage, search and the scanned message

On 1 April 2004, Google announced Gmail. Because of the date, many readers assumed it was a joke. The service offered one gigabyte of storage, at a time when the large free services offered a few megabytes.¹⁸ Google described it as a free, “search-based” webmail service. The company’s own account of its origin was that a user had complained about spending her time filing messages or trying to find them.¹⁹

Paul Buchheit, the engineer who led the project, had begun building it as an internal tool. Its design made a series of bets.

Storage instead of deletion

If storage is effectively unlimited, nobody needs to delete anything. Google’s infrastructure, built on large numbers of cheap commodity machines, let it offer storage at a cost its rivals could not easily match.

Search instead of sorting

Rather than asking users to file mail into folders, Gmail let them find any message by searching, and replaced folders with labels, so that a message could belong to several categories at once.

Conversations instead of messages

Gmail grouped a message with its replies into a single thread, so a discussion could be read as a whole.

Ads chosen by reading mail

Gmail paid for itself with text advertisements matched to the content of the message being read. Software scanned each message to choose the ads.

The fourth bet drew immediate objections from privacy groups, who argued that automated reading of private correspondence set a troubling precedent. Google replied that no person read the mail and that the process was similar to the scanning used to filter spam and viruses. Both statements described the same technology; the disagreement was about what the scanning meant. The question would return, more sharply, with artificial intelligence.

Google limited sign-ups through invitations at first, which made accounts scarce and desirable. Rivals responded by raising their own storage limits sharply, and Gmail’s gigabyte reset what users expected from every service. Gmail opened to everyone in 2007 and dropped its “beta” label in 2009. In October 2012, comScore’s desktop measurements showed Gmail narrowly passing Hotmail as the world’s largest email service, with about 288 million monthly visitors to Hotmail’s 286 million and Yahoo Mail’s 282 million; in the United States, Yahoo Mail still led.²⁰ In 2015 Google said Gmail had 900 million users, and in early 2016 more than a billion monthly active users.²¹ A figure of about 1.8 billion active accounts has been widely reported since around 2020, though Google has not published a recent official count.²²

In June 2017, Google announced that it would stop scanning consumer Gmail to personalise ads. Diane Greene, who ran Google Cloud, wrote that the change brought “Gmail ads in line with how we personalize ads for other Google products.”²³ Reporting at the time connected the decision to Google’s effort to sell its paid business service, G Suite, whose mail had never been scanned for ads; corporate customers had found the distinction confusing. Gmail continued to show ads, chosen from other signals. The scanning that filtered spam and malware continued, and it was about to become far more capable.

The spam war

When the internet opened to commercial use in the early 1990s, the ARPANET’s social rule disappeared, and nothing in SMTP replaced it. The economics were simple. Sending mail cost almost nothing, and every message could be sent to millions of addresses at once. Paul Graham estimated in 2002 that “the lowest rate seems to be about $200 to send a million spams.”²⁴ Even a tiny response rate was profitable. Because the cost of reading fell on the recipient, the sender had no reason to stop.

The responses to spam can be grouped by what they tried to change.

Making senders identify themselves or pay

At the World Economic Forum in January 2004, Bill Gates predicted that “two years from now, spam will be solved.” His proposals were authentication, to verify who was sending, and some form of payment, an electronic stamp that would make bulk sending expensive.²⁵ The idea had a respectable intellectual history. In 1992 the cryptographers Cynthia Dwork and Moni Naor had proposed “pricing via processing”: requiring the sender’s computer to solve a small mathematical puzzle for each message, cheap for a person sending a few messages, expensive for someone sending millions.²⁶ The idea later influenced Adam Back’s Hashcash and, through it, the proof of work used by Bitcoin.

No payment system was ever adopted for email. The obstacle was the same openness that made email universal. A charge only works if every server enforces it, and a federated network of millions of independently run servers could not be made to agree. A payment scheme would also have burdened legitimate senders, such as mailing lists and small organisations, and would have created a new party, whoever collected the payment, with power over who could be heard.

Making spam illegal

The United States passed the CAN-SPAM Act in December 2003, which took effect in 2004. It required commercial email to identify its sender, avoid deceptive subject lines and offer a way to unsubscribe.²⁷ Other countries passed their own laws, some stricter. The laws gave legitimate marketers rules to follow and gave regulators tools against the worst operators, but they could not reach senders in other jurisdictions, or the networks of hijacked computers that sent much of the spam. In practice the problem was solved, as far as it was solved, by software.

Learning to recognise spam

In August 2002 Paul Graham published an essay titled “A Plan for Spam.” It opened: “I think it’s possible to stop spam, and that content-based filters are the way to do it.” Graham described a statistical filter that counted how often each word appeared in a user’s spam and in their legitimate mail, and used those frequencies to estimate the probability that a new message was spam. On his own mail it missed “less than 5 per 1000 spams, with 0 false positives.”²⁴

Graham was not the first to apply Bayesian statistics to spam, but his essay showed that a simple version worked well, and filters based on it spread quickly into open-source tools such as SpamAssassin. His central insight was economic. Spammers could forge addresses and change their wording, but “the Achilles heel of the spammers is their message.” To make money they had to say something that ordinary mail did not, and a filter that learned the difference would always find them. He also identified the asymmetry that would govern all later filtering: “For most users, missing legitimate email is an order of magnitude worse than receiving spam.”²⁴

Statistical filters were the first large-scale use of machine learning that most people encountered, though few would have called it that. They were personal, because each one learned from its owner’s mail, and adversarial, because spammers changed their methods in response.

Making the network remember

Spam did not go away. Gates’s two years passed, and by around 2008 spam was estimated at more than 90 percent of all email traffic.²⁵ What changed was where it was stopped. Large providers could see patterns across billions of messages that no individual filter could see, and they began to judge mail by the reputation of the server and domain that sent it, as well as by its content.

To make reputation possible, the network needed a way to verify that mail really came from the domain it claimed. Three standards built that layer on top of SMTP without replacing it. SPF lets a domain publish a list of the servers allowed to send its mail.²⁸ DKIM lets the sending server sign each message cryptographically, so the receiver can check that it was not altered and that the domain approved it.²⁹ DMARC, developed from 2012 by a group of large mail providers and senders and published in 2015, tells receivers what to do when a message fails those checks, and sends reports back to the domain owner.³⁰

In October 2023, Google and Yahoo announced that from February 2024 anyone sending more than about 5,000 messages a day to their users would have to authenticate mail with SPF, DKIM and DMARC, offer one-click unsubscribe, and keep the rate at which recipients reported their mail as spam below 0.3 percent.³¹ With two companies setting the rules, practices that had been recommended for a decade became mandatory within months.

The trade-off is significant. Reputation systems made email usable again, but they work best for whoever sees the most mail. A small organisation running its own mail server today often finds its messages sent to spam folders until it builds a reputation, and building one is difficult without sending mail. The open protocol survives, but the practical ability to take part in it on equal terms has narrowed. The defence against spam pulled email toward a few large providers, which is the opposite of what federation was designed to do. Spam remains a large share of traffic: Kaspersky estimated it at 47 percent of global email in 2024.² Most users rarely see it, because the providers have become very good at keeping it away.

Machine learning learns the inbox

Once machine learning had proved itself on spam, the same approach was applied to a different problem. Spam filtering asks whether a message is wanted at all. The newer question was which of the wanted messages matter most.

Ranking

In August 2010 Google introduced Priority Inbox, which predicted which messages a user would consider important, using signals such as whom they wrote to most often and which messages they opened and answered.³² A paper by Douglas Aberdeen and colleagues described how the system learned a model for each user, combined with a global model, and updated it continuously.³³ In 2013 Gmail added tabs that automatically divided mail into Primary, Social and Promotions.³⁴ In 2019 Google reported that adding TensorFlow-based models to its existing filters blocked about 100 million additional spam messages a day, on top of the more than 99.9 percent of spam, phishing and malware it already stopped.³⁵

Each step moved the inbox further from a chronological list of everything received toward a ranked view of what the software predicted the user cared about. That was helpful, and it was also a transfer of judgement. The same message could appear prominently to one person and be hidden in a tab for another, according to a model neither person could inspect.

Replying

In November 2015 Google introduced Smart Reply in its experimental Inbox app. It used a sequence-to-sequence neural network to suggest up to three short responses to a message. The research paper describing it reported that the system was “responsible for assisting with 10% of all mobile responses” in the app.³⁶ The team had found that about a quarter of replies contained twenty words or fewer, and so could plausibly be predicted. The system needed safeguards, including a step that kept its suggestions varied and screened out inappropriate ones.

In May 2018 Google added Smart Compose, which predicted the rest of a sentence as the user typed. The engineering constraint was latency: because suggestions appeared on every keystroke, the model had to respond “ideally within 100ms for the user not to notice any delays.”³⁷ Smart Compose was a turning point that went largely unnoticed. Software was now proposing the words of the message, as well as sorting and answering mail, and many users accepted its proposals.

Writing and summarising

Large language models made these early features look modest. From 2023 Microsoft offered Copilot in Outlook for business customers, able to draft messages and summarise long threads, and in January 2025 it brought Copilot into consumer Microsoft 365 plans.³⁸ Apple’s on-device Apple Intelligence, introduced in iOS 18 in late 2024, added summaries of individual messages, a priority section for time-sensitive mail, and automatic categories similar to Gmail’s tabs.³⁹ In January 2026 Google announced what it called the Gemini era of Gmail, with AI Overviews that summarise threads and answer questions about the inbox, an improved Help Me Write, and an experimental AI Inbox that surfaces messages and tasks the model thinks need attention.⁴⁰

The direction is clear even if the details change every few months. The inbox is turning from a list of messages into a set of interpretations of messages: summaries, priorities, suggested actions, drafted replies. When both the sender and the recipient use these tools, a message can be drafted by one model and summarised by another, with the people involved approving rather than writing or reading. That possibility raises questions explored in the philosophy section below.

The receiver’s defences

Machine learning and related privacy features also changed the balance of power between senders and receivers. Marketers had long embedded invisible images, often called tracking pixels, in their mail; when the image loaded, the sender learned that the message had been opened, when, and roughly where. In June 2021 Apple introduced Mail Privacy Protection, which loads remote content through Apple’s servers in advance, hiding the recipient’s IP address and making opens impossible to detect reliably.⁴¹ Litmus, which measures email opens for marketers, notes that Apple’s protection now affects roughly 55 to 60 percent of all opens it tracks.⁴²

The inbox reinvented, and absorbed

From around 2010, a wave of start-ups and experiments tried to redesign the email client. Their histories show a recurring pattern, which is clearer when each decision is looked at from the company’s point of view at the time.

Sparrow (2010–2012)

A small French team built a fast, elegant mail client for the Mac and iPhone at a time when Gmail’s own mobile apps were limited. In July 2012 Google acquired the team, which joined Gmail; the apps remained available but stopped receiving new features.⁴³ For a five-person company selling inexpensive apps, joining the largest email provider offered more reach than staying independent, and for Google it brought designers who had shown how to build a good mobile mail client.

Mailbox (2013–2016)

Mailbox launched on the iPhone in February 2013 with a design built around swiping messages away and “snoozing” them until later. Demand was so high that it used a reservation queue; by the time Dropbox announced its acquisition in March 2013, it had taken more than a million reservations.⁴⁴ In December 2015 Dropbox announced that it would close Mailbox in February 2016, explaining that it was focusing on collaboration. Coverage at the time noted that the app had seen few updates and slipping usage.⁴⁵ Mailbox worked only with accounts at other providers, chiefly Gmail, which meant it depended on a platform owned by a company that could, and did, add similar features itself.

Acompli (2014)

Microsoft bought Acompli, a mobile mail and calendar start-up, in December 2014, reportedly for more than $200 million, and within weeks rebuilt its Outlook apps for iOS and Android on Acompli’s technology.⁴⁶ This is the clearest case of an incumbent buying its way into mobile quickly. It worked: the product that began as Acompli is still Outlook’s mobile app.

Inbox by Gmail (2014–2019)

Google’s own experiment, launched in October 2014, grouped mail into bundles, treated messages like tasks that could be marked done, and introduced Smart Reply. In 2018 Google announced that it would close Inbox in March 2019 in order to “focus solely on Gmail.” The product manager described Inbox as “a great place to experiment with new ideas,” many of which, including snoozing, nudges and Smart Reply, had already moved into Gmail.⁴⁷ Users who liked Inbox’s bundles found that not everything came across. From Google’s point of view, maintaining two mail apps for the same accounts divided effort; from the point of view of Inbox’s users, the experiment had been a product they relied on.

Facebook email (2010–2014)

In 2010 Facebook gave its users @facebook.com addresses as part of a combined messaging system; press coverage at the time speculated about a challenge to Gmail. In February 2014 Facebook closed the service, telling users that “most people haven’t been using their Facebook email addresses.”⁴⁸ A social network’s users already had email addresses elsewhere, and an address is only valuable if people give it out.

HEY (2020)

Basecamp launched HEY in June 2020 as a paid email service, $99 a year, built on a different idea of the inbox: unknown senders had to be approved before their mail appeared, and tracking pixels were blocked. Apple initially rejected an update to its iPhone app because HEY sold subscriptions only on the web rather than through Apple’s in-app purchase system, and the dispute became an early public test of App Store rules.⁴⁹ HEY remains in operation as a niche paid service. Its existence shows that some people will pay to change the terms on which strangers can reach them.

Superhuman (2017–)

Superhuman sold a fast, keyboard-driven client on top of Gmail and Outlook accounts for a monthly fee, later adding AI features. In July 2025 Grammarly acquired it, and in October 2025 Grammarly renamed its parent company Superhuman, with the mail app becoming Superhuman Mail.⁵⁰ The best-known independent mail client of the 2020s ended up as part of a larger AI writing company.

Looked at together, these histories show a structural pressure more than a series of poor choices. Most of these products were built on accounts hosted by Google, Microsoft or Apple, and their best ideas could be copied by the host. Snoozing, swiping, smart replies, bundled categories and keyboard shortcuts now appear in the major clients. The open protocol made it possible to build a new client without anyone’s permission. The concentration of accounts in a few providers made it hard for a new client to remain independent.

The giants that faded, sold or changed shape

The larger services faced different pressures, and their decisions also make sense in context.

Hotmail to Outlook.com

In July 2012 Microsoft introduced Outlook.com as the successor to Hotmail, with a cleaner design and integration with its Office brand. By May 2013 it had moved all of Hotmail’s active accounts, about 300 million, and roughly 150 petabytes of mail, to the new service.⁵¹ Microsoft’s decision was to stop competing as a stand-alone webmail brand and to tie consumer email to its productivity software, the area where it was strongest. Existing Hotmail addresses kept working.

Yahoo

Yahoo Mail was among the largest services through the 2000s. In 2016 Yahoo disclosed security breaches dating from 2013 and 2014, and in October 2017, after Verizon had completed its acquisition of Yahoo’s operating business, it said that all of its roughly three billion accounts had been affected by the 2013 theft.⁵² Verizon had already lowered its purchase price by $350 million after the earlier disclosures. In 2021 Verizon sold Yahoo and AOL to funds managed by Apollo Global Management for about $5 billion.⁵³ Yahoo Mail continues, with a far smaller share of attention than it once had.

AOL

AOL’s service introduced millions of people to email in the 1990s. Its dial-up internet access, once its core business, ended on 30 September 2025; AOL said that email addresses and other benefits of its plans were not affected.⁵⁴ Many older users still keep AOL addresses, which shows how sticky an email address becomes once it is in other people’s address books.

Lotus Notes

Notes was for many years one of the leading corporate mail and collaboration systems. IBM, which had bought Lotus in 1995, agreed in December 2018 to sell Notes, Domino and several other products to HCL Technologies for $1.8 billion.⁵⁵ Notes continues under HCL. Many large organisations had already moved their mail to Microsoft Exchange and later to cloud services.

Eudora

Qualcomm stopped developing Eudora in 2006, judging that it no longer fitted its main businesses at a time when Microsoft Outlook had become standard in companies. It funded an open-source successor built on Mozilla Thunderbird, which was discontinued a few years later.¹³

Windows Mail and the end of fetching

Microsoft ended support for the Mail and Calendar apps built into Windows on 31 December 2024, replacing them with the new Outlook for Windows.⁵⁶ In January 2026 Gmail stopped fetching mail from other accounts over POP and ended Gmailify, which had applied Gmail’s spam filtering and organisation to accounts at other providers.⁵⁷ Each change pushed users further toward reading each provider’s mail in that provider’s own app.

Who reads the world’s email now

Measuring email’s market is difficult because different measures count different things: accounts, active users, business mailboxes or opened marketing messages.

By accounts, Gmail is the largest consumer provider. Outlook.com, Yahoo Mail, Apple’s iCloud Mail and AOL follow, along with large regional providers such as Mail.ru in Russia and the many national telecom providers in Europe and Asia. In business email, Microsoft 365 and Google Workspace dominate.

By software used to open mail, Litmus’s measurements of more than 1.5 billion opens in September 2026 put Apple at 76.6 percent, Gmail at 16.1 percent, desktop Outlook at 3.5 percent and Yahoo Mail at 2.4 percent.⁴² These numbers measure marketing mail sent by Litmus’s customers, and Apple’s figure is inflated by Mail Privacy Protection, which preloads messages whether or not they are read. They show where commercial senders’ mail is displayed, not how many people use each service.

Two newer forces push against concentration. Privacy-focused services such as Proton Mail, founded in 2014 by scientists who had met at CERN, offer encryption under which, the company says, it cannot read users’ mail; Proton says it has more than 100 million accounts.⁵⁸ Governments have also begun to treat email as a question of sovereignty. In India, the National Informatics Centre has moved government mail to Zoho, an Indian company; a reply in parliament in 2025 put the number of official accounts moved at about 12.68 lakh (1.27 million).⁵⁹ In October 2025 the Home Minister, Amit Shah, announced that he had moved his own address to Zoho Mail.⁶⁰ Because email addresses are federated, an organisation or a country can change providers while keeping its domain, something no closed messaging system allows.

Understanding the philosophy

The history above is a series of engineering and business decisions. Behind them lie a few deep questions about strangers, trust, price, privacy, ritual and freedom. This section takes them one at a time. Each idea is explained in plain language, connected to the philosophical tradition it belongs to, traced through the history, tested against the strongest objection, and applied to email today. Except where a participant said so directly, the connections are interpretations offered to make the ideas clearer.

1. The right to knock: email as hospitality

The idea in plain terms

An email address is a promise that anyone may write to you. You do not have to accept a friend request, exchange numbers or join the same platform. A stranger on the other side of the world can reach your inbox. This is unusual, and email is almost the last large system that offers it.

The tradition

In 1795 the German philosopher Immanuel Kant published a short essay called Perpetual Peace, setting out conditions under which nations might stop going to war. One of them concerned strangers. Kant argued for what he called a right of hospitality: a stranger arriving in another land has the right not to be treated as an enemy. He was careful about its limits. The stranger has a right to visit, to present himself and seek contact, but not a right to be a guest, which requires an invitation. The host may turn him away, as long as this can be done without destroying him. Kant grounded the right in a plain fact: people share the surface of a round earth and cannot scatter forever, so they must find ways to tolerate one another’s approach.

Where it shows up

SMTP built something very close to Kant’s right of visit into the infrastructure of the internet. Any server may connect to any other and present a message. The receiver is free to refuse it, and modern filters refuse a great deal, but the attempt itself is permitted. Closed messaging systems took the opposite view: on most of them, a stranger cannot even knock without being let in first. HEY’s screening of unknown senders sits between the two, keeping the right to knock while making acceptance an explicit decision.

The strongest objection

Kant’s right of visit assumed that visitors were few and that each arrived in person. On the internet, a single visitor can knock on a billion doors at once at no cost. Spam is what happens when hospitality is offered without limit to people who can visit at scale. On this view, closed systems that admit only known contacts are not inhospitable; they are realistic.

What it means today

The tension has not been resolved; it has been managed. The spam filter is a way of keeping Kant’s right of visit while declining most of the visitors. Whether that compromise survives depends on whether the filtering stays fair to the stranger who has something real to say, such as the job applicant, the small business or the citizen writing to an official. The value of email is that it allows contact without prior permission. That value is lost if the filters admit only those who are already known.

2. Be liberal in what you accept: robustness and its price

The idea in plain terms

A system can be designed to reject anything that does not follow its rules exactly, or to accept anything it can make sense of. The first is safer; the second keeps working when things are imperfect. Email chose the second, and much of its history is the cost of that choice coming due.

The tradition

Postel’s robustness principle, “be conservative in what you do, be liberal in what you accept from others,” is an engineering rule, but it expresses an ethical stance that philosophers of interpretation would recognise as a principle of charity: when someone says something that could be read in several ways, read it in the way that makes most sense. Charity makes communication possible between people who do not share every convention. In 2023 the Internet Architecture Board published RFC 9413, “Maintaining Robust Protocols,” which revisited Postel’s principle after four decades. Its authors argued that tolerating flawed input can allow errors and abuses to accumulate until they become impossible to remove, because systems come to depend on the tolerance. Being liberal in what you accept, they suggested, can make a protocol fragile in the long run.⁶¹

Where it shows up

Email accepted messages whose sender could not be verified, headers that did not quite follow the format, and HTML that no two clients displayed alike. That generosity let email spread across incompatible systems. The same generosity let anyone forge a From line, which made phishing possible, and it made authentication impossible to add later without decades of work. SPF, DKIM and DMARC are, in effect, the network slowly becoming less liberal in what it accepts. The 2024 bulk-sender rules mark the point where large receivers began to refuse what they had always tolerated.

The strongest objection

Strictness has costs of its own. Had email required verified identity from the start, it would probably have been owned by whoever issued the identities, and it might never have spread beyond a few institutions. The protocols that lasted are often the forgiving ones. Generosity created the network that strictness is now being used to protect.

What it means today

The balance has shifted from accepting almost everything to accepting what can be verified. The question is who decides what counts as verified. When the decision is made by a few large receivers, robustness is replaced by permission, and the network becomes less open even though its protocol has not changed.

3. What a message should cost: signals that cannot be faked

The idea in plain terms

When something is free, it says nothing about how much the sender cares. A handwritten letter shows effort. A mass email shows none. Much of the struggle against spam is an attempt to restore some cost, of money, computation or reputation, so that receiving a message tells you something about the sender.

The tradition

In 1973 the economist Michael Spence published a paper on what he called job market signalling. Employers cannot see how capable a job applicant is, but applicants can acquire credentials, such as a difficult degree, that are cheaper for capable people to obtain. Because the signal is costly in this specific way, it carries information.⁶² Two years later, the biologist Amotz Zahavi proposed the handicap principle to explain features such as a peacock’s tail: a signal can be trusted only if it is costly in a way that a dishonest signaller could not afford.⁶³ Both ideas rest on the same insight. A signal that costs nothing can be sent by anyone, and therefore proves nothing.

Where it shows up

The Penny Post’s stamp was a small cost that kept letters meaningful. Email removed it. Dwork and Naor’s pricing via processing and Gates’s electronic stamp were attempts to reintroduce a cost that honest senders could bear and spammers could not. Neither was adopted, because the federated network could not enforce it. Reputation took the place of price. A domain with a long history of wanted mail has paid, in time and good behaviour, for a signal that a new spam operation cannot fake. Graham’s filters looked for a different kind of honest signal: the content a spammer cannot avoid sending.

The strongest objection

Costly signals favour those who can afford them. Spence’s degrees reward those with the money and time to earn them, and email reputations reward organisations that have been sending mail for years. A cost that deters spammers also deters the poor, the new and the unusual. The free message is democratic precisely because it says nothing about its sender’s resources.

What it means today

Generative AI reduces the cost of a convincing, personalised message almost to zero, undermining one of the last honest signals left: the effort of writing. When any message can be fluent and tailored, a well-written email no longer shows that a person spent time on it. The search for signals that cannot be faked is moving toward cryptographic proof of identity and toward the relationship itself: whether you have corresponded with this person before.

4. Search, don’t sort: order without categories

The idea in plain terms

Folders ask you to decide in advance where each thing belongs. Search lets you find it later without deciding. Gmail’s bet that people would rather search than sort changed how millions of people relate to their own records.

The tradition

In Everything Is Miscellaneous (2007), David Weinberger described three “orders of order.” The first is the arrangement of physical things, such as books on a shelf, where each object can be in only one place. The second is the arrangement of information about things, such as a library catalogue, which can list a book under several headings but still relies on a fixed scheme. The third is digital, where items and their descriptions can be rearranged at the moment of use, by anyone, for any purpose.⁶⁴ The older ideal of classification, which goes back to Aristotle’s attempt to sort all things into kinds, assumes that every item has a proper place. Weinberger’s third order abandons that assumption.

Where it shows up

POP-era email clients, with their nested folders, were a second-order system. Gmail’s labels, which let a message belong to several categories, and its search box, which ignored categories altogether, belonged to the third order. Priority Inbox, tabs and AI Inbox go further: the software creates the order, for each user, at the moment the inbox is opened.

The strongest objection

Sorting is also a way of thinking. Deciding where a message belongs is a form of thinking about it: is this a commitment, a reference, a piece of news? Search removes that moment of judgement. When a model sorts the inbox, the judgement is not removed but transferred, and the categories it chooses reflect its designers’ purposes, which may include advertising, as much as the user’s own.

What it means today

The AI inbox is the logical end of the third order: no folders, no fixed list, only an arrangement generated for this person at this moment. The question this raises is about authority. In a world of folders, the user decided what mattered. In a world of models, the model proposes and the user mostly accepts.

5. Who may read a letter: privacy as appropriate flow

The idea in plain terms

We tend to think of privacy as secrecy: something is private if no one else knows it. Email shows why that definition is too simple. Your mail provider must handle every message to deliver it, filter it and store it. The question is not whether anyone else touches your mail but what they are allowed to do with it.

The tradition

The philosopher Helen Nissenbaum proposed in 2004 that privacy is better understood as contextual integrity. Every social context, such as medicine, friendship, employment or the post, has norms about how information should flow within it. A doctor may share your test results with a specialist but not with your employer. Privacy is violated when information flows in a way that breaks the norms of the context in which it was shared.⁶⁵ On this account, the same act of reading can be appropriate or inappropriate depending on its purpose.

Where it shows up

The postal tradition established a strong norm: carriers move letters but do not read them. Email providers broke the letter of that norm from the start, since servers process every message, but mostly kept its spirit when they scanned only to deliver mail and to block spam and viruses. Gmail’s ad scanning in 2004 caused alarm because it introduced a new purpose, commercial targeting, into the context of private correspondence. Google’s decision in 2017 to stop it can be read as a return to the old norm. Apple’s Mail Privacy Protection blocks a different flow, from the reader’s behaviour back to the sender. AI summaries introduce yet another: the provider’s model now reads mail in order to tell you what it says.

The strongest objection

Users consent to these flows when they accept the provider’s terms, and many welcome them: spam filtering and summaries are services they want. Contextual integrity can seem to treat inherited norms as fixed, when norms change as technology offers new benefits. Perhaps the norm of the sealed letter is simply obsolete.

What it means today

Nissenbaum’s framework explains why the same model reading the same message can feel acceptable when it summarises the message for you and disturbing when it uses it for something else. The important questions about AI in email are about purpose and onward flow: whether what the model learns stays within the context of your correspondence, or travels into training, advertising or profiles.

6. Rituals of politeness: what automated replies automate

The idea in plain terms

Much of what we write in email is not information. It is courtesy: thanks, greetings, acknowledgements, “sounds good.” These small rituals show that we take the other person seriously. Smart Reply and its successors automate exactly these rituals, because they are predictable.

The tradition

The sociologist Erving Goffman studied the small rituals of everyday interaction. In his essay “On Face-Work” (1955, republished in Interaction Ritual, 1967), he defined “face” as “the positive social value a person effectively claims for himself by the line others assume he has taken during a particular contact.”⁶⁶ Much of ordinary conversation, Goffman argued, is work done to protect one’s own face and other people’s: greetings, apologies, compliments, careful refusals. These rituals are not empty; they hold social life together. They are predictable because they matter. Everyone knows what to say, and leaving it out would be noticed.

Where it shows up

The Smart Reply paper found that about a quarter of email replies were twenty words or fewer, which is why a model could suggest them.³⁶ Many of those short replies were Goffman’s face-work: “Thanks!”, “Got it”, “Looks great.” The model learned rituals because rituals are regular. Smart Compose extended the automation from replies to the polite connective tissue of longer messages, and large language models now generate entire courteous emails.

The strongest objection

If rituals are formulaic, automating them frees people for the parts of communication that are not formulaic. A thank-you suggested by software and approved by a person is still sent by that person, who chose to send it. The ritual is preserved and the effort is reduced. Nothing has been lost except tedium.

What it means today

The value of a ritual depends partly on the belief that the other person performed it. If recipients come to assume that courtesies are machine-generated, courtesies will stop carrying their meaning, and people will look for new signs of genuine attention. The more effectively software performs the work of protecting face, the less that work will count as care.

7. Exit, voice and loyalty: why people stay with email

The idea in plain terms

When you are unhappy with a product, you can leave it, complain about it, or put up with it. Email’s design makes leaving unusually possible, and that may be the main reason it lasts.

The tradition

In Exit, Voice, and Loyalty (1970), the economist Albert Hirschman described two responses to a decline in the quality of a firm, organisation or state. Exit means leaving, as a customer switches brands. Voice means staying and trying to change things by complaining or organising. Loyalty affects which one people choose: loyal members tend to use voice before exiting. Hirschman argued that healthy institutions need both, and that the possibility of exit makes voice more effective, because those in charge know that members can leave.⁶⁷

Where it shows up

Federation makes exit possible at several levels. A user can change mail apps without changing provider, as Sparrow, Mailbox and Superhuman users did. An organisation with its own domain can change provider without changing addresses, as the Indian government did when it moved to Zoho. The protocol does not depend on any company, so no single company’s decline can end it. Closed systems, where your identity belongs to the platform, make exit expensive. Leaving a closed messaging network means losing contact with everyone still on it.

The strongest objection

In practice, exit from email is harder than the protocol suggests. Most personal addresses end in a provider’s domain, such as gmail.com, so leaving the provider means changing your address. Large providers’ reputation systems and the closure of features such as POP fetching and Gmailify make it harder to combine services. The protocol permits exit, but the market makes it costly.

What it means today

Email’s persistence may come less from loyalty than from the possibility of exit. People stay because they could leave, and because their address, if they own the domain, can follow them. That suggests a practical lesson: an address under one’s own domain is a form of freedom that no closed platform can offer.

Bringing the threads together

The seven ideas describe one system from different angles. Email offers strangers the right to knock (hospitality), accepts what arrives generously (robustness), and therefore needs ways to tell meaningful messages from worthless ones (costly signals). Its providers handle every message, which raises the question of what they may do with what they see (contextual integrity). As the volume grew, software took over the work of ordering mail (Weinberger’s third order), and then the work of writing its rituals (Goffman’s face-work). Through all of it, the protocol has preserved the possibility of leaving (Hirschman’s exit), which may be why people still use it.

What email has been negotiating all along

It is tempting to describe the history of email as a series of products: CTSS MAIL, SNDMSG, Eudora, Hotmail, the BlackBerry, Gmail, Inbox, Copilot. But the products changed while the underlying problem stayed the same. Every decision covered in this essay, from the restriction Van Vleck and Morris placed on mass mailing in 1965 to the bulk-sender rules of 2024, has been an attempt to answer one question: on what terms may one person claim another’s attention?

The ARPANET answered with a social rule and an administrator. SMTP answered with unconditional acceptance. The spam war produced statistical filters and then reputation systems that answered on each user’s behalf. Gmail’s ranking, tabs and AI Inbox answer with predictions about what each person values. Generative models now take part on both sides, writing the claim and judging it.

Each answer moved some of the decision away from the person who receives the mail. That has made email bearable at a scale its designers never imagined, and it has given the companies that run the largest filters a quiet influence over who can be heard. Email’s openness is real, but it now depends on judgements made by a small number of providers, whether those judgements are made by rules, by reputation or by models.

Email’s critics have called it a poor product for decades, and in many ways they are right. It outlives its replacements because it is the one system in which the right to write to someone is not granted by a company, and the address can, in principle, belong to the person. Every new communication product starts by asking who is allowed in. Email started by letting everyone in, and has spent sixty years working out what that costs.

Selected bibliography

• Van Vleck, Tom. “The History of Electronic Mail.” Multicians.org. multicians.org
• Crocker, David H., John J. Vittal, Kenneth T. Pogran, and D. Austin Henderson Jr. “Standard for the Format of ARPA Network Text Messages.” RFC 733, November 1977. rfc-editor.org
• Postel, Jonathan B. “Simple Mail Transfer Protocol.” RFC 821, August 1982. rfc-editor.org
• Crocker, David H. “Standard for the Format of ARPA Internet Text Messages.” RFC 822, August 1982. rfc-editor.org
• Postel, Jon, ed. “DoD Standard Transmission Control Protocol.” RFC 761, January 1980. rfc-editor.org
• Borenstein, Nathaniel, and Ned Freed. “MIME (Multipurpose Internet Mail Extensions).” RFC 1341, June 1992. rfc-editor.org
• Templeton, Brad. “Reaction to the DEC Spam of 1978.” templetons.com
• Dwork, Cynthia, and Moni Naor. “Pricing via Processing or Combatting Junk Mail.” Advances in Cryptology: CRYPTO ’92, pp. 139–147. PDF
• Graham, Paul. “A Plan for Spam.” August 2002. paulgraham.com
• Kannan, Anjuli, et al. “Smart Reply: Automated Response Suggestion for Email.” Proceedings of KDD 2016. arXiv:1606.04870
• Chen, Mia Xu, et al. “Gmail Smart Compose: Real-Time Assisted Writing.” Proceedings of KDD 2019. arXiv:1906.00080
• Kitterman, Scott. “Sender Policy Framework (SPF).” RFC 7208, April 2014.
• Crocker, D., T. Hansen, and M. Kucherawy, eds. “DomainKeys Identified Mail (DKIM) Signatures.” RFC 6376, September 2011.
• Kucherawy, M., and E. Zwicky, eds. “Domain-based Message Authentication, Reporting, and Conformance (DMARC).” RFC 7489, March 2015.
• Thomson, Martin, and David Schinazi. “Maintaining Robust Protocols.” RFC 9413, June 2023. rfc-editor.org
• Newport, Cal. A World Without Email: Reimagining Work in an Age of Communication Overload. Portfolio, 2021.
• Kant, Immanuel. Perpetual Peace: A Philosophical Sketch. 1795. In Kant: Political Writings, ed. H. S. Reiss, trans. H. B. Nisbet. Cambridge University Press, 1991.
• Spence, Michael. “Job Market Signaling.” Quarterly Journal of Economics 87, no. 3 (1973): 355–374.
• Zahavi, Amotz. “Mate Selection: A Selection for a Handicap.” Journal of Theoretical Biology 53, no. 1 (1975): 205–214.
• Weinberger, David. Everything Is Miscellaneous: The Power of the New Digital Disorder. Times Books, 2007.
• Nissenbaum, Helen. “Privacy as Contextual Integrity.” Washington Law Review 79, no. 1 (2004): 119–157.
• Goffman, Erving. Interaction Ritual: Essays on Face-to-Face Behavior. Anchor Books, 1967.
• Hirschman, Albert O. Exit, Voice, and Loyalty: Responses to Decline in Firms, Organizations, and States. Harvard University Press, 1970.
• Seneca. Letters from a Stoic (Epistulae Morales ad Lucilium). Trans. Robin Campbell. Penguin, 1969.

Research limitations

Market figures for email are inconsistent and should be read as estimates. Radicati’s and Kaspersky’s numbers come from their own methods, which are only partly published; the Radicati figure was taken from the executive summary of its 2024–2028 report. Google has not published an official Gmail user count since 2016, so the widely repeated figure of 1.8 billion is reported here as a secondary estimate. Litmus measures opens of marketing email sent by its customers, and Apple’s share is inflated by Mail Privacy Protection; it is not a measure of how many people use each service. The comScore figures for 2012 counted only desktop visitors. Proton’s account numbers are the company’s own.

Accounts of the earliest email systems rely mainly on participants’ later recollections, especially Tom Van Vleck’s history and Ray Tomlinson’s interviews, rather than on contemporary records; Van Vleck dates the staff note proposing MAIL only approximately, and other sources give a different note number. The 1978 DEC message is described as the first spam by convention; earlier mass messages existed, including the CTSS message Van Vleck describes. Reported prices for acquisitions such as Hotmail, Mailbox and Acompli are press estimates, since the companies did not disclose terms. The estimate that spam exceeded 90 percent of traffic around 2008 comes from a security vendor’s report as quoted in the press.

The reasons given for product closures, such as Dropbox’s for Mailbox, Google’s for Inbox and Facebook’s for its email service, are the companies’ own public statements; press interpretations are identified as such. This essay analyses decisions in the context in which they were made and does not attribute motives beyond what the companies said. Details of the January 2026 Gmail changes come from press coverage of Google’s announcement, and some features were still in limited testing. Seneca and Kant are paraphrased rather than quoted, to avoid relying on a particular translation. The philosophical section connects email to Kant, Postel and RFC 9413, Spence and Zahavi, Weinberger, Nissenbaum, Goffman and Hirschman; apart from Postel’s principle, which shaped the protocols directly, these connections are the author’s interpretation, offered to make the ideas understandable.

Footnotes

1. The Radicati Group, Email Statistics Report, 2024–2028, Executive Summary (October 2024), radicati.com; Statista, citing Radicati, gives an estimate of 376 billion a day for 2025, statista.com. ↩

2. Kaspersky, annual spam and phishing report for 2024, as reported in “Kaspersky reports nearly 900 million phishing attempts in 2024,” CXOToday, cxotoday.com. ↩ ↩²

3. Tom Van Vleck, “The History of Electronic Mail,” multicians.org. ↩ ↩² ↩³

4. Van Vleck, “The History of Electronic Mail”; and accounts based on Tomlinson’s later interviews, for example “Ray Tomlinson, email inventor who picked the @ sign for addresses, dies,” PCWorld, March 2016, pcworld.com. ↩ ↩²

5. David H. Crocker, John J. Vittal, Kenneth T. Pogran and D. Austin Henderson Jr., “Standard for the Format of ARPA Network Text Messages,” RFC 733 (21 November 1977). ↩

6. Brad Templeton, “Reaction to the DEC Spam of 1978,” including Einar Stefferud’s note on the overflowing address list, templetons.com; and Abusix, “The Birth of Email Spam: Gary Thuerk’s 1978 ‘Green Card’ Incident,” abusix.com. ↩

7. Templeton, “Reaction to the DEC Spam of 1978”: Maj. Raymond Czahor’s message of 4 May 1978, and Richard Stallman’s messages of 8 and 9 May 1978 to MsgGroup. ↩ ↩²

8. Jonathan B. Postel, “Simple Mail Transfer Protocol,” RFC 821 (August 1982), §1 Introduction. ↩

9. Craig Partridge, “Mail Routing and the Domain System,” RFC 974 (January 1986). ↩

10. Jon Postel, ed., “DoD Standard Transmission Control Protocol,” RFC 761 (January 1980), §2.10 Robustness Principle. ↩

11. Joyce K. Reynolds, “Post Office Protocol,” RFC 918 (October 1984). ↩

12. Mark R. Crispin, “Interactive Mail Access Protocol: Version 2,” RFC 1064 (July 1988). ↩

13. Computer History Museum, “The Eudora Email Client Source Code,” computerhistory.org. ↩ ↩²

14. Nathaniel Borenstein and Ned Freed, “MIME (Multipurpose Internet Mail Extensions),” RFC 1341 (June 1992). ↩

15. Nathaniel Borenstein, “MIME & Me,” guppylake.com; Rutgers University, “Commemorating the 20th Anniversary of Internet Email Attachments,” rutgers.edu. ↩

16. “Sabeer Bhatia,” MIT Technology Review, Innovators Under 35, technologyreview.com; launch date and founders as reported in contemporary and later coverage. The acquisition price was not disclosed and is a press estimate. ↩ ↩²

17. HISTORY.com Editors, “First BlackBerry device, the 850, released,” This Day in History, 19 January, history.com; on RIM’s US market share in 2010–2011, Commoncog, “How BlackBerry Died,” commoncog.com. ↩ ↩²

18. Google, “Google Gets the Message, Launches Gmail,” press release, 1 April 2004, as reproduced in “It’s Gmail’s 20th birthday,” Business Insider, businessinsider.nl; “Gmail Launched 20 Years Ago on April Fools’ Day,” VOA Learning English (Associated Press), voanews.com. ↩

19. Google, “Google Gets the Message, Launches Gmail” (1 April 2004), paraphrasing Larry Page; see also “When Gmail Was First Announced, People Thought It Was an April Fools’ Joke,” ScienceAlert, sciencealert.com. ↩

20. comScore data for October 2012, as reported by GigaOM and summarised in “Gmail outsets Hotmail and becomes the largest e-mail service in the world,” Gigazine, 1 November 2012, gigazine.net. ↩

21. “Gmail Is Closing In on 1 Billion Users,” Time, May 2015, time.com; “Gmail Now Has More Than 1B Monthly Active Users,” TechCrunch, 1 February 2016, techcrunch.com. ↩

22. For example, “Gmail Launched 20 Years Ago on April Fools’ Day,” VOA Learning English (Associated Press), April 2024, which gives about 1.8 billion active accounts. ↩

23. Diane Greene, Google blog post of 23 June 2017, as reported in “Google Will Stop Scanning Your Gmail Messages to Sell Targeted Ads,” Fortune, 23 June 2017, fortune.com. ↩

24. Paul Graham, “A Plan for Spam” (August 2002), paulgraham.com. ↩ ↩² ↩³

25. “Gates: We’ll End Spam Within 2 Years,” ITPro Today, 28 January 2004, itprotoday.com; “Bill Gates’ Spam Prediction Misses Target,” InformationWeek, informationweek.com, which also cites Cisco’s 2009 Annual Security Report for the estimate that spam reached about 92 percent of email traffic around 2008. ↩ ↩²

26. Cynthia Dwork and Moni Naor, “Pricing via Processing or Combatting Junk Mail,” Advances in Cryptology: CRYPTO ’92, Lecture Notes in Computer Science 740 (1993), pp. 139–147, PDF. ↩

27. Controlling the Assault of Non-Solicited Pornography and Marketing (CAN-SPAM) Act of 2003, Public Law 108-187, signed 16 December 2003, effective 1 January 2004. ↩

28. Scott Kitterman, “Sender Policy Framework (SPF) for Authorizing Use of Domains in Email, Version 1,” RFC 7208 (April 2014), which replaced the experimental RFC 4408 of 2006. ↩

29. D. Crocker, T. Hansen and M. Kucherawy, eds., “DomainKeys Identified Mail (DKIM) Signatures,” RFC 6376 (September 2011), which replaced RFC 4871 of 2007. ↩

30. M. Kucherawy and E. Zwicky, eds., “Domain-based Message Authentication, Reporting, and Conformance (DMARC),” RFC 7489 (March 2015). ↩

31. Google, “Email sender guidelines,” Gmail Help, support.google.com; for a summary of the joint Google and Yahoo requirements effective 1 February 2024, see dmarcian, “Yahoo and Google: DMARC Required,” dmarcian.com. ↩

32. “Google launches ‘Priority Inbox’ for Gmail,” TechRadar, 31 August 2010, techradar.com. ↩

33. Douglas Aberdeen, Ondrej Pacovsky and Andrew Slater, “The Learning Behind Gmail Priority Inbox,” NIPS 2010 Workshop on Learning on Cores, Clusters and Clouds. ↩

34. Gmail announced the tabbed inbox in May 2013; see the Gmail blog, “A new inbox that puts you back in control,” 29 May 2013. ↩

35. Neil Kumaran, “Ridding Gmail of 100 million more spam messages with TensorFlow,” Google Workspace blog, 6 February 2019, workspace.google.com. ↩

36. Anjuli Kannan et al., “Smart Reply: Automated Response Suggestion for Email,” Proceedings of the 22nd ACM SIGKDD Conference (2016), abstract and §1, arXiv:1606.04870. ↩ ↩²

37. Yonghui Wu, “Smart Compose: Using Neural Networks to Help Write Emails,” Google AI Blog, 16 May 2018, research.google. ↩

38. Microsoft, “Copilot is now included in Microsoft 365 Personal and Family,” Microsoft 365 Blog, 16 January 2025, microsoft.com. ↩

39. “Mail app in iOS 18,” 9to5Mac, 9to5mac.com; and Knak, “Apple Intelligence and email,” knak.com. ↩

40. “Gmail Enters the Gemini Era With AI-Powered Inbox Overhaul,” TechRepublic, January 2026, techrepublic.com; “Gmail enters the Gemini era with powerful AI assistance features,” CyberInsider, cyberinsider.com. ↩

41. “7 major privacy updates announced at WWDC 2021,” Trusted Reviews, 7 June 2021, trustedreviews.com. ↩

42. Litmus, “Email Client Market Share,” data for September 2026 based on more than 1.5 billion opens, litmus.com. ↩ ↩²

43. “Google acquires popular iOS and Mac mail client Sparrow for its Gmail team,” The Next Web, 20 July 2012, thenextweb.com. ↩

44. “How Mailbox became an instant sensation,” Fortune, 12 February 2013, fortune.com; “Dropbox acquires email startup Mailbox,” CNN, 15 March 2013, cnn.com. ↩

45. “Failing to Find Users, Dropbox Will Shut Down Mailbox in February 2016 and Carousel in March,” TechCrunch, 7 December 2015, techcrunch.com, quoting the blog post by Drew Houston and Arash Ferdowsi. ↩

46. “Microsoft buys Acompli to enhance Outlook for iOS, Android,” Computerworld, 1 December 2014, computerworld.com. ↩

47. Matthew Izatt, Google blog post of 12 September 2018, as quoted in 9to5Google, 9to5google.com, and Digital Trends, digitaltrends.com. ↩

48. “Facebook ends @facebook.com email address service, citing low usage,” The Next Web, 24 February 2014, thenextweb.com. ↩

49. “Apple approves, then rejects Hey Email app from creators of Basecamp,” iMore, June 2020, imore.com; “Apple’s App Review Board confirms rejection of Basecamp’s Hey email service,” MacStories, macstories.net. ↩

50. “Grammarly rebrands to Superhuman, launches a new AI assistant,” TechCrunch, 29 October 2025, techcrunch.com; Michael Tsai, “Grammarly Acquires Superhuman,” 1 July 2025, mjtsai.com. ↩

51. “Microsoft Gives Hotmail a Facelift, Says It Will Transition It to Outlook,” NPR, 31 July 2012, npr.org; “Hotmail users migrated to Outlook,” The Register, 2 May 2013, theregister.co.uk. ↩

52. “Yahoo says all 3 billion accounts hacked in 2013 data theft,” Reuters, as published in Khaleej Times, October 2017, khaleejtimes.com. ↩

53. “Verizon sells Yahoo and AOL businesses to Apollo for $5 billion,” NBC News, 3 May 2021, nbcnews.com. ↩

54. “AOL is ending dial-up internet after 34 years,” ConsumerAffairs, August 2025, consumeraffairs.com. ↩

55. “IBM selling Lotus Notes/Domino business to HCL for $1.8B,” TechCrunch, 7 December 2018, techcrunch.com. ↩

56. Microsoft Support, “Windows Mail, Calendar and People are becoming new Outlook,” support.microsoft.com. ↩

57. “Gmail preparing to drop POP3 mail fetching,” The Register, 5 January 2026, theregister.com. ↩

58. Proton, “Proton Mail,” proton.me; figures are the company’s own. ↩

59. “Centre Moves 12.7 Lakh Official Email Accounts To Zoho,” Inc42, inc42.com, reporting a Lok Sabha reply. ↩

60. “HM Amit Shah switches account from Gmail to Zoho Mail,” IANS, 8 October 2025, ianslive.in. ↩

61. Martin Thomson and David Schinazi, “Maintaining Robust Protocols,” RFC 9413 (June 2023), rfc-editor.org. ↩

62. Michael Spence, “Job Market Signaling,” Quarterly Journal of Economics 87, no. 3 (1973): 355–374. ↩

63. Amotz Zahavi, “Mate Selection: A Selection for a Handicap,” Journal of Theoretical Biology 53, no. 1 (1975): 205–214. ↩

64. David Weinberger, Everything Is Miscellaneous: The Power of the New Digital Disorder (Times Books, 2007), chapter 1. ↩

65. Helen Nissenbaum, “Privacy as Contextual Integrity,” Washington Law Review 79, no. 1 (2004): 119–157. ↩

66. Erving Goffman, “On Face-Work: An Analysis of Ritual Elements in Social Interaction,” Psychiatry 18 (1955); reprinted in Interaction Ritual: Essays on Face-to-Face Behavior (Anchor Books, 1967), p. 5. ↩

67. Albert O. Hirschman, Exit, Voice, and Loyalty: Responses to Decline in Firms, Organizations, and States (Harvard University Press, 1970). ↩