Spam filters protect your attention

Spam is cheap for the sender and costly for the reader. One message takes seconds to delete, but first it makes you stop, read enough to judge it, decide, and recover what you were doing.

We do not see the true scale because good filters are invisible. In 2023, Google said Gmail was blocking nearly 15 billion unwanted emails every day. We notice the few that get through, not the flood that never gets a chance to interrupt us.

The problem is older than email as we know it

In 1975, RFC 706 described the “junk mail problem” on the ARPA network. A computer that wanted some messages had to accept and examine all of them. The cost was processor time, network capacity and delay. Replace the processor with a person and the diagnosis still fits.

Four years earlier, Herbert Simon made the wider point in “Designing Organizations for an Information-Rich World”. When information grows rich, attention grows scarce. He argued that a good information system should filter input into something that demands fewer hours of attention. Spam is that imbalance in its purest form: the sender can make another copy almost for free, while every reader spends a piece of a finite day deciding it is worthless.

From rules to probabilities

The first filters used rules and trust scores. Refuse a known bad machine. Shut a mail server that anyone could abuse. Match a suspect address or phrase. A 1999 guide from the Internet Engineering Task Force described accept and refuse lists, sender checks and filters for each user. These tools were fast and clear, but an attacker could change an address, misspell a word or move to a new server.

The rise of statistics changed the question. Instead of asking whether a message broke a fixed rule, a filter could ask how likely its words and other traits were to appear in spam. A 1998 paper on Bayesian junk-mail filtering showed how a filter could learn from examples. It also gave a different cost to blocking wanted mail and letting spam through. A spam filter could now learn, not just keep lists.

Machine learning moved the fight to patterns

As email grew, providers could learn from far more than words. Today’s systems read the text and also check who sent it, where it came from, its links, the rest of the mail run and user reports. No new wave made the old one useless. A sender check asks who sent a message. Past behaviour shows what that source did before. A content model asks what the message looks like. The final question comes from the owner: did I want this?

That last question matters because spam has a gray zone. A receipt is wanted. A forged receipt is not. A newsletter can be useful to one person and noise to another. The 1999 IETF guide already called for filters for each user because one set of rules could not suit everyone.

AI changes both sides

Generative AI can write decent text at low cost, in many languages and in endless forms. In 2024, the UK National Cyber Security Centre said AI was already making phishing more effective. It could write strong lures without the poor wording, bad spelling and weak grammar that once gave them away. Word lists help less when each message can be rewritten.

Defenders use AI for the same reason. A language model can read meaning in context rather than treat each word as a clue on its own. Models such as XLM-RoBERTa can do this across many languages. They are not an oracle. They do not replace trust, sender checks or feedback. They add a better reading of the message to the stack.

The bottleneck is now attention

AI can help one person draft a report, write code or plan a campaign faster. It cannot add another hour to the reader’s day. As tools make more, picking the right work matters more. The next work problem is not only how to make more. It is deciding which results deserve a person.

Spam taxes attention at the worst time: before you know the message is worthless. Work you choose can wait. Spam demands a small choice now. It makes the reader review the sender’s work for free. Then the reader must find the thread of what mattered before.

This is also why the score alone is not enough. Spam that gets through spends attention. Good mail sent to junk hides something that earned it and harms trust in the filter. A useful system has to be firm with clear noise, careful with wanted mail and able to learn the gray zone for each user.

A filter decides what earns a human

The best defense still has layers. Stop known sources of abuse before they send mail. Check that real senders are who they claim to be. Use modern models to read the text. Learn from the person who owns the inbox. Most of all, decide before an alert, preview or unread count makes that person decide instead.

Klar is built around that principle. A multilingual XLM-RoBERTa model classifies each message on your Mac, inside Apple Mail. The message is not sent to a cloud AI service. The goal is not to make the inbox intelligent for its own sake. It is to make less of the inbox demand you.

All posts