Artificial Intelligence (AI)

Artificial intelligence (AI) describes systems that perform tasks normally requiring human judgement — classifying, predicting, generating — by learning patterns from data rather than by following rules a person wrote out in advance.

Email was one of the first mass applications of machine learning, and it remains one of the clearest: spam filtering has used statistical classification since the early 2000s, long before the term AI came back into fashion.

How is AI used to filter spam?

Spam filtering is a classification problem, and it has been solved with progressively more elaborate models:

• Bayesian filters. The original approach, still in use in SpamAssassin and rspamd. The filter learns the probability that a message is spam from the tokens it contains, trained on the messages a user marked one way or the other. It is cheap, explainable, and adapts per-user.
• Supervised classifiers on engineered features. Rather than words alone, the model sees header structure, URL properties, sending patterns and authentication results. This is what most commercial filters run today.
• Neural models. Used mainly by large mailbox providers, for phishing detection in particular, where the signal is in the phrasing and the visual presentation rather than in individual keywords.

The economics are unusual. False positives cost far more than false negatives — a missed spam is an annoyance, a blocked invoice is a business problem — so production filters are tuned much more conservatively than raw accuracy figures suggest.

How is AI used for IP and domain reputation?

Reputation systems consume behavioural data about a sender — volume over time, ratio of valid to invalid recipients, spam trap hits, complaint rates, authentication results, the behaviour of neighbouring addresses in the same range and the same ASN — and produce a score that receiving systems use to accept, defer or reject.

Two properties of these systems matter to anyone running a mail server:

1. They are predictive, not just historical. A brand-new IP address with no history is not neutral; it is scored by the properties it shares with other addresses, which is why warming up a new sending IP is necessary at all.
2. They are opaque. A model-derived score cannot be traced back to a single rule, which is why “why was my mail rejected?” so often has no better answer than a generic 5xx and a postmaster page.

Where else does AI appear in email systems?

• Phishing and BEC detection, where the model looks for anomalies in a relationship — a first-time sender asking for a payment change — rather than for bad content.
• Natural language processing for routing, summarising and drafting replies in helpdesk and shared-mailbox tooling.
• Anomaly detection on authentication logs, flagging a mailbox whose sending behaviour has suddenly changed, which is usually the first evidence of a compromised account and of the abuse that follows.

What are the limits?

AI-based filtering is probabilistic, and that has consequences the marketing rarely mentions. Decisions are hard to explain, which makes disputes hard to resolve. Models drift as sending behaviour changes and need retraining. Adversaries adapt deliberately, since spam is one of the few classification problems with an opponent actively optimising against the classifier. And a model trained mostly on English commercial mail will behave differently on a German-language transactional stream.

None of this argues against using them. It argues for keeping deterministic controls — authentication checks, explicit blocklists, rate limits — underneath the probabilistic layer, so that the outcomes you care most about do not depend on a score.

Note on MailMum

The techniques above describe the broader email industry, not MailMum specifically. MailMum blocks or passes a connection based on explicit blacklist/whitelist rules an admin defines, not a machine-learning reputation score — every decision is traceable to a specific rule, logged in the audit trail, rather than an opaque model output.