AI Fixed the Typos. Phishing Training Needs to Catch Up.

For twenty years the standard phishing advice was to watch for bad spelling and awkward grammar. Real companies write properly, the reasoning went, so a message full of errors is probably fraudulent. It was easy to teach and for a long stretch it genuinely worked.

It does not work now. Attackers write with AI, and AI writes cleanly. The broken phrasing that used to give phishing away has disappeared, and the messages reaching your team read as well as anything from a legitimate supplier. They can also be written to match the voice of someone the recipient knows.

The consequence is worse than the messages simply being better. Your staff are still scanning for a signal that no longer exists, so a clean, well-written email now reads as reassuring. The old training has inverted into a liability.

Why it stopped working

The spelling tell worked because many operators were writing in a second or third language, and it showed. That constraint is gone.

The UK’s National Cyber Security Centre has stated plainly that generative AI now produces credible lures stripped of the translation and language errors that historically exposed them. The FBI has published the same assessment. Two agencies, arriving independently at the conclusion that the tell most people were trained on has expired.

What changed beyond the grammar

Three things, and the second should concern you most.

The writing is clean, because a machine produced it in seconds in whatever register the operator requested.

The messages are specific. An attacker can feed public information about your company into a model, scraped from your website, LinkedIn profiles, a press release, and get back something tailored: correct names, correct titles, a plausible reason for the contact. The generic “your account is suspended” has given way to a note to your bookkeeper that appears to come from a supplier they genuinely work with, references a project that genuinely exists, and asks to update banking details for the next invoice. The only false thing in it is who sent it.

And there is more of it, because the cost per message collapsed. IC3 gave AI its own section in the bureau’s annual report for the first time, attaching it to upward of 22,000 complaints and roughly $893 million in losses.

Your filter will not close this

Email security catches a great deal and should stay on. But consider what the filter has to work with. A well-written, personalized message asking a reasonable business question, with no malicious link and no attachment, gives it very little to object to. The payload is a request, and requests are what business email consists of. Both agencies expect more of these to reach inboxes, which puts the final check on a person who knows what to look at.

The signals that survived

If the writing no longer tells you anything, look at what the message is asking you to do. Those signals have not changed, because they are structural rather than stylistic.

Be alert when a message wants money, gift cards, or payment routed to an unfamiliar account. When it wants credentials, a one-time code, or personal information. When it manufactures time pressure. When it proposes changing banking details on an invoice or supplier record. When it carries an unexpected link or attachment. Or when the display name reads correctly but the underlying address does not match it.

Every one of those concerns the request rather than the prose. So the rule to hand your team fits in a sentence: when a message touches money, credentials, or how you pay someone, slow down before acting.

It is not confined to email

AI has done the same to voice. The FBI has cautioned that a usable clone can be built from a short audio sample, enough to leave a voicemail in the register of an executive or a relative pressing for an urgent payment. The countermeasure does not change: any call asking for money or access gets ended, and you dial back on a number you sourced yourself.

What to actually change

Stop telling staff to watch for bad spelling, and point them at the request instead.

Establish one non-negotiable rule for payment changes: every modification to banking details is verified by phone, on a number you sourced yourself, regardless of how urgent the message claims to be. That single control defeats most business email compromise.

Deploy phishing-resistant MFA or passkeys, so a stolen credential is harder to use. Make reporting easy and socially costless, because the person who checks and turns out to be wrong needs to feel fine about it. And revisit it periodically, since five minutes in a team meeting outperforms a poster nobody reads.

Getting your team current

MVP Network Consulting helps businesses across Buffalo and Western New York update staff training for how phishing actually looks now, and puts the technical controls behind it. If your last security briefing told people to look for typos, it is out of date. Get in touch.