AI phishing detection means using a trained machine learning model - not just fixed rules - to judge whether an email is fraudulent, and the biggest practical decision in that architecture is where the model runs. On-device AI phishing detection processes the email locally in your browser and nothing leaves your machine; cloud-based AI phishing detection sends the email content to a remote server for classification. Both can be accurate. Only one guarantees your inbox never touches someone else's infrastructure.
What AI phishing detection actually means
Most "phishing detection" on the market is still rule-based: check the sender domain against a blocklist, check whether a link's visible text matches its destination, check for urgency phrases. Those checks are fast and explainable, but they only catch what someone thought to write a rule for.
AI phishing detection adds a second layer: a model trained on thousands of real phishing and legitimate emails learns the statistical shape of a scam - word choice, structure, tone - and applies that learned pattern to a new email it has never seen. It doesn't need an exact keyword match. That's the entire value proposition, and it's also why the underlying architecture (where the model actually runs) matters more than most marketing copy admits.
Phishing itself covers a wide range of attacks, from mass-mailed scams to narrowly targeted spear-phishing, but the AI detection layer we're describing here is specifically about classifying the email vector - not website spoofing or voice-based social engineering, which use different detection methods entirely.
Why the architecture matters more for email than for other AI use cases
Email is uniquely sensitive content - invoices, account numbers, medical correspondence, legal threats, exactly the kind of thing a scam message impersonates. Sending that content to a cloud API to check if it's fraudulent means trusting a second party with the very data a phishing attacker was trying to steal in the first place. On-device processing sidesteps the question entirely: there's no server log to breach, no third party to trust, because the content never left your machine to begin with.
This is also why we treat Proton Mail users and cryptocurrency exchange users as a priority audience, not an afterthought. Both groups already made a deliberate choice to avoid handing their data to a third party. A phishing detector that quietly routes their email through a cloud API undoes that choice without asking. That's the stake behind the architecture decision below - not an abstract preference, but where your actual inbox content ends up.
On-device AI vs cloud AI phishing detection
The label "AI-powered" tells you almost nothing about where your email data actually goes. The architecture underneath is the real differentiator, and it splits into two models with genuinely different trade-offs.
| Aspect | On-device AI | Cloud AI |
|---|---|---|
| Where analysis runs | In your browser, via WebAssembly | On a remote server |
| Email content leaves the device | No | Yes, for every scan |
| Works offline | Yes, after the initial model download | No, requires a live connection every time |
| Latency | No network round trip | Depends on server response time |
| Model updates | Shipped with extension updates | Can be updated centrally, instantly |
| Best fit | Privacy-conscious individuals, crypto and Proton Mail users | Enterprise SOC teams needing central visibility across many accounts |
Cloud AI isn't wrong for every use case - a security team monitoring thousands of corporate mailboxes genuinely needs centralized visibility. But for an individual checking their own inbox, sending private email content to a third-party server to find out if it's a scam is a strange trade to make, especially when an on-device model can answer the same question locally.
How our on-device AI Deep Scan works
We built AI Deep Scan as a second opinion layered on top of heuristic checks, not a replacement for them. It runs a DistilBERT model - a distilled, lighter version of BERT - trained specifically for phishing detection, and it scores 99.58% in our test set. The model downloads once, roughly 67 MB, which takes 10 to 30 seconds depending on your connection. After that first download it's cached in your browser and runs completely offline through WebAssembly - no server call, no API key, no account.
When you click "Deep Scan with on-device AI" on a flagged or borderline email, the model reads the subject, sender, and body, and returns a verdict (safe, suspicious, or scam), a confidence percentage, and a two-to-three sentence explanation of its reasoning. It isn't reading the email blind, either - it also sees what the heuristic checks already found, so a suspicious link or spoofed sender factors into its own read of the message.
What AI catches that heuristic rules miss
Rule-based checks are precise but brittle. A rule that flags "your account has been suspended" catches that exact phrasing and nothing structurally similar written a different way. A trained model generalizes - it can flag a message with unfamiliar wording that still follows the statistical shape of a scam, because it learned the pattern rather than the phrase.
The reverse is also true, which is why we never let AI operate alone. We combine the two verdicts with one fixed rule: a strong heuristic finding - like an anti-phishing code mismatch, or several high-severity red flags - can't be overridden by an AI verdict of "safe." The result stays at least "suspicious" in that case. On the other hand, AI can raise a verdict to "suspicious" on its own if something reads wrong, but it can never push a verdict all the way to "scam" without the heuristic checks agreeing. Neither system gets a blind veto over the other.
What people get wrong about AI phishing detection
The most common misconception is that "AI-powered" automatically means smarter or safer than rule-based detection. It doesn't - it means different. A model is only as good as what it was trained on, and a model trained mainly on consumer phishing (fake shipping notices, fake account alerts) can miss a narrowly targeted business email compromise attempt that never uses a suspicious link at all, just a convincing request from someone posing as a colleague.
The second misconception is that on-device necessarily means less capable. Distillation techniques - shrinking a large model like BERT into a smaller one like DistilBERT while preserving most of its accuracy - exist specifically to close that gap. Academic work on browser-resident phishing detection has been moving in the same direction: lightweight transformer models that run in the browser instead of a remote classifier, precisely because latency and privacy both favor keeping the model local.
Limitations worth knowing before you rely on it
No phishing detection system - AI or rule-based - catches everything, and pretending otherwise does readers a disservice. AI Deep Scan is trained mainly on consumer phishing patterns, so it can miss a narrowly targeted business impersonation email that reads like ordinary internal correspondence. When our AI verdict disagrees with the heuristic findings - AI says safe, heuristics found a real red flag - we surface that disagreement explicitly and recommend trusting the heuristic result, rather than letting the AI's confidence quietly override it.
Phishing volume also keeps climbing, which is the reason detection architecture matters at all. The Anti-Phishing Working Group's Phishing Activity Trends Report has repeatedly logged phishing volumes in the range of 800,000 to over a million unique attacks in a single quarter - a scale no manual review process can keep up with, which is exactly the gap both heuristic rules and AI models exist to close.
Once you understand how AI phishing detection compares to older heuristic-only methods, the natural next question is what a phishing email actually looks like in practice - our guide on how to spot a phishing email breaks down the exact red flags, with real examples.
Final verdict - AI phishing detection
AI phishing detection is a genuine improvement over rules alone, but the architecture decision - on-device or cloud - matters as much as the model itself. On-device AI keeps your email content on your machine and works offline after setup; cloud AI trades that privacy for centralized management that mainly benefits enterprise teams, not individual inboxes. If you're checking your own email, on-device is the fit that doesn't ask you to trust a third party with the messages you're trying to protect. If you want to see the full picture of how detection methods stack up against each other, our best phishing protection comparison covers heuristics, AI, and manual review side by side.