How TrustLens Shield Works
What our AI can do — and just as importantly, what it can't. We'd rather you know the limits than trust us blindly.
Measured accuracy
We tested 34 real and fake scam messages, labelled by hand, against the live detection engine. Measured on 19 August 2026. This is not a marketing estimate — and it's not the same as the quick automated check we run for every code change either (that one uses a simplified stand-in model for speed, not the real one).
Precision
Of the content we flagged as a scam, how often we were right.
95.8%
target: above 90.0%
Recall
Of the actual scams in our test set, how many we caught.
100.0%
target: above 90.0%
Registry verification accuracy
Of registration numbers we could check against an official registry, how often our Verified/Unverified/Unauthorized call was correct.
100.0%
target: above 99.0%
Recall meets our own target in this test set, but no detector catches everything. We publish this number rather than a qualitative claim because it directly affects how much you should rely on a "low risk" verdict: treat it as one input, not a guarantee.
What "Unverified" means — and why we never say "safe"
A registration check can only ever return one of three answers:
- Verified — a positive, current, exact match on an official regulatory registry. This is the only case where we affirm anything.
- Unauthorized — a positive match on a regulator's own warning list.
- Unverified — everything else: no match, a registry that's unreachable or stale, or an ambiguous result. This is the default whenever we're not sure.
We never invent a fourth answer like "probably fine" or "safe." A green verdict always means a positive match we actually found — never the absence of a red flag.
Data sources
Every registration or warning-list check draws on one of the regulator sources below. "Live" means we fetch directly from the regulator; "synthetic" means we ship a representative sample while a live connector is still pending — never presented as real regulator data.
What's not covered
We cover 11 jurisdictions' primary securities regulator. We do not cover banking/deposit-taking regulators, insurance regulators, commodity/derivatives-only regulators, cryptocurrency-specific regulators, or any jurisdiction not listed above. A registration genuinely held with an uncovered regulator or jurisdiction will show as Unverified, not Verified — see "What Unverified means" above.
Deepfake detection is advisory only
When we scan a video or image and don't find signs of manipulation, that is not proof the content is authentic — deepfake detection is not foolproof, and a well-made fake can pass undetected. We always combine this signal with the text and regulatory checks; a clean media scan on its own never clears content that's flagged elsewhere, and it never overrides a registry match.
Platform availability
An installable Android app is available as a native build, alongside full access from any web browser.
There is no native iOS app. iOS is covered by this installable Progressive Web App instead — open this site in Safari, then Share → Add to Home Screen. This is a deliberate reach decision, not an unstated gap.
iOS Safari has historically placed tighter limits on background and offline behaviour for installed web apps than Android — treat offline features on iOS as best-effort, not guaranteed.
What this AI cannot do
- It misses real scams — see the recall figure above. A "low risk" result is not a certificate of safety.
- It can only verify a registration number against the registries we've connected — an entity genuinely registered somewhere we don't cover will show as Unverified, not Verified.
- It reads text extracted from images and audio, and that extraction can itself fail or be imperfect — a garbled extraction can lead to a garbled analysis.
- It is not financial, legal, or investment advice, and it does not replace verifying important claims directly with the official regulator or your bank.
- Its accuracy for languages beyond Hindi and English has not been separately measured — treat results in other languages with extra caution until we publish numbers for them too.