Transparency
How DiReFTY uses AI
AI finds it. People handle it. This page states plainly what DiReFTY's detection engines actually do - and, just as important, what they are deliberately not used for.
TL;DR
- Detection runs on AI - review authenticity scoring, impersonation matching, domain threat classification, and cross-channel correlation.
- Enforcement does not. Evidence packaging, platform routing, and escalation are handled by people.
- Model output is a prioritised signal, evidence-first, never an automatic verdict - a person reviews every finding before a platform is approached.
- No accuracy percentage is published without a methodology behind it, and no model name or signal weighting is published at all - competitive and confidentiality-sensitive detail that adds nothing a client can act on.
What We Use AI For
Review authenticity scoring
Every review is assessed on behavioural signals - timing clusters, reviewer history, language repetition, rating distribution anomalies - combined with a model-based authenticity judgement, weighted differently per business category.
Impersonation matching
Handle permutations, display-name variants, and copied bios are matched across platforms. A model then reads each candidate profile and judges whether it is actually impersonating you, or just shares a name.
Domain threat classification
Newly registered lookalikes are graded on infrastructure signals, page content, and visual similarity to your real site, then classified by threat type with the supporting evidence recorded alongside the verdict.
Cross-channel correlation
A counterfeit listing, an impersonation account, and a lookalike domain are often one operator. Signals are correlated across channels so a threat surfacing in one place is visible everywhere it spreads.
What We Do Not Use AI For
- Decide what to report, or to whom. Filing under the wrong policy ground gets a report dismissed, and a rejected first report makes the second one harder - that call is made by a person, every time.
- File a report, send a legal notice, or contact a platform on its own. Every enforcement action is initiated and reviewed by a person.
- Give legal advice, or make a defamation or infringement determination. That is a human judgement call, made with the evidence the detection layer surfaced.
- Produce a final verdict. Model output is a prioritised signal - evidence-first, scored on supporting evidence rather than name similarity alone - never an automatic yes or no.
- Predict a threat before it exists. Detection finds what is already there, early. It does not forecast what an attacker will do next.
No model does that reliably, and we do not pretend otherwise. Every AI claim on this page is one we could defend in a demo call with a technical buyer who asks "show me."
How We Handle Model Error
Every finding is evidence-first - scored on supporting evidence rather than name similarity or pattern-match alone - and a person reviews it before anything is reported to a platform. That review step is what catches a false positive before it becomes a wasted or wrongly-filed report, and what catches a false negative before it goes unmonitored. Detection speed is the control here, not a claim of perfect accuracy.
Why Detection Is Automated, and Enforcement Is Not
A brand's exposure is spread across thousands of listings, handles, and domain permutations, most of which are noise - that is a scale problem, and AI is genuinely good at scale problems. Deciding which policy ground to file under, what evidence to attach, and when to escalate a rejected report is a judgement problem, and platform reporting outcomes turn on getting that judgement right. DiReFTY runs each half on the side built for it. See the Detect → Respond → Enforce → Protect framework for how the two connect.
FAQs
AI questions
Which parts of DiReFTY actually run on AI?
- Four detection functions: review authenticity scoring, impersonation matching across platforms, domain threat classification, and cross-channel correlation. Each combines behavioural or infrastructure signals with a model-based judgement.
- Nothing outside detection runs on AI. Evidence packaging, platform routing, escalation, and any client-facing decision are handled by people.
Why isn't enforcement automated too?
- Because a wrong call has consequences a model cannot be held accountable for - the wrong policy ground gets a report dismissed, and a rejected first report makes the second one harder. Platform reporting outcomes depend on judgement, not just pattern-matching.
- Detection is the part AI is genuinely good at: scaling across volume no team could sweep manually. Judgement about what to do with a finding is not.
What happens when the model gets it wrong?
- Model output is a prioritised signal, never an automatic verdict. Every finding is evidence-first - scored on supporting evidence, not name similarity or pattern-match alone - and reviewed by a person before anything is reported to a platform.
- That review step is what catches false positives before they turn into a wasted or wrongly-filed report.
Does DiReFTY publish its accuracy numbers or model details?
- No, and we're explicit about why: an accuracy percentage without a published methodology is not a verifiable claim, and model names or exact signal weightings are competitive and confidentiality-sensitive detail that adds nothing a client can act on.
- What's published instead is what each detection function actually does and does not do - that's the standard this page holds itself to.
See what the detection layer finds for your brand.
Get Your Free Digital Risk Assessment - a focused review of your exposure, active threats, and response options.