Skip to content

Sensity AI

EnterpriseInstitutional

Vendor: sensity.ai | Type: Detector | Cost: Enterprise

TL;DR

The institutional-grade deepfake detection platform deployed by Rappler and the #FactsFirstPH coalition on the Doc Willie Ong eye-drop deepfake and the Brawner "Dark Eagle" deepfake; produces confidence scores, heatmaps, and forensic reports across video, audio, and image. Primary 1C.2 entry, with a documented vendor- versus-casework gap of roughly twenty percentage points.

What it does

Sensity ingests video, audio, and image inputs and produces multilayer analysis output: pixel-level facial-manipulation scores, acoustic deepfake probabilities, metadata reconciliation, and a behavioural component that scores anomalies in motion and lip-sync. The platform exposes its results as confidence scores plus heatmaps and as forensic-report artefacts intended for evidentiary use. The documented Rappler casework returns three separate scores per item when a deepfake combines facial manipulation, AI-generated objects in the visual frame, and audio manipulation: for the Doc Willie Ong eye-drop deepfake, Sensity returned 98% facial-manipulation, 75.5% AI-object-generation, and 94% audio-manipulation .

The platform aggregates several internal models in producing those scores; under Architectural Anchor 3 the toolkit treats the entire Sensity output for a single item as one detector signal class, not as multiple independent signals. The same wrapping rule applies when a fact-checker runs Sensity in parallel with Hive at 1A.1 or with InVID-WeVerify's deepfake tab at 1B.1: the multiple detectors together still count as one detector signal class for the purposes of Anchor 2's editorial floor on publishable claims.

When to use it

  • A Rappler-grade investigative pipeline needs to produce forensic- quality scores on a public-figure synthetic-content case where the case will be publicly debunked under the newsroom's name.
  • A defamation-defence file needs a vendor-supplied forensic report that a court or platform-appeals process will recognise as institutional-grade output.
  • A multimodal case (face + AI-generated object + manipulated audio) needs separate per-modality scores rather than a single composite verdict.
  • An institutional partner is funded specifically for the licence cost and is producing analysis that a regional fact-check coalition can cite.

Independent accuracy

Vendor claim vs independent assessment

Vendor claim: Sensity markets a headline of approximately 98% accuracy and the Rappler workflow returned 98% facial-manipulation, 75.5% AI-object-generation, and 94% audio-manipulation on the Doc Willie Ong eye-drop deepfake

.

Independent finding: in the documented Brawner "Dark Eagle" deepfake (a separate political-impersonation case from the same actor pool covered by Rappler / #FactsFirstPH), the same Sensity workflow returned 79.3% AI-generated, roughly twenty percentage points below the headline figure . The toolkit's editorial position is that the Brawner figure is the more operationally honest reference point: it is what Sensity actually returns on contested SEA political content rather than on a vendor-favourable demonstration.

No independent SEA-specific benchmark identified as of May 2026. Sensity is included in the Deepfake-Eval-2024 anonymised commercial pool but no provider-level score is publicly separable from that benchmark; the best commercial video detector in that pool reached only 0.78 accuracy / 0.79 AUC on 2,036 in-the-wild 2024 deepfakes, which is a category-level upper bound rather than a Sensity-specific number . The 98% headline should therefore be read as the Sensity ceiling on favourable inputs and the 79.3% Brawner figure as a documented field reading.

Limitations

Limitations

  • Vendor headline of 98% accuracy stands in tension with the same operation returning 79.3% on the Brawner case (verbatim retention required by C2 limitation faithfulness rule).
  • Enterprise-only access; high cost barrier.
  • No independent tool-specific benchmark publicly available; Sensity sits in the Deepfake-Eval-2024 anonymised commercial pool where only the pool's best score (0.78) is public

. - Performance is not stress-tested on the full SEA face distribution or on the codec-compressed material common in WhatsApp / Facebook / TikTok pipelines. - Black-box proprietary platform; the score breakdown is vendor-defined and not openly reproducible at the level of academic benchmarks.

Privacy and threat model

Sensity is a hosted enterprise platform: an analyst submits the suspect file to Sensity's cloud or on-premises deployment for processing. The analyst's organisation is bound by the contract terms governing data retention, access logging, and disclosure. For routine work on already-public material (a viral TikTok clip, a screenshot from an open Facebook page), the upload is low-risk because the content is already public and the chain of custody is documented through the published source.

For source-identifying material (a clip provided by a whistleblower, audio with diagnostic background, or anything where the suspect file itself names a source), the upload is the risk regardless of the verdict that comes back. Decision 7's paired honest-gap policy and the Architectural Anchor 1 framing apply: in surveillance-environment contexts (Sri Lanka, Laos), classify the file before the upload, and where the classification is source-identifying, route to an offline forensic alternative (Sherloq) and through an institutional partner outside the surveillance jurisdiction rather than uploading directly.

Source-protection override (S1 — source-identifying upload risk)

Uploading source-identifying frames or audio to Sensity exposes the source to a vendor cloud and to the contract-governed retention and disclosure regime. Mitigation steps:

  1. Classify the suspect file as public, sensitive, or source-identifying before any upload.
  2. For source-identifying material, do not upload to Sensity directly. Route the case through an institutional partner outside the surveillance jurisdiction, or use offline forensics (Sherloq) for the modules that can be run locally.
  3. If the case requires Sensity's specific output, strip identifying context (crop, blur, redact audio) on the analyst's machine before submission of the redacted version.
  4. Document the chain of custody for the redacted file in case the original is later required for evidentiary purposes.

Source-protection override (S9 — cross-border data transfer)

Sensity is a vendor-cloud platform with contract-governed retention. Uploads cross into the vendor's jurisdiction regardless of where the analyst sits. Mitigation steps:

  1. Read the active contract's retention and disclosure terms before any Sensity submission; record the version with the case file.
  2. Confirm that organisational data policy permits the upload under the contract terms; some newsroom and partner agreements restrict cross-border transfer of sensitive material.
  3. For surveillance-risk countries, prefer the institutional- partner submission path so the analyst is not the named account holder on the platform.

Country and platform applicability

  • Indonesia: language-agnostic; relevant in principle for the Prabowo / Sri Mulyani 2025 deepfake cluster where institutional-grade detection would have applied; no documented Indonesian newsroom direct deployment.
  • Laos: language-agnostic for image and video; the cloud-upload privacy posture is wrong for a Lao surveillance- environment workflow without partner-mediated routing.
  • Malaysia: language-agnostic; relevant for the King / Anwar Ramadan-aid 2026 cluster context if institutional access is available.
  • Philippines: primary documented use. Rappler / #FactsFirstPH workhorse on the Doc Willie Ong eye-drop deepfake (98% / 75.5% / 94%) and Brawner "Dark Eagle" (79.3%); part of the Rappler / #FactsFirstPH multi-tool pipeline alongside Hive AI and InVID-WeVerify .
  • Sri Lanka: language-agnostic for visual modules; surveillance-environment routing through institutional partners is the right path; direct Sri Lankan newsroom use not documented.
  • Thailand: language-agnostic; AFP regional partnership may include institutional access; Thai PBS uses SynthID Detector as the documented provenance-first path on the Anutin / Mauerberger case rather than Sensity.

Platform applicability: works on any video, audio, or image file the user can submit; relevant across Facebook, Facebook Groups, TikTok, YouTube, WhatsApp (forwarded clips). Documented Rappler use covered Facebook- distributed political-impersonation content.

How to access

Enterprise sales contact via the Sensity website. The platform is not available in a free or self-service tier; access requires a commercial agreement. Cloud deployment and on-premises options exist; the on-premises route is the path used by national-security or law-enforcement institutional partners with strict data-residency requirements.

Cost (current as of 2026-05)

Enterprise pricing in the $50,000 to $200,000-plus per year range . Specific numbers move with the vendor's pricing; verify through Sensity's sales process before committing to a procurement. The bracket is included to make the structural reason for Sensity's institutional-tier placement concrete: at this cost, the tool is realistic for broadcaster-grade or coalition-grade operations (Rappler /

FactsFirstPH, AFP regional partnerships) and out of reach for the

typical individual SEA newsroom unless the budget arrives through a specific institutional grant.

Quickstart

  1. Confirm institutional access. The realistic path for a frontline SEA newsroom is partner-mediated rather than direct.
  2. Classify the suspect file per the danger admonition above before any upload.
  3. Submit the file through Sensity's API or web interface as the licence agreement specifies.
  4. Inspect the per-modality scores: facial-manipulation, AI-object-generation, audio-manipulation, plus the heatmap visualisation.
  5. Treat the multiple modality scores as one detector signal class per Anchor 3, not as multiple independent confirmations.
  6. Pair the Sensity output with at least one non-detector signal class (a reverse-image hit, a provenance manifest, an archived-version search, a tipline-database match) before any publishable claim, per Anchor 2.
  7. Retain the forensic report artefact for chain-of-custody if the case may face a defamation defence or a platform appeal.

In the toolkit's workflow

Cautious-detector pillar tool per Architectural Anchor 1. Sits in 1C.2 Institutional deepfake platforms as primary, ahead of Reality Defender (alternative for broadcaster-grade deployments) and DeepfakeBench (academic evaluation harness, not a deployable detector).

Standard combinations:

  • With Hive AI at 1A.1 / 1A.3 as the documented Rappler / #FactsFirstPH pipeline; the two detectors together count as one detector signal class under Anchor 3, not as two.
  • With InVID-WeVerify at 1A.2 / 1B.1 for source- history work alongside Sensity's detector pass.
  • With Meedan Check at 2B.1 for the #FactsFirstPH coalition's claim-database backbone.
  • With TRIED Benchmark at 1C.3 for sociotechnical evaluation of Sensity's output on the same case before the finding is published.

This tool's verdict is one weak signal class per Anchor 1; never publish a binary claim from it alone. Counted as one detector signal class under Anchor 3, regardless of how many of Sensity's internal modality scores contribute to the verdict.

Decision-tree references: T5 escalation escalates to Sensity at T5.6 (image professional) and T5.7 (video professional) when the case warrants forensic-grade scoring. T6 source-protection routes the upload through the S1 sub-routine before any submission.

Conflict resolution behaviour

When Sensity's verdict disagrees with another detector (Hive, Deepware, InVID deepfake tab), Sensity carries equal-or-stronger evidence weight at the institutional tier on documented Rappler casework but no stronger weight than its own field reading: the 79.3% Brawner figure is what Sensity actually returns on contested SEA political content and is the operational reference point. When the Sensity verdict disagrees with a non-detector signal (provenance manifest, reverse- image hit, archived earlier instance, or a tipline-database match), the non-detector signal wins because Anchor 2 mandates two non-detector signals before any strong public claim. Inside an institutional pipeline, Sensity's forensic-report artefact carries chain-of-custody weight that Hive's free-tier output does not, but that weight applies to the report's evidentiary status, not to the underlying probability.

Override notes

Override notes

  • Multi-detector wrapped as one class (g3-multi-detector- wrapped-as-one-class): Sensity's internal modality scores (face, object, audio) and any parallel detector run (Hive, InVID's tab) together count as one detector signal class under Anchor 3. Workflow section above invokes this rule explicitly.

  • Vendor wrapping pair (c1-pair-vendor98-vs-Brawner79.3): the 98% headline against the Brawner 79.3% field reading is rendered in full in the Independent accuracy admonition above, with the Deepfake-Eval-2024 anonymised-pool framing surfaced for context.

  • Verbatim Brawner counter-example (c2-verbatim-Brawner- counter-example): the Brawner 79.3% figure is preserved verbatim in Limitations rather than softened or merged into the headline figure.

  • Enterprise pricing (d1-score-1-enterprise): D1 score of 1 rendered concretely in Cost as the $50K-$200K+ per year range; this is the structural reason the tool sits at institutional tier rather than at PV or FLT.

  • Cloud-upload mitigation (d4-mitigation-pass-cloud-upload- S1): the S1 source-protection sub-routine is bound into the danger admonition above, with concrete pre-upload mitigation steps for surveillance-environment work.

Sources

  • Sensity. Sensity AI — forensic deepfake detection platform. Sensity (Amsterdam), 2024. sensity.ai.
  • Sensity. Sensity Platform — detection product page. sensity.ai/platform.