Skip to content

TrueMedia.org (Georgetown McCourt School)

FreeProfessional

Vendor: truemedia.org | Type: Detector + non-detector | Cost: Free

TL;DR

The revived TrueMedia.org, now hosted at Georgetown University's McCourt School of Public Policy, is a non-partisan ensemble deepfake detector with a mission of radical transparency about detection accuracy and uncertainty. Currently in closed beta; request access at truemedia.org. The toolkit's escalation option for institutional partners on audio, video, and image deepfake casework.

What it does

TrueMedia accepts uploaded social-media content (image, audio, video; text veracity to come) and runs the input through an ensemble of AI detectors and human verification, returning a deepfake-probability verdict alongside transparent reporting of accuracy metrics, uncertainty levels, and known limitations. The original TrueMedia, founded by Dr Oren Etzioni, shut down in January 2025; Georgetown's McCourt School of Public Policy revived it in April 2025 through the Massive Data Institute and the Tech & Public Policy program, where Dr Lisa Singh and Michelle De Mooy oversee the initiative.

The toolkit ships TrueMedia as the escalation-option entry in 1B.3 audio deepfake forensics for two reasons. First, the ensemble approach (multiple AI detectors plus human verification) is closer to the toolkit's own architectural anchors than any single-model detector: every decision determined by a model is checked by a human, and the platform reports uncertainty, not a single binary verdict. Second, the mission framing of radical transparency (no claim of perfect accuracy; reliability measured, documented, and continuously evaluated) matches the toolkit's own detector-as-weak-signal commitments and its honesty about SEA-language audio detection limits.

The closed-beta access status is the operational constraint. Until beta access is granted, TrueMedia is a referenceable institutional option, not a deployable tool for SEA newsrooms.

When to use it

  • An institutional partner with closed-beta access has a high-stakes deepfake case (audio, image, video) and wants ensemble verdict plus documented uncertainty, not a single-vendor score.
  • A research collaboration with Georgetown's MDI / TPP needs a deepfake-detection backbone that ships with transparent methodology for academic citation.
  • A regional partner with IFCN signatory status or formal academic affiliation can credibly request beta access for a documented case pipeline.
  • A workshop demonstrates the difference between vendor-style closed-source detector products (Sensity, Reality Defender) and a mission-driven open-methodology platform. TrueMedia is the cleanest demonstration available.

Independent accuracy

Vendor claim vs independent assessment

Vendor claim: TrueMedia explicitly does not claim perfect accuracy ("No detection system achieves perfect accuracy. We do not claim otherwise.") and frames itself as reporting performance metrics, uncertainty levels, and known limitations, not headline numbers.

No independent SEA-specific benchmark identified as of May 2026. The closed-beta status limits public independent evaluation; TrueMedia's own self-reporting commits to publishing performance and uncertainty metrics, but no third-party SEA- language audit has been published in the toolkit's research base. Verify directly when beta access is granted.

Limitations

Limitations

  • Currently in closed beta (May 2026); not deployable for SEA newsrooms without granted access. Verify access status before relying on this tool in any case (see truemedia.org).
  • Ensemble approach inherits the limitations of its constituent detectors; per Anchor 3, multi-detector wrappers count as one detector signal class regardless of how many internal models contribute to the verdict.
  • SEA-language coverage and SEA-specific audio performance are not independently documented; the platform's radical-transparency commitment is to publish metrics, but the disclosed metrics base is not yet SEA-specific.
  • Original TrueMedia shut down in January 2025; the Georgetown revival relies on continued institutional commitment from the McCourt School. Bus-factor risk is institutional, not single-developer, but it is real.

Privacy and threat model

The detection portal at detect.truemedia.org and the public site at truemedia.org operate on Georgetown University infrastructure (truemedia.georgetown.edu per the privacy policy). Uploads route through Georgetown's data-handling practices. Verify the current data-flow practices in the privacy policy at the time of access; the policy was last updated 11 March 2026.

For source-identifying material, the upload itself is the risk regardless of the verdict. In Lao and Sri Lanka surveillance work, classify the source file before any upload and document the disclosure.

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

Submitting source-identifying frames, audio, or video to the TrueMedia portal routes the file to Georgetown University infrastructure under the published privacy policy. Mitigation steps:

  1. Classify the suspect file as public, sensitive, or source- identifying before any submission.
  2. For source-identifying material, do not invoke TrueMedia directly. Route the case through an institutional partner outside the surveillance jurisdiction, or use the source- history modules of InVID-WeVerify first.
  3. If a TrueMedia verdict is operationally necessary, strip identifying context (crop, blur, redact audio) on the local machine before submission of the redacted version.
  4. Document the submission in the case record per the closed- beta access conditions.

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

TrueMedia operates from US jurisdiction on Georgetown University infrastructure. Mitigation steps:

  1. Check the active privacy policy at time of access (last updated 11 March 2026); record the version with the case file because policy text drifts.
  2. Confirm organisational policy permits the upload under US university data-handling terms.
  3. Prefer EU-jurisdiction alternatives (InVID's deepfake tab via CERTH for redacted material with documented 30-day retention) when the source is in a surveillance-risk country.

Country and platform applicability

  • Indonesia: no documented Indonesian newsroom deployment recorded; tool applies in principle if beta access is granted to a partner like CekFakta or MAFINDO.
  • Laos: no documented use; ensemble approach language-claimed but not Lao-tested.
  • Malaysia: no documented use; ensemble approach language-claimed.
  • Philippines: no documented Rappler / VERA Files deployment recorded; the closed-beta status is the operational constraint.
  • Sri Lanka: no documented use; ensemble approach language-claimed but not Sinhala or Tamil tested.
  • Thailand: no documented use; tonal-language performance is the explicit "no evidence" gap from the same Deepfake-Eval-2024 framing as Hiya.

The closed-beta status means SEA documented use will accumulate only after beta access is widely granted to regional partners.

Platform applicability: works on social-media content regardless of source platform; the platform-specific compression caveats from the DW audio audit apply equally here.

How to access

The public site at truemedia.org carries the "Request Beta Access" call-to-action; the detection endpoint at detect.truemedia.org requires a beta password. Request access through the form on the public site; the For Teams page describes API access, organisation accounts, and team management features that ship with the platform when beta access is granted.

Access status as of 2026-05-10

Web-fetch verification (Exa search, 2026-05-10) confirms: revival under Georgetown McCourt School (April 2025 announcement); current site at truemedia.org; detection endpoint at detect.truemedia.org behind a sign-in; closed-beta "Request Beta Access" CTA active. Privacy policy at truemedia.org/privacy-policy updated 11 March 2026; published domain references include truemedia.georgetown.edu. Verify again before institutional reliance.

Cost (current as of 2026-05)

Free (when beta access is active). Non-profit, non-partisan; Georgetown McCourt School institutional commitment underwrites operational costs.

Quickstart

  1. Visit truemedia.org and click "Request Beta Access".
  2. Submit the access request with your organisational affiliation and case-pipeline description.
  3. After access is granted, sign in at detect.truemedia.org.
  4. Upload the suspect image, audio, or video file.
  5. Read the ensemble verdict alongside the disclosed uncertainty metrics; the platform's radical-transparency commitment means the uncertainty band is the operationally important field, not the single-number probability.
  6. Pair with non-detector signals: provenance check via Content Credentials Verify or SynthID Detector; source-history work via InVID-WeVerify. Per Anchor 2 the toolkit's editorial position is unchanged: two non-detector signals are required before any strong public claim, regardless of how transparent the detector platform is.

In the toolkit's workflow

Cautious-detector and source-history pillar mixed-with-declaration tool per Architectural Anchor 1. TrueMedia's ensemble approach combines model verdicts (cautious-detector signal) with human verification (source-history-style signal). Sits as escalation-option at 1B.3 audio deepfake forensics; usable across image, audio, and video casework when access is granted.

Standard combinations:

  • With Hiya at 1B.3 – Hiya is the press-button deployable option; TrueMedia is the institutional escalation when beta access is available. Per Anchor 3 both count as one detector signal class.
  • With Deepfake Total at 1B.3 – the audit-leader (DW 7-of-10); TrueMedia is the ensemble-with-transparency alternative. Same Anchor 3 framing.
  • With Content Credentials Verify and SynthID Detector at 1A.4 for non-detector provenance pairing.
  • With Sensity at 1C.2 for institutional contrast. Both are enterprise-tier; Sensity is closed-source vendor; TrueMedia is open-methodology academic.

This card declares signal-class per output: when the verdict comes from the ensemble of model detectors, the signal class is detector (one weak signal under Anchor 1, one signal class under Anchor 3 regardless of internal model count). When the verdict incorporates human verification, the signal carries weight closer to a source-history signal; the card and the user must declare which part of the verdict they are reading from for any given case.

Decision-tree references: T3 audio triage – TrueMedia is the institutional escalation branch at T3.8 (audio detector run) behind beta gating. T6 source-protection – same upload-classification discipline as Hiya.

Conflict resolution behaviour

When TrueMedia's ensemble verdict disagrees with Hiya or Deepfake Total on the same input, TrueMedia carries slightly stronger evidence weight by virtue of the human-in-the-loop verification layer (the platform commits that "every decision determined by a model is checked by a human"). The toolkit's editorial position is that no audio detector verdict alone is publishable, including TrueMedia's; per Anchor 2 a non-detector signal (caller verification, platform-of-origin check, manual voice-comparison) must independently corroborate before any publication. When TrueMedia disagrees with a non-detector signal, the non-detector signal wins.

Override notes

Override notes

  • Revival confirmed but beta-gated (e2-revival-confirmed-but-beta-gated): the E2 anti-criterion (abandoned) was triggered when TrueMedia shut down in January 2025 but is overridden by the Georgetown McCourt School revival in April 2025, web-fetch-verified May 2026. The card enters at 1B.3 as escalation-option (D3 fail at First-Line Triage acceptable at Professional Verification per the E2 revival override).

  • Closed-beta disclaimer (closed-beta-disclaimer-required): "Currently in closed beta; request access via truemedia.org." Rendered as the leading sentence in the How to access section above; the access-status note records the May 2026 verification.

  • Cautious-detector with multi-pillar framing (g1-cautious-detector-with-multi-pillar-framing): the ensemble + human verification design genuinely serves both cautious-detector and source-history pillars; the workflow section names this declaratively.

  • Anchor-aligned mission radical transparency (anchor-aligned-mission-radical-transparency): TrueMedia's mission framing aligns with the toolkit's Architectural Anchor 1 (workflow over directory, transparency over headline accuracy) and Anchor 2 (no strong claim from detector signal alone). The Limitations and Independent Accuracy admonitions reference this alignment in framing; the caveats stand regardless.

Sources