Skip to content

Tool cards — pivot view

These are the sixty-five tools the toolkit references, organised by where they sit in the workflow. Each entry links to a one-page card with what the tool does, how to access it, language coverage, known limitations, and regional relevance. Most tools sit in one cell. A few span several and are listed at every cell where they apply, with a single canonical card. Three reference cards anchor the rest: InVID-WeVerify as the multi-cell, multi-pillar exemplar; MAFINDO Kalimasada as a single-country tipline exemplar; TRIED Benchmark as an institutional-reference exemplar, a framework for evaluating detectors instead of a deployable tool. Use these as orientation for what a tool card looks like before reading the cell-by-cell pivot below.

Pillar 1 — Detect and triage

1A — First-Line Triage

1A.1 — Image

A five-minute phone read on whether a forwarded image carries any AI signal at all. The cell is for the screenshot-from-WhatsApp moment when the question is whether to keep working with the file or set it aside; three tools cover it – a multimodal commercial detector, an uncertainty-declaring open alternative, and a non-detector provenance check, so that no single output reads as a verdict. Cases that warrant a thirty-minute desk pass escalate into 1B.1.

1A.2 — Video

The same triage band as 1A.1, but for moving image, where detector class confidence sits structurally lower – independent benchmarking measured best-in-class at 0.78 accuracy on 2,036 in-the-wild deepfakes, and SEA-platform- compressed video has no public benchmark. Two tools only, by deliberate honest-gap policy: a multi-pillar plugin whose reverse-image and metadata modules often resolve the case before the deepfake tab is ever opened, plus one press-button detector. Cell-level discipline is binding here – a detector verdict on its own never carries a public claim.

1A.3 — Audio (constrained)

A single multimodal entry at first-line tier, by deliberate honest-gap policy. The audio detector class is broken across SEA languages per the DW Innovation November 2025 audit; tonal-language phonetics, codec-compressed audio, and low-resource scripts defeat the press-button options that exist. Use Hive AI (Hive Moderation / Hive Detect / Hive AI Detector) for a quick lay-read on language-agnostic synthesis artefacts only, and escalate to 1B.3 the moment the case warrants more than five minutes.

  • Hive AI — detector — primary (cross-cell from 1A.1)

1A.4 — Provenance / watermark (architecturally important)

The only Pillar 1 first-line band where non-detector signals are the primary mode. When a file carries C2PA Content Credentials or a SynthID watermark, the question shifts from "is this AI-generated?" to "what does the manifest say about who made this and how?" – a path that aligns directly with Anchor 1's provenance-first framing. The three verifier tools cover end-user verification, Google-DeepMind-watermarked content, and forensic inspection of the manifest itself. The cell-level gap is manifest survival across SEA hardware (Realme, Vivo, Oppo, Xiaomi, Transsion) and across WhatsApp/Facebook/TikTok strip-on-upload – not independently documented.

1B — Professional Verification

1B.1 — Multi-tool plugins (workhorse hub)

The desk-hour workhorse of Pillar 1, where a working fact-checker assembles a verifiable source-history record before any escalation. Four privacy-postured options let the user match tool to threat model: a cloud-mitigated default, an offline desktop alternative for surveillance-risk environments (Sri Lanka, Laos), a local-WASM browser option for source-protection-binding cases, and the command-line standard for batch or scripted work. Every card here carries an S1-mitigation note by cell-level binding. Entry from 1A FLT escalations; bridges down to 1B.4 for ELA work and up to 1C.2 for enterprise-grade deepfake platforms.

  • InVID-WeVerify — non-detector + cautious-detector mixed — primary (cross-cell)
  • ExifTool — non-detector — alternative
  • Sherloq — non-detector — alternative (offline desktop)
  • MetaDataKit — non-detector — alternative (browser-WASM)

1B.2 — AI text detection (contracted)

Contracted on purpose. Users arrive at this cell asking whether they can prove a piece of text was generated by an LLM; the toolkit's plain answer is "not reliably, especially in non-English." Two cards stand here as cautionary cases – one defensible current vendor, one clearest cautionary tale per the Stanford 2023 and Perkins 2024 evidence – both under a binding "do not use as standalone evidence" frame. Productive work on textual cases routes to 2B.2 claim extraction instead.

1B.3 — Audio deepfake forensics (structural ceiling)

The productive ceiling for audio work in this toolkit. A Lao political voice clip, a Thai scam call, or a Tagalog candidate-impersonation audio that warrants a thirty-minute desk pass lands here. Three viable detectors plus the DW Innovation editorial reference; under Anchor 3 the three count as one detector class signal, and under Anchor 2 each case still needs a non-detector pair – caller verification, platform-of-origin check, voice-comparison interview – before any public claim. Cross-language detector unreliability is binding at the cell level.

1B.4 — Metadata / ELA forensics

Where ELA, EXIF, and metadata-spoofing checks happen once a multi-tool pass has surfaced an image worth deeper inspection. Three tools by threat-model band: a command-line standard for non-destructive metadata extraction (no upload, low-risk), an offline desktop alternative when the source file cannot leave the machine, and a cloud ELA option for low-sensitivity material where S1 classification clears web upload. The cell-level discipline is plain: metadata is spoofable, and most images encountered on WhatsApp, Facebook, or TikTok have already lost their EXIF before they reach the user – that absence is itself a signal to read, not a tool failure.

  • ExifTool — non-detector — primary (cross-cell from 1B.1)
  • Sherloq — non-detector — alternative (cross-cell from 1B.1)
  • FotoForensics — non-detector — alternative

1B.5 — News-reliability scoring

When the question stops being "is this content AI-generated?" and becomes "is the outlet that published this reliable?", users reach for source-reliability scoring. The cell pairs a multilingual fact-check claim explorer (the default first stop, returning existing fact-check work and ClaimReview metadata across Bahasa Indonesia and more) with a paid reliability service for the 16 languages it covers (Indonesian, Tagalog, Thai included) and an EU DisinfoLab knowledge layer for institutional readers. Sinhala and Lao outlets sit outside NewsGuard coverage, so Sri Lankan and Lao workflows rely on the local fact-check ecosystem (Hashtag, Watchdog, FactSeeker) alongside the Google explorer.

1C — Institutional-Level Analysis

1C.1 — CIB detection (largest single landscape)

Entry to Pillar 1's institutional tier, and the single largest landscape in the inventory at 38 entries. This is where the question becomes "is there a behavioural pattern across accounts, posts, and timing?" – the analytical move from organic claim to coordinated operation, and one that requires Python proficiency or paid enterprise access. The four-tool ladder runs from open-source primary (locally on user-supplied data) through two academic R/Python alternatives to the enterprise escalation option; the standard institutional pipeline pairs CIB output with 2B.2 extraction and 2C.1 monitoring. SEA-specific worked-example casebook absence is the cell's named gap.

1C.2 — Enterprise / open-source deepfake platforms

The deepfake-detection ceiling of Pillar 1. This is where a Rappler-grade investigative pipeline lands when it needs to certify whether a video case is AI-generated – the Doc Willie Ong eye-drop case (98% / 75.5% / 94%) and the Brawner 79.3% work are the documented precedents. Three institutional-grade options cover documented-deployment confidence, broadcaster-grade alternative, and the academic evaluation harness; under Anchor 3 all three still count as one detector signal class no matter how many ensemble verdicts they aggregate internally. Vendor accuracy figures come from anonymised pools where independent scores are not separable – cell-binding caveat on every card.

  • Sensity AI — detector — primary (cross-cell awareness with 1A.1, 1A.2, 1B.1 escalations)
  • Reality Defender — detector — alternative
  • DeepfakeBench — non-detector + reference benchmark — alternative

1C.3 — Explainable AI forensics

Where a probability score stops being enough and the question becomes whether the verdict can withstand a defamation suit or a legal challenge – the toolkit's explainability layer for evidence-grade reporting. The cell carries a GradCAM/SHAP/LIME-on- EfficientNet primary (heatmaps that can be embedded in a published forensic report), a reliability-map alternative whose declared uncertainty band is itself the signal under Anchor 1, and the WITNESS evaluation framework as institutional reference for methodology review. Explainability is built and benchmarked on English-language and Western-face data; Sinhala and Lao defamation contexts still require manual interpretation alongside the visualisation.

  • XAI-Deepfakes — detector + explanation-output — primary
  • TruFor — non-detector + reliability-map — alternative (mixed-with-declaration)
  • TRIED Benchmark (WITNESS) — institutional reference (third reference card; institutional- reference exemplar)

Pillar 2 — Counter

2A — AI-Assisted Workflows

2A.1 — Transcription / summarisation / translation

The pure productivity entry of Pillar 2. The cell is for handling a Lao voice memo, a Thai parliamentary transcript, or a Tagalog Facebook livestream when the source has to stay local enough that S1 sensitivity classification holds. An open-source local primary covers routine work; paid Google Cloud Translation is the only documented Lao path, accepted with the Google-server caveat; the journalist-pipeline option covers multi-source archival cases where Google-server upload is acceptable. Whisper's low-resource performance for Sinhala and Lao is named honestly in the cards; fallback to multilingual figures would mask the gap.

2A.2 — Reverse-image / geolocation enhanced

Where most non-text claims start to resolve, long before any detector touches the file. A reverse-image hit or an archival capture is one non-detector source-history signal under Anchor 2, combining with claim-extraction (2B.2) or fact-check-database (1B.5) results for any institutional reading. Four tools cover the standard one-click plugin (cross-cell from 1B.1, handling reverse-image queries plus archived versions in one move), a Faktisk-validated AI geolocator, an Amnesty-maintained YouTube data viewer, and the Bellingcat archiving companion for rapid-archiving discipline. No cell-level honest-gap of note – the tools work language-agnostically.

2A.3 — Prebunking / media-literacy (rich)

One of the few cells where SEA already has gold-standard examples in the inventory. Users reach for these tools in training contexts, not on artefacts in active fact-check work – a workshop in Indonesia runs Tempo Detektif Deepfake directly, and a trainer in Thailand, Sri Lanka, or Malaysia adapts the Cambridge inoculation-game architecture because no SEA-localised inoculation game exists outside Indonesia yet. The IREX/RAND 25% multi-source-checking improvement and the Jigsaw Indonesia campaign are cited as evidence that prebunking effects are real and measurable. Localised effectiveness measurement outside Indonesia is the cell's named gap.

2A.4 — Newsroom workflow management (INCLUDE MARGINALLY)

Included marginally as an editorial choice: an adjacent pointer, not a primary recommendation row. Two SEA-grounded examples – a Tempo archival-query assistant for Indonesia and the MAFINDO Satgas Pemilu 2024 Election Task Force as a workflow design pattern – show how Indonesian newsrooms have orchestrated AI assistance around verification work. The cell is deliberately short, because the inventory outside these two is dominated by non-deployable Innovation Challenge prototypes; the toolkit does not pretend that workflow-management AI is yet a deployable category in SEA. Bridge to 2B.1, where the MAFINDO tipline itself lives.

2B — Collaborative Verification

2B.1 — Multilingual tiplines (rich)

The heart of Pillar 2 – the place where the claim of durable, strategic value is anchored, because the user base does the work and AI helps process the intake. Match platform to country: WhatsApp + Bahasa points to Kalimasada; LINE + Thai points to Cofact (the only LINE-native option in the inventory); a multilingual deployment with self-hostability points to Meedan Check (140K+ users, CekFakta and #FactsFirstPH backbone); Malaysia's multi-language plus Tamil case points to Sebenarnya AIFA with the public-facing exception-pass and the government-operator independence caveat attached. Lao has no tipline in the toolkit and the cell says so directly; Sri Lanka remains a form/email/WhatsApp patchwork distribution, not a Meedan-style backend.

  • Meedan Check — non-detector + tipline-platform — primary
  • Cofact Thailand — non-detector + tipline — alternative (LINE-native)
  • Kalimasada (MAFINDO) — non-detector + tipline — alternative (Indonesia WhatsApp)
  • Sebenarnya.my AIFA — non-detector + tipline — alternative (Malaysia, government-operated with independence caveat; B1 exception-pass)

2B.2 — AI claim extraction (rich)

Where tipline volume scales beyond manual processing and AI extraction keeps the operation functional. The largest inventory landscape after CIB (42 entries); five cards cover SEA's language asymmetry. Sri Lanka work routes to the Sinhala-asymmetric-strong primary (Watchdog/LIRNEasia Dissect) first, then to the pan-Indian Tamil tool with the Sri-Lanka adaptation gap noted; Indonesia work routes to MAFINDO's Bahasa database with the multilingual XLM-R fallback; cross-country and Lao material routes to Alegre because XLM-R is the only multilingual non-vendor option that covers Lao at all (with a low-resource disclaimer pass). Tamil Sri-Lanka-specific adaptation gap (X-CLAIM is pan-Tamil, not SL-adapted) and Lao claim-extraction absence are the cell's binding named gaps.

  • Dissect (Watchdog SL / LIRNEasia) — non-detector — primary (Sinhala asymmetric-strong)
  • Meedan Alegre — non-detector — alternative (multilingual XLM-R, Lao fallback path; matching engine of Check)
  • Yudistira (MAFINDO) — non-detector — alternative (Bahasa Indonesia)
  • ClaimBuster — non-detector — alternative (English baseline, b2-disclaimer-pass)
  • X-CLAIM — non-detector — alternative (Tamil pan-Indian; Sri-Lanka-specific adaptation gap binding caveat)

2B.3 — LLMs in fact-check workflows

The LLM-as-tool layer of Pillar 2: useful, hallucinatory, and moving fast. Journalists handling Bahasa, Filipino, or Thai material run SEA-LION locally for hallucination-mitigated draft outputs, then verify every claim downstream against 2B.2 extraction or 1B.5 reliability scoring before publication; lay users in Laos, Malaysia, Philippines, and Thailand have the mobile app option under a binding hallucination caveat; Malay-first work in Malaysia escalates to MaLLaM (Mesolitica) as the Malay-specific option. SEA-LION v4 (AI Singapore, March 2026) closed the previously-paired Lao real-gap in this cell; Sinhala remains absent from the major SEA LMs and is handled outside this category via the Dissect / LIRNEasia stack noted in 2B.2.

  • SEA-LION — non-detector — primary (regional foundation model; explicit Lao coverage in v4 March 2026; Sinhala absent)
  • AI Fact Checker App — non-detector + mobile — alternative (LA / MY / PH / TH app-store localisation)
  • Google Pinpoint — non-detector + LLM-pipeline — alternative (cross-cell from 2A.1; cloud-LLM journalist-pipeline mode)
  • MaLLaM (Mesolitica) — non-detector — escalation-option (Malay-specific, Mesolitica)

2C — Large-Scale Intelligence

2C.1 — Automated claim monitoring

Cross-platform monitoring infrastructure across mixed cost tiers, with civil-society alternatives explicit. The cell demonstrates that monitoring exists at multiple price bands and that free academic and SEA-civil-society options are genuine alternatives, not consolation prizes.

  • Information Tracer (Rolli) — non-detector — primary (paid mid-tier cross-platform tracer)
  • Media Cloud — non-detector — alternative (free academic news-source corpus)
  • Sinar Project iMAP — non-detector — alternative (free open-source SEA-specific network monitoring; MalaysiaNow / Malaysia-Today MCMC blocking documented)
  • FactFlow AI (Newtral) — non-detector — alternative (Telegram-specific platform specialist; Animal Político 10M-message processing)

2C.2 — Network analysis

The mapping and visualisation layer downstream of CIB analysis (1C.1) or claim-monitoring (2C.1). Maltego is the OSINT gold standard with a free Community Edition; Gephi is the free open-source visualisation alternative; Meta Content Library is the IFCN-gated data source for institutional partners with the access path.

  • Maltego — non-detector — primary (OSINT gold standard, freemium with free Community Edition)
  • Gephi — non-detector — alternative (free open-source desktop visualisation; standard pairing downstream of Maltego, CooRTweet, Mango Tree)
  • Meta Content Library — non-detector — escalation-option (IFCN-signatory only; journalists excluded; combined Sebenarnya operator-identity caveat + Graphika cite-as-external-source quickstart redirect)

2C.3 — Cross-lingual narrative tracking + DISARM

Framework references rather than deployable tools; the TRIED-pattern voice register applies. ABCDE is the toolkit's surface analytical vocabulary per Decision 2; DISARM is the institutional annex; Online Operations Kill Chain is the operational sequencing complement; DISARM Navigator + STIX2 is the technical tooling layer for institutional CSIRT and government-partner deployment.