Mal-uses of AI-generated Synthetic Media and Deepfakes: Pragmatic Solutions Discovery Convening, June 2018

Mal-uses of AI-generated Synthetic Media and Deepfakes: Pragmatic Solutions Discovery Convening, June 2018 is a report published by WITNESS on July 2, 2018, summarising discussions and next-step recommendations from a multi-disciplinary convening of thirty specialists held on June 11, 2018, co-organised with First Draft (a project of the Shorenstein Center on Media, Politics and Public Policy at Harvard Kennedy School). The report is the first formal civil-society intervention to frame deepfakes and AI-generated synthetic media as a human rights documentation problem — not merely a disinformation or electoral challenge — arguing that declining barriers to creating sophisticated audiovisual manipulations pose an emerging systemic threat to the evidentiary value of video recorded by human rights defenders and journalists in the field. WITNESS, founded in 1992 by Peter Gabriel, Human Rights First, and the Reebok Human Rights Foundation after the Rodney King video demonstrated video's capacity to expose hidden abuses, brought to the problem a 25-year institutional stake in the integrity of eyewitness footage — a stake no primarily election-integrity or media-industry framing of the deepfake threat could fully represent.

The June 2018 convening

The convening assembled thirty participants spanning independent and company-based technologists, machine learning specialists, academic synthetic-media researchers, human rights researchers, and journalists, operating under Chatham House Rules. The timing placed it at the opening of the publicly visible deepfake era: the term "deepfake" had emerged on Reddit in late 2017 when a user began circulating face-swapped non-consensual pornography using open-source deep learning tools, and commercial face-reenactment products were entering the market. WITNESS framed the convening not as a crisis response to existing harm but as a proactive, early-stage intervention: a deliberate effort to establish a civil-society and practitioner framework before synthetic media techniques scaled to the point where reactive responses would be the only option. The companion blog series published alongside the report on July 30, 2018 functions as the convening's public-facing synthesis — a two-part analysis that covers the threat landscape and a survey of solutions respectively.

Threat taxonomy

The WITNESS threat framing identifies five categories of disruption that AI-generated synthetic media enables. Reality edits — face-swaps, lip-sync alterations, voice simulations — directly undermine the evidentiary value of video that human rights documentation depends on: footage that was authentic at capture can be credibly contested as synthetic after the fact, degrading the evidentiary floor. Credible impersonations enable targeted attacks on individuals — placing a person's face or voice in a fabricated scene of violence, sexual content, or political statement — with disproportionate exposure among women and journalists already subject to coordinated harassment. News remixing weaponises the credibility built into real-news format and presentation, attaching synthetic content to trusted institutional containers. Plausible deniability inverts the prior relationship between video and accountability: where authentic footage of an atrocity once served as irrefutable evidence, the mere existence of convincing deepfakes gives perpetrators a rhetorical escape route — the "liar's dividend" in which real footage is dismissed as synthetic. Coordinated falsehood campaigns via computational propaganda use synthetic media as a force multiplier for bot networks and disinformation operations, particularly in contexts where misinformation and information disorder are already widespread. The report places the greatest concern at the intersection of the first and fourth categories: the human rights documentation community's entire model of accountability rests on the assumption that video evidence is harder to fabricate than to capture; once that assumption erodes, the structural position of documentary footage in legal and advocacy contexts changes in ways that disproportionately harm the communities whose protection depends on it.

Solutions framework

The solutions survey catalogues sixteen distinct solution areas and organises them across three registers. The technical layer includes authentication-at-capture tools such as WITNESS's own ProofMode (which adds cryptographic signatures and metadata at the moment of recording), cross-source spatial validation using multiple camera angles to establish ground truth, manual forensics (examining lighting consistency, reflection patterns, audio coherence), deep-learning detection trained on synthetic datasets, and adversarial obfuscation techniques that embed pixel-level perturbations preventing unauthorised training on an individual's visual data. The institutional layer covers journalist verification toolkits integrated into existing platforms such as First Draft and Bellingcat, collaborative journalism networks for coordinating verification during crises, platform detection at upload and sharing stages with contextual UI signals to users, commercial tool regulation requiring manipulation software to embed machine-readable forensic indicators, and industry-community dialogue structures ensuring affected populations guide development. The normative and legal layer addresses media literacy investments, quality incentive programs (analogous to bug bounties rewarding accurate identification), and new legal frameworks modifying defamation, copyright, and intermediary liability regimes to respond to synthetic media. The convening's integrating principle across all three registers is that no single solution works at all scales and in all contexts: a media-literacy intervention effective in a high-bandwidth urban media environment provides limited protection in a low-connectivity rural context where shared video circulates via WhatsApp without metadata or provenance chain.

Global South centring

The report's distinctive methodological commitment is its insistence that the communities most exposed to malicious synthetic media must be centred in solution design, not treated as beneficiaries of solutions engineered for the Global North. WITNESS's framing draws on its own organisational experience working with human rights defenders in over 60 countries: in many of those contexts, misinformation and information disorder are already rife, meaning that the marginal harm of adding synthetic media to an already-degraded information environment falls entirely on communities whose coping resources are least developed. The solutions survey raises a specific tension that election-integrity framings of deepfakes typically miss: platform takedown approaches — already problematic in the cases of Facebook's handling of Myanmar/Burma content and YouTube's suppression of Syrian documentation — carry a structural risk of suppressing legitimate human rights documentation in the name of removing synthetic content. An error in detection becomes an erasure of evidence; the cost of false positives is borne by the documented communities, not the platforms.

Position within the corpus

Mal-uses of AI-generated Synthetic Media and Deepfakes is the corpus's first WITNESS publication entry, closing the AI-synthetic-media / deepfakes civil-society-documentation slot in the publications layer. It arrived at the critical early moment — mid-2018, roughly six months after "deepfake" entered public discourse — when the field was almost entirely occupied by two framings: election integrity (the spectre of synthetic political speech) and non-consensual pornography (the immediate, gendered harm). WITNESS's intervention introduced a third framing that neither of those covered: the threat to the evidential infrastructure of human rights accountability, the long-duration documentary record that organisations like WITNESS had spent decades building and whose value rests entirely on the assumption that authentic video is harder to fabricate than to capture. By framing the problem this way, the report positioned synthetic media as a structural challenge to advocacy and accountability — not a novel crime type or a political-speech problem — and connected it to the corpus's wider thread on AI-enabled erosion of documentation, surveillance, and accountability systems. The convening itself seeded the "Prepare, Don't Panic" initiative that WITNESS sustained through regional workshops in Brazil, South Africa, and Southeast Asia (2019–2020), extending the same multi-stakeholder, practitioner-centred methodology to regional contexts where the threat model differs materially from the US/European focus of most synthetic-media policy work.