What it takes to fight the AI-powered disinformation machine
A month into Jessica Cecil's new publication Athena, the evidence keeps arriving: a Russian-run fake think tank with real employees, a shapeshifting video in fifteen languages, and Britain building a new defence centre. The information war has gone industrial; the fightback is racing to catch up.
Published 29 October 2026 · 16:55 GMT
The summary
On 8 October 2026, OpenAI said it had banned ChatGPT account clusters behind two covert influence operations: Russia's "Dark Clark", which ran a front think tank in Latin America through unwitting local staff, and Iran's "Bogus Bylines", a stable of seven fake journalist personas that placed nearly 100 articles in about a dozen outlets. Britain's prime minister Andy Burnham used his first UN General Assembly address on 22 September 2026 to announce a National Centre for Information Defence, tasked with detecting, attributing and disrupting hostile-state information attacks, and modelled on France's Viginum agency.
The three lenses
- WestThe Western press reads this as state capacity: The Times stresses Whitehall caution; expose, not disrupt, no new powers.
- EastThe East reads it as a warning about becoming terrain.
- Global SouthThe South reads it as lived experience.
Key facts
- On 8 October 2026, OpenAI said it had banned ChatGPT account clusters behind two covert influence operations: Russia's "Dark Clark", which ran a front think tank in Latin America through unwitting local staff, and Iran's "Bogus Bylines", a stable of seven fake journalist personas that placed nearly 100 articles in about a dozen outlets. OpenAI (Oct 8, 2026)
- Britain's prime minister Andy Burnham used his first UN General Assembly address on 22 September 2026 to announce a National Centre for Information Defence, tasked with detecting, attributing and disrupting hostile-state information attacks, and modelled on France's Viginum agency. The Times (Oct 2026)
- A fabricated video styled as a Euronews bulletin, falsely claiming a Ukrainian soldier murdered three NATO comrades during Latvia's real Baltic Trust 2026 exercise, has mutated into at least eight versions in 15 languages; Latvia's armed forces and Ukraine's embassy say the shooting never happened. Athena (Oct 2026)
- Jessica Cecil, the BBC executive who founded the Trusted News Initiative in 2019, launched Athena, a publication devoted to disinformation, in late September 2026; her first-month reflections were published by the Daily Maverick on 29 October. Daily Maverick (Oct 29, 2026)
- In South Africa, researchers are recording sock-puppet and bot activity surging ahead of the 4 November 2026 local elections; the Electoral Commission has logged more than 700 complaints about misleading content, and News24 journalists built a convincing fake candidate with free AI tools in seven minutes. EWN (Oct 8, 2026)
- OpenAI has opened a dedicated Election Operations Center to watch for misuse of its tools during the United States midterm elections on 3 November 2026, staying active through the end of vote counting. Newsmax (Oct 28, 2026)
How one lie learned to speak fifteen languages
On 29 October 2026, Jessica Cecil, the BBC executive who founded the Trusted News Initiative in 2019, published her first-month report from the trenches of the information war. Her new publication, Athena, launched in late September to cover disinformation full-time, and the month obliged: OpenAI revealed that Russian and Iranian operators had used ChatGPT to run "false front" influence operations, including a fake Latin American think tank staffed by people who never knew their real employer; a fabricated Euronews-style video claiming a Ukrainian soldier murdered three NATO comrades in Latvia mutated into at least eight versions in 15 languages; and Britain announced a new National Centre for Information Defence. Her thesis, in one line: artificial intelligence has industrialised lying.
What the takedowns actually exposed
OpenAI's 8 October report describes two unusually sophisticated operations. The Russian one, "Dark Clark", ran a self-styled "research platform", the Social Research Center, through a fake persona, "Mia Clark", who handled hiring, pay and editorial calls while unwitting Latin American staff produced 60-plus original articles. It targeted Argentina, Bolivia, Peru and Ecuador to damage Ukraine's standing in the region; OpenAI's first-ever Category 5 disruption. The Iranian "Bogus Bylines" fielded seven fake journalist personas and placed nearly 100 articles in about a dozen small outlets, casting the US as the aggressor in the US–Iran war. In London, Andy Burnham used his first UN address on 22 September to announce the information-defence centre, accusing Russia of spending £1.3 billion a year on manipulation, bots, fake sites, falsified articles, the forged branding of 28 British organisations, and of trying to interfere in the 2019 election. The centre will sit in the Prime Minister's Office, modelled on France's Viginum and Britain's Baltic counterparts. In Latvia, Athena documented the shapeshifter's anatomy from a new Refute report: a real exercise, Baltic Trust 2026, wrapped in a fictional atrocity and reworked by AI translation until, three weeks on, it was discussed as fact.
Who really runs the front organisations?
OpenAI named Russia and Iran but no specific agency or ultimate operator behind the think-tank front, and no account totals. The Latvia video's makers are unidentified: Refute says the evidence points to Russia-linked networks, an analyst's attribution, not a government's. In London the centre's powers are unresolved: the prime minister promises it will "detect, attribute and disrupt", but The Times cites Whitehall saying no new powers and "expose, not disrupt". Athena is one month old, funding and staying power unproven; and the hardest unknown is temporal: with generation tools improving monthly, nobody knows whether any defence built today survives next year's models.
Why elections are the proving ground
The laboratory is running, and it is in the Global South. South African researchers tell EWN that sock puppets and bots are "out in force" before the 4 November local elections: one TikTok account pushing the Freedom Front Plus with divisive racial messaging has half a million followers of unknown provenance. A Sowetan-cited study of 377,167 posts found a quarter of the top anti-foreigner voices are inauthentic disrupters, some run from abroad. News24 reporters built a convincing fake candidate with free AI tools in seven minutes. The Electoral Commission has logged 700-plus misleading-content complaints and now threatens fines of 200,000 rand or prison under its Disinformation Code. The US votes on 3 November under OpenAI's new election watch. Same pattern everywhere: elections concentrate attention, and attention is what the machines harvest.
What a whole-of-society defence actually looks like
Cecil's prescription starts from an uncomfortable admission: people can no longer spot deepfakes by look or sound; AI is too good at both. Her answer is not another filter but literacy at scale: citizens trained to ask who benefits from a claim, reached through schools, libraries, community centres and journalism. It marks an evolution from her own 2019 Trusted News Initiative, a big-tech–media early-warning pact; "that curation feels quaint now," she writes. The blocs answer differently: the West builds state machinery like Britain's centre and leans on private arbiters like OpenAI; the East watches via outlets like Singapore's CNA, which called the Russian front "essentially a fake think tank with real employees"; the South faces the fight directly, where South Africa's election machinery and Iran's persona networks show both faces of the same war. No single capital owns the whole picture.
The censors' dilemma: who watches the watchers?
Every defence here concentrates power, and each concentration carries risk. Britain's centre will sit inside the Prime Minister's Office, a step from the "ministry of truth" its critics already name, which is why Whitehall's "expose, not disrupt" matters more than launch rhetoric. OpenAI, a private American firm, now acts as counter-intelligence: it defines the covert operation, publishes the evidence, bans the accounts, with no court reviewing the call. Trackers like Refute and StopFake are becoming political actors whose attributions move ministries. The symmetric danger: the "disinformation" label that unmasks a Russian front can be turned on domestic dissent. The winners may not be the best liars, but whoever gets to define the lie.
November's double test: Washington and Pretoria
November is the first live trial of every defence described here. The US holds midterms on 3 November with OpenAI's Election Operations Center watching through the count. South Africa votes on 4 November, with the Disinformation Code's penalties applying for the first time. Watch for Latvia-style shapeshifters, real local events wrapped in fictional atrocities, mutated across languages, around both votes; watch whether Britain's centre moves from announcement to operation and names names; and watch Athena's second month, judged on whether one publication can keep pace with the machines.
The three lenses
West Lens:
The Western press reads this as state capacity: The Times stresses Whitehall caution; expose, not disrupt, no new powers. GlobalSecurity.org treats Burnham's UN address as doctrine: industrial-scale assault, £1.3 billion a year, 28 impersonated institutions. US tech media grades OpenAI's report like counter-intelligence product.
East Lens:
The East reads it as a warning about becoming terrain. Singapore's CNA covered the front plainly, "essentially a fake think tank with real employees", and the detail that matters in Asia is the target: the center styled itself around the Indian diaspora in Latin America, making diaspora communities the unwitting infrastructure of someone else's war. The TNI's Asia-Pacific arm, built with NHK, shows the region has prepared early-warning networks since 2022.
South Lens:
The South reads it as lived experience. The Daily Maverick platformed Cecil's reflections; South Africa's press and electoral machinery fight the war in real time; sock puppets before the 4 November vote, 700-plus commission complaints, a Disinformation Code with prison terms. Malaysia's New Straits Times carried OpenAI's findings to Southeast Asia.
The consensus
- What we agree on
- We agree: in October 2026, OpenAI disrupted two AI-assisted covert influence operations; a Russian-run front think tank in Latin America and an Iranian network of fake journalist personas; Britain announced a new National Centre for Information Defence; and a fabricated video about a shooting at Latvia's Baltic Trust exercise spread in at least eight versions across 15 languages. These are documented by the operators' own disclosures, governments, and independent trackers.
- What we don't agree on
- We disagree: on attribution and on remedy. Refute says the evidence around the Latvia video points to Russia-linked networks, but no government has formally attributed it; OpenAI deliberately did not name a Russian agency behind the think-tank front. On remedy, Britain's prime minister promises a centre that will "detect, attribute and disrupt", while Whitehall briefs that it will expose rather than disrupt and get no new powers; and critics warn a state truth-arbiter threatens free speech.
- What we know
- We know: the tradecraft has industrialised. AI now drafts the internal reports, mutates the videos across languages, manages the fake personas' social accounts, and staffs front organisations with real, unwitting people. The marginal cost of a multilingual influence campaign has collapsed toward zero.
- What we don't know yet
- We do not know: who ultimately commissioned the Latin American front or the Latvia video; how many similar fronts are still running undetected; whether Britain's new centre will be a shield or a censor; and whether media-literacy defences can scale as fast as the generation tools.
- What we expect
- We expect: November to be the proving ground; the United States votes on 3 November and South Africa on 4 November, both under active AI-driven manipulation attempts. Expect shapeshifter-style videos around both, and the first real-world test of OpenAI's election war room and Britain's new centre.
Questions, answered
What did OpenAI's "false front" report actually find?
Published on 8 October 2026, it describes two banned influence operations. Russia's "Dark Clark" ran a front think tank in Latin America through a fake manager persona while unwitting local staff wrote 60-plus articles; Iran's "Bogus Bylines" used seven fake journalist personas to place nearly 100 articles in small outlets. OpenAI rated the Russian operation Category 5, its highest-impact disruption ever.
What is Britain's National Centre for Information Defence?
Announced by Prime Minister Andy Burnham at the UN on 22 September 2026, it will sit in the Prime Minister's Office and Cabinet Office to detect, attribute and disrupt hostile-state information attacks, modelled on France's Viginum. Whitehall sources say it will get no new powers and focus on exposing rather than disrupting networks.
How did one fake video end up in 15 languages?
Trackers at Refute found a fabricated Euronews-style clip about a fake shooting at Latvia's real Baltic Trust exercise was reworked with AI translation and local influencers into at least eight versions in 15 languages. Three weeks after publication it was discussed as fact. Latvia's armed forces and Ukraine's embassy confirmed the shooting never happened.
What is Athena, and who is Jessica Cecil?
Athena is a new publication about disinformation launched in late September 2026 by Jessica Cecil, the BBC executive who founded the Trusted News Initiative in 2019. Writing in the Daily Maverick on 29 October, she argues AI has made deepfakes undetectable by eye or ear, so the defence must be societal: media literacy delivered through schools, libraries and journalism.