How cheaper AI changes the cost of attacking a reputation, and how to prepare.
Listen to this article · 18 minutes
Read by an AI voice.
Generative AI is cutting the cost of building disinformation and coordinated influence campaigns. When each attempt costs less, an attacker can try more often, in more languages and against more targets.
A convincing campaign against a company once required a team. Researchers investigated the target, writers developed allegations, designers produced supporting material and operators created and maintained the accounts that carried the message. Reaching audiences in several countries added translation, local knowledge and coordination.
Generative AI makes much of that work faster and cheaper. A small team can now produce articles, reports, social media posts, images and translations that previously required more people. It can create material for fabricated organisations, develop online personas and support activity across several platforms.
Disinformation, impersonation and manufactured public support existed long before ChatGPT. AI reduces the labour needed to produce and maintain them.
For organisations, that changes the nature of the risk. Attackers can produce more material, experiment with different narratives and continue publishing even when their initial efforts attract little attention. Meanwhile, investigating and responding to a suspected attack still requires evidence, time and judgement.
Constructing the appearance of a credible institution, campaign or public movement traditionally involved several connected activities. Generative AI can now assist with each.
| Function | Before generative AI | With generative AI |
|---|---|---|
| Research and analysis | Researchers gathered background, identified vulnerabilities and monitored activity. | AI can summarise research, analyse material and help identify patterns. |
| Content production | Writers, translators and designers created articles, reports, posts, graphics and supporting material. | AI can draft, translate, localise and produce visual content. |
| Personas and coordination | Operators developed online identities, maintained accounts and coordinated activity across platforms. | AI can generate persona material, adapt content and support account workflows. |
| Community engagement | Operators wrote replies, joined discussions and attempted to create the appearance of public support. | AI can generate replies and supporting content in large quantities. |
OpenAI's investigations have documented operators using ChatGPT for social media posts, replies, images, logos, translations and account-management tasks. These tasks once needed separate people.
In August 2026, OpenAI disclosed an influence operation built around an organisation called the International Burke Institute.
The institute presented itself as an Israel-based community of experts. Its website carried academic articles, named prominent scholars and published a Burke Sovereignty Index that favoured Russia while criticising Western countries.
To an unfamiliar visitor, the institute had many of the features associated with a legitimate research organisation. It had an institutional identity, named experts, a body of published work and a framework that appeared to give its conclusions analytical weight.
Much of its credibility was borrowed. OpenAI examined a sample of 36 articles linked to experts on the institute's website and found that 34 had been copied from elsewhere on the internet. Some were years old and others were attributed to the wrong authors, including real academics who had not written them. The site also listed prominent scholars without establishing their involvement.
The Council on Foreign Relations identified current and former experts who had been listed as affiliated with the institute without their knowledge. UNICEF had also been falsely identified as a partner.
OpenAI assessed the operation as very likely originating in Russia. ChatGPT had been used to create promotional social media posts, replies and supporting material, including content in English and German.
OpenAI said the articles on the site and its country sovereignty reports were not generated using its models. AI instead supported a wider operation built around conventional techniques including impersonation, copying published work and fabricating institutional credibility.
This illustrates how the technology can make an existing deception easier to promote and sustain. Operators can combine authentic material, fabricated identities and AI-generated content to present a coordinated narrative as credible and independent.
The Burke operation also shows a reputational risk for third parties. By borrowing the names of genuine academics and institutions, the operators attempted to give their own material authority while potentially damaging the reputations of those whose identities had been used.
The institute had an institutional website, a substantial collection of published material and several channels through which that material could be promoted, including Telegram channels with roughly 10,000 to 20,000 followers. It had been established in February 2025 and was exposed by OpenAI in August 2026.
The underlying methods were already being used against companies before generative AI became widely available.
In 2022, cybersecurity company Mandiant documented an influence campaign by DRAGONBRIDGE, a pro-China network that targeted three rare-earth companies.
Among its activities, the network used fabricated social media accounts to oppose a rare-earth processing facility in Texas planned by Australian company Lynas Rare Earths.
Some accounts posed as local residents concerned about environmental damage and public health. Their posts attempted to make opposition appear to originate from people who would be directly affected by the facility.
The case showed how an organised influence network could exploit a genuine area of public concern to target a business.
AI makes the production of similar material easier. An allegation can be adapted into different formats, translated for local audiences and supported with images and social media replies without requiring separate specialists for every task.
The underlying tactic remains the same. Fabricated participation is designed to make a coordinated campaign look like an independent public reaction.
In June 2026, OpenAI described another operation, which it called Data Center Bandwagon.
The activity came from a cluster of ChatGPT accounts assessed as likely originating in China. OpenAI believed the operators were associated with a private Chinese technology company undertaking work for provincial-level government clients. The campaign targeted debates about data centres and electricity costs in the United States.
Operators generated comments and images arguing that rising demand from AI infrastructure was pushing electricity costs onto ordinary Americans. Some of the material was based on reporting from a legitimate regional newspaper, and some posts linked to legitimate news stories about electricity-market capacity auctions.
The accounts distributing this material were assessed as likely inauthentic. The wider operation also used fabricated American identities, including workers, students, mothers and investors.
ChatGPT helped generate comments, edit graphics and support account-management workflows. Operational reports examined by OpenAI described plans to maintain credible personas, encourage interactions between accounts and preserve the appearance of organic engagement.
Concerns about electricity prices and data centre development can be legitimate. Residents may oppose industrial development for good reasons, and businesses should expect criticism.
The deception in this case was the attempt to manufacture the appearance of independent public participation.
For an organisation facing a contentious planning application, product controversy or environmental dispute, that creates a difficult distinction. An increase in online opposition could represent genuine public concern, coordinated political activity, fabricated engagement or a mixture of all three.
The content alone may not settle the question. An operator can attach a misleading narrative to a real controversy and cite authentic reporting as supporting evidence. Determining whether the apparent public reaction is genuine requires investigation beyond the claims themselves.
The potential market for these techniques extends beyond state-backed campaigns.
Citizen Lab has documented the activities of BlackCore, an Israeli company that marketed influence and cyber capabilities to clients. Its advertised services included activities described as Discourse Dominance, Organic Engagement and Counter Operations.
Researchers identified hundreds of fake social media accounts across platforms including Facebook, TikTok and Instagram. They also documented AI-generated profile pictures and the use of generative AI to produce comments designed to fit the context of online discussions.
Commercial providers can package specialist influence capabilities into services that clients can purchase. An operation may combine fabricated identities, supporting websites, content production, online engagement and monitoring.
AI allows some of this work to be produced and adapted with fewer resources, potentially widening access to capabilities once associated primarily with specialist political and state operations.
For businesses, this makes the identity of a potential attacker harder to assume. A hostile campaign might originate with a geopolitical actor, a commercially motivated organisation or another party purchasing influence services.
Where engagement was reported in the documented operations, it was low.
OpenAI's investigation into Tech and Tariffs, another operation assessed as likely originating in China, found that AI-generated cartoons and comments about US technology policy and trade tariffs attracted little observable attention.
OpenAI also identified a linked network on X that spread false claims about ChatGPT user data being compromised. It assessed the network as likely part of the same operation. OpenAI found no evidence that the false claims spread beyond X or were amplified by authentic accounts with substantial audiences.
The International Burke Institute likewise had relatively low social media engagement, while Mandiant assessed that DRAGONBRIDGE had achieved limited success.
The cases show the work AI was used for, including posts, replies, images, translations and account management.
In a labour-intensive operation, every additional article, translation, image or social media interaction consumes resources. Failure imposes a cost that limits how many further attempts an operator can make.
When AI reduces those production costs, failed experiments become less expensive. An operator can test several messages, abandon those that attract no attention and continue producing others.
The defender faces a different cost structure. A company cannot automatically dismiss an allegation because the accounts distributing it appear inauthentic or have small audiences. It may need to investigate whether a document is genuine, whether an apparent employee exists, whether accounts are coordinated and whether a claim is reaching customers, journalists or other stakeholders.
That investigation requires time, evidence and judgement, even if the material under examination took very little effort to create.
An attacker can afford repeated unsuccessful attempts when producing another costs little. A defender must decide which attempts warrant attention without knowing in advance which, if any, will reach an audience that matters.
Individual false claims are often easier to investigate than an apparent body of corroborating evidence.
A reader encountering one anonymous allegation can question its source. The same allegation can appear more credible when several apparently independent websites, social media accounts and research organisations repeat it.
That is the purpose of many coordinated influence operations. They attempt to give one narrative the appearance of support from several directions.
The Burke operation used an institutional website, copied academic work, attributed experts and social media accounts. Data Center Bandwagon combined fabricated personas with material drawn from genuine reporting. BlackCore's documented activity involved networks of accounts designed to participate in apparently ordinary online conversations.
In each case, the mechanism depended partly on obscuring relationships between sources. Ten accounts repeating the same claim may lead back to one operator.
Generative AI makes it easier to expand the amount and variety of supporting material. That raises a further concern as people increasingly use AI systems to research organisations, assess claims and form an initial understanding of unfamiliar subjects.
AI-generated answers can draw on published material and may present information from several sources in a single response. Where those sources are unreliable or part of a coordinated operation, the apparent breadth of agreement can be misleading.
The documented cases do not establish that these operations succeeded in systematically manipulating AI answers. They do, however, provide examples of how fabricated authority and repeated claims can enter the information environment from which both people and AI systems draw conclusions.
A reputational attack need not reach millions of people to become a problem. It may be enough for misleading material to be encountered by a journalist researching a story, a customer evaluating a supplier or an investor conducting preliminary checks.
The audience that matters may be small, provided it includes people whose decisions affect the organisation.
Organisations cannot win on volume. Their stronger position comes from accurate evidence, trusted sources and an established process for dealing with questionable claims. Four preparations matter.
Map vulnerabilities. Identify the issues around which a hostile narrative could most readily develop. These might include executive conduct, workplace practices, environmental performance, product safety, data security or major infrastructure projects.
A company should understand where inaccurate claims could exploit public uncertainty, missing information or differences between what the organisation says and what outsiders can verify. Genuine criticism and legitimate activism must remain distinct from fabricated or deceptive activity.
Keep evidence ready. Maintain accessible primary documents, accurate corporate information, technical explanations, historical records and clear statements of policy.
This information should be current, internally consistent and available to the people responsible for responding. Where appropriate, it should also be published through authoritative channels so that journalists, customers, search engines and AI systems can find it.
Monitor narratives. Monitoring should extend beyond counting mentions of the organisation. Teams need to understand which claims are emerging, which sources are promoting them and whether apparently independent accounts or websites are connected.
AI can assist by summarising material, comparing competing claims and identifying patterns that merit investigation. Search results and AI-generated answers can also be sampled to understand how an organisation is being represented, although individual outputs do not establish what audiences generally believe.
Agree the response in advance. Decide who assesses suspected impersonation, fabricated documents, synthetic media and coordinated activity. Establish who has authority to approve a response, what evidence must be preserved and when an issue should be escalated to legal, security or senior management teams.
Not every misleading claim deserves a public rebuttal. Responding can give an obscure allegation more attention than it would otherwise receive. In other circumstances, failing to correct a falsehood may allow it to circulate without challenge.
The decision should follow an assessment of the evidence, the likely audience and the consequences of both responding and remaining silent.
Generative AI reduces the production effort behind several parts of an influence operation. It makes content adaptation, repeated attempts and sustained publishing easier to organise with fewer resources.
Where engagement was reported, it was low. That is why the lower cost of repetition deserves attention. Operators can continue testing narratives and audiences without bearing the same production costs that once constrained these activities.
Organisations face a different challenge. Establishing whether an allegation is credible still requires evidence to be located and assessed. Suspected deception must be investigated, and decisions about whether to respond still carry reputational consequences.
AI can help defenders work more efficiently, but it cannot replace the credibility of the underlying evidence or the judgement required to use it.
A practical starting point is to identify the three issues around which a hostile narrative could most readily develop. For each, the organisation should know what evidence it holds, whether that evidence is current, where it can be found and who is responsible for assessing a disputed claim.
Adelbury's AI Reputation Audit shows how AI tools describe an organisation and which sources sit behind those descriptions, which gives this preparation a baseline.
Stephen is the founder of Adelbury. He has spent the last twenty years working in strategic digital communications for some of the largest agencies and brands.
