AI and the risks to cognitive security

Notes from the Sofia Information Integrity Forum
By Katy Ling, Client Partner

Some of the most useful conversations I've had about AI this year happened in Sofia, Bulgaria, with people who have spent years watching information operations up close.

I was there last week for the Sofia Information Integrity Forum, speaking on a panel about AI and cognitive security. The term sounds abstract, but the idea is simple: protecting how people come to believe what they believe. More and more, that process starts with an answer from an AI model, and very few people check where the answer came from.

From false claims to narratives

Over three days, the conversation kept coming back to narratives rather than individual false claims. A single fake story can be checked and corrected. A narrative is a frame repeated across enough sources that it starts to feel like common sense, and it survives any single claim being debunked. The techniques used to shape opinion about governments are also used against companies and individuals, often timed around a deal or a court case. That was the subject of my own talk, and it is a growing part of our work at Dablam.

AI is where narratives now settle

A claim placed on an obscure website doesn't stay there. It is crawled, indexed and retrieved until it becomes part of a chatbot's answer, and along the way the details that would make a careful reader suspicious drop out, including who wrote it and the fact that ten "sources" are really one. Peter Benzoni of the Institute for Strategic Dialogue (ISD) compared that final step to ventriloquism: the model speaks someone else's words in its own fluent voice. Because the answer sounds authoritative, few people question it, and by then there is little left to investigate.

False information is also hard to undo once people have seen it. Jeremy Fannin's research at EDHEC Business School found that deepfakes damage the reputation of the person targeted even when viewers are told the video is fake. If that holds for a video people know is fabricated, it should worry anyone thinking about AI answers that people have no reason to doubt.

Learning from Central and Eastern Europe

Countries closest to Russia have lived with its information operations for decades. Across the region, manipulation is treated as a security issue, and there is a strong push to teach people how to judge what an AI tells them. It is the information literacy many of us were taught for the early internet, then social media, now updated for AI. In the UK, AI answers are still mostly treated as a search or marketing question, rather than a risk that senior leadership needs to manage.

What this means for companies and the people who lead them

Find out what the AI models currently say about you.  Whether you are mentioned matters less than the narrative they build around you, including which controversies they bring up and whether that picture improves or worsens over time. Few people check where an AI answer came from, so for many readers that narrative becomes what they believe about you.

Trace where it came from. Several outlets repeating the same claim can look like confirmation to the LLMs, when they are really copies of one original or part of the same inauthentic network.

Act early. Corrections rarely undo the damage, so the strongest position is an accurate, authoritative record in the sources AI draws on. For a private individual or a young company with little written about them online, there is not much else for a model to use, so a hostile narrative can quickly become the main thing it finds.

In Sofia, protecting the information people rely on is treated as part of national security. Companies and the people who lead them should give the same care to what AI says about them.

Related News & Insights

AI and the risks to cognitive security

Read more

Multiple truths, one answer

Read more
curved architecture graphic

LLM Poisoning: Russia's New Playbook and the Risk to Your Reputation 

Read more
Analysis icon on the gradient background blue to orange diagonally

Trust is not a software problem

Read more

What Managing Reputation in AI Actually Looks Like in Practice

Read more

The first read on you is now an AI model

Read more

Protect your reputation in AI-Search & LLM

Manage high-stakes issues and build a digital reputation that creates competitive advantage

See how
Copyright © 2026 Dablam Ltd (company number 15115628) | All Rights Reserved

14-15 Southampton Place, Holborn, London WC1A 2AJ
linkedin facebook pinterest youtube rss twitter instagram facebook-blank rss-blank linkedin-blank pinterest youtube twitter instagram