The world is at a point where artificial intelligence tools are no longer just helpers for writing emails but are becoming active participants in geopolitical struggle. According to a recent report by OpenAI, LLM models (large language models) are now the primary tool for what is known as astroturfing – creating a false impression of mass support or opposition to a particular topic.
Digital Disinformation 2.0: How Does AI Manipulation Work?
Traditional bot farms, which we knew from past years, had their limits. They were easily detectable due to repetitive style and grammatical errors. Generative AI, however, completely removes these barriers. According to the OpenAI report, Chinese entities used ChatGPT to create high-quality content that was spread on the social network X (formerly Twitter).
The manipulation focused on two key topics:
- Political Economy: Generating slogans criticizing the tariff policy of US President Donald Trump.
- Environmental Concerns: Creating visual content (images) that depicted future AI data centers as the main culprits of energy waste and environmental destruction.
An interesting aspect is the multilingualism of these operations. ChatGPT was not only used for English content but also for posts in Japanese and Italian, indicating an effort to spread disinformation globally, not just within the USA. This shift from local disinformation to global campaigns is a warning sign for security experts worldwide.
Comparison of Capabilities: OpenAI vs. Competition
In the fight against these manipulations, model providers are at the forefront. While OpenAI strives for active detection and blocking of these activities, competitors have different approaches:
- ChatGPT (OpenAI): Offers robust systems for monitoring abuse. The GPT-4o model is available in the Czech Republic for free (with limits) or as part of a Plus subscription for 20 USD/month (approx. 460 CZK).
- Claude (Anthropic): Known for its higher degree of "ethical barriers" and security, making it a more difficult tool for generating hateful or manipulative content. Claude 3.5 Sonnet is available as part of a subscription for 20 USD/month.
- Gemini (Google): Google uses its extensive knowledge of the internet ecosystem to detect bots, giving it an advantage in identifying content dissemination within search and YouTube.
Impact on the European Scene and the Czech Republic
Although the report focuses on the USA, for us in Europe and specifically in the Czech Republic, this news is very relevant. It's not just about distant geopolitical games, but about a technological weapon that can be aimed at our voters or domestic industry.
The EU AI Act, the most important artificial intelligence regulation in the world, already contains mechanisms for addressing systemic risks. One of them is the obligation to identify and mitigate disinformation generated using AI. The European Union places great emphasis on transparency – if content is generated by AI, it should (ideally) be clearly marked. However, this OpenAI case shows how difficult it is to catch these manipulations in real-time before they spread.
For the Czech user, this means we must increase our digital literacy. If you see a sudden influx of highly emotional and visually strong images on social networks (e.g., on X or Facebook) concerning controversial topics – such as energy or politics – it is necessary to consider whether an automated system is behind them.
How to Recognize AI Disinformation? Practical Tips
- Analysis of visual details: Although images generated by DALL-E 3 or Midjourney are constantly improving, they often show minor anomalies in background text or object geometry.
- Source checking: If a post is published by dozens of anonymous accounts simultaneously with identical or slightly modified text, it is a sign of astroturfing.
- Emotional charge: AI is great at evoking immediate emotion (fear, anger). If content immediately "pulls" you towards a radical opinion, be cautious.
Conclusion
While OpenAI states that the impact of Chinese manipulations has been limited so far, the very existence of these operations confirms that LLM models have become a key pillar of modern information warfare. For companies and individuals in the Czech Republic, this means that protecting digital identity and critical thinking are more important than ever before. With the advent of new models, this battle between disinformation generators and their detectors will become increasingly intense.
Can ChatGPT directly generate false news that looks true?
Yes, technically it is possible. Although OpenAI has set up safety filters (guardrails) to prevent the generation of harmful content, advanced users try to bypass these barriers using so-called "jailbreaks" or sophisticated prompt engineering to make the model create disinformation.
What is the difference between a bot and an AI agent in the context of manipulation?
A classic bot is a simple script that repeats the same texts. An AI agent is much smarter; it can react to other users' responses, lead debates, change arguments according to context, and create unique content, which makes it a much more dangerous tool for disinformation.
Is it possible to detect these manipulations in the Czech Republic using common tools?
For the average user, there isn't one "detector," but a combination of critical thinking and fact-checking through trustworthy media is the best defense. For companies, there are specialized cybersecurity tools that analyze anomalous behavior on social networks.
