Listen to this article:
In the world of technological development, we are accustomed to a constant battle for market share among giants like OpenAI, Google, or Anthropic. However, on the issue of AI safety, these companies have for the first time stood together on the front line. Their warning is not based on speculation about a distant future, but on a real shift in how digital information can now be translated into biological reality.
Erosion of Knowledge Barriers: How AI Is Changing Biology
Historically, the production of biological weapons was limited not only by the availability of materials, but primarily by extremely high expertise requirements. Designing an effective pathogen required years of study in genetics and biology. As reported by qz.com, AI models have the potential to completely destroy these "knowledge barriers."
Large language models (LLMs) such as GPT-4, Claude 3.5 Sonnet, or Gemini 1.5 Pro excel at information synthesis. They can search through thousands of scientific studies and find connections that might elude a human scientist. The danger lies in the fact that the model can help a user design a DNA sequence that is specifically modified to evade existing detection mechanisms.
A real-world example is 2017, when Canadian scientists were able to recreate the extinct horsepox virus using DNA purchased by mail for $100,000. With the rise of AI, the "design" process of this virus becomes much faster and more accessible.
The Synthetic DNA Problem and Failing Screening
Today's technology allows companies to "print" (synthesize) custom genetic sequences on demand. The problem is that not all of these companies perform thorough checks on their customers or the sequences being ordered.
Research conducted by Microsoft has shown that there are AI tools for protein design that can generate dangerous sequences in such a way that current software filters cannot recognize them. This creates a dangerous gap between digital design and physical manufacturing.
Dario Amodei (Anthropic), Sam Altman (OpenAI), and Demis Hassabis (Google DeepMind) argue that regulation must address two points:
- Customer screening: Companies producing DNA must know who is buying material from them.
- Order screening: Every ordered genetic sequence must be compared against databases of known pathogens.
Practical Impact for Europe and the Czech Republic
Although the initiative is primarily aimed at the U.S. Congress, its effects will be felt worldwide, including the Czech Republic and the EU. In the European context, this topic directly concerns the EU AI Act (Artificial Intelligence Regulation), which already classifies certain systems as high-risk.
For the Czech scene, this has several dimensions:
- Biotechnology research: The Czech Republic has a strong tradition in medicine and biotechnology (e.g., research within university centers or pharmaceutical companies). If these regulations become stricter, research institutions will have to invest more in control and compliance processes to ensure their DNA orders are not blocked.
- EU regulation: The European Union often uses American initiatives as a basis for its own standards. If strict screening is adopted in the U.S., it is highly likely that European regulators will require similar transparency from biological material suppliers as well, within the framework of critical infrastructure security.
- Tool availability: AI model developers (OpenAI, Google) are already implementing safety filters ("guardrails") that refuse to answer queries related to biological weapons production. These filters are available in both Czech and English and work globally.
Comparison of AI Safety Approaches
Different companies approach the mitigation of these risks differently, which is also evident in their technical parameters:
- Anthropic uses the concept of Constitutional AI, where the model has built-in principles (a constitution) that forbid it from generating dangerous content.
- OpenAI focuses on intensive "red teaming" — simulated attacks by experts who try to make the models violate rules, so that weaknesses can be fixed before reaching the public.
- Google DeepMind integrates safety layers directly into the development of multimodal models (Gemini), which can analyze not only text but also visual data from laboratory tests.
Can AI actually create a new virus?
Not directly. AI has no physical body or laboratory. However, it can act as a "super-expert" that provides precise instructions on how to design or modify a virus to be more dangerous or difficult to detect to someone who has access to laboratory equipment and a DNA synthesizer.
How will this affect ordinary biotech companies?
Companies will have to prove their legitimacy when ordering genetic materials. This may mean more bureaucracy and the need to use certified suppliers that meet strict screening protocols. For small startups, this could mean higher operating costs.
Is this issue relevant for Czech AI developers as well?
Yes, because developers must implement these safety guardrails directly into their models. If a Czech startup is developing its own LLM, it must expect that regulations (including the EU AI Act) will require evidence that the model does not assist in the creation of dangerous substances.
