Opinion | The real threat AI poses to journalism isn't deepfakes. It's DeepSeek.

Public anxiety about artificial intelligence in 2025 misses the mark. Deepfakes grab headlines, but journalism faces a quieter threat. DeepSeek, the model making headlines and developed in China, reflects an evolution in information control that restructures what readers can know and journalists can report.

Traditional censorship blocks content. DeepSeek transforms it. The system filters discussions of events like Tiananmen Square, but goes further by actively reshaping historical narratives and omitting key information on topics not aligned with core socialist values. This capability extends beyond Chinese borders with its rapid adoption. As of this writing, it has literally moved the financial markets. While all models carry inherent biases and limitations, DeepSeek highlights its restrictions in its terms of service. 

The 2024 elections revealed this shift. According to the Microsoft Threat Analysis Center, Chinese influence operations focused on down-ballot Republican candidates and members of Congress that advocate for anti-Chinese policies, including campaigns against Rep. Barry Moore, Sen. Marsha Blackburn and Sen. Marco Rubio. Actors parroted antisemitic messages, amplified accusations of corruption and promoted opposition candidates. High-level political figures also shared manipulated media. Truth eroded not through obvious forgeries but through accumulated small changes to the information landscape.

At scale, newsrooms feel the economic impact of this as well. Search results shift to reflect the broadening, but restrictive, narrative. Context changes over time. When algorithms optimize for engagement or efficiency, they create blind spots in coverage. The process happens so gradually, and at such a scale, that the true nature of past events fades away.

The journalism community lacks adequate protections against these shifts. Professional standards for AI development remain undefined, and much of the work that has been done to generate broad mandates have already been repealed. Few mechanisms exist to test for bias or enforce transparency about AI's role in news production. Without such safeguards, the foundation of factual reporting weakens.

DeepSeek signals a turning point. As AI systems gain sophistication, their influence over public discourse grows more profound. The future of journalism depends on whether newsrooms can maintain independence from technologies that subtly reshape reality.

What is Nota doing about this? As we scale to provide assistive AI to thousands of news outlets across the country, we’re mindful of the threat that model bias represents and our outsized position within this space. While we continue to strengthen our data and model policies, over the last 24 hours we’ve also placed an explicit “cannot use in any capacity” on the DeepSeek models. That means we cannot download, use, benchmark or deploy DeepSeek or any of its (now over 500) variations in any capacity, by any employee of Nota. 

Moving forward, we are developing model analysis and validation tools, training our editorial staff to spot geo-political abnormalities in outputs, and working on more reasoning checks and validation loops of our proprietary model, Polaris. 

We have the good fortune that DeepSeek is explicit in its use of bias and redaction, but future models will not be transparent, and the opportunity that generating a highly adopted model fine-tuned to a geopolitical agenda cannot be lost on other nation states keen to have their worldview exposed in every AI output. 

This is another next level of information control, one that will become proliferant and harder to analyze as success is shown and the models evolve in obfuscation and sophistication.

Every company is now a media company. Let's ensure that human oversight remains central to a free and informed society.

Josh Brandau is CEO and co-founder of Nota.