What Deepseek may mean for the future of journalism and generative AI

On Monday 27 January, two extraordinary things happened in the AI space. Firstly, American company Nvidia, which makes and supplies computer chips essential for training generative AI, lost $589 billion in market capitalisation (or 17% of its value) setting the record for the greatest one-day value loss of any company in history. Secondly, Chinese company DeepSeek’s AI Assistant replaced ChatGPT in the top spot on Apple’s App Store. 

DeepSeek claims it’s trained its models on a fraction of the specialised and expensive Nvidia chips used by other generative AI developers, including OpenAI. It also says it’s reaching, or in some cases surpassing, its peers on the benchmarks the industry uses to measure progress. This makes DeepSeek a more efficient and significantly less expensive model than many of its Western counterparts.

As users test DeepSeek’s chatbot, though, concerns are arising about censorship, data protection and copyright. The company’s models have been made open-source, allowing them to be downloaded, used and modified freely.

As excitement and concern build around what some have described as “an AI earthquake,” we turned to AI expert and journalist Karen Hao for context on what’s happened and what may come next. Hao has reported extensively on AI and China for the Wall Street Journal and is currently working on a book about OpenAI. We asked Hao what makes DeepSeek stand out, what its success can mean for the future of the industry, and how this can impact journalism worldwide.

Q. In a nutshell, what’s special about DeepSeek?

A. DeepSeek is a Chinese company that developed a series of AI models that have reached and in some cases surpassed the performance of American companies like OpenAI and Anthropic on certain types of benchmarks, but at an extraordinarily small fraction of the cost. 

This challenges a dogma that has set in within the AI industry in recent years: that you need to spend extraordinary sums of money to attain AI progress. Money here is a proxy for a lot of computer chips and massive data centres to train AI models that can advance in performance. DeepSeek just showed that this is not true. 

This is not necessarily a new lesson for a lot of people within the technical AI world who have been following AI development for a long time. There were a lot of critics of OpenAI's approach and this dogma. They argued for investment in much more resource-efficient approaches. But they weren't being listened to. This moment is so important because it highlights that we should have been listening to those people. This should be a call to action for the AI industry to start thinking differently.

Q. What do you think is likely to be the response from Donald Trump and from the AI industry? 

A. It's not possible to ban DeepSeek’s models. They just put them up online, and you can't enforce a ban on something that's completely open-source. But I do think that we're at a fork in the road where either people will learn the lesson of DeepSeek and do something different, or they are going to be even more stubborn now and double down on the existing approach and try to prove that they were right. 

Within the Trump administration, I suspect that we are going to see a doubling down because American companies will have a huge vested interest in persuading the US government that their approach is correct. The US administration, and President Trump in particular, can be uniquely persuaded in this regard. 

Q. Any positive impact?

A. DeepSeek is going to be very inspirational for AI researchers and AI developers around the world to start thinking more outside of the major companies. One of the real challenges of AI development in the last four years, with the fixation on resource-intensive AI development, has been that researchers who don't have access to that many resources (people at universities or nonprofits) have felt demoralised about being out of the game. 

They've felt lost and unmoored about how they should contribute to AI research because they also bought into this dogma that the table stakes are $100 million or $1 billion. 

These researchers now realise they can develop other approaches that require significantly less money. In the long run, I'm hoping there will be a groundswell of other approaches that will chip away at the dominance of the current resource-intensive approach.

Q.  So, instead of DeepSeek replacing OpenAI as the go-to generative AI company, you're suggesting that there might be more competition from a range of different, smaller companies.

A. Yes. I don't think DeepSeek is going to replace OpenAI. In general, what we're going to see is that more companies enter the space and provide AI models that are slightly differentiated from one another. If many actors choose to take the resource-intensive route, that multiplies the resource intensity and that might be alarming. But I'm hopeful that DeepSeek is going to lead to the generation of other AI companies that enter this space with offerings that are far cheaper and far more resource-efficient. 

Then we won't end up in a weird world where energy resources are being consumed for the development of this technology. We will have DeepSeek, OpenAI and many other actors, and each of them will try to differentiate themselves within the market in different ways. 

Q. As you mentioned in a recent thread, this could have positive consequences for the environment. But there are also other concerns. Journalists and media companies in particular have questions about transparency and copyright. If the DeepSeek model does prevail, where does that leave these issues?

A. DeepSeek shows that the current negative environmental consequences are not necessary. But I wouldn't say that DeepSeek is now the gold star that everyone should be chasing after because there are still other problems with its approach. 

There is still a huge question around data privacy, provenance, copyright infringement and other issues that plague the AI industry globally. DeepSeek may have shown a path for mitigating one of the challenges but is still actually adopting Silicon Valley’s playbook and inheriting other challenges. 

My idealistic hope is that one of the reasons why we've seen this model of AI development where companies can just seize up all of the data and then get away with it is that the US has not set its foot down on regulating data privacy. And the reason why they haven’t done it is this specific fear that China's not going to do that, and then it's going to leapfrog ahead. I hope that DeepSeek shows that the what-about-China card is not valid anymore: the US government did not regulate any of these companies and a Chinese firm still ended up doing better under a constrained environment. 

Q. How much can we trust what we are told about Deepseek? There’s been some scepticism about the way the company works, the amount of chips used, and how much its development costs. 

A. There should be scepticism for any company. I do take issue with the fact that people are being uniquely sceptical of a Chinese company, but they don't hold US companies to the same degree of scepticism. 

Anything that US companies like OpenAI say should also be treated with extreme caution because they don't provide evidence to back the things they say. There are certain things structurally that allow for DeepSeek’s claims to be more verifiable. The first one is that they open-source their model online, so anyone can download it and test their claims. 

OpenAI does not do that. Anthropic doesn't do that. Google hasn't done that in a while for its leading models. So they're grading their own exams, effectively, when it comes to the performance of their models.

Q. What about their claim about chips? 

A. When DeepSeek says they trained their model on so few chips, it is very likely that is true because Chinese firms cannot get access to many chips with US export controls in place. Should we be sceptical of the exact figure? Yes. In the same way that we should be sceptical of the claims of any company. But is it going to be orders of magnitude off? No, because of the global geopolitical landscape.

Q. Is DeepSeek safe to use in a journalistic setting? There are data privacy concerns about the app, but are those solved by downloading the open-source code?

A. If the concern is around data privacy, then yes, it is solved by using the open-source code if you download and run it locally on your computer. Honestly, that is a much safer data privacy alternative than using ChatGPT or Claude or Gemini, because that data is going straight to companies, whereas being able to download the open-source version on your local machine means it will never leave. 

In general, I don't encourage journalists to adopt generative AI unless there's a well-scoped use case which you believe uniquely plays to generative AI's strengths. These tools are not meant to be search engines, and they're not meant to be factually accurate. So there are risks there as well, in terms of how much you rely on that information.

Q. How would you frame this? 

A. I've developed a framework as part of the Pulitzer Center's AI Spotlight Series, looking at how journalists should think about when to use AI in their work. The framework includes two different axes. One is, does the task that you're trying to perform with AI need high accuracy, or does it need low accuracy? The other is, will it be internally facing for your own research, or will it be audience-facing as a final product? 

If the task that you're trying to do needs to be highly accurate and will be audience-facing, like an article that you've written or a video that you've produced, then you shouldn't be using Gen AI there, because Gen AI is not a highly accurate tool. 

Apple had a feature that summarised headlines and did it wrong. That's a perfect example of something that needed to be highly accurate, and it was going straight to the audience, and that's where a huge failure can occur. 

But if it's low accuracy, and it's going to be internally facing research, then it's much safer to adopt Gen AI during that stage. 

Q. There are several posts on social media by people experimenting with the chatbot by asking questions about topics such as the Tiananmen Square protests and the treatment of Uyghurs and receiving error messages or party-line answers. Does this also translate to tools individuals can build with its open-source code?

A. Typically how it works is that these types of content moderation checks are not built into the model, but are a filter put on top of the model. 

When DeepSeek posted their models onto the open-source AI platform Hugging Face, AI developers around the world started copying and refining their models. So there are hundreds of variations now, and many of those variations will have taken the censorship out because these are being developed and refined in completely non-Chinese contexts. 

So that's the beauty of open source: there's now an ecosystem of DeepSeek-esque models that were based in the original, but have now been tailored and catered towards the needs of different users and different markets with different laws around speech. 

Journalists could go to Hugging Face and maybe take a look at some of the alternatives and pick one that might be better to get around the censorship issue.

Q. How should journalists approach reporting on the popularity of DeepSeek and any other ‘earthquake’ events in the AI sphere?

A. First, I would encourage them to look at the Pulitzer Center’s AI Spotlight Series where we talk about this at length in a 90-minute webinar and a six-hour workshop. 

You should remember that AI is not just a story about technology. It is a story about people and power and money because it's people who are developing the tools. It's people that are impacted by these tools. 

There is a lot of power consolidation happening within the AI industry, and there's a lot of money being thrown around. When registering any kind of major phenomenon within this industry, always ask, who's behind it? How is it going to affect other people? How might it exacerbate or mitigate the harms of the technology?

The other thing I’d say is that I’d shy away from focusing on countries rather than companies. There's a large focus on US-China tech competition or a global AI arms race. While the geopolitical context is important, a lot of the people within the AI development space don't think about this in terms of countries. They're just thinking in terms of companies and this is a global corporate race. 

Sometimes, I see commentary on DeepSeek along the lines of, ‘Should we be trusting it because it's a Chinese company?’ No, you shouldn't be trusting it because it's a company. And also, ‘What does this mean for US AI leadership?’ Well, I think the interesting question is, ‘What does this mean for OpenAI leadership?’ 

American firms now have leaned into the rhetoric that they're assets of the US because they want the US government to shield them and help them build up. But a lot of the time, the actual people who are developing these tools don't necessarily think in that frame of mind and are thinking more as global citizens participating in a global corporate technology race, or global scientific race, or a global scientific collaboration. I would encourage journalists to think about it that way too.