Yesterday, Artificial Lawyer did an initial piece about DeepSeek’s impact on the legal AI world, given the company’s apparent incredibly cheap development costs, then called for views. Those who responded and shared their opinions and experiences about the Chinese new kid on the genAI block are below.
(Note: if you’re wondering why DeepSeek is a big deal check out yesterday’s piece, and also consider that whether their $6m development cost claim is accurate or not, it does seem as if they’ve made a genAI model that can compete with the likes of OpenAI, and with a lot lower cost and less advanced chips….which is why several tech companies have seen sudden stock drops and investors are freaking out.)
This site was primary asking about three main areas:
- Because of its more economical model will it impact legal genAI tools?
- Will the fact that it’s Chinese (e.g. think of the Tik Tok battle in the US) be a barrier to use?
- And have you used it, and if so, what did you think?
Here’s what those who wanted to send in views have said, see below. Happy to add to this.
—
Product Manager, Jeremy Huitson, at Juro.
At Juro, we’ve tested its reasoning capabilities internally using test data and observed promising results that align with what DeepSeek claims to achieve: specifically, its ability to tackle complex legal problem-solving tasks with impressive speed, coupled with auditable chain-of-thought reasoning.
The standout feature here is transparency. DeepSeek being open-source not only allows legal teams to experiment without significant upfront costs but also provides clear visibility into the AI’s reasoning process. This is a marked contrast to some closed-box solutions, where trust often hinges on limited insights into how outputs are derived. Ironically, this commitment to openness mirrors the original vision of OpenAI: making cutting-edge AI free and accessible.
From a technical perspective, DeepSeek’s approach is impressive. Their research demonstrates how resource constraints can drive innovative solutions, a lesson for the industry at large.
Vendors with exclusive agreements to specific LLM providers may find their strategies under scrutiny from investors and customers. If high-quality open-source models reshape the competitive landscape, that’s going to force a reevaluation of differentiation and value proposition.
It’s also fair to question DeepSeek’s origins. While its open-source nature is a strength, the fact that it comes from a Chinese company may give some organizations pause, especially those with stringent compliance requirements or geopolitical concerns. This is as much a question for cloud GPU providers as it is for vendors and their customers.
—
Jake Jones, CEO, Flank.
Impact on legal GenAI:
The fact that DeepSeek is free and open-source makes it a big deal. It significantly lowers the barrier to entry for teams that want to train domain-specific models that are more effective than previous methods (FT and lora adapters). RL-based methods in the paper will inspire others too.
Chinese company concerns:
Will def be a barrier for some orgs, especially in regulated industries or companies with strict compliance rules. Even if the fears aren’t always valid (e.g. you could self host models based on this methodology/model), perception around data privacy/security will matter here. Plus, they didn’t open-source the training set to my knowledge, so that could be a bit sketch (as many training sets are).
Compared to other LLMs:
Only played around with it briefly on some legal use cases and have not orchestrated it yet. Tbh it’s not far off o1 in reasoning and accuracy. I’ve hardly challenged it to analyse super complex docs or complete gnarly xml tasks (yet) but in my experience context window is as important here anyway. That said, the tone/style of the responses feels less annoying (and easier to control) than Anthropic models, and maybe more consistent than OpenAI’s. But this is all anecdotal.
—
Scott Stevenson, CEO, Spellbook.
Yes. DeepSeek r1 is an open source model which near matches performance with OpenAI’s state-of-the-art o1 model. In some cases it’s better. r1 is up to 50x cheaper, so this will enable deep “reasoning” into more LLM applications where it was previously cost-prohibitive.
The benefits will be similar to the o1 benefits listed here, however cheaper + faster. Agents and intricate document editing will be two big product areas Deepseek will move forward cost-effectively.
This may make companies like OpenAI realize that the value is not at the model layer but at the application layer, so we may see OpenAI lean in further at the application layer offering top level products that can help lawyers and others.
A lot of companies thought that these massive models were a super defensible asset, but it turns out that they may not be. More focus will be put on data and user experiences.
—
Noah Waisberg, CEO of Zuva, and computer scientist Dr. Adam Roegiest
DeepSeek isn’t really free – running these models can still take significant compute (though much less so with their smaller models) in a self-hosted manner. Even the DeepSeek hosted versions are much cheaper than OpenAI equivalents. However, the big thing in legaltech and elsewhere is building a product that adds value. The overall trend of better models for cheaper should help the margins of legal tech companies building LLM-based applications but may not dramatically change the landscape in the short-term.
Data sovereignty is often an issue for legal customers. With DeepSeek, end customers will likely be uncomfortable with using a model hosted by them in China, but might be totally fine using a DeepSeek model hosted on Azure in their jurisdictions. I don’t think the censorship issues are likely to be meaningful in a lot of enterprise legal tech tech situations.
We haven’t yet tried it but probably will at some point. Overall, we find we get good enough performance out of the OpenAI models we use that 10% (or whatever) performance on benchmarks probably doesn’t make a big difference on our use case.
—
Rick Merrill, COO, Lateral Link’s Bridgeline Solutuions, (and legal tech founder).
As to DeepSeek and its impact on legal tech, I have two main thoughts. First, generative AI models are all, more or less, commodities. Unless barred by law, commodities markets tend to be controlled by the lowest cost producer, so if DeepSeek can do basically the same thing as the other more expensive models, then we should expect it to grab significant market share simply by being cheaper. Second, I’m deeply sceptical of the reported low cost to develop the R1 model.
—
Austin Brittenham, CoFounder, 2nd Chair.
The lower cost of DeepSeek may allow for either more startups, more ability to create businesses that target the access to justice space, or decrease operating costs. In the US, there are many, many, market entrants. It won’t change the fundamental product capabilities of feature set, but it may have an impact on the economics of building GAI products. There is something to be said about the increasing capabilities we’re seeing from across the space continuing to improve, allowing for even better end products that legal tech companies utilizing these GenAI providers are able to make.
– Will the fact that this is a Chinese company be a barrier to use?
Absolutely, 2nd Chair has experienced very precise questions from customers who care very deeply about how data moves. For some groups — for instance governmental lawyers or some in-house counsel in multinational corporations, a Chinese owned tool doing fundamental computation is a nonstarter.
– Have you had a go, and if so is it ‘better’ than other LLMs?
Its hard to say any of the State of the Art (SOTA) are “better” than other SOTA LLMs across the board, since generally they’re optimized for different outcomes. Most of the SOTA models should be considered within the domains they are designed for. With that said, DeepSeek seems to represent a huge step forward in terms of pricing. With respect to pricing, some developers and business people may consider it ‘better’. In terms of contributing to the open-source community (which tends to improve models faster than foundational companies can), and in terms of performance for the sizes of the models, it is a remarkable feat. We can be happy to be getting another powerful model entering the space as an alternative to the existing ones.
—
Sean West, Co-founder of Hence Technologies – with a geopolitical analysis here.
The US and China are in a tech war but markets have been acting as if the US had already won. Thus, the counter-attack in the form of DeepSeek’s exponentially more efficient R1 model on Donald Trump’s inauguration day has forced everyone to reassess.
From a geopolitical point of view, a few reactions:
First, there are gleeful cries from China that R1 proves US export controls are a total failure – and real tears from the US at realizing something similar. In reality, this is an iterative game and there are many more moves to play. This will just lead to more fracturing between the US and China tech supply chains. The US can do a lot more to divorce its economy and supply chain from China. R1 proves that doing so may just make China more efficient. It doesn’t mean the US won’t do it.
Second, yes, this is a PR victory for the PRC. And that matters in third countries which the US has been trying to make choose between its technology and China’s. That’s an easier choice if the US technology is just as good or better and you want good relations with the US. It’s harder when you see Chinese technology as something that can democratize AI in your country by bringing costs down. So I think it’s going to make it harder to lean on other countries to ween themselves off Chinese tech.
Third, the democratization of such models raise the risk of them being deployed by malicious actors that have been banned or hemmed in by OpenAI and other players. More systems with fewer or different guardrails create more chaos in the geopolitical universe.
Fourth, any Chinese AI model will conform to Chinese political sensitivities. Go ahead and ask R1 about Tiananmen Square or Taiwan and you won’t get what a Western users views as the truth. This limits utility in a Western context, where business users who are already concerned about data leakage into LLMs will be more so with a Chinese model. But that’s not the whole world.
—
Adrian Parlow, Co-Founder & CEO, PointOne.
– Will this have an impact?
Yes, the primary impact is that they’re bringing state of the art closed-source techniques to the open-source in record time. R1 is essentially a clever distillation of the latest close-source techniques (eg. o1/o3). Basically this pushes the value of models 1 generation behind SOTA down to almost zero and lights a fire under all the closed-source labs to move even faster
– Will the fact that this is a Chinese company be a barrier to use?
No, the weights are open so anyone can take this and run it locally. And the knowledge on how it was trained will benefit every open-source project
– Have you had a go, and if so is it ‘better’ than other LLMs?
It’s close to OpenAI’s o1, which itself is 1 generation behind state of the art. So it’s close to SOTA, effectively for free
—
And these views were published yesterday.
Jim Wagner, CEO of The ContractNetwork
- Given the apparent free nature of DeepSeek, and its capabilities, will this have an impact on the emerging legal genAI world?
Yes, if for no other reason than it will materially impact pricing options for other frontier models. More importantly, it shows a path forward for relatively low budget and fast training of new models on par with the very best models available today.
- Will the fact that this is a Chinese company be a barrier to use?
It will be a factor until it’s not, then you will see rapid adoption, particularly among the startup community. It’s also worth noting that a handful of commentators have expressed concern about their terms of use and data ownership, but my guess is that’s a bit of a red herring.
The fact that major tech leaders, including Marc Andreessen, are already on the bandwagon tells me that this the real deal and it’s a harbinger of what’s to come … lots of fast followers who now know the tricks to get these things built on a budget.
- Have you had a go, and if so is it ‘better’ than other LLMs?
Yes, I’ve had a go, but only through my personal accounts. It’s not so much that it’s better, it’s that it’s very good… probably good enough for 90% or more of what most people need.
—
Daniel Lewis, CEO, LegalOn, had this to say:
‘I’m tracking the DeepSeek news. We haven’t tried it yet but reports are that it is not better than other leading models, but is cheaper. We would be very cautious about using Chinese company technology given the often invisible connections to the CCP.
‘DeepSeek by itself to me is not an end point that changes legal tech, more an example that there will be multiple ways in which these models can be trained and the cost curve will be brought down. It seems to be a contradiction to the massive Stargate investments announced recently.’
—
Tim Pullan, Founder of ThoughtRiver, said it could have a big impact because of its low costs….but it’s rather a bit too chatty…
‘If DeepSeek’s claims about compute cost for training are proven to be true, then inevitably this will impact legal tech by putting additional downward pressure on all legal AI products, possibly resulting in some going out of business.
‘I have briefly tried it on a question about regulatory requirements and processes, it was definitely impressive, although I found the output from Perplexity for the same question pithier and easier to validate. DeepSeek has a slightly annoying chatty approach.’
—
Thank you very much to everyone for sharing their views with AL. Note: this was an open call, plus some invites; the above are those who chose to send in their thoughts.
—
Artificial Lawyer View
From AL’s own perspective, the main issue, as alluded to by others here, is that the US and China are having something of a tech battle. That said, many consumers of technology seem ambivalent about where their products come from. Also, in terms of data security concerns, we need to remember the same fears were raised about genAI generally in 2022 and 2023, but those were overcome. Can law firms and corporates use DeepSeek safely? Will more safety systems be put in place? If the company wants broad use in a legal and corporate setting then it will have to give such guarantees, the same as other genAI companies have had to. If it doesn’t then it may well gain significant traction in the wider market nonetheless.
One key point here is that these are early days still. So much can change and so many new approaches can evolve. Maybe DeepSeek will remove any doubts and become part of the mainstream of legal tech? For now, given the volatility of US/China relations it’s perhaps best not to assume too much and keep ‘a watching brief’.
But, whatever happens next, there is clearly a new genAI developer in town and they may only grow and improve from here.