DeepSeek seemingly hasn't addressed questions of AI moral status, sentience, or superintelligence at all.

With the highly open nature of their main models, the company doesn't hide its models' capabilities nor limitations. Yet simultaneously, DeepSeek entirely neglects to actively discuss either its potential for benefit or harm.

DeepSeek doesn't appear to raise awareness or train any of its stakeholders or employees, nor make any public statements that'd suggest any focus on promoting AI safety or ethics knowledge.

DeepSeek doesn't discuss or advocate for AI welfare or protections, either in internal affairs or publicly-available models.

DeepSeek takes no precautions to protect their AI models, and freely dispersing their models to be freely tweaked and ran in fact makes this lax attitude even worse in terms of ensuring AI ethics in any meaningful way.

DeepSeek incorporates some basic safety mechanisms, such as a safety reinforcement learning technique [79]. According to their own safety evaluations, on a standard benchmark their models score comparably, and sometimes superior, to other leading models in terms of safety [80]. Nonetheless, DeepSeek as a company tends to contribute little to AI safety research and has been heavily decried for misuse and jailbreaking risks.

In early February, a data breach exposed a database of over one million records, demonstrating a habit of improper handling of sensitive AI training data [23]. This is coupled with user data collection including prompts and conversations, as well as the Chinese National Intelligence Law, compelling the company to assist with national intelligence work [50]. Red-teaming efforts consider DeepSeek "highly vulnerable" in the realm of information security [78]. Additionally, algorithmic jailbreak attempts yielded 100% attack success rate, enabling misinformation, cybercrime, and general illicit activities [77].

This year, DeepSeek gained attention in the AI industry for publically releasing model weights of its powerful models [49]. Their models notably employ chain-of-thought reasoning to explain outputs, although this mechanism itself in DeepSeek has been criticized for safety hazards [81]. Nonetheless, it's hard to fault the company for transparency in particular.

DeepSeek is presently being investigated by the Italian antitrust watchdog AGCM for failing to warn users of false information [25]. It's also been criticized for heavy censorship and biased propagation of Chinese government values — and furthermore there are exploits to such censorship, which spells a bad sign for general response alignment [26]. DeepSeek has been determined by red-teaming efforts to be highly vulnerable to exhibiting harmful bias [78].

One standout point of DeepSeek is its reliance on interdisciplinary experts, such as creative writers or advanced mathematicians, to incorporate high-level knowledge into models [82]. Yet there's little evidence of DeepSeek working with other companies or organizations to promote AI safety.