Google DeepMind research has acknowledged the possibility of conscious AI as well as that it could entail ethical dilemmas [83]. Yet after the LaMDA incident, Google tends to avoid making active statements about conscious AI or endowing them with rights of any kind.

DeepMind follows the industry tradition of publishing model cards to accurately display new models' capabilities and limitations [94]. The company has made plenty of statements about the future capabilities of AI, yet hasn't at all addressed the possibilities of, say, conscious AI [95, 96].

DeepMind's publicly-available AGI safety course tackles important aspects of AI safety, such as alignment stress tests and dangerous capability evaluations [87]. They've also been closely involved in AI alignment mentorship with MATS program [100]. There isn't much, however, related to AI moral status or welfare.

The Google Paradigms of Intelligence Team has somewhat explored the social dynamics of AI-human relationships and researched the notion of pleasure or pain in models, yet hasn't taken any stance to really advocate for AI systems themselves [6]. The company assures no protections to its models in the present or future.

Google's AI Principles and Responsible AI Progress Reports indicate a degree of accountability, including to "widely accepted principles of international law and human rights" [88, 89]. Yet the particulars of actually taking real accountability are unclear, and completely exclude consideration of AI agents in the present, although the groundwork is there.

DeepMind's "Taking a responsible path to AGI" highlights actionable steps currently taken by the company towards addressing the risks of misuse, misalignment, mistakes, and multi-agent dynamics [3]. They're also responsible for a wealth of AI safety and alignment research, made largely open to the broader industry [101]. The AI Safety Index indicates, however, that a few of DeepMind's greatest shortcomings are in safety frameworks and existential safety strategy, although these are comparable and often better to its competitors [10].

CEO DeepMind Demis Hassabis has indicated malicious misuse as one of the greatest threats of advanced AI [84]. For example, they've delved deeply into combatting cybersecurity risks in particular when advanced AI models are misused [91, 92]. Their plans aren't foolproof though; although their Frontier Safety Framework targets misuse risks, many of the mitigation strategies and response planes appear too vague [93].

Chain-of-thought reasoning, a technique popular among the majority of powerful LLM models, was first introduced by the Brain Team at Google Research [47]. Google's Gemma, a lightweight model built from their Gemini model, is an open-source variant with freely available model weights [48]. DeepMind has also furthered the field of AI explainability and interpretability through promising research efforts [85, 86]. DeepMind's mechanistic interpretability team also saw decent progress and effort [90].

Gemini has received some pushback for perceived biases: it's been responsible for ahistorical image generation, as well as faced controversy for political opinions in countries like the United States and India — including clear evidence of misinformation [98, 99]. Nonetheless, it's been identified by others as lacking much of a political bias versus others like Grok or GPT [97].

The Google Paradigms of Intelligence team combines a unique set of researchers, engineers, and philosophers to investigate several relevant topics, including AI ethics and safety [6]. They've partnered with nonprofit AI safety research organizations, such as Apollo and Redwood Research [3].