AI Lab Watch

Categories

Companies

Resources

Blog

About

			*
Weighted score	28%	22%	20%	4%	3%	3%	1%
Risk assessment	49%	34%	43%	1%	1%	1%	0%	28% weight
Scheming risk prevention	2%	8%	2%	2%	2%	2%	2%	21% weight
Boosting safety research	68%	55%	35%	28%	0%	15%	8%	13% weight
Misuse prevention	9%	4%	5%	0%	0%	0%	0%	12% weight
Prep for extreme security	2%	4%	0%	0%	0%	0%	0%	12% weight
Information sharing	35%	13%	32%	0%	28%	0%	0%	8% weight
Planning	14%	26%	0%	0%	0%	1%	0%	6% weight

Up to date as of May 21; not updated for Claude 4.

Overall score

Information sharing

Incident reporting

Weighted 40% of category

	0%	No, but Anthropic published a short report focused on misuse in influence operations
	0%	DeepMind doesn't do this
	10%	No, but OpenAI does publish information about misuse attempts and how it tries to disrupt them more
	0%	Meta doesn't do this
	0%	xAI doesn't do this
	0%	Microsoft doesn't do this
	0%	DeepSeek doesn't do this

Talk about extreme risks

Weighted 15% of category

	60%	Anthropic and its leadership sometimes talk about extreme risks, the alignment problem, and especially the prospect of dangerous AI appearing soon more
	30%	DeepMind's leadership expresses concerns about extreme risks from AI, but Google's does not, and DeepMind and Google themselves never articulate concerns about extreme risks more
	25%	OpenAI and its leadership used to talk about extreme risks but mostly haven't done so recently, especially catastrophic risks from misaligned AIs more
	0%	Meta doesn't do this more
	20%	CEO Elon Musk talks about extreme risks, including risks from misalignment, and how powerful AI might appear soon. xAI itself does not
	0%	Microsoft doesn't do this more
	0%	DeepSeek doesn't do this

Describe worst-case outcome

Weighted 10% of category

	0%	Anthropic doesn't do this more
	0%	DeepMind doesn't do this
	0%	OpenAI doesn't do this more
	0%	Meta doesn't do this
	0%	xAI doesn't do this
	0%	Microsoft doesn't do this
	0%	DeepSeek doesn't do this

Don't publish some capabilities research

Weighted 35% of category

	75%	Anthropic avoids publishing capabilities research on dangerous paths, but details are unclear more
	25%	DeepMind publishes capabilities research on dangerous paths, but at least it doesn't publish much on its Gemini models more
	70%	OpenAI avoids publishing capabilities research on dangerous paths but doesn't have a policy on this and hasn't articulated this principle
	0%	Meta publishes pretraining research and doesn't plan to stop; see e.g. Llama 3
	70%	xAI has not published reports on its models, but it doesn't have a policy on this
	0%	Microsoft publishes pretraining research and doesn't plan to stop; see e.g. Phi-4.
	0%	DeepSeek publishes pretraining research and doesn't plan to stop; see e.g. DeepSeek-V3.