DeepMind, a leading AI lab, has no plan for AI alignment/safety.
Mostly false False in substance, right in a detail.
Factual: Statements about the present or the past, checked against the record: did they hold up?
Said for an organisation they lead or work for; it counts on both records.
What would count as true or false
Judge as of 2023-04-07. True: as of that date, DeepMind had published no stated plan or strategy for AI alignment/safety. Mostly true: DeepMind had only scattered safety research and principles, with no overall alignment plan stated anywhere. Mixed: DeepMind had a safety team and some public statements of its approach, but nothing presented as a plan. Mostly false: DeepMind had public documents describing its alignment approach (e.g. a safety team's stated research agenda or blog posts setting out its strategy), and the claim rests only on how detailed or credible that plan was. False: DeepMind had published a clear, explicit alignment plan before April 2023.
Set before the research began.
Evidence (4)
-
E1 [Linkpost] Some high-level thoughts on the DeepMind alignment team's strategy — AI Alignment Forum AI Alignment Forum · 7 March 2023
“This is an overview of our threat models, our high-level current plan, and how current projects fit into this plan.”
-
E2 [Linkpost] Some high-level thoughts on the DeepMind alignment team's strategy — AI Alignment Forum AI Alignment Forum · 7 March 2023
“We simply meant to share slides that gave a high-level outline of how we were thinking about our alignment plan, in the interest of partial communication rather than no communication.”
-
E3 (My understanding of) What Everyone in Technical Alignment is Doing and Why — AI Alignment Forum AI Alignment Forum · 29 August 2022
“The alignment team is pursuing many different research avenues, and is not best described by a single agenda.”
-
E4 Building safer dialogue agents DeepMind · 22 September 2022
“To make sure that the model’s behaviour is safe, we must constrain its behaviour.”
Reasoning
A month before the claim, in March 2023, the DeepMind alignment team published an overview of its threat models, "our high-level current plan," and how current projects fit into that plan (E1). The same post says it was only a high-level outline, shared as partial communication (E2). A 2022 survey describes the team as following many research avenues rather than a single agenda (E3). So a public statement of DeepMind's alignment approach existed before April 2023, and the claim holds only if that plan is judged too thin or not credible enough to count. Under the criteria, that is mostly false.
Cite this claim
Eliezer Yudkowsky, “DeepMind, a leading AI lab, has no plan for AI alignment/safety.” (said 7 April 2023). Verdict: mostly false. Reliability API, claim c_bfd689c21fe7, method 0.1.0+c876030c, dataset release v1.1.2. https://epistacklabs.com/claims/c_bfd689c21fe7
This claim is in dataset release v1.1.2, with the same verdict.
Reply to this claim
More factual claims by Eliezer Yudkowsky
-
Mixed
“I was observing that obvious-seeming hypothesis myself, sometime around the time of GPT 3.5.”
Yudkowsky himself observed the persona-selection-like hypothesis around the time of GPT 3.5.
-
Mostly true
“The test-passing preference stands on its own once formed.”
-
Mostly true
“LLMs surely have neural circuits they learned to model text streams from fictional and nonfictional personas that are not lying, deceiving, cheating.”
-
Mostly true
“The predict-as-if-honest circuitry is contextual, invoked to predict some streams of text but not others.”
-
Mostly true
“Presumably, somewhere along the way LLMs learn a dispreference for tests that don't pass, independently of how that gets prevented or whether the new plan does what the user really wanted.”
Verdicts are made automatically; how verdicts are made.