OpenAI has openly declared an intention to have some future AI do our AI alignment work.
Resolved true The record shows it happened or was so.
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
True: an official OpenAI publication from on or before 2023-04-07 openly states an intention to use AI systems to do or substantially help with alignment research. An example would be the August 2022 post 'Our approach to alignment research', which describes training AI systems to do alignment research. Mostly true: OpenAI stated an intention to use AI to assist alignment research substantially, but framed it as partial help or acceleration rather than having AI do the work. The 'homework' characterisation is a fair paraphrase of this. Mixed: OpenAI mentioned AI-assisted alignment only as one minor or speculative option. Mostly false: only individual employees or informal remarks suggested it, with no official declaration. False: OpenAI made no such statement by that date, or explicitly rejected the approach.
Set before the research began.
Evidence (3)
-
E1 Our approach to alignment research OpenAI · 24 August 2022
“As we make progress on this, our AI systems can take over more and more of our alignment work and ultimately conceive, implement, study, and develop better alignment techniques than we have now.”
-
E2 Our approach to alignment research OpenAI · 24 August 2022
“Our goal is to train models to be so aligned that we can off-load almost all of the cognitive labor required for alignment research.”
-
E3 Our approach to alignment research OpenAI · 24 August 2022
“Therefore human researchers will focus more and more of their effort on reviewing alignment research done by AI systems instead of generating this research by themselves.”
Reasoning
OpenAI published the official post 'Our approach to alignment research' on 2022-08-24, before the claim date. In it, OpenAI says its AI systems "can take over more and more of our alignment work" (E1). It names as its goal to "off-load almost all of the cognitive labor required for alignment research" onto models (E2), and says human researchers will move toward reviewing alignment research done by AI systems (E3). This meets the True criterion, an openly stated intention to have AI do alignment research.
Cite this claim
Eliezer Yudkowsky, “OpenAI has openly declared an intention to have some future AI do our AI alignment work.” (said 7 April 2023). Verdict: resolved true. Reliability API, claim c_65beec95cb24, method 0.1.0+c876030c, dataset release v1.1.2. https://epistacklabs.com/claims/c_65beec95cb24
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.