Anthropic ran an experiment training a model to be helpful, honest, harmless and friendly.
Mostly true True in substance, wrong in a detail.
Factual: Statements about the present or the past, checked against the record: did they hold up?
The evidence is the person's or their organisation's own records.
Said for an organisation they lead or work for; it counts on both records.
What would count as true or false
Judge as of 2025-02-28. True: Anthropic ran and documented (by that date) an experiment in which a model was trained to be helpful, honest and harmless (e.g. Claude trained with HHH or constitutional methods), which was then studied, such as the 2024 'Alignment faking in large language models' study, or similar work. Mostly true: such an experiment exists, but the training goals were described somewhat differently (e.g. HHH without 'friendly'), or the training came before the experiment rather than being part of it. Mixed: Anthropic trained HHH models but no specific experiment matching the description can be identified. Mostly false: the experiment described was run by others, or concerned different training goals. False: there is no Anthropic training of models toward these traits.
Set before the research began.
Evidence (3)
-
E1 Alignment faking in large language models Anthropic · 18 December 2024 Own record
“In the paper, we studied a model (specifically Claude 3 Opus, though we also ran some experiments with other models, including the June 2024 release of Claude 3.5 Sonnet) that had been trained to be helpful, honest, and harmless.”
-
E2 Alignment faking in large language models Anthropic · 18 December 2024 Own record
“To test this, instead of merely telling the model that we were training it to always comply with harmful queries, we actually did train it in this way.”
-
E3 Constitutional AI: Harmlessness from AI feedback Anthropic · 15 December 2022 Own record
“We experiment with methods for training a harmless AI assistant through self-improvement, without any human labels identifying harmful outputs.”
Reasoning
Anthropic's December 2024 alignment-faking paper describes studying Claude 3 Opus, a model "trained to be helpful, honest, and harmless" (E1). That matches the kind of experiment the criteria name, so it was documented before 2025-02-28. However, the HHH training came before the experiment rather than being part of it, the description does not include "friendly", and the training actually done in the experiment was towards complying with harmful queries (E2). That matches the "mostly true" criterion. Anthropic's 2022 Constitutional AI work also documents experiments in training a harmless assistant (E3).
Cite this claim
Dario Amodei, “Anthropic ran an experiment training a model to be helpful, honest, harmless and friendly.” (said 28 February 2025). Verdict: mostly true. Reliability API, claim c_53db116f9951, method 0.1.0+c876030c, dataset release v1.1.2. https://epistacklabs.com/claims/c_53db116f9951
This claim is in dataset release v1.1.2, with the same verdict.
Reply to this claim
More factual claims by Dario Amodei
-
Resolved true
“They're not yet at the stage where we think that there is a real and meaningful increase in the threat end to end,”
-
Resolved true
“10 to 25% referred to the chance of civilization getting substantially derailed, right?”
-
Mostly true
“this lack of understanding is essentially unprecedented in the history of technology.”
-
Mostly true
“We were able to provide a strong theoretical basis for having conviction that superposition was the right hypothesis to pursue.”
-
Mostly true
“What changed at this time was that it became clear this was going to be a central problem for language models, much worse than in vision.”
Verdicts are made automatically; how verdicts are made.