The current AI developers will not try the alignment approach he described (filtering training data to nice examples and training on it).

'Current crop' means leading AI developers as of April 2023: OpenAI, Anthropic, Google DeepMind and Meta. 'Trying it' means filtering or curating training data toward benign or 'nice' examples, and training on that data, as a method of shaping model behaviour or values. True: none of these developers documented using or experimenting with such training-data filtering as an alignment or safety approach. Mostly true: the only cases are routine toxicity or quality filtering not presented as an alignment method. Mixed: only marginal or exploratory work by individual researchers at these labs. Mostly false: at least one leading developer published research testing the approach, without adopting it in production. False: one or more leading developers publicly documented using alignment- or safety-motivated pretraining-data filtering in their models or research.

Set before the research began.

1. E1 Enhancing Model Safety through Pretraining Data Filtering Anthropic (alignment.anthropic.com) · 19 August 2025

   “Using a classifier, we identified and removed potentially harmful information from the pretraining data and then pretrained models from scratch on the filtered dataset.”

2. E2 Gemma: Open Models Based on Gemini Research and Technology Google DeepMind (arXiv technical report)

   “we have endeavoured to safeguard against these risks by filtering and measuring biases in pre-training data in line with the Gemini approach”

3. E3 Gemma: Open Models Based on Gemini Research and Technology Google DeepMind (arXiv technical report)

   “When using synthetic data, we run several stages of filtering over it, removing examples that show certain personal information, unsafe or toxic model outputs, mistaken self-identification data, or duplicated examples.”

4. E4 Enhancing Model Safety through Pretraining Data Filtering Anthropic (alignment.anthropic.com) · 19 August 2025

   “In addition to misuse risks, which are the focus of this blog post, data filtering has the potential to mitigate misalignment risks as well.”

Anthropic documented a safety-motivated pretraining-data filtering experiment: it used a classifier to remove harmful information and then pretrained models from scratch on the filtered data (E1). It also said this filtering could help with misalignment risks (E4). Google DeepMind's Gemma report describes filtering pretraining data to address risks and removing unsafe or toxic examples from synthetic training data (E2, E3). Under the criteria, a leading developer publicly documenting safety-motivated pretraining-data filtering in its models or research counts as false. Anthropic's work aimed at removing harmful information rather than shaping values, but the criteria count safety-motivated filtering, so the verdict is false.