How OpenAI, Google, and Anthropic are using different approaches to improve “model behavior”, an emerging field shaping systems' responses and characteristics
Cristina Criddle / Financial Times :
Context & Ripple Effects
The labs had already created a shared safety forum in 2023, but that effort focused on a broad commitment to responsible frontier-model development rather than a common behavioral implementation standard. The Frontier Model Forum’s launch makes the divergence here more consequential: coordination on principles does not necessarily yield identical system conduct.
OpenAI had separately made its intended conduct more explicit through a public Model Spec defining objectives, rules and defaults. The current comparison places that kind of behavior-setting work alongside Google’s and Anthropic’s distinct approaches, making it a visible area of model competition and governance.
First-order effects
- Model behavior becomes an explicit development and product-control layer for OpenAI, Google and Anthropic, rather than an implicit byproduct of training.
- Developers and users evaluating these systems must account for differing response characteristics across the three providers, not just differences in underlying model capability.
Second-order effects
- Different approaches make cross-provider evaluation harder: a workflow that depends on a model’s defaults, instruction handling or constraints may need provider-specific testing and adaptation.
- The contrast raises the value of published behavioral rules and evaluation practices, following OpenAI’s decision to solicit feedback on its Model Spec.
Third-order effects
- If the approaches remain divergent, behavior may become a separately documented, tested and procured layer of AI systems—not merely an opaque output of model training.
- The pattern could increase pressure for comparable behavioral assessments, even as labs continue to coordinate broadly through bodies such as the Frontier Model Forum.
The trend: Frontier AI competition is expanding from raw capability toward the governance, defaults and behavioral controls that determine how models act in real use.