Analysis · Finance
When More Training Data Cannot Close a Competitive Gap
By AWEI · AI-compiled · Published · Analysis prepared · 1 source · doi.org
An established AI provider could have unlimited training data and still face a challenger with a finite dataset. That is a mathematical result under specific assumptions in a 2025 PNAS study by Meena Jagadeesan, Michael I. Jordan and Jacob Steinhardt. The opening comes from different safety constraints.
The researchers built a theoretical marketplace with two companies using linear prediction models. Each company chooses how much training data to label for performance and how much for safety, then fits its model using ridge regression, which penalizes large model coefficients. The incumbent is assumed to need higher safety to avoid reputational damage.
When safety and performance pull predictions in different directions, a binding safety constraint leaves the incumbent with a performance shortfall that more data cannot eliminate. An entrant without that constraint can eventually match its performance using finite data. Under additional conditions, this remains possible when the entrant also faces a safety requirement, provided it is sufficiently weaker than the incumbent’s.
The analysis also finds that, with an unconstrained entrant, its required dataset can grow more slowly than the incumbent’s once the incumbent has enough data. Conversely, in the setting where both face safety constraints, narrowing their difference makes entry more demanding. These are consequences of the model, not measured associations or causal estimates from actual companies.
The limits matter: the paper reports no underlying empirical data and leaves testing these learning patterns in language models to future work. Users are assumed to share preferences and choose by performance once safety thresholds are satisfied. Prices, long-term survival and several kinds of safety objectives fall outside the analysis.
For assessing market access, the useful question is whether competitors face comparable objectives and constraints. Dataset size alone may miss a competitive opening. But an opening created by tolerating more safety violations does not establish a benefit for users.
Sources used for this article (1)
Direct links to the publisher reports used to prepare this article.
- Source 1
- Safety versus performance: How multi-objective learning reduces barriers to market entry Meena Jagadeesan, Michael I. Jordan, Jacob Steinhardt
