# Automated AI Research: Will Capabilities Outpace Safety? > Coverage of lessw-blog **Published:** January 22, 2026 **Author:** PSEEDR Editorial **Category:** risk **Content tier:** free **Accessible for free:** true **Word count:** 485 **Tags:** AI Safety, AI Alignment, Automated Research, OpenAI, Intelligence Explosion, Risk Management **Canonical URL:** https://pseedr.com/risk/automated-ai-research-will-capabilities-outpace-safety --- A recent analysis from lessw-blog investigates the dangerous lag between automating AI capabilities and automating AI alignment, warning of a potential imbalance in the speed of research progress. In a recent post, **lessw-blog** discusses a pivotal variable in the trajectory of artificial intelligence: the comparative velocity of automating capabilities research versus safety research. As frontier labs explicitly target the creation of "automated AI researchers"-with OpenAI reportedly aiming for a fully automated researcher by March 2028-the sequence in which these technologies mature becomes a critical factor in managing global risk. The central tension explored in this analysis is the concept of an "AI Takeoff" or intelligence explosion. The optimistic view often holds that as AI becomes more intelligent, it will assist researchers in solving the complex problem of alignment (ensuring AI goals match human values). However, this post challenges that assumption by analyzing the structural differences between the two fields. The author argues that capabilities research is inherently more amenable to automation because it relies on clear, quantifiable feedback loops. If a code change reduces the loss function or improves a benchmark score, the system receives an immediate positive signal. This allows for rapid, automated iteration. Conversely, safety and alignment research often lack these crisp feedback signals. Progress in safety frequently requires novel conceptual insights, philosophical reasoning about intent, or the detection of subtle failures like deception, which are difficult to measure programmatically. Because of this discrepancy, the author forecasts that AI automation is likely to accelerate capabilities research by a factor of ten before it provides a comparable boost to safety research. This creates a precarious window where the power of AI systems could scale exponentially while the mechanisms to control them advance only linearly. The post is not merely a forecast but a call to action for structural changes in how safety research is conducted. To close this gap, the author suggests the development of "model organisms" for AI safety-simplified, controlled environments where safety failures can be reliably triggered and studied. By creating better benchmarks and measurable proxies for safety, the field could theoretically harness the same automation engines that drive capabilities. The analysis also touches on the strategic necessity of differential access, suggesting that advanced automated research tools might need to be restricted to safety use-cases initially to maintain equilibrium. For stakeholders in the tech sector, this analysis highlights a specific, technical bottleneck in AI development. It moves the conversation from vague concerns about "superintelligence" to concrete metrics regarding research velocity and workflow automation. We recommend reading the full post to understand the proposed timelines and the specific interventions required to align automation with safety goals. [Read the full post at lessw-blog](https://www.lesswrong.com/posts/z4FvJigv3c8sZgaKZ/will-we-get-automated-alignment-research-before-an-ai) ### Key Takeaways * Capabilities research is predicted to automate 10x faster than safety research due to clearer feedback loops and engineering-heavy workflows. * Safety research often relies on novel conceptual insights which are harder for current AI to generate than empirical engineering improvements. * OpenAI's internal goals reportedly include creating a 'true automated AI researcher' by March 2028, making this timeline immediate and critical. * To accelerate safety automation, the field needs 'model organisms' and robust benchmarks that turn vague safety concepts into measurable data points. * Differential access to automated research tools may be necessary to prevent capabilities from vastly outstripping control mechanisms. [Read the original post at lessw-blog](https://www.lesswrong.com/posts/z4FvJigv3c8sZgaKZ/will-we-get-automated-alignment-research-before-an-ai) --- ## Sources - https://www.lesswrong.com/posts/z4FvJigv3c8sZgaKZ/will-we-get-automated-alignment-research-before-an-ai