Over the past few months, headlines have highlighted both the benefits and harms of AI applications. From helping filmmakers enhance their audio for accurate language depictions like the Oscar-nominated film The Brutalist, to streamlining the documentation process for healthcare providers. However, while advances in AI could enable creative solutions to audio and visual issues, AI has poses risks and harm to hundreds of women and girls who are targets of deepfake sex crime. The same AI-driven transcription tools used in hospital settings for visits manufacture “clumps of text and whole sentences,” creating a significant risk for patients receiving sensitive care. And last year, another AI chatbot was involved in a teenager’s suicide.
These stories reveal important reality. Real life is related to the deployment of AI applications. Our understanding of these issues comes primarily from ad hoc research reports rather than direct information. In fact, despite widespread adoption, despite growing influence, there is surprisingly limited data on how these systems work in the real world after deployment.
“Real life depends on the deployment of AI applications.”
Recent reports documenting the impact of the underlying model highlight the need for change. While most industry and policy-making efforts are focused on ensuring that foundation models are “safe” to deploy, companies need to take the lead in assessing the real-world impact and implications of post-development models.
Changes start in the industry
Some companies have already shown what this leadership looks like. The human economic index, which aims to understand the impact of AI assistant Claude on the labor market and economy, is an example of how model providers can share usage information and provide insight into how their models affect a particular industry. By sharing anonymized usage data, Anthropic provides policymakers and researchers with new information to assess the economic impact of AI.
Meta is also helping the bar rise. Their research on sustainable AI impacts is a prime example of a model provider that allows and shares research on social impacts. By investigating the environmental impact, we were able to analyze the carbon footprint of AI models from both the AI and the hardware development lifecycle. Meta researchers were able to identify ways to optimize AI models to reduce AI’s overall carbon footprint.
These examples are promising, but they are still exceptions and not rules. Therefore, more companies need to follow in examining and monitoring the impact of AI after deployment.
“Despite the clear advantages, impact documentation is not yet an industry norm.”
Why do companies have to lead?
Foundation Model Providers design, train, release and update models that power AI applications, providing unique visibility into how models work in the real world. Being at the heart of the AI ecosystem means we are responsible for voluntarily documenting and sharing these insights in the absence of regulatory oversight.
As outlined in our report, by reviewing, collecting, aggregating and sharing post-deployment impact information, we offer four key benefits for actors across the value chain of the underlying model, such as:
Amplification of social benefits: Documenting post-development impacts raises awareness of the benefits of the underlying model and improves stakeholder literacy while building trust. Risk Management and Mitigation: Documenting post-development impacts enables stakeholders to identify, assess, and realize the potential or realized adverse effects of AI systems on society. Evidence-based policy development: Documenting post-deployment impacts provides critical data for policymakers to develop and implement effective, balanced regulatory and governance frameworks that protect people while considering implementation costs. Advance documentation standards through shared learning: Multi-stakeholder collaboration to share post-development impact impact documents can help establish best practices and drive the industry towards a standardized approach.
They can’t do it alone
Although model providers must lead this effort, it is a common responsibility to document the impact of AI. Other actors across the AI value chain also play important roles, including application developers, researchers, policy makers, and civil society.
However, the government is changing priorities, with some focusing on promoting the development and deployment of AI systems in their regions and other developments towards deregulation. These shifts slow the pace of regulatory development, but AI continues to develop rapidly, making it difficult to advance global governance. This regulatory uncertainty makes voluntary initiatives and research not only useful but essential. Industry-driven transparency practices can help establish consistent industry standards that reflect what works well based on real-world use cases and inform regulatory efforts.
“Regulatory uncertainty makes voluntary initiatives and research not only useful, but essential.”
Where to go from here
AI landscapes have already undergone another evolution with the emergence of AI agents and the emergence of systems that can take action in virtual environments with minimal surveillance, and their ability to understand their impact remains limited. Understanding the impact of these systems on our society, and the emerging impact of agents in media integrity, labor, economics, and public policy is one of our priorities for 2025.
A shift in AI policies focused on promoting and deregulation in AI systems requires industry stakeholders to shape sectors and influence other actors to foster shared responsibility ecosystems. Multi-stakeholder collaboration is required to deepen our understanding of the social impact of the foundation model, but change begins in the industry. Read our full report to learn how you can lead impact documents and help shape a safer and accountable ecosystem.
