AI Inside: Five Years of AI Ethics at Adobe
For Wednesday, February 5, 2025
In the latest episode of AI Inside, Jeff Jarvis and I had the pleasure of speaking with Grace Yee, Senior Director of Ethical Innovation at Adobe, about the company's approach to AI ethics and responsible innovation. Our conversation revealed how Adobe has maintained consistent ethical principles since establishing its AI ethics committee in 2019, built around three core tenets: accountability, responsibility, and transparency.
One of the most fascinating aspects of our discussion centered on Adobe's practical approach to AI development and content creation. Yee explained their careful consideration of AI training data, noting that Adobe uses "licensed content from our Adobe stock collection" and "public domain content whose copyright had expired." She emphasized the importance of human oversight in AI creation, stating that "it's so important for humans to be involved in the loop" rather than blindly accepting AI-generated content.
The conversation took an interesting turn when we discussed how AI tools are enhancing rather than replacing creative professionals. As Yee pointed out, "creativity is a very human characteristic," and Adobe's AI features are designed to make it "easier for our creators to express themselves." She illustrated this with the example of Photoshop's generative fill feature, which has transformed time-consuming manual tasks into quick, intuitive processes while freeing up creators to focus on their artistic vision.
Meta's Open Source AGI Ambitions Meet Reality Check
Meta's bold announcement about creating open source AGI has made way for their new Frontier AI Framework. This framework introduces a multi-risk system to manage AI development risks, with high-risk scenarios covering cybersecurity breaches and chemical/biological attacks requiring limited internal access and restricted release. For critical-risk scenarios involving potentially catastrophic events, development would be halted with strict leak control protocols.
Google's Shifting AI Stance
In a notable policy shift, Google has quietly removed its "Applications we will not pursue" section from its AI principles, as spotted by Bloomberg. The company's new "Responsible AI: Our 2024 report" emphasizes collaboration between companies, governments, and organizations to create AI that protects people, promotes global growth, and supports national security. The updated stance maintains commitments to mitigating harmful outcomes and unfair bias while aligning with human rights principles.
Deep Research Launch
OpenAI released Deep Research, a new research agent tool that can navigate the open web, gather information sources, and generate comprehensive reports. Initially available to Pro members at $200/month, with Plus and Teams access coming soon, the tool promises to compress research tasks that typically take 30 minutes to 30 days into just 5-30 minutes. The system features recursive searching capabilities and includes citations, though accuracy remains a consideration.
O3-Mini Release
Responding to DeepSeek's advances, OpenAI has released o3-mini, their first reasoning model, available free of charge. While the full o3 model is still pending, this smaller version offers generous usage limits compared to Pro users. An AI safety report led by Yoshua Bengio indicates the model has outperformed human experts in certain areas, raising significant implications for AI risk assessment.
Federal AI Integration
Thomas Shedd, Technology Transformation Services director, is spearheading an "AI-first strategy" in Washington. The initiative includes developing AI coding agents across agencies, creating a centralized contract database for AI analysis, and automating government tasks. This transformation suggests imminent workforce reductions alongside increased automation.
Proposed China AI Restrictions
Senator Josh Hawley has introduced a bill that could significantly impact AI development relationships with China. While not specifically naming DeepSeek, the legislation aims to prohibit U.S. persons from advancing AI capabilities within China and prevent the import of Chinese AI technology. Violations could result in severe penalties: up to 20 years imprisonment, $1 million fines for individuals, and $100 million for businesses.
Anthropic's New Defense Against AI Jailbreaks
Anthropic has developed a promising new defense system targeting jailbreak attacks on Large Language Models. The system implements a sophisticated filter for both input and output, scanning for potential attempts to circumvent restrictions. After rigorous testing involving 183 users in a bug bounty program and analysis of 10,000 jailbreak prompts, the results are impressive - successful attacks plummeted from 86% to just 4.4%.
However, there are some trade-offs. The system occasionally blocks legitimate questions, particularly in fields like biology and chemistry. Additionally, it comes with a 25% increase in compute costs.
AI Consciousness: A Framework for the Future
In a significant move for AI ethics, Stephen Fry along with more than 100 experts have outlined five crucial principles for responsible research into AI consciousness. These principles emphasize understanding AI consciousness to prevent mistreatment and suffering, implementing proper constraints and safeguards, and ensuring gradual progress with expert involvement. The framework also stresses the importance of balancing transparency with safety, while maintaining careful communication that avoids overconfident claims about consciousness and acknowledges uncertainties.
Yann LeCun's Vision for AI's Next Leap
Meta's AI pioneer Yann LeCun, recently honored with the Queen Elizabeth prize for engineering alongside six other engineers, has shared his perspective on AI's future trajectory. LeCun points out that current systems remain significantly limited, particularly in areas involving physical world interaction like domestic robotics and fully automated vehicles.
Setting realistic expectations, LeCun emphasizes that the goal isn't yet to match human-level intelligence: "We're not talking about matching the level of humans yet. If we get a system that is as smart as a cat or a rat, that would be a victory".