Now loading...
Anthropic is reversing its strategy regarding access to its advanced AI model due to backlash over its earlier decision to limit certain requests. On Tuesday, the company unveiled Fable 5, part of its powerful Mythos series, which was initially withheld from the public in April due to concerns about its potential to outsmart cybersecurity measures and its overall danger level.
This latest release is being positioned as safer, with Anthropic claiming that its safety precautions are now robust enough to manage the risks associated with this capability. Dianne Na Penn, the company’s head of product management, highlighted increased confidence in these safety measures, which were described as essential for user protection.
However, the unveiling of Fable 5 was met with backlash when it was discovered that a specific safeguard in the system documentation indicated a process whereby certain requests from advanced AI developers would be downgraded to a less advanced version of the model without alerting the user. This led to a significant outcry from the AI research community, who argued that this limitation could hinder progress in AI development.
Notable figures in AI research, including Jeremy Howard from Fast.ai, voiced their concerns, stating that restricting access to top-tier models would ultimately stifle advancements. He urged for a system that allows broader access while preventing misuse of technology at the frontier of AI research.
In response to the criticism, Anthropic announced adjustments to Fable 5’s safeguards, making it clear when a downgrade occurs. Moving forward, the company stated that flagged requests will be redirected transparently, allowing users to see when their queries are not processed as expected. Additionally, specific requests will still be moderated in line with the company’s terms preventing competition from emerging AI systems.
Anthropic cited national security concerns as part of the rationale for these limitations. The company expressed its intent to prevent adversaries from enhancing their AI capabilities at the expense of U.S. interests, particularly highlighting the strategic advantage held by the U.S. in state-of-the-art technology.
This development sheds light on the growing intersection of AI safety and national security discourse. Earlier this year, Anthropic found itself in a contentious situation with the Department of War, which sought full access to its models but faced pushback from Anthropic related to the potential misuse of AI technology for military purposes.
Compounding these issues, Anthropic recently filed for an IPO, creating additional scrutiny on its commitment to safety and transparency in its technological endeavors. The company admitted to shortcomings in its previous approach, acknowledging that it did not strike the right balance between safety and accessibility, and issued an apology for the misstep.
