Tool Removes AI Watermarks from Anthropic's Claude Models
Hoskinson has released a tool to remove invisible watermarks from text generated by Anthropic’s Claude models. The move follows recent changes made by Anthropic that added these marks to AI-generated content in an effort to comply with Europe’s AI Act transparency code.
The new tool was developed by Charles Hoskinson, founder of Cardano. He views this release as a way to address potential problems related to ownership and co-authorship if users rely on these models without understanding their limitations. According to Hoskinson, the watermark could create issues that affect claims of ownership or serve as evidence of joint authorship.
The tool operates in three distinct modes: Clean removes predictable trailers; humanize rewrites text for a more natural sound; and orchestrate halts the process when Claude is being used. However, there’s an important limitation – if users run the rewritten content through Claude again, the watermark will be added back on.
One of the tool’s key limitations comes from its code structure: it doesn’t carry much of the watermark due to limited word/phrase replacement opportunities in code syntax. This compromise affects how effective the tool can be despite not posing major risks itself – for now, users have to weigh benefits against these drawbacks.
Hoskinson has released his tool under the Apache 2.0 license, which allows others to copy, modify, sell or include the code in other products as long as original copyright and license notices are preserved. This move also includes a patent license that prevents Anthropic from stopping people from creating modified versions of this code.
The release of Hoskinson’s tool has sparked concerns about ownership and co-authorship related to AI-generated content. For businesses, particularly those relying on tools like Claude for image generation or writing assistance – often referred to as ‘AI generated images’ in the industry – this is a pressing issue that needs attention. The use of such models may have unintended consequences.
Hoskinson emphasizes potential risks associated with relying on these tools, including ownership and co-authorship disputes that could arise if users don’t grasp model limitations. He’s aiming to make progress toward humanizing AI by making its processes more transparent – a goal Anthropic itself is tackling through initiatives like the Claude transparency code update.