An economics model says AI may raise research volume faster than quality, while workplace adoption and a watermark remover built in five hours show review and oversight struggling to keep pace.
Artificial Intelligence··Evening
Saved time may produce more research rather than better research
A theoretical economics study by researchers at Princeton and the University of Washington argues that time saved by language models makes each researcher's remaining hours more valuable. In the model, faster early screening makes researchers more selective but less thorough on the projects they retain. Faster writing and formatting allow weaker papers to reach submission and fill the review pipeline; quality rises only when AI speeds up the analytical core. Supporting findings include a METR study in which developers using AI tools felt 24 percent faster while taking 19 percent longer to finish their tasks.[1]
Employee use has moved ahead of workplace rules
An analysis in the Dutch economics journal ESB finds that, in every sector, the share of employees who say they use tools such as ChatGPT and Copilot exceeds the share who say their employer actively supports that use. The gap is above 15 points in most fields and approaches 30 points in government and education. A De Nederlandsche Bank survey last October found active employer support reported by only 25 percent of workers. When approved tools and rules lag, staff may use consumer accounts on corporate networks, allowing company data, client information and source code to leave the organisation.[2]
A watermark remover shipped in five hours
An entrepreneur told Business Insider that, after Anthropic announced an invisible text watermark, he built and published the first version of a removal tool on GitHub in about five hours. A post introducing the project on 11 August drew more than 2 million views on X and spread to LinkedIn. The tool checks for the watermark, generates small changes that preserve meaning and repeats the process until the mark disappears; it applies a similar loop to images. Its developer argues that a method based on statistical word patterns can falsely flag someone who used a model for only one line of a paper.[3]
