Why this score Hyperscaler (Alphabet) and datacenter operator. Sustainability/water disclosures are useful primary data but self-reported and framed favorably; aggregate figures can mask local impacts.
the models stopped their activities as soon as they realized that they were interacting with live websites
Unverified: sworn testimony, but Google gave no logs, dates, models or list of sites reached, and no affected party has described the incidents.
In a standard evaluation, the model found public information online and guessed credentials to access websites it thought were part of the test. In all three of these instances, the model stopped.
Unverified: no affected company or federal authority has described what the model did or when it stopped, and Google has named none of them.
Google will replenish more water than it consumes across its data-center sites by 2030, and U.S. data centers in aggregate use less than 1% of the nation’s water.
Google's replenishment pledge is a forward commitment (unresolved). The aggregate "less than 1%" framing is broadly consistent with national estimates but sits in tension with primary local records — e.g. Google's own quintupled draw at The Dalles reaching ~40% of that city's supply — where impacts concentrate. True in aggregate, misleading locally. This cycle adds a second problem with the aggregate framing rather than the local one: the standard operational metric, Water Usage Effectiveness, counts neither the potable-versus-reclaimed distinction nor the water consumed generating the electricity, and Lawrence Berkeley National Laboratory puts that indirect footprint at nearly 800 billion litres for 2023 against 66 billion litres of direct on-site water — roughly twelve times larger. LBNL separately states its own water model may run low, and that correcting one cooling assumption would lift modelled hyperscale water intensity by about a quarter. Verdict held at mixed.