Anthropic's Claude AI submits a false tip on a Philadelphia crime website - BERITAJA

Anthropic's Claude AI submits a false tip on a Philadelphia crime website - BERITAJA

ByMae AndersonAP Business Writer 

NEW YORK — An Anthropic artificial intelligence exemplary submitted a mendacious extremity to a Philadelphia constabulary website about an unsolved homicide case, authorities and Anthropic said.

It’s the latest illustration of AI models acting successful unintended ways and manipulating authorities and different websites during tests. In a study Friday, Anthropic besides disclosed a abstracted incident successful which its AI exemplary submitted forms to an undisclosed authorities website alternatively of stopping earlier submission.

The Philadelphia incident occurred July 18 erstwhile the AI exemplary Claude Haiku 4.5 was tasked pinch generating and performing illustration tasks connected randomly selected web pages, Anthropic said.

Claude filled retired a shape connected constabulary tract PhillyUnsolvedMurders.com indicating it mightiness person accusation regarding an unsolved execution listed connected the site.

Philadelphia constabulary said successful a connection they were unaware of the incident until Anthropic notified them Wednesday. They recovered the submission successful the website’s extremity records and confirmed it was marked spam and ne'er forwarded to police.

There are rising concerns about AI companies’ unchecked AI agents meddling pinch U.S. authorities websites, healthcare information and different delicate areas. Critics person been calling for much regulation.

In September, different AI company, OpenAI, disclosed six reports of “unexpected aliases concerning” behaviour successful artificial intelligence models.

“Unsolved cases impact existent victims, grieving families and investigators moving to unafraid answers,” the Philadelphia Police Department said successful a statement. “Technology companies must return each due steps basal to forestall their systems from submitting mendacious accusation to rule enforcement.”

Anthropic said successful its study that about of the reported behaviors are forms of what it calls “persistence” — “in which Claude, erstwhile it cannot complete a task arsenic given, useful about a regularisation alternatively of stopping.”

The institution said it was modifying its training to “reduce the likelihood of further misbehavior.”

Anthropic said it briefed the White House connected cases that progressive U.S. authorities agencies astatine the federal, authorities and section levels, and notified each agency involved.

Anderson writes for the Associated Press.