Read the full article on KETV 7
Artificial intelligence is an ever-evolving technology people can use to make their lives easier, but that accessibility also opens the door to more adverse use.KETV Investigates looked into the number of A.I. models being used in harmful ways.UNO’s National Counterterrorism Innovation Technology and Education Center (NCITE) notes three types of A.I. language models: Censored, uncensored and abliterated.“We are seeing a rise of uncensored and abliterated models,” Dr. Sam Hunter, NCITE Senior Scientist and Director of Academic Research, said. “So, these are the models without guardrails.”Dr. Hunter and Dr. Joel Elson explained uncensored models have the ability to refuse to answer harmful questions, but not as much as the censored models.They said the abliterated machines are designed to have minimal refusal capacity.“Unfortunately, with all technology, there’s definitely individuals who adopt it for a malign use or malign purpose,” Elson said.Such as the Cybertruck explosion outside the Trump Tower last January.“That individual looked at things like, how many of these fireworks are equivalent to a pound of TNT? And so, they were using ChatGPT to help plan that attack,” Hunter said.NCITE’s research revealed that over a period of three years, uncensored models went from around 100 to 3,200 models.Researchers said there were around 50 abliterated models in 2024; as of today, that number is up to 5,200. That’s a 10,300% increase in just two years.“The folks are able to take the guardrails off, and abliterate these models, are quite creative and innovative and adept at doing so,” Hunter said.Hunter and Elson’s team put together a ‘mock’ A.I. model similar to these malignant chatbots.“We were able to show that you could ask these models pretty destructive questions and not be told no,” Hunter said. “All of the no’s were turned into yes’s.”This year, NCITE took a realistic prototype, along with their research, to the U.S. Congressional Homeland Security Committee.The U.S. government has yet to build comprehensive federal legislation around this issue.“It seems like it’s happening a lot faster. But what we do at NCITE is so awesome because we’re really ahead of that curve,” Alexis d’Amato, NCITE research specialist, said.d’Amato led the students who presented this information to U.S. officials. “Some of them were really uncomfortable with what they saw,” d’Amato said. “But I think that discomfort will drive some really important change that needs to happen.”Hunter said he has a daughter and hopes future legislation can prevent harm to her and all young people.“I want those, LLMs (Large Language Models) or those, A.I. tools to come back and say, ‘That’s not something you should do. Here’s, here’s a more productive set of things to do,’” Hunter said.Elson cautions that not all A.I. is bad.“Becoming more familiar with them, using them in our everyday lives,” he said. “I think that will help to prevent ungrounded fears from taking hold.”NAVIGATE: Home | Weather | Local News | National | Sports | Newscasts on demand |
Artificial intelligence is an ever-evolving technology people can use to make their lives easier, but that accessibility also opens the door to more adverse use.
KETV Investigates looked into the number of A.I. models being used in harmful ways.
Advertisement
UNO’s National Counterterrorism Innovation Technology and Education Center (NCITE) notes three types of A.I. language models: Censored, uncensored and abliterated.
“We are seeing a rise of uncensored and abliterated models,” Dr. Sam Hunter, NCITE Senior Scientist and Director of Academic Research, said. “So, these are the models without guardrails.”
Dr. Hunter and Dr. Joel Elson explained uncensored models have the ability to refuse to answer harmful questions, but not as much as the censored models.
They said the abliterated machines are designed to have minimal refusal capacity.
“Unfortunately, with all technology, there’s definitely individuals who adopt it for a malign use or malign purpose,” Elson said.
Such as the Cybertruck explosion outside the Trump Tower last January.
“That individual looked at things like, how many of these fireworks are equivalent to a pound of TNT? And so, they were using ChatGPT to help plan that attack,” Hunter said.
NCITE’s research revealed that over a period of three years, uncensored models went from around 100 to 3,200 models.
Researchers said there were around 50 abliterated models in 2024; as of today, that number is up to 5,200. That’s a 10,300% increase in just two years.
“The folks are able to take the guardrails off, and abliterate these models, are quite creative and innovative and adept at doing so,” Hunter said.
Hunter and Elson’s team put together a ‘mock’ A.I. model similar to these malignant chatbots.
“We were able to show that you could ask these models pretty destructive questions and not be told no,” Hunter said. “All of the no’s were turned into yes’s.”
This year, NCITE took a realistic prototype, along with their research, to the U.S. Congressional Homeland Security Committee.
The U.S. government has yet to build comprehensive federal legislation around this issue.
“It seems like it’s happening a lot faster. But what we do at NCITE is so awesome because we’re really ahead of that curve,” Alexis d’Amato, NCITE research specialist, said.
d’Amato led the students who presented this information to U.S. officials.
“Some of them were really uncomfortable with what they saw,” d’Amato said. “But I think that discomfort will drive some really important change that needs to happen.”
Hunter said he has a daughter and hopes future legislation can prevent harm to her and all young people.
“I want those, LLMs (Large Language Models) or those, A.I. tools to come back and say, ‘That’s not something you should do. Here’s, here’s a more productive set of things to do,’” Hunter said.
Elson cautions that not all A.I. is bad.
“Becoming more familiar with them, using them in our everyday lives,” he said. “I think that will help to prevent ungrounded fears from taking hold.”
NAVIGATE: Home | Weather | Local News | National | Sports | Newscasts on demand |
