Showing posts with label need to verify AI outputs. Show all posts
Showing posts with label need to verify AI outputs. Show all posts

Saturday, August 23, 2025

PittGPT debuts today as private AI source for University; University Times, August 21, 2025

MARTY LEVINE, University Times; PittGPT debuts today as private AI source for University

"Today marks the rollout of PittGPT, Pitt’s own generative AI for staff and faculty — a service that will be able to use Pitt’s sensitive, internal data in isolation from the Internet because it works only for those logging in with their Pitt ID.

“We want to be able to use AI to improve the things that we do” in our Pitt work, said Dwight Helfrich, director of the Pitt enterprise initiatives team at Pitt Digital. That means securely adding Pitt’s private information to PittGPT, including Human Resources, payroll and student data. However, he explains, in PittGPT “you would only have access to data that you would have access to in your daily role” — in your specific Pitt job.

“Security is a key part of AI,” he said. “It is much more important in AI than in other tools we provide.” Using PittGPT — as opposed to the other AI services available to Pitt employees — means that any data submitted to it “stays in our environment and it is not used to train a free AI model.”

Helfrich also emphasizes that “you should get a very similar response to PittGPT as you would get with ChatGPT,” since PittGPT had access to “the best LLM’s on the market” — the large language models used to train AI.

Faculty, staff and students already have free access to such AI services as Google Gemini and Microsoft Copilot. And “any generative AI tool provides the ability to analyze data … and to rewrite things” that are still in early or incomplete drafts, Helfrich said.

“It can help take the burden off some of the work we have to do in our lives” and help us focus on the larger tasks that, so far, humans are better at undertaking, added Pitt Digital spokesperson Brady Lutsko. “When you are working with your own information, you can tell it what to include” — it won’t add misinformation from the internet or its own programming, as AI sometimes does. “If you have a draft, it will make your good work even better.”

“The human still needs to review and evaluate that this is useful and valuable,” Helfrich said of AI’s contribution to our work. “At this point we can say that there is nothing in AI that is 100 percent reliable.”

On the other hand, he said, “they’re making dramatic enhancements at a pace we’ve never seen in technology. … I’ve been in technology 30 years and I’ve never seen anything improve as quickly as AI.” In his own work, he said, “AI can help review code and provide test cases, reducing work time by 75 percent. You just have to look at it with some caution and just (verify) things.”

“Treat it like you’re having a conversation with someone you’ve just met,” Lutsko added. “You have some skepticism — you go back and do some fact checking.”

Lutsko emphasized that the University has guidance on Acceptable Use of Generative Artificial Intelligence Tools as well as a University-Approved GenAI Tools List.

Pitt’s list of approved generative AI tools includes Microsoft 365 Copilot Chat, which is available to all students, faculty and staff (as opposed to the version of Copilot built into Microsoft 365 apps, which is an add-on available to departments through Panther Express for $30 per month, per person); Google Gemini; and Google NotebookLMwhich Lutsko said “serves as a dedicated research assistant for precise analysis using user-provided documents.”

PittGPT joins that list today, Helfrich said.

Pitt also has been piloting Pitt AI Connect, a tool for researchers to integrate AI into software development (using an API, or application programming interface).

And Pitt also is already deploying the PantherAI chatbot, clickable from the bottom right of the Pitt Digital and Office of Human Resources homepages, which provides answers to common questions that may otherwise be deep within Pitt’s webpages. It will likely be offered on other Pitt websites in the future.

“Dive in and use it,” Helfrich said of PittGPT. “I see huge benefits from all of the generative AI tools we have. I’ve saved time and produced better results.”"

Saturday, May 17, 2025

Anthropic’s law firm throws Claude under the bus over citation errors in court filing; The Register, May 15, 2025

Thomas Claburn, The Register; Anthropic’s law firm throws Claude under the bus over citation errors in court filing

"An attorney defending AI firm Anthropic in a copyright case brought by music publishers apologized to the court on Thursday for citation errors that slipped into a filing after using the biz's own AI tool, Claude, to format references.

The incident reinforces what's becoming a pattern in legal tech: while AI models can be fine-tuned, people keep failing to verify the chatbot's output, despite the consequences.

The flawed citations, or "hallucinations," appeared in an April 30, 2025 declaration [PDF] from Anthropic data scientist Olivia Chen in a copyright lawsuit music publishers filed in October 2023.

But Chen was not responsible for introducing the errors, which appeared in footnotes 2 and 3.

Ivana Dukanovic, an attorney with Latham & Watkins, the firm defending Anthropic, stated that after a colleague located a supporting source for Chen's testimony via Google search, she used Anthropic's Claude model to generate a formatted legal citation. Chen and defense lawyers failed to catch the errors in subsequent proofreading.

"After the Latham & Watkins team identified the source as potential additional support for Ms. Chen’s testimony, I asked Claude.ai to provide a properly formatted legal citation for that source using the link to the correct article," explained Dukanovic in her May 15, 2025 declaration [PDF].

"Unfortunately, although providing the correct publication title, publication year, and link to the provided source, the returned citation included an inaccurate title and incorrect authors.

"Our manual citation check did not catch that error. Our citation check also missed additional wording errors introduced in the citations during the formatting process using Claude.ai."...

The hallucinations of AI models keep showing up in court filings.

Last week, in a plaintiff's claim against insurance firm State Farm (Jacquelyn Jackie Lacey v. State Farm General Insurance Company et al), former Judge Michael R. Wilner, the Special Master appointed to handle the dispute, sanctioned [PDF] the plaintiff's attorneys for misleading him with AI-generated text. He directed the plaintiff's legal team to pay more than $30,000 in court costs that they wouldn't have otherwise had to bear.

After reviewing a supplemental brief filed by the plaintiffs, Wilner found that "approximately nine of the 27 legal citations in the ten-page brief were incorrect in some way."

Two of the citations, he said, do not exist, and several cited phony judicial opinions."