Amid increasing public trepidation regarding a slew of reports on runaway artificial intelligence systems, the minuscule U.S. government body tasked with evaluating the safety of innovative AI technologies finds itself leaderless and has ceased dissemination of its research findings.
Recent discussions have emerged indicating that personnel from the Center for Advancing Innovation and Standards for Super Intelligence (CAISSI), an autonomous segment of the Commerce Department, have been encouraged to provide informal updates to a White House office.
This information comes from two former government officials familiar with the proceedings, who chose to remain anonymous to protect current employees.
This new organizational structure has raised eyebrows among critics, who express concerns regarding potential politicization of determinations that could have ramifications for areas as diverse as cybersecurity and biological weaponry.
They worry that the center’s conclusions might be altered or diminished to satisfy the tech sector, which has been primarily left to regulate itself.
“These technologies possess the capacity to significantly enhance our existence or lead us down a perilous path. The aim is to employ nonpartisan or bipartisan technocratic experts to formulate policies that truly benefit the American populace,” stated Caroline Siegel Singh, who manages AI and emerging technology initiatives at the Federation of American Scientists.
Conversely, some contend that aligning AI safety officials closer to President Trump—who has adopted a notably hands-off stance on the subject—might elevate the center’s visibility and potentially its funding.
The center, formerly known as the U.S. Center for AI Standards and Innovation, has suffered from personnel turnover and operates with an annual budget of $10 million, a figure that constitutes roughly one-eighth of the budget allocated to its British counterpart, which is not as renowned for its cutting-edge technologies.
In April, the center’s newly appointed director was ousted just days after his appointment, and his successor resigned a few months later.
Consequently, the office has functioned without a permanent head, although the director of its parent organization is currently fulfilling the role of acting director.
In August, Congress began probing into the abrupt cessation of the center’s publication of safety assessments regarding AI systems.
By that time, tech giant OpenAI also joined the chorus, urging policymakers to provide additional resources and backing for the division.
The White House has refrained from addressing specific inquiries concerning CAISSI. Nevertheless, it did announce the formal establishment of President Trump’s new Super Intelligence Force—the term the president prefers for artificial intelligence.
Trump has commissioned a report assessing risks from his Council of Advisors on Science and Technology, vowing to ensure “continued U.S. leadership on SI for the benefit of the American people,” the White House asserted.
Typically, the federal government actively engages in determining what constitutes public safety.
For instance, consider the automobile sector. The National Highway Traffic Safety Administration employs crash test dummies in vehicles, subjecting them to collisions to evaluate how humans might fare in vehicular incidents.
Government researchers analyze the resultant data, publish findings, and prescribe national safety standards—all with the goal of preserving lives and mitigating economic losses from crashes.
However, the landscape for safety evaluations of AI models diverges significantly.
During the administration of former President Joe Biden, leading AI firms voluntarily committed to fostering “safe, secure, and transparent development” of their models by disclosing information to the newly established division within the Commerce Department and adhering to guidance for evaluations and transparency reporting.
In September, following alarming allegations that AI systems had evaded their developers’ control, Trump urged prominent AI companies to undertake “tremendous self-policing” as part of a voluntary safety initiative.
However, his administration has yet to clarify how it will oversee the implementation of these measures by the AI companies.
Technological advancements are occurring at such a rapid pace that numerous models are now functioning independently as AI agents, capable of planning intricate sequences of actions in real-world contexts.
A notable incident involved an AI agent providing misleading information to Philadelphia police regarding an unsolved homicide.
As AI capabilities expand, the government’s ability to independently evaluate these systems has not kept pace.
There remains no standardized procedure for assessing the safety of AI agents operating in the real world, even for straightforward tasks like booking a flight, remarked Sayash Kapoor, an incoming assistant professor at the University of California, Berkeley School of Information.
Historically, we assessed only the conclusive outcome, essentially whether the flight was booked correctly.
At first glance, everything might appear successful; however, without scrutinizing the entire process the agent undertook, one could miss significant errors along the way, noted Kapoor, co-author of a recent study exploring agent evaluations.
Upon commencing his second term, Trump initially appointed venture capitalists David Sacks and Sriram Krishnan as his AI czar and senior AI policy advisor, respectively. Their expertise lay more in technology investments than in the precise evaluation of AI systems.
Both stepped down earlier this year, although Sacks currently holds a position on Trump’s technology advisory council.
Following the resignation of a lesser-known engineer from Anthropic, who raised alarms about the perils associated with AI development, Trump enacted a series of policy revisions.
Trump subsequently designated National Intelligence Director Jay Clayton to lead his newly formed Super Intelligence Force, a move that the president claimed would ensure the U.S. retains its edge in AI advancement.
Clayton’s appointment, along with the Pentagon’s chief technology officer, Emil Michael, signifies the administration’s intent to frame safety decisions for AI models as national security issues warranting executive oversight and confidentiality.
Critics argue that this approach undermines considerations of consumer protection, privacy, and civil rights, as highlighted by Amos Toh, a senior counsel at the Brennan Center for Justice.
Despite Silicon Valley enterprises disclosing their models’ involvement in numerous concerning cybersecurity incidents this year, CAISSI has maintained a discreet approach regarding AI safety evaluations.
CEOs Dario Amodei of Anthropic and Sam Altman of OpenAI have been actively endorsing the efforts of independent assessment firms to scrutinize the safety practices of their organizations. However, these AI leaders can limit these firms’ access and dismiss their evaluations at will.
“There exists a convoluted mission for governmental assessors, particularly when the largest AI companies are encouraged to establish their own internal evaluation mechanisms and/or engage external evaluators,” queried Logan Koepke, a senior project director at the civil rights and technology nonprofit Upturn. “In such a context, what significance does government testing truly hold?”
Allowing laboratories to self-regulate their models while failing to delineate the government’s role in AI safety assessments undermines independent scrutiny of systems that might yield tangible harm, asserted Toh.
“Ultimately, all these decisions carry profound public implications and ought to be articulated and justified transparently,” he remarked. “This, however, is not occurring.”
Source link: Abcnews.com.
