The Logistic Regression of the AI Safety Institutes, UK/US By David Stephen
How do you keep AI safe and how important is AI safety should have been the most important questions of any AI safety institute. The questions are not how to keep the AI model of a company safe by evaluating it before release, or that AI should simply be safe because of awareness or being trendy.
If any AI safety institute that meant business paid attention to the news—of misuses of AI for fake images affecting schools, fake AI voices used for impersonation against loved ones, fake videos used against businesses, wrong information from AI used for nefarious purposes, malware generation and so forth—it should have been clear that the approach to making AI safe has to be general, with tens of different methods at any time, in direct and indirect forms.
How do you keep AI safe?
AI safety should have led AI regulation, such that tools for safety may not need to be enforced for people to use them in a jurisdiction, but that people would adopt those tools because of their necessity. Simply, AI regulation that would work would be AI safety tools, as products, against unsafe and unaligned AI models or outputs. Though tough, there was not even an explored path [known] towards this outcome, making regulations go forward, then it became unpopular because of the vacuum left by innovation.
What AI safety institutes should have centered on, also, is that any unsafe AI from any source, is a risk wherever it is accessible. This means that it is important to work on AI model safety, output AI safety, platform AI safety, and jurisdiction AI safety. Model safety would apply to models. Output safety would be against misused AI outputs. Platform safety would be on platforms where AI outputs may appear, or where AI models can be found like social media, search engines, ISPs, company network, and so forth. Jurisdiction AI safety could be a county, a state, a nation, in ways that may not easily be bypassed.
The vision for safety and the approach could have been a driver to pursue technical answers, with the likelihood of minor to major success that would have been solid enough to withstand political pressures or a hard pivot.
The UK and the US AI safety institutes did not seem to have those, and they both seem to be on the verge of inconsequentiality. Already, the UK AI safety institute has been renamed to the UK AI security institute, changing its mandate since the prior mandate had little direction or anything of magnitude to show. A report on TIME, Inside the U.K.’s Bold Experiment in AI Safety, detailed a humiliating experience for the UK AI safety institute, where AI labs would not allow them access to the weights of the models in evaluation, meanwhile, DeepSeek, a model of a possible adversary, opened their weights to everyone. The UK AISI said it was a mistake to ask for the weights, maybe, but the lack of direction in assuming that evaluating a few models meant AI safety was a total waste, showing of a lack of scope on what their mission was about.
The US AI safety institute was an absolute nod off effort given the opportunity to make a major difference. There was the AI safety consortium that nothing ever came out of. Lots of organizations joined. They did not show or explain on the list page what their ongoing projects in AI safety or alignment were, and those that were working on AI safety would have done so, regardless of being a part of the consortium.
The US AISI had an agreement to evaluate the models of OpenAI and Anthropic, two companies that for all the misuses of AI, in the last two years, did not provide general or industry-wide answers, just like the rest. This means that even if those models were fairly safe, what would it matter if some families were put through a bad experience in deception of fake AI audio of a loved one or the several other scalable possibilities of misuses?
The US AISI did not have regulation staving technical plans, where it would not have been absolutely necessary to have laws or litigation for AI—that cannot punished—but to seek technical tools in directions of models, outputs, platforms and jurisdictions.
AI safety is now disdained, seen as hamstringing and unnecessary, because two major institutes did work that was not enough. There are several ways that the UK and US AI safety institutes would have made major progress that would have led the labs, industry, and regulators, but they did not. It was almost like they were seeking what to do, with the assumption that they had all the time, not knowing that changes that would be cataclysmic were ahead.
AI is not the kind of field where risks, threats, misuses, and vulnerabilities are nearly fixed, like a drug whose worse is fairly known or nearly constant. It is where those things evolve, so fast that pursuing after them requires constant updates and adjusted approaches just to ensure that society does not lose the fight.
There is a recent report on Axios, Scoop: NIST prepares to cut AI Safety Institute, CHIPS staff, stating that, “These cuts mean that the U.S. AI Safety Institute, which has been working on ensuring emerging AI models are trustworthy, is gutted, along with most staff working at NIST’s Chips for America program. NIST is preparing to cut 497 people, according to sources familiar, which includes: 74 postdocs 57% of CHIPS staff focused on incentives 67% of CHIPS staff focused on R&D Context: NIST is facing an uncertain future. AISI lost its leader earlier this month and its staff were left out of the AI summit in Paris last week.”
It is now mostly be up to the AI labs to pursue AI safety, as well as several non-profits. The authority that UK and US AISIs would have commanded is now in a vacuum, for leading paths. The other AI safety institutes around the world would unlikely be able to fill the void, given that those who inspired them already faltered. Those ones, too, may not even be rated by the AI labs. The opportunity wasted is a loss for human society.
There is another recent report on Axios, How AI safety is dying in government, stating that, “Rep. Jay Obernolte said earlier this month that he’ll reintroduce legislation to rename the AI Safety Institute to the Center for AI Advancement and Reliability.”
David Stephen currently does research in conceptual brain science with focus on the electrical and chemical signals for how they mechanize the human mind with implications for mental health, disorders, neurotechnology, consciousness, learning, artificial intelligence and nurture. He was a visiting scholar in medical entomology at the University of Illinois at Urbana Champaign, IL. He did computer vision research at Rovira i Virgili University, Tarragona.
See more breaking stories here.
