The recent spike in warnings about artificial intelligence from industry leaders has triggered questions about whether their calls for “AI safety” reflect genuine fear for humanity or disguise an orchestrated push for more power.
asserted
calls → trigger → power
In September, an Anthropic employee resigned, releasing a public statement condemning his employer and OpenAI, where he formerly worked, citing allegations that “the people building AI earnestly believe that it could kill us all by the end of the decade.”
uncertain
it → resign → decade
That resignation seemingly sparked a renewed commitment from Anthropic, along with OpenAI, and a host of other tech leaders to enact “guardrails” ensuring AI safety.
asserted
resignation → spark → safety
What exactly those reforms could mean, and who the guardrails protect, however, has become increasingly disputed.
uncertain
guardrails → mean → who
As panic spreads across the country, Congress is reaching a bipartisan consensus that change is needed.
asserted
change → spread → consensus
But critics, from Wall Street expert Michael Burry and former Meta chief AI scientist Yann LeCun to President Donald Trump’s top AI adviser, see a trap in which enacting certain changes could allow a select group of tech leaders to gain antitrust waivers and squelch competition, allowing them to consolidate power.
uncertain
them → see → power
For example, LeCun believes the push is designed to sideline U.S. companies that favor more open-source AI models.
asserted
that → believe → models
Anthropic and OpenAI are among industry leaders that have developed closed AI models, in contrast to Meta, Nvidia, and other companies that favor open-weight models.
asserted
that → develop → models
Open models are typically free and allow users more control over the model’s core inputs, as they can download and customize them to individual or business preferences.
asserted
they → allow → preferences
In recent years, they have been gaining steam in the United States, where closed models have enjoyed dominance.
asserted
models → gain → dominance
When pressed on whether LeCun’s fears are true, Kevin Frazier, a professor at the University of Texas at Austin School of Law, told the Washington Examiner it is difficult to determine which guardrails would strike the balance between preserving robust competition and ensuring AI safety.
asserted
guardrails → press → safety
“There’s a tension between policies that restrict access to the core inputs of AI or heighten the thresholds for releasing AI models and a more competitive ecosystem,” the AI innovation and law fellow said.
asserted
fellow → ’ → models
“That said, competition for the sake of competition is not good policy.
asserted
competition → say → competition
The aim should be to have as many reliable, useful AI models as possible available to meet a variety of users’ interests, needs, and values.
asserted
models → have → interests
Given the current state of the science of AI, which is highly uncertain, it’s difficult to know which policies adequately strike that balance.”
asserted
policies → give → balance
One by one, technology leaders, from Anthropic’s Dario Amodeo to SpaceXAI CEO Elon Musk, have backed industry-wide coordination on AI, in a call recently supported by Trump.
asserted
leaders → back → Trump
They have framed collaboration on the world’s most powerful technology as needed to implement safety controls.
asserted
They → frame → controls
Frazier agreed that coordination is necessary, but argued that concentrating power long-term in the hands of a few people is “troubling.”
asserted
concentrating → agree → people
“Under the status quo — one in which AI labs have lured academics and experts from the Ivory Tower to the Silicon Tower and those same labs have near-exclusive access to the latest tools, forming a sort of tokenocracy — there’s a need for collaboration on the science of AI because they are the entities with the talent and the resources,” he said.
asserted
he → lure → talent
“However, this troubling concentration of power should only be justified in the short term and under very specific constraints.
asserted
concentration → justify → constraints
We need to invest more in universities and research entities that can offer independent expertise and provide external bases of knowledge and guidance.”
asserted
that → need → knowledge
Business goals versus addressing the problem
Vahid Behzadan, a professor at the University of New Haven’s Department of Computer Engineering and Computer Science, said he believes industry-wide talks of coordination tend to “coincide with their business goals rather than actually addressing the problem.”
asserted
talks → address → problem
He said the AI safety debate over the years has often “somehow ended up promoting products and services rather than advancing actual implementation of safety guidelines or controls.”
asserted
debate → say → guidelines
Behzadan stood by his belief that collaboration is generally beneficial and that “some level of federal regulation is going to be required to ensure at least the basic degrees of safety” for AI.
asserted
level → stand → AI
However, he conceded that more guardrails on the industry would likely spell trouble for smaller start-up labs and those developing and using open models.
asserted
guardrails → concede → models
New licensing requirements, or mandating permissions to develop or lease models, “could become a costly barrier,” restrictions on downloadable models “could become an issue,” and smaller developers might avoid releasing “useful models because their exposure is unpredictable
uncertain
exposure → mandate → models
,” he said.
“This is not a new argument.
asserted
This → say → ?
Every time regulations are proposed by the bigger names in AI, this argument comes up,” Behzadan noted during an interview.
asserted
Behzadan → propose → interview
“The bigger names have significantly more resources, both in funding and access to talent and manpower to implement changes in their workflow, mandatory or otherwise, and making those changes mandatory would essentially equal the cost on a smaller developer or a smaller research lab.
asserted
changes → have → developer
“Essentially, smaller labs with either budgets may be constrained or may be budgeted out of the competition because of the more stringent regulations and compliance requirements,” he added.
uncertain
he → constrain → regulations
Industry calls for reforms have included warnings that independent oversight of models is needed through third-party auditors.
asserted
oversight → include → auditors
But some critics have raised concerns that such calls do not mark a genuine demand for enhancing AI safety.
asserted
calls → raise → safety
They have speculated that Anthropic’s suggestion that the prominent third-party evaluator METR could help make AI safer may be a ploy for a “regulatory capture machine” because of concerns that METR is financially tied to the company.
uncertain
METR → speculate → company
Frazier said that there are valid concerns that third-party reviewers “have a concerning level of personal and financial connection to many labs.”
asserted
reviewers → say → labs
He suggested fixing the issue by randomizing selections of auditors, mandating that auditors only remain at one place for a finite period of time “to avoid them acclimating to the culture and becoming too close to employees,” and making them subject to periodic review by a government authority to ensure they have the requisite expertise to do the work in question.
uncertain
they → suggest → question
China plays catch-up …
asserted
China → play → up
Concerns remain that scaling back AI development could give the reins of the technology to China, allowing it to surpass the U.S. as the world’s dominant superpower.
uncertain
it → remain → superpower
Such accusations have been prominently voiced in the Trump administration, including by David Sacks, the president’s AI adviser.
asserted
accusations → voice → Sacks
Sen. Ted Cruz (R-TX) and Rep. Brett Guthrie (R-KY) are among those in Congress who have raised fears that China is helping to orchestrate the latest wave of panic about AI safety in the U.S. so it can win the technology race, echoing allegations raised in investigations by the Bitcoin Policy Institute and OpenAI.
asserted
it → raise → Institute
Specifically, China could be weaponizing the open-closed model debate in the U.S. to push its own agenda.
uncertain
China → weaponize → agenda
…and 18 more, not listed.
The Aporia
