Hey and welcome to Eye on AI. On this version…The Pentagon struggle with Anthropic raises three essential questions…OpenAI raises $110 billion in new funding…Meta experiments with an AI procuring assistant…LLMs can determine pseudonymous web customers at scale…knowledge facilities on the entrance traces within the Iran battle.
A very powerful story in AI in the intervening time, for sure, is the struggle between the U.S. Division of Warfare and Anthropic. For those who haven’t been following the drama, you’ll be able to compensate for the story by studying protection from me and my Fortune colleagues right here, right here, right here, right here, right here and right here.
This story raises at the least three essential questions: who ought to have management over how AI is utilized in a democratic society? How ought to that management be exercised? What ought to the implications be for a corporation that disagrees with the federal government’s coverage?
No matter you consider OpenAI CEO Sam Altman and his determination to brush in and signal a take care of the Pentagon—together with a contractual obligation to permit the army to make use of OpenAI’s AI fashions “for any lawful purpose” that Anthropic had refused to comply with—Altman accurately recognized what’s at stake on this struggle.
In an “Ask Me Anything” session on X over the weekend, Altman mentioned:
A very essential level: we’re not elected. We have now a democratic course of the place we do elect our leaders. We have now experience with the know-how and perceive its limitations, however I believe you have to be fearful of a personal firm deciding on what’s and isn’t moral in a very powerful areas. Appears effective for us to determine how ChatGPT ought to reply to a controversial query. However I actually don’t need us to determine what to do if a nuke is coming in the direction of the US.
This was the crux of the Pentagon’s said objection to Anthropic’s present contract. The army didn’t suppose it was proper to have a personal firm dictating insurance policies to an elected authorities.
AI strikes lightning quick, Congress at a snail’s tempo
Most Individuals would possibly agree with the Pentagon’s place—in precept. Besides it’s difficult, in observe, by three issues. First, AI know-how is transferring extraordinarily quick, however the mechanisms of democratic management—laws, Congressional oversight, elections—transfer extraordinarily slowly. Within the three years since ChatGPT debuted, Congress has not handed any federal AI laws. The Trump Administration has dismantled restricted AI rules put in place by its predecessor, whereas additionally performing to punish states that cross their very own AI rules.
So whereas many individuals would possibly agree that insurance policies on the federal government’s AI use must be set by elected officers, there may be the sensible concern of what to do when these elected representatives fail to behave. The thought of making an attempt to reach at AI coverage by contractual negotiations between labs and authorities is a poor substitute for true democratic governance, nevertheless it is perhaps higher than no governance in any respect. The controversy over Anthropic’s Pentagon contract ought to be a get up name for Congress to behave.
Second, the pattern amongst governments over the previous a number of a long time has been to interpret present legal guidelines broadly in an effort to increase the ability of the federal government to make use of know-how to surveil its residents. (The story has been one of many government department regularly clawing again surveillance powers it misplaced by Congressional motion following the scandals that emerged with Watergate and the Church Committee hearings within the mid-Nineteen Seventies.) Many actions of the army are additionally cloaked in secrecy that makes democratic oversight and accountability troublesome. This fixed pushing on the boundaries of what the regulation will enable has made the general public distrustful of the federal government’s intentions. So it’s not stunning that some folks at this level may very well have extra religion in a seemingly well-intentioned and good, however unelected, know-how government, akin to Anthropic’s Dario Amodei, to do the precise factor and set the precise insurance policies.
Lastly, there may be the problem that many Individuals have with this particular authorities. The Trump administration has repeatedly taken unprecedented actions to punish home dissent, usually on flimsy authorized justifications, or with no authorized justification in any respect, and has repeatedly deployed the army domestically to intimidate or punish perceived home opposition. It has additionally launched a number of army actions abroad with little to no authorized justification. So is it any marvel that many query whether or not this specific administration ought to be given the ability to make use of AI for something its personal attorneys imagine is authorized?
Is the nationalization of AI inevitable?
Even in case you suppose the Pentagon is right that democratic governments, not non-public corporations, ought to determine on how AI is used, the following query turns into how that management ought to be exercised? Altman put his finger on the last word query hanging over the business: if frontier AI is a strategic know-how, why doesn’t the federal government merely nationalize it? In any case, many different breakthroughs with massive strategic implications—from the Manhattan Challenge to the house race to early efforts to develop AI—had been government-funded and largely government-directed. As Altman mentioned, “it has seemed to me for a long time it might be better if building AGI were a government project,” although he added it “doesn’t seem super likely on current trajectory.”
The Pentagon’s present method comes near nationalization by different means. One choice the DoW threatened was utilizing the Protection Manufacturing Act, a Chilly Warfare-era regulation, to compel Anthropic to ship an AI mannequin on its most popular phrases—a kind of delicate nationalization of Anthropic’s manufacturing pipeline. And the retaliatory determination to label Anthropic a “supply chain risk” is designed partially to intimidate different AI corporations into accepting the Pentagon’s most popular contract phrases, which once more appears nationalization-adjacent.
What ought to the price of dissent be in a democracy?
Lastly, this brings us to the query of what an applicable punishment ought to be for an AI firm that refuses to comply with the federal government’s most popular contract phrases. As Dean Ball, an AI coverage knowledgeable who labored briefly for the Trump administration on its AI Motion Plan, has mentioned, the federal government appears inside its rights to cancel its $200 million contract with Anthropic.
However the determination to go a lot additional and label Anthropic a “supply chain risk” strikes on the coronary heart of personal property rights and free speech in a liberal democracy. The designation—which was supposed for use in opposition to applied sciences that would assist a international adversary sabotage essential protection methods—had by no means earlier than been utilized to a U.S. firm and by no means earlier than been used to punish an organization for not agreeing to contract phrases that U.S. army desired. The choice, Ball has mentioned, quantities to “attempted corporate murder,” since underneath the SCR designation any firm doing enterprise with the Pentagon could be barred from any industrial relationship with Anthropic. If that interpretation stands—and plenty of authorized students have mentioned it is not going to—it may very well be a mortal blow to Anthropic, which will depend on promoting to giant Fortune 500 corporations that additionally do work for the Pentagon for income, cloud computing infrastructure, and enterprise capital backing. Ought to the punishment for arguing with the federal government be the dying of your small business? That actually appears un-American.
Altman has claimed he struck his take care of the Pentagon partially to de-escalate the strain between the federal government and AI corporations, saying that “a close partnership between governments and the companies building this technology is super important.” Whereas I’m uncertain of Altman’s true motives, I agree with him on this final level. At a time when AI doubtlessly threatens unprecedented adjustments to the economic system and society, fomenting mistrust and battle between the federal government and the folks constructing superior AI methods looks as if a fairly unhealthy concept.
FORTUNE ON AI
Anthropic’s Claude overtakes ChatGPT in App Retailer as customers boycott over OpenAI’s $200 million Pentagon contract—by Marco Quiroz-Gutierrez
Iran has the intent—and more and more the instruments—for AI-powered cyberattacks—by Sharon Goldman
Unique: CrowdStrike and SentinelOne veterans elevate $34M to deal with enterprise AI’s governance hole—by Beatrice Nolan
OpenAI’s Pentagon deal raises new questions on AI and mass surveillance—by Beatrice Nolan
The week the AI scare turned actual and America realized perhaps it isn’t prepared for what’s coming—by Nick Lichtenberg
AI IN THE NEWS
Meta is testing an AI procuring assistant. That’s in response to a narrative in Bloomberg, which says the social media big is hoping to create an AI procuring instrument that may rival ecommerce choices being integrated into OpenAI’s ChatGPT and Google’s Gemini. The Meta characteristic, now rolling out to some US net customers, gives product suggestions in a carousel format with photos, costs, model particulars, and transient explanations, and tailors solutions primarily based on inferred knowledge akin to location and gender, although purchases have to be accomplished on exterior service provider websites. CEO Mark Zuckerberg has framed the transfer as a part of Meta’s push towards “personal superintelligence,” hinting that future agentic procuring instruments might deepen ties between its AI merchandise and its promoting ecosystem.
Considering Machines loses two extra founding workforce members. Christian Gibson and Noah Shpak, two members of the founding workforce on the high-profile AI “neolab” based by former OpenAI CTO Mira Murati, have quietly left to affix Meta. That’s in response to Enterprise Insider. Their departures add to a broader wave of exits from the San Francisco-based firm, which raised a $2 billion seed spherical at a $12 billion valuation however has struggled to retain key personnel as rivals like Meta and OpenAI poach engineers.
EYE ON AI RESEARCH
AI CALENDAR
March 2-5: Cell World Congress, Barcelona, Spain.
March 12-18: South by Southwest, Austin, Texas.
March 16-19: Nvidia GTC, San Jose, Calif.
April 6-9: HumanX 2026, San Francisco.
BRAIN FOOD
As AI turns into more and more essential to preventing wars, do knowledge facilities turn out to be prime targets? That’s what some persons are asking after Amazon reported that two of its AWS knowledge facilities within the UAE and one in Bahrain had been struck by Iranian missiles or drones, taking them out of service. The assaults pressured customers to modify to providers hosted in additional distant areas and resulted in non permanent service outages. It additionally could have launched extra latency into cloud-based purposes.
It isn’t recognized precisely why the Iranians struck the information facilities. It may very well be that they had been merely making an attempt to disrupt web providers as a method of punishing Gulf States that hosted U.S. army bases. However Yanis Varoufakis, the economist and former Greek finance minister, was amongst these speculating that Iran hit the amenities in an effort to disrupt the U.S. army’s use of Anthropic’s Claude AI fashions.
Regardless of the Pentagon labeling Anthropic a “supply chain risk” and saying the army would stop utilizing Anthropic’s Claude “immediately,” the Wall Avenue Journal and Axios have reported that the army is utilizing Claude for assist with goal processing as a part of Operation Epic Fury, its battle in opposition to Iran. Additionally it is recognized that at the least a number of the categorised networks the army runs Claude on are hosted by AWS.
So it stands to purpose, Varoufakis and others speculate, that Iran attacked the information facilities in an effort to disrupt the U.S. army’s use of Claude. It’s not clear whether it is true on this case, however it’s prone to be true in future conflicts that knowledge facilities, even these very distant from the entrance traces, will turn out to be targets due to how essential AI is turning into to battle preventing.
