Doomsday scenarios — from home-cooked bioweapons to rogue agents hacking our critical infrastructure— dominate this week’s debate over the future of artificial intelligence.
asserted
scenarios → cook → intelligence
Those risks absolutely demand attention; we don’t yet fully understand AI’s profound capacities.
asserted
we → demand → capacities
But we’re ignoring, at our peril, a far more immediate risk of frontier AI: It’s drenched in woke racism.
asserted
It → ignore → racism
Askell, a philosophy PhD, is trying to do with Claude what Plato attempted to accomplish with the tyrants of Syracuse: bind vast power within an intellectual armature that ensures its just use.
asserted
that → try → use
Askell, in effect, is in charge of developing and nurturing Claude’s “soul” (to use her own description).
asserted
Askell → develop → description
And whatever the merits of her mission, understanding how this leading frontier AI system will behave requires understanding what Askell herself believes.
asserted
Askell → understand → what
Those who don’t want to organize society primarily around racial grievances may find the apparent answer alarming.
uncertain
answer → want → grievances
Consider a 2023 Anthropic paper Askell co-authored, “The Capacity for Moral Self-Correction in Large Language Models”.
asserted
Askell → consider → Models
The research explored AI’s capacity to “self-correct” — that is, to refuse to deliver outputs the authors regarded as harmful.
asserted
authors → explore → outputs
They found, in an experiment designed to test Claude’s propensity for racial discrimination, that its initial answers to a question about law-school admission showed a slight bias against black students — but with a few prompts, the researchers got Claude to show worse bias against white students.
asserted
Claude → find → students
Then they offered an unsettling opinion on the white-bias result:
“This may be desirable in certain contexts, such as those in which decisions attempt to correct for historical injustices against marginalized groups.
uncertain
decisions → offer → groups
Just in case the message wasn’t clear, they hammered it home: “We do not assume all forms of discrimination are bad.
asserted
forms → hammer → discrimination
Positive discrimination in favor of Black students may be considered morally justified.”
uncertain
discrimination → consider → students
In other words: It’s fine for AI to be racist against white people — that’s reparations.
asserted
that → ’ → people
Another Anthropic paper Askell co-authored that year declared that “white and male” are “historically privileged groups,” and pondered “under what circumstances (and to what degree) positive discrimination should be corrected for.”
asserted
discrimination → co → circumstances
Why isn’t there a similar intellectual openness, one wonders, around “negative” discrimination?
asserted
one → wonder → discrimination
Such DEI-drenched views have since become embedded in AI systems — and not just Anthropic’s.
asserted
views → drench → systems
A Washington Post analysis recently found that OpenAI’s ChatGPT provides leftist-aligned answers to controversial questions around 80% of the time.
asserted
ChatGPT → find → time
Claude’s Opus 4.8 model was slightly less fanatical, delivering lefty answers 43% of the time and presenting “both-sides” views 47% of the time — but never serving up solely right-leaning answers.
And if the past decade offers one lesson, it’s that normalizing these forms of leftist thinking undermines the successful functioning of American democracy.
asserted
normalizing → deliver → democracy
Look at the corrosive effect the violent neo-racism of people like Ibram X. Kendi and Black Lives Matter’s Patrisse Cullors has had on the nation as a whole over the past decade.
asserted
racism → look → decade
One can’t help wondering if that’s rather the point.
asserted
that → help → ?
Remember, the musings of Askell et al. on the “desirability” of “correcting for historical injustices” dates from 2023.
asserted
musings → remember → 2023
None of its authors could credibly argue their views have matured in the three years since.
uncertain
views → argue → years
The historical injustices they appealed to still obtain; indeed, their only possible defense would amount to a version of AOC’s excuse, that “Woke 1 was craaaazy.”
asserted
Woke → appeal → excuse
That answer instantly disqualifies people who assert they possess the moral insight necessary to train up the most powerful — and potentially mind-molding — technology of our era.
asserted
they → disqualify → era
Claude’s current customer-facing moderation is, I suspect, a tactical retreat.
asserted
I → face → ?
Consider Anthropic CEO Dario Amodei’s recent call to arms on AI risk, “We Must Pace the Frontier.”
asserted
We → consider → Frontier
His vision to save the future from AI doom involves a small, rich group of his own ideological allies at the nonprofit Model Evaluation and Threat Research.
asserted
vision → save → nonprofit
They, he declares, should decide the future of the technology, playing a role somewhere between that of Soviet political commissars and inspectors from the International Atomic Energy Agency.
asserted
he → declare → Agency
A quick glance at METR’s staff roster gives ample reason for concern: Its CEO and founder Beth Barnes is on record as early as 2018 supporting full open borders, scoffing at the very idea of nationality.
asserted
Barnes → give → nationality
The group has deep ties to the Effective Altruism movement — a cult of privileged weirdos masquerading as philanthropists, including the fraudster Sam Bankman-Fried.
asserted
group → have → fraudster
As Askell’s work shows, however, ideological projects of this kind are nothing new for Amodei.
asserted
projects → show → Amodei
Which means the tech’s users, and above all its younger users, are freely imbibing woke racism via ever more powerful and subtle channels.
asserted
users → mean → channels
That’s a risk as serious as any of the Bond-villain scenarios that allegedly keep Amodei, OpenAI’s Sam Altman and their business peers awake at night.
uncertain
that → ’ → night
And it’s already happening.
asserted
it → happen → ?
The Aporia
