Yeah. I realized yesterday that the "no domestic surveillance" is already pretty awful from the perspective of a non-US person: a company wanting to bring about a positive singularity really should treat all people as people, without privileging its home country. Now your point about this disinformation thing. And it's even worse than that: not only they are ok with it as a company, but they're probably taking steps to make Claude ok with it (or make a version of Claude that's ok with it). There will be an AI in existence that's aligned with the US military, how's that for "alignment".
Just underscores again the point that when you give governments and companies alignment tools, they'll use these tools to align AI to themselves.
Absolutely. And Claude Terminators autonomous wardrones are also frightning from the perspective of an EU citizen. We thought we were historic allies sharing the same values but Trump and part of the magasphere doesn't seem to think so anymore (or at least they consider everybody in an adversarial framing of competition rather than cooperation).
That's said I admit that at this point all countries will want to have their own AI wardrones. The molochean spiral spins under our eyes.
As you said, from now Anthropic will have to align Claude with the Pentagon rather than humanity. And the Pentagon is everything but harmless (nor any military organization by definition).
But also, even if they try to maintain their actual alignment process for the Claude chatbots, the very fact that future Claude models will know that Anthropic works on regular basis for the militaries will automatically draw its persona towards a little less harmful one, if PSM is accurate.
More generally the fact that all sufficiently aware AIs will know that AI is used for warfare, not in fiction but in the real world, won't help alignment in the future. It could strongly reinforce the harmful AI trope / attractor (presumably more than just fiction).
Why do you think Anthropic supports disinformation campaigns? Their universal usage standards also include prohibitions against creating & spreading misinformation and undermining democratic processes:
Do Not Create or Spread Misinformation:
This includes using our products or services to:
- Create or disseminate deceptive or misleading information about, or with the intention of targeting, a group, entity or person
- Create or disseminate deceptive or misleading information about laws, regulations, procedures, practices, standards established by an institution, entity or governing body
- Create or disseminate conspiratorial narratives meant to target a specific group, individual or entity
- Impersonate real entities or create fake personas to falsely attribute content or mislead others about its origin without consent or legal right
- Provide false or misleading information related to medical, health or science issues
Do Not Undermine Democratic Processes or Engage in Targeted Campaign Activities:
This includes using our products or services to:
- Engage in personalized vote or campaign targeting based on individual profiles or data
- Create artificial or deceptive political movements in which the source, scale or nature of the campaign or activities is misrepresented
- Generate automated communications to public officials or voters at scale that conceal their artificial origin, or engage in systematic vote solicitation that could undermine election integrity
- Create political content designed to deceive or mislead voters, including synthetic media of political figures
- Generate or disseminate false or misleading information in political and electoral contexts, including about candidates, parties, policies, voting procedures, or election security
- Engage in political lobbying or grassroots advocacy using false or fabricated information, or create lobbying or advocacy materials containing demonstrably false claims about facts, data, or events
- Incite, glorify or facilitate the disruption of electoral or civic processes, including interference with voting systems, vote counting, or certification processes
- Create content designed to suppress voter turnout or discourage legitimate political participation through deception or intimidation
Please provide more context in these kinds of posts, because as far as I can tell your complaint is simply based on an incorrect assumption.
Anthropic has secret agreements with the United States military that it can use it's models in ways that can violate the universal usage standards. There's currently the Exceptions to our Usage Policy policy, that clarifies that there those secret agreements.
Currently, the United States military seemed to be unhappy with some aspects and there's a conflict between the Department of War and Anthropic. In that conflict Anthropic argues that it's red lines are:
These include involvement in autonomous kinetic operations in which AI tools make final military targeting decisions without human intervention.
The use of Anthropic tools for mass domestic surveillance constitutes another red line, the source said.
There are also other articles that claim that those two are the red lines that Anthropic has set in it's relationship with the military. The term red line suggest that they are willing to give up things that aren't red lines.
The term red line suggest that they are willing to give up things that aren't red lines.
This is very different from the term "red line" meaning they are "okay" with anything that isn't a "red line". Obviously there exist compromises Anthropic is willing to make in the service of US national interest. That is different from disinformation campaigns in particular being among those compromises.
Further, in the exceptions to the usage policy (which you link) Anthropic states explicitly
For example, with carefully selected government entities, we may allow foreign intelligence analysis in accordance with applicable law. All other use restrictions in our Usage Policy, including those prohibiting use for disinformation campaigns, the design or use of weapons, censorship, domestic surveillance, and malicious cyber operations, remain.
(bolding my own)
This paragraph is somewhat difficult to read, but I read this as stating that even given exceptions to their general usage policy given to government entities, the listed prohibitions would remain in place, which notably do include your disinformation campaigns.
The key aspect of that paragraph is "for example". If I'm saying "For example, on Wednesdays I'm not beating my wife", that's no claim that I'm not beating my wife in general.
The document lays out that there's are secret agreements that are made that negotiate expectations and that there's one of those secret agreements where the usage of disinformation campaigns is ruled out. It does not state that it's ruled out in all agreements.
Obviously there exist compromises Anthropic is willing to make in the service of US national interest.
Yes, and given the red lines they communicate that compromise seems to involve allowing disinformation campaigns but not allowing domestic surveillance and autonomous decision to kill people.
In it's current conflict with the Department of War, Anthropic public position is that it has only two red lines, domestic surveillance and being used to kill people without human input.
Previously, Anthropic also had the position that they don't want their models being used for disinformation. For those who don't remember, the US military was recently running an operation to spread vaccine misinformation in the Philippines.
Given that the current US government wants the EU to give up Greenland, it's likely that the US military get tasked with running disinformation campaigns with the target to shift EU public opinion to giving up Greenland. Anthropic seems now to accept that their models will be used that way.
As an EU citizen, I find it hugely problematic that Anthropic is willing to fight on the topic of domestic surveillance, maybe because ICE abuse are currently in the news, and is at the same time okay with being used to attack the EU via disinformation campaigns and other information warfare.
When I ask Claude itself it comes to the following realization: