AI safety has been all over the news recently.The US Department of War asked Anthropic to allow use of its AI for "any lawful use," and the AI company ultimately refused.While this is seen by many as a positive step, it still seems the case that AI companies only really have a conscience when it suits them.
Related I self-host my own private ChatGPT with this tool Running your own AI isn't only easy, it is a lot more cost effective if you already have a gaming PC.Posts 2 By Nick Lewis The removal of Anthropic's safety pledge Anthropic's 'don't be evil' moment The whole timeline of this story is something of a whirlwind.On February 14, the Wall Street Journal reported that Claude, Anthropic's AI chatbot, was used by the US military during its operation targeting Nicolás Maduro in Venezuela, although it was not revealed how the AI was used.
Just over a week later, Secretary of War Pete Hegseth summoned Anthropic CEO Dario Amodei to the Pentagon for negotiations, before issuing a public ultimatum for Anthropic to allow the DoW "unfettered access" to Anthropic's AI models "for any lawful use," with a deadline of Friday of the same week.In curious timing, the very same day the ultimatum was issued, Anthropic updated its Responsible Scaling Policy (RSP).The 2024 version of the document had previously stated that if the company developed an AI model that required the highest level of security—such as a model that could help individuals create/obtain and deploy chemical, biological, radiological, and nuclear weapons—it would "pause training until we have implemented the ASL-3 Security Standard and established it is sufficient for the model." The updated version released on February 24 no longer includes this pledge to pause training.
The document argues that if one developer paused to develop adequate safety standards while others didn't, it would allow the AI companies with the weakest protections to "set the pace." In other words, if Anthropic stuck to its original safety policy, other AI companies that didn't have the same scruples might get too far ahead.The implication is that not letting other companies get too big a lead in the AI race is more important than safety.Anthropic refused to cave to the DoW OpenAI immediately jumped in Despite the timing of the changes, which seemed to indicate that Anthropic was ready to bow to the DoW's demands, the deadline for the ultimatum came and went, with Anthropic refusing to budge on its two "bright red lines." These were that Anthropic's AI models should not be used for fully autonomous weapons, and that they should not be used for mass domestic surveillance.
This wasn't good enough for the DoW, suggesting that the US government may want the option to use AI for fully autonomous weapons and mass domestic surveillance, both of which have worrying implications.Anthropic was officially designated as a "Supply Chain Risk to National Security," and the government has entered a six-month window of phasing out the federal use of Claude, although Anthropic argues that the designation is limited to DoW contract use.It didn't take long for the DoW to find another AI partner, however.
Just a few hours after Anthropic was deemed a Supply Chain Risk, OpenAI announced a deal with the DoW to use its AI models.The backlash was almost immediate.The announcement was quickly followed by an online trend of people ditching ChatGPT and moving to alternatives such as Anthropic's Claude.
OpenAI's deal isn't as safe as it seems The agreement is open to interpretation Bizarrely, OpenAI stated that its deal with the DoW included red lines that meant OpenAI models could not be used for mass domestic surveillance or to direct autonomous weapons, the exact issues that made the DoW's deal with Anthropic fall apart so dramatically.However, the devil appears to be in the details.OpenAI's agreement contains the following: Consistent with applicable laws, including the Fourth Amendment to the United States Constitution, National Security Act of 1947, FISA Act of 1978, the AI system shall not be intentionally used for domestic surveillance of U.S.
persons and nationals.An Anthropic spokesperson told CBS News on February 26 that the DoW has offered a compromise deal that "was paired with legalese that would allow those safeguards to be discarded at will." The phrase "consistent with applicable laws" in OpenAI's deal appears to be doing a lot of heavy lifting; OpenAI claims it strengthens guardrails, but some critics argue that reliance on existing legal frameworks creates loopholes that could still allow the government to use OpenAI's tech for domestic surveillance.Anthropic isn't against fully autonomous weapons The issue may be one of blame Anthropic's refusal to bow to the ultimatum and OpenAI's willingness to step in prompted an online backlash against OpenAI, with the cancellation of ChatGPT accounts quickly becoming a trending topic.
Many people posted screenshots of their cancellation page, with a lot of users opting to make the switch to Claude in light of Anthropic's seemingly more ethical stance.However, the reality may be a little more nuanced.In an interview with CBS News, Amodei stated that "we are not categorically against fully autonomous weapons." His issue is that currently, AI models aren't reliable enough to be used in fully autonomous weapons.
In other words, AI models that can and do hallucinate probably shouldn't be given full control of deadly weapons.Subscribe to our newsletter for AI safety insight Get clear, evidence-based coverage—subscribe to the newsletter for in-depth analysis of AI safety, ethics, and corporate behavior.Track how policies, government deals, and model risks shape the debate so you can follow decisions with informed context.
Get Updates By subscribing, you agree to receive newsletter and marketing emails, and accept our Terms of Use and Privacy Policy.You can unsubscribe anytime.A more cynical interpretation is that the line in the sand, therefore, is nothing to do with a moral stance.
It's because Amodei doesn't want Anthropic's AI models being used in fully autonomous weapons that are at some point likely to fail.In theory, the government could then turn around and blame Anthropic for the outcome of those weapons failing, which wouldn't be good for the company's bottom line.Other AI companies are as bad or worse Moral AI companies don't seem to exist OpenAI and Anthropic are at the center of this current storm, but they're far from the only AI companies with questionable ethics.
It's alleged that Meta torrented more than 80TB of pirated books in order to train its AI models, with claims that internal communications indicate that the decision was given the green light by Mark Zuckerberg in full knowledge that this was pirated content.Elon Musk's Grok came under fire for allowing users to generate sexualized non-consensual images of real people, including high-profile figures, using its publicly available image generation tools.The company restricted the tool following investigations by national and multinational bodies.
The other major player is Google Gemini, operated by a company that famously removed "Don't be evil" from the opening pledge of its code of conduct and relegated it to a comment at the end.Google uses AI-generated overviews in its search results, and these have included suggestions such as putting glue on pizza and eating one small rock per day for the mineral content.Despite the significant health risks these overviews have posed, Google continues to display them.
Sadly, the reality is that AI companies only have a conscience when it's convenient.If that conscience could let their competitors get ahead or cost them significant amounts of money, it appears that money will always take precedence.AI models continue to be developed with minimal safety guards in place.
When those AI models have been trained on data that includes every single book and movie in which AI becomes self-aware and tries to destroy humanity, what could possibly go wrong?
Read More
