November 2 2022, I participated in the Google AI event in Novel York. One of the topics was responsible artificial intelligence. When I listened to the management staff about how they adapted their technology to human values, I realized that the plasticity of AI models was a double -edged sword. Models can be improved to say, minimized prejudices, but also to enforce a specific point of view. Governments may demand manipulation to censor unwanted facts and promote propaganda. I predicted it as something that an authoritarian regime can utilize, such as China. Of course, in the United States, the Constitution would prevent the government from spoiling the products of AI models created by private companies.
On Wednesday, Trump’s administration published its AI manifesto, far -reaching Action plan For one of the most essential problems that the country is facing – and even humanity. The plan generally focuses on defeating China in the AI supremacy race. But one part seems more synchronized with a Chinese textbook. On behalf of the truth, the US government now wants AI models to observe the definition of the word Donald Trump.
It will not turn out that the intentions were clearly found in the 28-page plan. Instead, he says: “It is necessary for these systems to be built from scratch with freedom of word and statements, and the US government policy does not interfere with this goal. We must make sure that freedom of speech blooms in the AI era and that AI acquired by the federal government objectively reflects the truth, not social engineering.”
It’s all right to the last sentence that raises the question – according to whom? And how exactly is the “Social Engineering Program”? We get a clue on this subject in the next paragraph, which is instructed by the Trade Department to look at the principles of AI from the time of Biden and “eliminate references to disinformation, diversity, equality and inclusion and climate change”. (Strange huge part written in the published plan) climate change Is social engineering? As for the truth, in information sheet The White Dom plan says: “LLM will be truthful and priority treat historical accuracy, scientific research and objectivity.” Sounds good, but comes from the administration that Limits the history of America To “raise” interpretations, he denies climate change and considers the claims of Donald Trump about being the greatest president of America as an objective truth. Meanwhile, this week, Trump’s social account published AI from Obama in prison.
In a speech advertising a plan in Washington on Wednesday Trump He explained the logic Behind the directive: “The American nation does not want to wake up Marxist madness in AI models,” he said. Then he signed the contract executive order entitled “preventing AI in the federal government”. Determining that “the federal government should hesitate before regulating the functionality of AI models on the private market”, he declares that “in the context of federal orders, it is obliged not to obtain models that devote the truth and accuracy to ideological programs.” Because all huge AI companies apply to government contracts, the order seems to be backdoor to ensure that LLM at all show the interpretation of the history of the White House, sexual identity and other problems with the balmy button. In case of doubt as to what the government considers to be a violation, the order gives several paragraphs demonizing artificial intelligence, which supports diversity, causes racial prejudice or values gender equality. Pogo alert-Trump executive prohibition prohibiting ideological prejudices is a glaring exercise in a top -down ideological attitude.
Marks Madness
It is necessary to determine from companies how to deal with these requirements. This week I talked to Openai engineer working on model behaviors, who told me that the company is already striving for neutrality. They said that in a technical sense, meeting government standards, such as being anti-irritation, should not be a huge obstacle. But this is not a technical dispute: it is constitutional. If companies such as anthropic, OpenAI or Google decide to minimize racial prejudices in their LLM or make a conscious choice to ensure that models’ reactions reflect the dangers of climate change, the first amendment probably protects these decisions as using “freedom of word and expression” advertised in the AI action plan. Government fines refusing government contracts to companies exercising this right is the essence of interference.
You may think that companies building artificial intelligence are fighting by citing their constitutional rights in this matter. But so far no huge technology company has publicly opposed the Trump administration plan. Google famous Support for a White House for his animal problems, such as increasing infrastructure. Anthropic has published a positive blog post On the subject of the plan, although he complained about the suddenly abandoning the sturdy control of the White House exports at the beginning of this month. Opeli says he is already close to achieving objectivity. Nothing about maintaining your own freedom of expression.
In action
Restraint is understood because in general AI’s action plan is Bonanza for AI. While the Biden administration ordered the control of great technology, Trump’s plan is a huge hefty green lithe for the industry, which he considers a partner in the national fight to defeat China. It enables AI entitled to bother the environmental reservations in the past when constructing massive data centers. AI research supports, which will be flowing into the private sector. There is even a recipe that limits some federal funds for states that try to regulate AI themselves. This is a consolation reward for a failed part of a recent budget account that would prohibit state regulations for a decade.
However, for the rest of us, the “anti-fed” order is not so easily rejected. AI is more and more often a medium in which we receive our messages and information. The founding principle of the United States was the independence of such channels from government interference. We saw how the current administration returned the parent companies of the media giants such as CBS in apparently compromise their Journalistic rules favor corporate goals. By expanding this “anti-impression” program to AI models, it is not unreasonable to expect similar accommodation. Senator Edward Markey has Written directly Alphabet, Anthropic, OpenAI, Microsoft and the finish line calling them to fight order. “Details and plan to implement this executive ordinance remain unclear,” he writes, “but this will create significant financial incentives for large technology companies … To make sure that their chatbots AI do not issue speech that would upset Trump’s administration.” In a statement for me he said: “Republicans want to use the government of the government to make chatgpt sound like Fox & Friends. “
As you can suspect, this view is not made available by the White House team working on the AI plan. They believe that their goal is true neutrality and that taxpayers should not pay for AI models that do not reflect objective truth. Indeed, the plan itself indicates a finger in China as an example of what happens when the truth is manipulated. It will instruct the government to examine border models from the People’s Republic of China in order to determine “adapting to Chinese talks and censorship of the communist party.” Unless the corporate ruler of AI does not receive the spine, the evaluation came American Frontier models can well reveal the lock alignment to conversations and censorship points in the White House. But you may not find out by asking about the AI model. He woke up too.
This is the edition Steven Levy’s Backchannel Newsletter. Read the previous reports of Steven Leva HERE.