Can Trump defeat ‘woke AI?'

Boston Globe

2 days ago

'Woke' is right-wing shorthand for a variety of liberal projects aimed at achieving racial and gender fairness, often using means that conservative voters reject, such as racial preferences in hiring and college admissions.
The Trump administration believes that these values have been embedded in the large language models (LLMs) that power many popular AI products, such as ChatGPT, leading them to produce information outputs that are slanted with liberal biases.
Advertisement
There's considerable evidence that this is true. Multiple studies by scholars at US and foreign universities have found that when asked political questions, the leading AI systems often favor more liberal perspectives on issues like abortion, climate change, or immigration.
Get Starting Point
A guide through the most important stories of the morning, delivered Monday through Friday.
Enter Email
Sign Up
In addition, there are high-profile examples of AIs generating false information in an apparent effort to reflect racial and ethnic diversity. Last year, a Google AI image generator depicted Black people when asked for images of Vikings and showed Black men and Asian women as World War II German soldiers.
Of course, there's also evidence that AI is sometimes biased against minorities, women, and gay people. But this isn't a high priority for the Trump administration.
Advertisement
Instead, it's mainly worried about AIs that are trained to promote diversity, equity, and inclusion, or DEI. Hence, its new executive order seeks to purge DEI from all artificial intelligence systems used by the federal government.
'President Trump is protecting Americans from biased AI outputs driven by ideologies like diversity, equity, and inclusion (DEI) at the cost of accuracy,' said a statement issued by the administration.
But Samir Jain, vice president of policy at the Center for Democracy and Technology, a tech-oriented political advocacy group, said the effort gets off to a bad start by mandating a ban on AI systems trained in DEI principles.
'The order itself is inherently contradictory,' said Jain, because eliminating DEI content from the training data will simply create a different form of bias.
For example, he said, suppose the federal Equal Employment Opportunity Commission, which enforces civil rights laws, relies on an AI chatbot for researching racial or gender discrimination cases. If the chatbot is purged of DEI-related content, it might miss relevant court cases or academic research. 'Then that tool is no longer as useful,' Jain said.
Massachusetts Democratic Senator Ed Markey went further, arguing that the Trump AI plan is unconstitutional. In a letter, Markey urged the heads of leading AI companies to resist the proposal.
'Republicans are using state power to pressure private companies to adopt certain political viewpoints, in this case by pressuring the Big Tech companies to ensure that responses from AI chatbots meet some unspecified, vague definition of ideological neutrality,' Markey said.
Andrew Hall, professor of political economy at
Advertisement
For instance, the order states that government workers using AI should be able to request ideologically-slanted information if they see fit. Thus, an AI would be barred from automatically flagging evidence of racism in government contracting. But a federal worker would still be free to ask the AI to seek out such evidence.
Still, purging all political bias from AI chatbots is probably impossible.
'Any model inherently reflects the priority viewpoints of the model builders,' said Jain. 'There's a real question whether there's anything you could call objective AI.'
Hall agrees that political biases can never be completely purged from AI chatbots. But he notes that not all biases are bad. A chatbot ought to be biased against Nazi ideology, or lynchings, for example.
The big challenge comes when dealing with less extreme controversies, where people of good will harbor major disagreements. How can an AI be trained to present a balanced point of view?
Hall offers a possible solution. In a recent research paper, he concludes that people are good at spotting left-wing bias in AI-generated information, regardless of their own political views.
'Americans view the bulk of LLM output on hot-button political issues to be left-slanted,' said Hall. 'Even Democrats say this, on net.'
His research also found that when people perceive an AI's output as unbiased, they are more inclined to trust it.
Hall says that this discovery opens the door to 'a thoughtful approach that puts the American public in charge.'
The leading AI bots could have their output regularly reviewed by panels of ordinary people, who'd grade the content for biases. Bot makers could tweak their output accordingly.
Advertisement
Whatever method might be used by AI vendors to comply with the executive order could be equally applied to commercial and consumer versions of their products. That could mean that in a few years all of us will be using AI systems that don't lean quite so far to the left.
Hiawatha Bray can be reached at