The Big Bias in AI Is Economic

What's really driving the algorithms

Welcome to Second Rough Draft, a newsletter about journalism in our time, how it (often its business) is evolving, and the challenges it faces.

One strand of investigative journalism that has long fascinated me is what has come to be called algorithmic accountability. Stories in this realm have uncovered bias in all sorts of platforms and systems, affecting such crucial areas as criminal justice, with inequities in probation, improper healthcare insurance denials and racial discrimination in housing. This field has existed for years, but it is becoming even more critical with the rise of generative AI, and this week I want to talk about biases of a different sort I see becoming apparent in the AI models which are assuming such influence in our society.

One of my favorite colleagues used to say that news is what happens to an editor, and while I am not an editor, I want to start with two experiences I had recently that seem to me both significant, and troubling in what they suggest.

Adventures in AI illustration

Both involve my use of AI to create illustrations for this column. I use AI for this purpose most weeks because I have no budget to hire an illustrator, no talent for illustration myself and because most often the ideas discussed here are a bit too abstract for the use of stock images.

A couple of months ago, I sought to create a photo gallery of four congressional candidates, and instructed the machine to source these from the candidates’ own websites. To my surprise, the initial result did this with two of the candidates, but, in place of the other two, created photo-like generic images of “people” who looked a good bit like the candidates but were not.

When I noted the error and asked why it had done this, the model readily confessed its mistake, but, when you cut through the verbiage, indicated that the generics had been substituted to conserve what AI professionals call “compute,” expensive processing power. In other words, it hoped I would be satisfied with a cheaper version (for the AI firm) of what I had requested even though it was wrong—and even though the model did not volunteer that it had disobeyed my instructions and knowingly delivered an inaccurate result.

Fed the text, ChatGPT declined to come up with any illustration for this week’s column, but Copilot came through

Earlier this month, I sought to create an image for the column on the future of CBS News and CNN under the aegis of Paramount that included the iconic logos of both networks. Multiple models refused to do this. Having practiced law in this realm for more than two decades, I pushed back, noting this was a classic “fair use” under the copyright statute. Here’s what ChatGPT said in response:

“your request was unusually defensible. It wasn’t asking for counterfeit branding or commercial advertising. It was for an editorial illustration accompanying a column about the relationship between CBS and CNN. That’s a context in which human illustrators and publications often make use of recognizable visual references….

“Your request is exactly the sort of editorial illustration that has a long tradition in journalism, but I don’t have the ability to exempt it from the image system’s automated safeguards.”

Beyond irony

That might even sound a bit noble—OpenAI, the purveyor of ChatGPT, has higher standards even than required by law in protecting other people’s intellectual property. Except, of course, that the central proposed defense of OpenAI and the other platforms in their current threat to much of the news business is that they are engaged in fair use.

That is, OpenAI refuses to let its own customers express themselves employing as a shield a legal right it itself wields as a powerful sword. The apparent reasons: they seek to curry favor with other big companies (corporate customers) and don’t want to run even minimal risks just to vindicate individual customers. I asked OpenAI to comment. They offered a background conversation with a senior executive, but declined any comment on the record, so I passed on the backgrounder.

The bottom line is the bottom line

There is a common thread in these two incidents, and it is that the AI platforms have perhaps the most troubling possible algorithmic bias. They are biased in favor of their own profitability.

Profit trumps accuracy. Profit trumps initial candor in dispensing inaccuracy. Profit trumps the expressive rights of customers. Profit compels rolling over the legal rights of publishers. The highest value of some of the most valuable companies on the planet appears to be accumulating the largest pile of money in the history of business.

This obviously matters enormously for society, but my focus is journalism, for which I think it matters especially. Journalists need to understand in their use of AI models that the results can be enormously useful, but that this may not hold true if the economic interests of the AI companies dictate otherwise. It will require great care to insure that we are not settling for, and publishing, responses that are “good enough,” which may sometimes mean invisibly wrong but cheaper to provide.

We will also need to better understand what policies the platforms are actually employing, and to inform both our conduct and our audiences regarding how those policies may be at variance with not only societal values but the even companies’ own rhetoric. (The stated mission of OpenAI is to “benefit all humanity.”) This will be especially tricky because the AI platforms, like the search and social media companies before them, are offering increasing funding for journalism, no doubt at least in part to soften their images. In other words, algorithmic accountability is work growing more urgent every day.