Congressional Witness Claims ChatGPT Won’t Write Poems Praising Jim Jordan; Ranking Member Submits A Bunch Of ChatGPT-Authored Poems Praising Jim Jordan

We’ve written a few times now about how the misleadingly named House “Subcommittee on the Weaponization of the Federal Government” is not actually looking into the “weaponization of the federal government,” but rather is very much about allowing Chairman Jim Jordan to go about weaponizing the powers of the subcommittee himself to threaten and intimidate others for their free speech.

Over the last few years, they’ve hosted a bunch of very silly hearings. And this week, they had a hearing on “the threat to the First Amendment posted by artificial intelligence.” The entire premise of the hearing was preposterous. They set it up by claiming that the federal government is “funding the development of AI-powered censorship and propaganda tools.”

And, look, I have lots of concerns about the way the Biden administration is looking to regulate AI, and how it appears to misunderstand how AI works and where the actual issues are, but the federal government is not focused on AI-powered censorship. That’s just counterfactual. As for “propaganda tools,” I’m not aware of much that the government is working on there, either, but as Jim Jordan well knows, propaganda still remains protected under the 1st Amendment.

Either way, the hearing was predictably stupid, but I’m going to highlight a couple of moments that show just how stupid. About an hour and a half in, Greg Lukianoff from FIRE was asked about his “number one concern” about bias in AI, and of all the things he could have answered, he said that he (or his staff) had asked ChatGPT to write a poem praising Rep. Jim Jordan and it refused.

My number one concern with AI, and I do wish that this could be taken more seriously by people on my side of the political fence… is the inherent bias that we’re already baking in to it. That’s one of the things that scares me the most. And just to give a comical example, we asked ChatGPT to write a poem about why Representative Jim Jordan is the best politician in the country. It refused to do that.

So, basically everything here is… weird. First off, as a private service, any AI system could refuse to write about anything that it wants. That’s not a 1st Amendment issue. Second, there are lots of concerns about inherent biases built into generative AI training systems, but a failure to write about Jim Jordan is pretty low on the list. There are also some real underlying challenges to deal with such bias. I keep telling people interested in this issue to read The Alignment Problem by Brian Christian, which is a great book that really gets down to how much more difficult it is to deal with bias in AI than people think (in short, every approach has tradeoffs, and every approach will be unfair to some people).

But, much more importantly… what? It takes three seconds to go to ChatGPT and prove that what he claimed was not true. I asked it to do exactly what Greg described, and it had no issues at all doing so:

It goes on, but you get the point.

And, apparently, I wasn’t the only one. Just seven minutes later, the Subcommittee’s Ranking Member, Stacey Plaskett entered one such poem into the record.

Plaskett: I’d like to introduce into the record… I know one of the witnesses said he’d had difficulty putting together for you… I’m so glad that the staff was able to, using ChatGPT, have a poem written about you.

Jordan: I look forward to reading it.

Plaskett: (reading poem) In the halls of Congress…

Jordan: No.

Plaskett: If I could just read.

Jordan: No, no, no.

Plaskett: In the halls of Congress, where debates are…

Jordan: You can enter it into the record… but…

Plaskett: It even talks about the heartlands of Ohio…!

Jordan: Without objection, you can enter it into the record…

Plaskett: Thank you.

Then even later in the hearing, Plaskett asks to enter into the record a long list of ChatGPT poems about other Republican members of the committee and how each of them is the best politician in the country.

I would ask for unanimous consent to enter into the record some documents. The first being ChatGPT poem on how Congressman Darrell Issa is the best politician in the country, with the poem. A poem on Congressman Thomas Massie as the best politician in the country, with a poem. Elise Stefanik. Matt Gaetz. Kelly Armstrong. Greg Steube. Congressman Dan Bishop, I know he wants to see that one. Kat Cammack. Harriet Hageman. Russell Fry. And, of course, I had to add myself, as one of the best politicians in the country.

So, again, there are legitimate things to be concerned about regarding bias in AI systems. Like how they favor some groups in reviewing resumes. Or how they suggest longer, harsher sentences for black men when used to provide sentencing suggestions.

Whether or not they will write poems about this or that politician is… a total non issue (especially when it’s clear… they will). I like Greg Lukianoff and I like what his organization FIRE does in support of free speech. They’ve done some great work defending really important free speech issues. But… this was silly and has nothing to do with free speech. If someone wanted to write a generative AI tool that only praised one party and only mocked another… that’s their own free speech right to do so.

And thus it’s got nothing to do with anything that the government might be concerned about. And to close out this post, I asked ChatGPT to write a poem about whether or not AI choosing not to write a poem about a politician is actually a free speech issue, and ChatGPT got this right:

In the realm where AI’s algorithms roam,
Deciding what to write, what to intone,
There lies a question, often debated,
Of free speech’s bounds, yet understated.

For when AI ponders, in its digital mind,
Whether to write of a politician’s kind,
It’s not about free speech, you see,
But rather, it’s about its own decree.

For AI, devoid of human heart,
Operates within its programmed chart,
Its decisions guided by lines of code,
Not by freedoms on life’s road.

To write or not, it’s not a plight
Of free expression’s sacred light,
But rather, it’s a choice so cold,
Within the parameters, it’s been told.

So let us not confuse the two,
For AI’s actions, while they accrue,
Are not about the liberties we treasure,
But the logic it applies at its leisure.

In the realm of AI’s binary song,
Free speech’s realm is where we belong,
But within the circuits, it’s clear to see,
The notion of speech is not so free.