That's a nice Think Tank you have there...
...be a shame if someone were to automate it.
I’m sure many think tankers are steadfastly resisting AI prose. But one blue chip institution has recently been called out for publishing bad AI writing. And I know of one very high-profile pundit who uses agents to produce their high-demand assessments. I won’t embarrass anyone, or their lawyers, by naming and shaming. There will be much, much more of this stuff. It’s not just think tanks: AI content is sweeping through intellectual life - including journalism, consulting and even, whisper it, academia. Talk is cheap, and it’s getting ever cheaper.
Well, surely we can do better than just whacking a prompt into ChatGPT and filling out an expenses invoice. We’ve just entered the era of agentic AI, where the bots are capable of heading out into the world and doing their own thing. So, I thought I’d create a virtual think tank staffed by said bots. If you’ve got a policy problem, if no one else can help you, and if you can find us1, then boy do we have a policy paper for you.
Welcome to Project Kennan, my pipeline for producing AI generated Think Tank reports.
Here’s how it works. You give me your desired ideological bent - choose your own, or for maximum convenience, pick from a drop down listing current institutions: CNAS, Brookings, whatever. Then ask your question. And that’s it. My team of bots gets to work, while you take an early lunch.
The pipeline looks like this:
A thing of beauty, hey?
It’s an ensemble approach, where we first create multiple agents and have them adopt suitable personas, complete with agendas and backstories. These then conduct their own research using a variety of internet tools, and sketch out some initial ideas. Then they come to the editorial meeting to debate these together, and draft a report. That initial report gets red teamed, and passed, along with the transcript of their discussion, to the editor in chief. Lastly, once they’re done, there’s a final AI-copyedit to excise any AI writing ‘tells’. No pesky em-dashes here.
Here’re a couple of examples to illustrate what’s possible. I asked the same question - should the US launch an amphibious assault on Iran? But got some very different takes by changing the profile of the think tank. I’ll give you a little flavour of the action, rather than the full tick-tock, which is extensive.
The fun bit for me is seeing the bots debate. Here’s the scene in the HeritAIge Institute’s editorial meeting:
It goes on for quite a while, and can get quite punchy. Eventually, the chair steps in to sum things up and gets down to drafting
Over at the AI-EI hawkish conservatism is in the air, as usual. But even they thought a land assault into Iran was nuts: Here’s a snippet from their report:
Down the street at the CAP (the Center for AImerican Progress, of course) the focus was on the impact on the American economy, and the need to engage Congress, the EU and the UN:
And so it goes. Next, I’m building in an image generator to produce photo realistic pics of my AI think tankers. I’m sure you’ll see our hot takes soon, quoted in The New York Times or The Economist.
We’re in the Los Angeles underground.