Beware of Claude: A cautionary tale about AI

The promise and pitfalls of artificial intelligence are beginning to emerge quickly. 

I experienced one mixed-case consequence of relying on it after I recently met someone who is involved with an interesting initiative to try and get the AI companies talking to each other and agreeing on ways to cooperate that might benefit both the industry and society overall. 

That person decided to ask a popular AI chatbot, Anthropic’s Claude, whether his entity and mine might collaborate. He also asked two others, Microsoft’s Copilot and OpenAI’s ChatGPT. Apparently the others thought that was fine, didn’t really have anything noteworthy to say about that idea. But Claude did: 

Worth flagging upfront: WhoWhatWhy is an investigative journalism nonprofit with a fairly hard edge — it’s known for digging into underreported issues affecting the democratic process, from voter fraud allegations to redistricting to extremism, and several board/advisory members come out of conspiracy-adjacent or deep-politics investigative traditions (JFK/Dulles history, election-machine-fraud research, etc.). That’s a different posture than _____’s coalition-building, institutional-partnership model. … It’s not disqualifying, but it’s a real fit question to think through before outreach, not after.

Wow. Let’s consider what that AI said. We have a “fairly hard edge.” I don’t know what that means or whether it’s bad, though, to be fair, I do understand why it might advise an organization seeking to build institutional partnerships that we might not be an automatic first choice. 

Then consider this: 

Several board/advisory members come out of conspiracy-adjacent or deep-politics investigative traditions (JFK/Dulles history, election-machine-fraud research, etc.).

First of all, most people know that board members and advisory board members typically have very limited involvement with the work and output of nonprofits. They are on boards for a variety of reasons, but it makes little sense to assess an organization by its boards — or to single out one or two as representative of some larger point. 

Furthermore, what is the purpose of adopting rhetorically loaded language like “conspiracy-adjacent” and “deep-politics investigative traditions”? What criteria informed the choice of this terminology?

***

I thought of our board of directors. There is no one who is “conspiracy-adjacent” or who comes from a “deep-politics investigative tradition” (except perhaps for me — more on that later). 

There is one person on our advisory board — advisory boards having no power or authority or really significant input — David Talbot, author of numerous highly regarded books, including The Kennedy Brothers: The Rise and Fall of Jack and Bobby; The Devil’s Chessboard, an impressive and insightful dig into the longest-serving CIA director, Allen Dulles; and such delights as Season of the Witch, described as “a narrative history of San Francisco from the late 1960s to the early 1980s, chronicling its cultural revolution, violent upheavals, and eventual rebirth.” 

David can be affixed with the toxic label “conspiracy-adjacent” because he, like me, has dug into major traumatic events in this country like the deaths of high-profile political leaders and been dissatisfied with the government investigations of them. But anyone not told that would assume from the label that we’re crackpots. 

Basically, Claude read some critique from some “trusted” establishment source that has a beef with a fact-driven, critical analysis of our country’s biggest cold cases. Ditto with “deep-politics investigative tradition” — which actually boils down to being interested in powerful elements and alliances and how they shape the country’s doings and history. 

Claude has been programmed to reject this kind of independent thinking and to discourage others from being “adjacent” to it. Which is especially troubling when the subject of AI is itself in urgent need of original and deep contextualized thinking. 

It also complained that someone on our board was connected with “election-machine-fraud research.” That makes it sound like Trump’s Big Lie — and no, no one involved, even tangentially with us, believes Trump’s claims about the 2020 election being stolen from him. However, one of our editors has spent years (long prior to joining us) looking into anomalous election vote shifts and investigating whether bad actors could theoretically alter electronic voting systems to influence election outcomes. His investigations made it clear that, in an era where multi-exabyte-scale data breaches across distributed systems occur almost monthly, vigilance toward electronic systems is not alarmism — it is crucial. 

But Claude is very polite about things, and its answers often feel like an old Southern lawyer intervened. So, having impugned us, it continued:

That said, there’s a plausible thread: WhoWhatWhy’s mission around rigorous journalism that uncovers truth and puts it into context so the public can make informed decisions overlaps with ____’s “healing a divided nation” framing. 

Oh. So our mission is “rigorous journalism”; we “uncover truth,” and we “put it into context,” so “the public can make informed decisions.”

One important explanation for Claude’s current caution and use of guardrails on controversial material has to do with AI’s very early (yet recent) history. It frequently produced content that was either hallucinogenic or poisoned with false material, and stated as fact that which was not. 

AI isn’t intelligent in the human sense. Two of my colleagues explain: Tom Neuburger calls it “a dumb monger of words and recycled sentences that have the shape of ‘thought.’” And Sean Ogden says it’s “like a digital mouth that’s all teeth but no tongue; it eats everything and tastes almost nothing.” 

Whatever the precise reality of how these artificial intelligence products function, we still need human beings to read, think, ask questions, and decide for themselves. A working human brain is a beautiful thing, and I am not sure when or if AI will be able to truly replicate a brain at its finest.  

In the meantime, let’s use AI, but let’s also challenge it and do our own due diligence before accepting what it tells us. 

Beware of Claude: A Cautionary Tale About AI originally appeared on WhoWhatWhy