Election Lies and Dodgy Chatbots

Chatbots — A Dodgy Stochastic Parrot

To continue from last time, I’ll pivot from lying Chatbots to dodgy Chatbots. To be specific, I am still keying off of Garance’s fine article about public general-use Chatbots that are emitting falsehoods that deceive and harm voters. So, first let me be specific about lying Chatbots vs. dodgy Chatbots.

Of course, no Chatbot (or other piece of software) can actually lie, because there is no consciousness and no intent to deceive. “Emit falsehoods” is an accurate, but bland description of some Chatbot misbehavior, so I use “tell lies” as a little rhetorical license. The same is true when I say Chatbots are born fabulists as a shorthand for a human-built tool that is a stochastic fabulist. Either way, the true origin of whatever deception is happening, is the humans who built the bots, and regret the lies and fabulism but don’t seem to actually care enough to stop it.

Those Dodgy Chatbots

By contrast, a dodgy Chatbot is one in which the humans who made the tool actually do care enough about some inaccuracies and falsehoods, so much so that they put a muzzle on the bot, trying to ensure that the bot dodges questions on specific hot topics, in the attempt to prevent harm. In the case of elections, that harm is what I call DIY disinfo, where people ask general purpose chatbots for accurate information about elections, and sometimes get junk answers that can prevent the voter from trusting the vote, or worse, not voting at all — and information which the voter might pass along to others, performing DIY misinformation as well.

To explain more, I’ll start with another headline improvement. This time it’s: “Anthropic Takes Steps to Prevent Election MisInformation,” which misleadingly sounds like Anthropic is part of a team that’s out there in the battle against election mis/dis/info. Not exactly! I’d say “Anthropic Gags Its Election Liar Chatbot” except that (though more accurate) it is far too harsh for a Gen-AI company that is trying to prevent its Chatbot from harming voters.

What really happened is this: Anthropic, like all the big Chatbot companies, publicly acknowledges that its Chatbot Claude is prone to hallucination, inaccuracies, and emitting what I regard as outright lies in a few cases. And because of that, they have a thing called “Prompt Shield” which (I suppose) started as a pretty simple keyword filter for user prompts, where words like “bomb” or “suicide” in user input would trigger the good old “I am not programmed to respond in that area.” Nowadays, Prompt Shield might be a bit more sophisticated, and in the case of elections it is a bit more helpful — redirecting users to an election information website with information that’s too hot for Claude to handle.

So … 10 out of 10 for muzzling the lying Chatbot, but minus 💯 for letting folks be misled by it in the first place. Yet, given the sad fact of these highly useful services, their operators follow the fire-aim-ready model:

1. Put a dangerous tool in the hands of large numbers of people,

2. Wait for serious harm to occur, and then

3. Put in place some reactive hack that might prevent that specific harm, but doesn’t make the tool any less effective.

And it is not just Anthropic. Chatbots dodging hot topics is now a trend, with Google’s Gemini being muzzled in a way that’s similar to Anthropic’s Claude and Prompt Shield

Simple Answers

Having added “dodgy Chatbots” to my previous reflections on Chatbots and DIY election disinformation/misinformation, there are some simple answers to the questions posed by Garance’s article:

1. Why is this happening? These tools are darn handy! Publicly available more-or-less free tools will be used by ordinary people for nearly anything. With the AI company’s fire-aim-ready model, there will always be a host of unintended consequences. Public harm to voters and elections is merely one example.

2. Why is it predictable? Chatbot’s election lies, like all the baked-in fabulism, are a basic part of the technology and the tools that these AI Titans have made public; such public access is in their own self interest, and they are clearly and publicly saying that these tools will and do make stuff up. Public harm is an inevitable part of public access.

3. What do the AI titans say? That they are working on improvements, but their public tools will always be capable of misbehavior. Of course, we can see that some of the “misbehaviors” are massive screw-ups, and the vendors respond to the worst by continuing to extend the list of verboten topics where the dodgy bots claim to know nothing. 

4. Is anybody trying to do anything to reduce the harm? Yes, the vendors themselves, with their muzzles that cause their bots to dodge questions on hot topics. But that’s a simple hack. The underlying tech is more complex, and is getting more complex as vendors try to improve safety with more complex countermeasures. But that takes time, and isn’t guaranteed to prevent (for example) election lies. So in the meantime, the quickest response is the blunt hammer.

5. What would be better? There isn’t an obvious path to “better” for every kind of Gen-AI use, but I see a critical differentiator that’s only now starting to be broadly understood: the difference between general use and specific use. Gen-AI tech is enormously flexible and pretty powerful, getting more so all the time. Being good to great at a wide range of tasks has the price of fabulism, and it probably always will. 

By contrast, we see a lot of these harmful uses in a specific context: a person looking for specific information about a specific topic. They’re treating a chatbot as a more convenient search engine, seeking authoritative information about a critical topic (health, finance, democracy, et al) that should have zero tolerance for lies. Of course it is temptingly convenient to do so, but why is it a problem for critical topics? The LLMs have a whole lot of information about a whole lot of things — but not all true, much less authoritative. Often, the results people get are factoids, but unlike search results of yore …

• They don’t get the source material;

• They don’t get to independently evaluate the source;

• They don’t get to see the full range of information; and

• They don’t get to decide on relevance and veracity.

Let’s Face it: Gen-AI, LLMs, and Chatbots are a lousy way to search for relevant and accurate information about a specific and important topic where there is little margin for error.

What Would be Better, Actually?

What would be better? An AI-based interactive agent that:

• Is built specifically to help people with answering questions about a specific important topic;

• Provides only information that’s derived from authoritative sources for that topic;

• Augments its response with citations to the sources;

• Politely declines to answer questions that are outside of the topic area;

• Never responds with with potential nonsense from an LLM; and

• Is operated by people with expertise in the topic area and can keep up-to-date the “knowledge base” of information from authoritative sources, the sole base for the agent’s responses.

Sound fanciful? Next time I will say more about such an agent, but…

<spoiler alert> WE ARE ALREADY WORKING ON IT.</spoiler alert>