In my last two posts, I wrote about the problem of falsehood-prone (a.k.a. lying) Chatbots giving out bad information on elections, and about dodgy chatbots that specifically refuse to say anything about elections and other specific hot topics. I answered some who/what/why questions and ended with the question of “What would be better?” This time, I offer a more complete answer. 

First, let me introduce a phrase-term of art: domain-specific natural-language agent, which is the unwieldy name for what would actually be better. You can read lots about that in a previous 6-part series, or just start here: what I’ll call “Agent” for short in this post, is kind of like a Chatbot for exactly one topic, but with extra goodies for accuracy, interaction, and for trust and safety.

From General to Specific

There are two important preliminary steps to answering the question: “What would be better?”

The first step is to acknowledge a serious mismatch with using a general purpose and unsafe tool for a specific purpose.

Existing Gen-AI tools are great at a lot of different things, but not one of them is this: helping a person navigate a specific body of important existing information, similar in some ways to site-search or enterprise search (neither ever having been a raging success). People use Chatbots and other Gen-AI tools for this purpose, not because they are accurate, but because they are convenient, with a friendly (conversational) natural language interface and with answers/responses crafted specifically for each question/prompt. 

That can be a problem when the topic is important. Which is unfortunate, when many people don’t immediately recognize the mismatch between such a specific purpose, and several other general purpose uses for Chatbots in which yes, there may be falsehoods. These range widely from a general writer’s assistant (“Give me a 5 paragraph essay on why Gen-AI can’t avoid lying about elections”) to very, very different “co-pilots” for software development and other technical tasks. By contrast, a topic-specific Agent would be the tool for an ordinary person wanting answers to questions about Topic-X — answers that are not made up, definitely based on an authority on Topic-X, and don’t include AI generated lies, fables, or other LLM nonsense.

From Requirements to Engineering

The second step is to go from this concept of specificity, a key realization, to technical  requirements, specifications, and engineering. For every topic that general Chatbots are specifically dodging, there is a potential for an Agent specific to that topic. Of course, elections are our topic, and it’s no coincidence that elections is a topic of significance that Chatbots are failing at.

However, the techniques and the engineering should be applicable to any topic space. For any such Agent, a bounded set of authoritative information is required (which can be revised over time along with changes in the real world that it describes), and which — if “searchable” via natural language processed Q&A — would provide real value to real, ordinary people. 

We call that collection of information a “Knowledge Base,” which is compiled from authoritative primary sources, and is the sole basis for an Agent’s response to a question (or prompt) that is in-scope of the topic. This bears worth repeating: the only “knowledge” from which the Agent can draw on for responses is the bounded Knowledge Base or “KB”.

So, in the case of elections a KB would include:

• Election law; election case-law; legal scholarly works from accredited institutions; regulations, statutes; and operating policies, processes, protocols, procedures, practices, etc.

But here is what a KB would not include:

• Articles from news, magazines, or independent commentary platforms; any social media postings; blogs; emails; and non-official, non-government-sanctioned online sources of information.

Now, not every specific topic is matched in the real world by authoritative sources. However, for election and election administration information, voters, and other election stakeholders there certainly is authoritative sources (as I outlined above) — and as it happens after 17-years, the OSET Institute knows a lot about the topic of elections and elections administration, and we have in-house experts who can assemble an authoritative KB. Stepping back, I expect that the same is true in many commercial settings, and government civilian and military settings as well.

However, regardless of the setting and/or the specific KB, there are a common set of methods and technologies for creating an Agent that can rapidly search a KB (far faster than any human, and on any single topic) for the relevant information needed to craft a natural language answer to a natural language question.

Key Characteristics

OK, so then what would such an Agent be like? To extend the answer from last time, here are some characteristics for an AI-based interactive Agent built specifically to help people with answering questions about a specific important topic, as represented in a specific Knowledge Base (KB):

• Generate responses that provide only information that is in a KB that was derived from authoritative sources for the topic; 

• Augment each response with citations to the sources;

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

• Use a language base model for crafting responses, but never respond with potential nonsense from an LLM. 

And last but not least, such an Agent would have to be operated by a “Pit Crew” that includes people with expertise in the topic area who can keep the KB up-to-date in order to be a timely basis for the Agent’s responses.

We believe that the technical course to such a system — for elections certainly, but possibly also for a range of other topics — is feasible, because we’ve already started down the path.

If you want to learn more about that course of development, here are two options for you…

1. Scan through previous 10 or so postings here for a full account of the background and requirements; or

2. Stay tuned here as we continue, because we’re actually working on this. And send us a direct message (with text in the body: “Ella?”) and you’ll be the first to know when we’re about to announce something formally.

Of course, you can always leave a comment or ask a question here (below this posting). 🤓