10 Defining Principles of Transparent AI

Robots should document and log the exact way that every sentence is understood.

Before introducing the 10 defining principles of Transparent AI, I would like to relate to the question as to why we need Transparent AI in the first place.

Several fields exist which are way too challenging for the present state of the art, but at the same time are way too simple to be dependent on (presently unachievable) Artificial General Intelligence.

Conversational Commerce is an obvious example of such a field which is neither trivial nor impossible. It does not need Artificial General Intelligence. I would like to argue that Transparent AI may just have all that is needed for Conversational Commerce.

The first defining principle of Transparent AI is the Principle of Full Disclosure: for any sentence or phrase that is being processed in a dialog, the full logical form is provided (not necessarily shown to the end user, but is made available and logged).

Therefore, the primary criteria of success is not simply *guessing* the correct answer, but demonstrating a full understanding of what is being said.

The robot documents how exactly every processed linguistic message, past or present, is understood.

The second defining principle of Transparent AI has to do with the fact that conversational agents always have to deal with other such agents. A multi-agent environment is the natural habitat of linguistic agents.

Each agent exchanges messages with other agents, and must have a “dossier” on each of its conversational partners. A “multi-agent” setting requirement goes to the very essence of natural language as a coordination tool.

The third defining principle of Transparent AI has to do with employing Presupposition Calculus. The Logical Form reported for each linguistic signal must state not only the denotation of a sentence, but also must include the set of relevant presuppositions.

Presuppositions are an integral part of the semantic interpretation of every sentence. If a contradiction is discovered with the assumed set of presuppositions, this set must be correspondingly accommodated.

The fourth defining principle of Transparent AI is the requirement that every transparent conversational agent must, in addition to its own field of specialization, be conversant in Second Order Predicate Calculus and elementary set theory.

In literature, combining symbolic reasoning with machine learning is occasionally discussed. No conversational agent can get away with failing to understand elementary logic — it is a basic component of the normal use of language.

The fifth defining principle of Transparent AI is Dialog Memory. Any previous text in the ongoing conversation must be instantly incorporated into the ontology of nameable items that could be referenced in the following messages.

The sixth defining principle of Transparent AI insists on utilizing up-to-date linguistic models. It is unacceptable to use archaic linguistic theories that today would never be used in a Ph.D. thesis in, say, MIT or Harvard.

The syntax and semantics must be reported in modern, scientific terms from an academic point of view. Consequently, Transparent AI agents must have access to a scientific linguistic parser — easier said than done.

In particular, the linguistic parser used must be able to recover “understood” material in ellipsis and other kinds of gaps.

The seventh defining principle of Transparent AI requires in particular that Logical Form supports bound variables. This is necessary in order to support quantifiers. As is well known, quantifiers play a very central role in the semantics of natural language.

The eighth defining principle of Transparent AI has to do with “meta-learning”. The parameters (such as rates, weights, biases, activations, and the like) are not off limits as a topic of discussion. We can ask a robot: “How did you come to this conclusion?”

The ninth defining principle of Transparent AI is about “terraforming”. A linguistic agent can be visualized as a robot moving within, interacting with, and ultimately contributing to the building of a Universal Knowledge Graph.

This Universal Knowledge Graph is basically the same as a global Semantic Graph, with one important difference — its items should be in a vector form, ready to be incorporated on demand into the agents’ state vector.

The tenth defining principle of Transparent AI is about using language as language as opposed to various auxialliary tasks (including translation).

For too long, the cutting edge research was able to claim success while simply performing only auxialliary tasks, without even attempting to use language as language, in the form of an ongoing dialog.