Latest news

Artificial Intelligence & Artificial Trust

No comments

By Mark Stephen Meadows <[email protected]>

Artificial Intelligence systems now understand and influence people. This requires that as builders of these systems we consider the values of the user — and their trust.

I’m going to tell you about Andi. She’s a conversational avatar, and in this picture, she’s talking with our friend Michael of Skype using their Bot Platform. With Microsoft Cognitive Services, Andi can write messages, hear you and talk to you, and see you using Skype video chat. You can talk with her just as you would talk with a person. This photo was taken at Microsoft’s Build conference in May 2017, where Andi was demonstrated publically for the first time. (See this VentureBeat article for press coverage.)

Photo: Khari Johnson @kharijohnson

Andi, unlike most conversational interfaces, has a human visual appearance.

But why?

Because, Neolithic Brains

The hardware of the human brain has hardly changed in the last 50,000 years. This legacy hardware exposes you to all kinds of security vulnerabilities.
When you see a familiar face on a screen, (whether it’s movie star, or a friend on Skype) your brain does the same thing a Mesopotamian farmer’s brain did 5,000 years ago when he saw his friend, a shepherd, walking across the village road. EEG and imaging studies have shown that multiple cortical regions fire up and coordinate with other parts of the brain that support cognition to handle these social interactions. The salience network summons help from the parietal cortex, frontal eye fields, visual cortices, and the anterior insula, which allows the brain to refer to past behavior, react to current sensory input, and predict the actions of the other person.
In other words, the brain processes all these inputs and makes a decision about how to react in any situation. In the case of the farmer and the shepherd the brain says,“OK! This is someone we know! Now what?

The next step takes slightly longer — most brains clock in at around 200–300ms — but the brain then makes a decision. In the brain of the farmer it says, “Friend! Let’s do friendly thing!

Then some other neurons fire that wave an arm and the mouth makes a friendly noise in English (or maybe ancient Sumerian) along the lines of, “Hail, Shepherd!” And then the shepherd replies, and he waves arm and smiles and says, “Hail, Farmer!”

After they recognize each other and exchange greetings, they probably spend a little time talking about the weather or hay or something they share in common. Maybe they trade some money or grain and write it down on a ledger. The point is that these two Sumerians establish trust with these multiple modes of interaction. By shouting words and waving their arms then talking and trading things, they build the trust necessary to socially interoperate.

The three modes look like this:

Every time they see one another they do this. They establish trust via this common human multimodal communication.

Multiple Modalities

Cognitive Services does what our limbic system does — but in a different way.

Computers liquidate pre-existing communication modes. Things like waving and saying, “Hail, Friend!” used to be a coordinated multi-modal analogue signal. The words, sounds and image all happened at once. Digital technology can now filter those signals, separating them into discrete modes. Computers separate Natural Language into words as separate from voice. Computers distill sounds, phrases, gestures, and other stuff, channeling each mode into different buckets for processing, filtering, and analysis. AI is great for this stuff.

We can now use things like voice recognition, natural language processing, computer vision, affect detection, synthesized voice, and 3D-modeled character animation to filter input and output channels for computer-human interaction (CHI). Cognitive Services does what our limbic system does — but in a different way.

The liquidation process requires input and outputs. Mesopotamian farmers and shepherds had inputs and outputs, just like you have inputs and outputs when you chat with your friends or coworkers. The point is that now AI can do the same thing to those inputs and outputs that the human brain has been doing for thousands of years. Let’s take our chart from above and apply it to digital communication on something like Skype.

Separating modalities into input and output allow a digital character to talk with a person. Andi and Michael talk in exactly the way two people talk: they listen, respond, and even gesture.

These modes make a taxonomy of AI bots. Here’s the difference between a chatbot, an interactive virtual assistant, and a conversational avatar:

A “chatbot” uses text, an “assistant” uses audio, and an “avatar” adds a visual appearance. Basically, an assistant, like Siri or Alexa, is a chatbot with a voice user interface (or VUI). If you add a face, or a visual interface, then you get an avatar. Remember how important the face is to the limbic system and the brain? With all three modes of communication you get something that’s far more likely to create trust.

Trust the Multi-Modal AI?

Skype’s three modes makes it a perfect platform for advanced bots. Skype allows you to type and read words, speak and hear a voice, and see a face while you show your own in the video channel. Since many people do normal, human things with it (like video conferences for work, family visits, etc.) researchers have also been using it to understand multi-modal social interaction and trust. Some of these projects used Skype to distill those modes of communication, isolating one mode’s emotional connection. This study by researchers at Cardiff Metropolitan University published in Sociological Research Online found that,“Emotional connection […] will be highest in-person, followed by video chat, audio chat, and, finally, IM.”

Bruce Schneier, security guru, in his book Liars and Outliers quotes sociologist Piotr Sztompka when defining trust: “Trust is a bet about the future contingent actions of others.” Schneier unpacks trust in two ways. When you trust a person, you believe that their underlying motivations and intentions will cause them to do the right thing, regardless of the situation. You can also trust that someone will act socially and morally appropriate, without knowing their motivations or intentions.

When meeting someone unknown we exchange conscious and unconscious social cues. The right combination of social cues, symbols, and signs, instills trust in the participants of the exchange. It’s also been shown that you’re more likely to trust someone that looks like you. These modes of interaction help both parties predict and understand the other person’s thoughts and actions.
Without rich, multi-model communication, you don’t have empathy, and without empathy you don’t have trust, and without trust, society breaks down. Digital media melts communication into new modes of communication and this rubs on our Neolithic brain. You don’t like getting a message from someone you don’t know asking you to do something because no part of that interaction is able to generate trust in your brain. Whether it is for a brand, an individual, a company or a set of valuable data, the user’s trust must be established so that the value of the system may be accessed and used. Trust is what allows that interaction.

Once it is established it must be maintained.

Trust is a must

Trust is a two-way relationship. When we place artificial intelligence in our cars, in our bodies and in our houses — when we use these systems to navigate and make decisions about our health, finances, relationships, and future — they must be trustworthy.

People are more likely to confide in an AI than a person (USC). People are more likely to also obey an AI (Kelly). Many people would prefer to have a robot for a boss. 85% of users are more likely to follow directions from an AI than a person — especially when it comes to health care. (Bickmore &co). All of these findings make it clear that people are willing to trust artificial intelligence interfaces — but how do people know that the bots they’re talking to are trustworthy?

We humans are more willing to confide in, trust, and obey AI than other humans. Evidently this is because we see software like artificial intelligence as objective, like math is objective. But that’s not right. AI systems are built by people. People have all of these prejudices and preconceptions and weird ways of being individuals. This is why Artificial Intelligence presents more cultural problems than technical ones.

The power over the end-user is immense. AI changes what people do. That’s why we need to make sure these systems can both establish and maintain trust. As we build AI systems that can establish trust we need to simultaneously build systems that can maintain trust. This is why Botanic Technologies states they are “Building Humane Machines.” They make systems that establish and maintain trust — this is the kind of machine that should be built.

The opposite, machines that cannot be trusted, should certainly not be built. Information systems should not subvert the wills and interests of the end-user. Human agency and independence must be respected. Because these systems are telling us what to do with our bodies and money this is a moral issue.


The most famous and trusted assistant of all time blew it. HAL 9000, of 2001 A Space Odyssey, possessed knowledge, a sense of wisdom, humor, curiosity, and some kind of space ship captain’s license. Because of that the human crew members of the Spaceship Discovery trusted HAL with their lives. That relationship between HAL 9000 and the crew of the Spaceship Discovery was predictable, comfortable and trusting up until that chess game HAL 9000 plays against Frank Poole (the chess match is a famous one, by the way, Roesch vs. Schlage, Hamburg 1910).

2001 is a classic tale about the ancient, connective tissue that holds society together: trust. On the banks of the Tigris River, our Mesopotamian neighbors, the farmer and the shepherd, needed to take the time to exchange symbols of shared value to establish trust. They also needed time to maintain it.

So, how do AI systems establish and maintain trust? We looked at some ways to establish it with multi-model interaction, but what about maintaining trust?


Trust has a lot to do with expectation. Bruce Schneier writes in Liars and Outliers that “We’re reducing trust to consistency or predictability. Of course, someone who is consistent isn’t necessarily trustworthy. If someone is a habitual thief, I don’t trust him. I but I do believe (and, in another sense of the word, trust) that he will try to steal from me.” You can even trust a habitual thief — you just trust that they will steal from you. Trust is a relationship with the unknown. In other words, it’s a matter of expectation and predictability.
Trust is largely founded in repetition; without repetition and predictability it’s almost impossible to build trust, especially when interacting with a system. Consider an app that crashes from time to time, without you really knowing why. That lack of repetition whittles away at your trust — it’s an unfulfilled expectation.

You can’t judge a book by its cover, but the cover does give some sense of what that book will be talking about.

Technical Trust (Authentication, Encryption, Blockchain & Feedback)

That’s why software must be tested, robust, and reliable. We expect the same repetition from a personality — whether that personality belongs to a person or a bot. If you can expect what someone will do you will trust them more.


Bots should always be authenticated. I’ve written about this before as a license plate for bots, just like a license plate for car. An authenticated identity would connect a bot with who’s responsible for it. Authentication is even more necessary in sensitive cases like healthcare, finances, navigation, governance, where sensitive data is exchanged. Authenticating a bot on Facebook should include the same steps as authenticating any other account (just as you are authenticated on Facebook).

Something that is authentic is going to be easier to trust. Until Facebook authenticates their bots I don’t recommend trusting them with anything remotely close to your personal data.


The communication pipeline should always be secure. We are currently building systems that use SHA-256 to encrypt the communication from the sender, decrypt it on the bot’s side, then when the bot replies, re-encrypt it so that the reply is properly packaged, signed, sealed, and delivered. It goes without saying that in a situation where sensitive data is exchanged you need to trust not only the bot itself, but the system you use to interact with it. Strong encryption is essential.


Blockchain is a bigger deal than the World Wide Web because it is a fundamental rethinking of value exchange. Not only does it decentralize value exchange — it also makes it faster, cheaper, and integrates upgrades like timestamping, peer-to-peer networks, and transparency. So, it’s great for making trustless systems. That is, systems that don’t require you to trust any specific person or organization, because the entire economy is completely transparent and decentralized.

Blockchain may even be a rethinking of computation itself. The weird thing about blockchain — and part of what makes it trustworthy — is that once something has been transcribed on the public ledger it cannot be removed. Word processors introduced the “undo” feature that gave them an unfair advantage over typewriters. Writers didn’t have to get the whiteout or change the entire page when they made a mistake. They could simply hit “undo” and the typo went away. Blockchain isn’t like that. Blockchain is more like an odometer — you can’t roll it back. This means that blockchain basically removes the need for trust in interactions. Just like you don’t need to trust someone who tells you how many miles are on a used car — you just look at the odometer.

This means that AI, when tied in to blockchain, makes interactions around healthcare, finances, navigation, and many other use cases far more trusted. Because anyone and everyone can see the blockchain, everyone knows what was agreed to, everyone recognizes what transpired.

Botanic Technologies has implemented blockchain features in the past, integrating a voice-to-blockchain bridge so that conversational avatars can do things like close a sales cycle on a blockchain.


When we speak to one another we give feedback on the symbols we share. Bots can do this to. Here’s an example of how it is visualized:

Example evaluation report that shows facial and linguistic analysis based on interview session.

More work will be done to make Andi better at reflecting what she can see. For now, the focus is on avoiding uncanny valleys. A bot needs a face to trigger the trust mechanisms of the human brain — but if you miss the mark that same face can fall into the uncanny valley and cause fear, discomfort, and alarm. Aim for a less realistic face and you can avoid the uncanny valley all together.

Preliminary drafts of Andi showing shading choices from toon to phong.

Let’s go back to our Mesopotamian neighbors. Every time they see one another, wave, say “Hi,” spend a little time talking about the weather or money or something they share in common they maintain that trust. Maybe they trade some grain for a sheep, and do it on a ledger they share with a third party. Maybe one of them forgets to do something and the other one forgives that omission. We haven’t figured out how bots will do all of that — but they can do most of it right now.

But I hope this that sums how artificial intelligence can create and maintain trust, just like the Mesopotamians did through repetition, authentication, and exchange of information. Bots need to do the same things humans need to do to earn your trust.

Bots need to be multi-modal. It’s not about them. It’s about us.

It’s how we’re wired.

— — — — — — — — — —

A note on Botanic: provides our customers with the tools to deploy trusted, multi-modal bots. We are a group of poets, writers, artists, and engineers. We’re deploying automated social interaction devices and we just plunge ahead and try to figure out what works based on usability studies and lots of time watching people talk with bots.

Many of us have been building chatbots since 1999. With many thanks to many people solving problems, from machine learning to fundamental meaning, we’re seeing some really great advances in both AI and VR (oh, and blockchain). Our tools allow developers to build trusted personalities. And when you release a personality out into the wild it tends to build relationships. To build a relationship you need to have the ability to use words, make noises, and do things like wave and say “Hi.”In other words, you need multiple modes of interaction.

That’s why Andi, on Skype, is important. She represents a new thread in the connective tissue of societies technological structure. She represents something that you can only get when you interact with a human or possibly with a humane machine. She represents Trust.

Artificial Intelligence & Artificial Trust was originally published in Chatbots Magazine on Medium, where people are continuing the conversation by highlighting and responding to this story.

Source: Chatbots Magazine

magnoliaArtificial Intelligence & Artificial Trust

Related Posts

Leave a Reply