INTERACTION WITH AI MODULE 2 Session 2 Asbjørn Følstad, SINTEF
Interaction with AI module 2 Session 1 & 2: The user and interaction design perspective on interaction with AI Asbjørn Følstad Interaction design Four sessions Session 3 & 4: The machine learning perspective on interaction with AI Morten Goodwin 2
Themes 3&4 2 1 Interaction with AI - overview User-centred design of AI Chatbots interacting with AI i natural language Intelligent agents - AI system as partner Interaction with AI in social contexts Platforms and frameworks for interaction with AI Machine learning Explainable AI 3
Chatbots principles of conversational interaction design 4
5 Conversations as the object of design
First step towards a successful chatbot: A good use case Desirable Possible 6
Need to understand conversational processes Speech acts We speak to perform something Conversational implicature Cooperative principle and maximes of conversation 7
I'd love a cup of coffee Need to understand conversational processes Speech acts We speak to perform something Conversational implicature Cooperative principle and maximes of conversation I'll make you some The utterance (locutionary) The speech act (illocuationary) The effect of the speech act in the listener (perlocutionary) 8
Need to understand conversational processes Speech acts Utterances to perform something Conversational implicature Cooperative principle and maximes of conversation 9
Cream or sugar? Need to understand conversational processes Speech acts Utterances to perform something Conversational implicature Cooperative principle and maximes of conversation Black, please Make your contribution such as it is required, at the stage at which it occurs, by the accepted purpose or direction of the talk exchange in which you are engaged 10
Need to understand conversational processes Speech acts Utterances to perform something Conversational implicature Cooperative principle and maximes of conversation 11 Erica Hall (2018) Conversational Design
Need to understand conversational processes Speech acts Utterances to perform something Conversational implicature Cooperative principle and maximes of conversation Maxim of Quantity: Be as informative as required Maxim of Quality: Speak what you believe is the truth. Maxim of Relation: Be relevant Maxim of Manner: Be clear and unambiguios 12
Maxim of Quantity: Be as informative as required Maxim of Quality: Speak what you believe is the truth. Maxim of Relation: Be relevant Maxim of Manner: Be clear and unambiguios 13
Maxim of Quantity: Be as informative as required Maxim of Quality: Speak what you believe is the truth. Maxim of Relation: Be relevant Maxim of Manner: Be clear and unambiguios 14
Maxim of Quantity: Be as informative as required Maxim of Quality: Speak what you believe is the truth. Maxim of Relation: Be relevant Maxim of Manner: Be clear and unambiguios 15
Maxim of Quantity: Be as informative as required Maxim of Quality: Speak what you believe is the truth. Maxim of Relation: Be relevant Maxim of Manner: Be clear and unambiguios 16
A consistent persona for the conversational agent - Victorian servant? - Engaging partner? - Helpful customer service agent? - and many other options 17
A consistent persona for the conversational agent - Victorian servant? - Empathic friend? - Helpful customer service agent? - and many other options 18
A consistent persona for the conversational agent - Victorian servant? - Empathic friend? - Helpful customer service agent? - and many other options 19
20 The importance of a good self presentation
Conversation repair Prevent errors by expecting variations Design dialogue so as to identify and mitigate interpretational errors Provide helpful hints, prompts or alternative directions Be prepared to help at any time Fail gracefully 21 https://developers.google.com/actions/downloads/design-principles-quick-reference.pdf
User-centred design of AI tentative design principles 22
23 AI with incredible advances due to progress in machine learning
De fleste datasett som kan brukes til å trene dype nett til å bli gode på en eller annen funksjon, er kjempestore. [ ] Et nytt triks som mange bedrifter nå bruker, er å først designe en tjeneste som mange brukere vil ha. [ ] Gjennom bruken av tjenesten gir brukerne fra seg verdifulle data som igjen brukes til å trene nett for å gi brukerne nye funksjoner de setter pris på. 24 Bjørkeng, P. K. (2018). Kunstig intelligens den usynlige revolusjonen. Vega.
En datafelle er begrepet som brukes om Teslas særegne tilnærming til datainnsamling. Google-eier Alphabet og deres selskap Waymohar nå brukt ti år bare på å samle inn data om mange nok ulike trafikksituasjoner. (menneskelige codrivers i selvkjørende biler) Elon Musk og Tesla hadde slett ikke 10 år til overs. I stedet etablerte de en datafelle. [ ] Her eneste nye Tesla-eier er nå med i dette gigantiske datainnsamlingsprosjektet (shadowmode) 25 Bjørkeng, P. K. (2018). Kunstig intelligens den usynlige revolusjonen. Vega.
Learn Improve Fuelled by large data sets Dynamic Mistakes inevitable Data gathering through interaction VS. 26
Learn Improve Fuelled by large data sets Tentative design principles in response to characteristics to be discussed 27
Learn Improve Fuelled by large data sets Learning system - design for change Explain dynamic character (?) Show capabilities Be clear on limitations set expectations right 28
Learn Improve Fuelled by large data sets Learning system - design for change Explain dynamic character (?) Show capabilities Be clear on limitations set expectations right 29
Learn Improve Fuelled by large data sets Learning system - design for change Explain dynamic character (?) Show capabilities Be clear on limitations set expectations right 30
Learn Improve Fuelled by large data sets Learning system - design for change Explain dynamic character (?) Show capabilities Be clear on limitations set expectations right 31
Learn Improve Fuelled by large data sets Mistakes inevitable - design for uncertainty Inform and adapt depending on certainty level (?) Make recovery easy Learn from mistakes 32
Learn Improve Fuelled by large data sets Mistakes inevitable - design for uncertainty Inform and adapt depending on certainty level (?) Make recovery easy Learn from mistakes 33
Learn Improve Fuelled by large data sets Mistakes inevitable - design for uncertainty Inform and adapt depending on certainty level (?) Make recovery easy Learn from mistakes 34
Learn Improve Fuelled by large data sets Mistakes inevitable - design for uncertainty Inform and adapt depending on certainty level (?) Make recovery easy Learn from mistakes 35
Learn Improve Fuelled by large data sets Mistakes inevitable - design for uncertainty Inform and adapt depending on certainty level (?) Make recovery easy Learn from mistakes 36
Learn Improve Fuelled by large data sets Mistakes inevitable - design for uncertainty Inform and adapt depending on certainty level (?) Make recovery easy Learn from mistakes 37
Learn Improve Fuelled by large data sets Mistakes inevitable - design for uncertainty Inform and adapt depending on certainty level (?) Make recovery easy Learn from mistakes 38
Learn Improve Fuelled by large data sets Data wanted design for data capture Accommodate gathering of data from users but with concern for the risk of being gamed Make users benefit from data Privacy by design 39
Learn Improve Fuelled by large data sets Data wanted design for data capture Accommodate gathering of data from users but with concern for the risk of being gamed Make users benefit from data Privacy by design 40
Learn Improve Fuelled by large data sets Data wanted design for data capture 41 Accommodate gathering of data from users but with concern for the risk of being gamed Make users benefit from data Privacy by design https://www.technologyreview.com/s/610634/microsofts-neonazi-sexbot-was-a-great-lesson-for-makers-of-ai-assistants/
Learn Improve Fuelled by large data sets Data wanted design for data capture 42 Accommodate gathering of data from users but with concern for the risk of being gamed Make users benefit from data Privacy by design https://www.buzzfeednews.com/article/nicolenguyen/ amazon-fake-review-problem#.fjmkl3ynd
Learn Improve Fuelled by large data sets Data wanted design for data capture Accommodate gathering of data from users but with concern for the risk of being gamed Make users benefit from data Privacy by design 43
Learn Improve Fuelled by large data sets Data wanted design for data capture Accommodate gathering of data from users but with concern for the risk of being gamed Make users benefit from data Privacy by design 44