Rethinking How AI And Humans Interact To Get The Best Of Both

Summary

このトークでは、MITメディアラボのPatty教授とその指導の下で博士課程を進めるPatが、人工知能（AI）と人間の相互作用について語ります。Pattyは、AI研究者として30年以上前からMITのAIラボで研究を始め、後に人間に輔助する技術に興味を持ち、メディアラボに移籍しました。彼らは、AIが人間の能力を高める手段として、健康、学習、高齢者の記憶など、実際の問題を解決するのに役立つと信じています。しかし、AIの開発は人間のデザイン問題でもあり、人々がAIと協調して働く方法を研究し、適切なインターフェースを設計することが重要だと述べています。また、AIが人間の能力を高めるためには、人々がAIと批判的に考えながら対話する必要があると強調しています。

Takeaways

• 🎓 講演者は、MITのAIラボからメディアラボに移り、人間を強化するテクノロジーに興味を持つようになったと語っている。
• 🧠 AI開発者はAGI（人工総知能）を目指しているが、講演者は人類を支援するAIの開発に重点を置くべきだと主張している。
• 👥 メディアラボでは、AIを活用して学習、健康、高齢者の記憶などの問題に対処するプロジェクトを進めている。
• 🤖 講演者は、ソフトウェアエージェントが人々を支援し、情報の管理や利用を効率化できると提唱している。
• 🔍 過去には、ソフトウェアエージェントと直接操作の間で議論があり、講演者はエージェントがより適切な情報提供を行うと主張している。
• 📚 AIは単なる技術問題ではなく、人間のデザイン問題と信頼性の問題でもあると講演者は述べている。
• 📉 AIの開発が産業界に集中し、資金がAGIへの投資に吸収されていることに講演者は懸念を表明している。
• 👨‍🏫 AIは教育の分野でも重要な役割を果たす可能性があり、講演者はAIが学習を支援するツールとして機能するべきだと考えている。
• 🧐 講演者は、人々がAIと協働する際に批判的思考を促すことが重要だと示唆している。
• 📊 AIの開発には、心理学、デザイン、社会学、政治、法律など、多岐にわたる専門知識が必要であると強調している。
• 🚫 講演者は、人々の関与と議論がAIの将来についての重要な議題であると主張しているが、現在はそれが不十分であると懸念している。

Q & A

Patがメディアラボに移る理由は何ですか？

Patは、人間の側面に焦点を当てた研究に興味を持ち、MITのAIラボからメディアラボに移りました。彼は知能機械を開発するのではなく、人々を強化し、より良いレベルでパフォーマンスさせることに興味がありました。

現在、AI開発者たちが追求している目標は何ですか？

現在、多くのAI開発者はAGI（人工一般知能）を追求しています。彼らはAGIを達成することで世界がより良い場所になると考えています。

Patは人間の能力を強化するインターフェースについてどう考えていますか？

Patは、AI技術を利用して人々の問題を解決し、健康、学習、高齢者の記憶などに関するプロジェクトに取り組んでいます。彼はAIが人々をより良いレベルでパフォーマンスさせると信じています。

Ben Schneidermanとの議論において、Patはどのような立場を取っていますか？

Patは、ソフトウェアエージェントが人間の能力を強化するべきだと主張しており、Ben Schneidermanとは直接操作について議論しています。

PatはなぜAIと人間の相互作用が重要だと考えていますか？

Patは、AIは単なる技術問題ではなく、人間のデザイン問題であると考えています。人々の信頼、理解、そしてデザインに関する多くの問題を解決する必要があると主張しています。

AIが発展しても人間のスキルは維持されるでしょうか？

Patは、人々がAIに過度に依存し、自分のスキルを失う可能性があると懸念しています。AIは人々のスキルを補助するものであり、人間のスキルを置き換えるべきではありません。

教育においてAIはどのような役割を果たすでしょうか？

AIは教育において、人々の学習プロセスを支援し、より効果的な学習を促進するべきだとPatは考えています。AIは人々に答えを与えるのではなく、問題解決プロセスに関与し、人々の思考を刺激するべきです。

PatはAIの将来についてどのように考えていますか？

Patは、AIが教育、健康などに貢献する可能性に興奮していますが、AGIを追求するという目標には懸念を持ちます。AIの将来についての議論が不足していると感じており、多様な分野の専門家の意見を取り入れるべきだと主張しています。

人々がAIと対話する際に注意すべきこととは何ですか？

人々はAIと対話する際に、AIが提示する情報に過度に依存しないように注意する必要があります。AIは人々の思考を刺激し、問題解決プロセスに関与するべきであり、単に答えを与えるものではないとPatは考えています。

AIが発展するにつれて、人々の生活はどのように変わるでしょうか？

AIは人々の生活をより良いものに変えることができますが、AGIを追求するという目標は人々の利益を最優先に考えていないとPatは警告しています。AIは人々の能力を補助し、健康、学習、その他の分野で貢献するべきです。

Outlines

00:00

🤖 AI研究からメディアラボへ: パティの旅

パティはMITのAI研究者としてキャリアをスタートし、その後、メディアラボに移って人間とAIの相互作用に焦点を当てるようになりました。AI技術を用いて人間を補強し、知能を向上させるという視点が、今日ますます重要であると述べています。特に、企業がAGI（汎用人工知能）を追求する一方で、人々の利益に貢献する技術開発の必要性を強調しています。

05:01

📱 現在のAIインターフェースの課題

現在使用されているスマートフォンやラップトップなどのデバイスのインターフェースは、人々が情報を探すために注意を逸らさなければならない点で理想的ではないとパティは述べています。AIエージェントはユーザーのコンテキストを理解し、関連情報を提供することで、人々がその瞬間に集中できるように支援する役割を果たすべきだとしています。

10:01

📰 AIと人間の相互作用の重要性

AIは単なるエンジニアリングの問題ではなく、人間のデザイン問題でもあります。特に、人々がAIを過信しすぎる傾向があり、それが結果としてパフォーマンスを低下させる可能性があるとパティは述べています。彼女は、AIシステムと人間の協力を促進するために、ユーザーが批判的に思考するようなインターフェースの設計の重要性を強調しています。

15:04

🔬 過去のAIと現在の教訓

1980年代後半から90年代にかけてのAIブーム時に開発されたエキスパートシステムが成功しなかった理由の一つは、人間の要因が軽視されたためです。パティは、現在のAI開発でも同様の問題が繰り返される可能性があると懸念しています。彼女は、AGIの開発がすべての資金を吸い上げることに対する懸念を示し、AIの未来について広範な議論が必要であると述べています。

20:05

🌟 AIの未来: 人間中心のアプローチ

パティは、AGIではなく他の形態のAIが教育や健康の分野で大きな利益をもたらす可能性に期待しています。彼女は、AI開発において人間中心のアプローチを取ることの重要性を強調し、企業がAIを仕事のプロセスに統合するための小規模な実験を行うべきだと提案しています。さらに、AI開発にはエンジニアリングだけでなく、心理学やデザイン、社会学などの多様な専門知識が必要であると述べています。

Keywords

💡人工知能（AI）

人工知能とは、人間のように思考し判断を行う能力を持つコンピュータシステムのことを指します。このビデオでは、AIが人間の能力を増強し、学習や健康などの分野で人々の生活を改善する可能性に焦点が当てられています。例えば、スクリプトではAIチューターを使った学習の民主化や、AIシステムを使った精神健康や身体健康的サポートなどについて話されています。

💡メディアラボ

メディアラボは、MIT（マサチューセッツ工科大学）の研究機関の一つであり、先端技術を通じて人間の生活を改善することを目的としています。ビデオでは、メディアラボのfluid interfacesグループがAIを活用して人間の能力を増強するプロジェクトに取り組んでいると紹介されています。

💡人間の側面

このビデオでは「人間の側面」という言葉が使われており、AIが単に技術的な面にとどまらず、人間の健康、学習、記憶力などへの影響を重視するアプローチを意味しています。スクリプトでは、AIが人々の生活の様々な問題に対処するのに役立つと強調されています。

💡AGI（人工総脳）

AGIとは、人間と同じかそれ以上の知能を持つAIのことを指します。ビデオでは、AGIを目指すことが業界の主流であると批判されており、その代わりにAIを通じて人類の利益を追求することが重要だと主張されています。

💡インタフェース

インタフェースとは、人間にAIを使いやすくし、効果的にコミュニケーションを取るための方法やシステムのことを指します。ビデオでは、AIと人間が効果的に協調するインターフェースの設計が重要であると強調されており、人間がAIと批判的に思考し、判断するプロセスに関与することが促されています。

💡ソフトウェアエージェント

ソフトウェアエージェントとは、ユーザーの目標や情報を把握し、プロアクティブに情報を提供するAIシステムのことで、ビデオではその重要性が強調されています。スクリプトでは、ソフトウェアエージェントがユーザーの状況に応じて適切な情報を提供し、人々によりリアルタイムで関連性の高いサポートを提供することができると述べています。

💡人間のデザイン問題

人間のデザイン問題とは、AIが人間に適切に適応し、効果的に機能するためにはデザインと人間の要素を考慮する必要があるという考え方を指します。ビデオでは、AIの開発には単なる技術的問題ではなく、心理学、デザイン、社会学、政治、法律など多岐にわたる分野の専門知識が必要であると主張されています。

💡信頼性

信頼性は、人々がAIシステムを信頼し、その判断やアドバイスに依存する能力を指します。ビデオでは、人々がAIに過度に頼りすぎて自己判断をやめるリスクがあると示唆されており、AIと人間の間の信頼関係の重要性が強調されています。

💡ユーザースタディ

ユーザースタディとは、人々がAIとどのように相互作用し、どのような反応を示すかを研究する活動のことを指します。ビデオでは、企業がAIを導入する前に小規模なユーザースタディを通じて実際の使用状況を理解することが推奨されており、その重要性が説明されています。

💡教育と学習

教育と学習は、AIが人間の能力を増強する分野の1つとしてビデオで強調されています。スクリプトでは、AIチューターが学習をより効果的に行う手助けをして、人間の学習プロセスを支援することができると示されています。

Highlights

Pat and Professor Patty discuss the transition from AI research to focusing on human augmentation at MIT Media Lab.

The importance of using AI to augment human intelligence rather than replacing humans is emphasized.

Current AI development is criticized for focusing too much on achieving AGI (Artificial General Intelligence).

The need for AI to be used for benefiting humanity and people is advocated.

Projects that address health, learning, and memory for the elderly using AI are highlighted.

The debate between software acting as agents to augment humans and direct manipulation interfaces is discussed.

The potential of software agents to allow people to be more present by proactively providing relevant information is explored.

The importance of human design in AI development and the risks of overlooking it are underlined.

Studies showing people's over-reliance on AI and the resulting degradation of performance are mentioned.

The need for interfaces that engage users in critical thinking with AI is suggested.

The implications of AI in education and the importance of not just providing answers but engaging learners are discussed.

The interdisciplinary nature of AI development, involving psychology, design, sociology, and legal aspects, is emphasized.

The concern that the pursuit of AGI could lead to another AI winter is expressed.

The call for involving target users in AI development to prevent past mistakes from recurring is made.

Recommendations for those interested in a human-centered approach to AI include conducting user studies and involving various disciplines.

The potential benefits of AI in education, health, and other areas are acknowledged, with a warning against the singular focus on AGI.

An invitation for those interested in augmenting human intelligence with AI to visit the media lab's group is extended.

Transcripts

yeah so this session is going to be

about human AI interaction and or the AI

interface and um I many my name is Pat

I'm a PhD candidate at the media lab and

right here is my mentor Professor Patty

Mars who is also the director of the MIT

media lab fluid interfaces group so it's

great uh to have this conversation with

you Patty um so the first question I

have for you is um you start off as an

AI researcher at MIT and then you

transition to the media lab to work on

more of the human side of this kind of

question can you maybe talk a little bit

about your journey from the AI lab to

the media lab I did my PhD in AI over 30

years ago and uh in

Belgium and um moved to the AI lab what

was then called the AI Lab at MIT now

called seale uh because that was of

course the mecca for AI research at that

time and still um but after a couple of

years years at the AI Lab at MIT I

decided that I did not necessarily want

to develop intelligent machines that one

day could replace us uh surpass us that

I was much more interested in using

these same

Technologies to augment people to make

people more intelligent uh perform at a

better level and so on so that's

actually why I moved from the the AI lab

to the media lab and I think that that

point of view is more relevant than ever

today um AI developers today mostly

these days in companies like open Ai and

so on are especially focused on and

obsessed with this goal of AGI if only

we achieve AGI then the world will be a

better place I think that that is a very

nice leave goal and um it's really

unfortunate that all the money is going

towards that goal right now we should

think about why we um or what goal we

set for ourselves with AI and I think

surpassing people um in intelligence and

replacing people is one of the worst

goals we could set for ourselves so I

really have been advocating for decades

that we should instead focus on how do

we use these Technologies to benefit

Humanity benefit people yeah so in your

work you you focus on on human and

augmenting human sort of intell

intelligence what are some of the sort

of interface or what are some of the

project that you have done in this area

that address this yeah so I'm very

excited about the potential for these

Technologies to help people with issues

such as health learning

um even memory for the elderly and so on

so in our group we mostly take um real

problems out there um and try to think

about how these AI Technologies can be

um um used to basically benefit people

for example democratizing learning

making learning more effective by using

um AI Tutors or helping people with

mental health um or physical health by

using um AI systems that get to know

them uh get to know um their particular

issues ETC and that can assist them

right yeah I think at the time when you

came up with this idea of intelligence

system and and and you know human

computer interaction there was a huge

debate uh in the community between you

um and Ben Schneiderman I think you

advocate for the idea of you know

software acting as agent that augment

human being where Ben was talking about

direct manipulation maybe how how does

that work and and how is that irrelevant

today so yeah it's a little funny um

back in 97 I published a Scientific

American article on software

agents and so over 30 years ago um uh or

no not quite but

almost and um I argued that people were

getting more and more information uh all

the time and that they would need

software agents that knew about their

goals about uh their life about their

information that could help them with

keeping track of all this data making

data available um in real um on right uh

basically

proactively uh for people I still think

that the interfaces that we use today uh

to access the information World um these

smartphones laptops and so on the whole

style of interaction is actually not at

all um perfect or ideal because it

requires that we take away our attention

from the situation that we're in the

people we're talking to to then like um

look up information that may be relevant

or um and things like that so I've

always argued that software agents can

take a large role in really allowing to

be more present in the moment because an

agent can actually bring um up

proactively information that is relevant

to whatever is happening right now like

an uh my my phone should know that I'm

here on stage in an interview and that

it shouldn't be buzzing right now or

maybe if there is something particularly

relevant to what I'm saying maybe I'll

get just one word or something hint from

an agent saying don't forget to mention

X or something so I think there's a we

an opportunity to really reink our

relationship with devices by using AI

systems and in a way this type of

interaction is very proactive in the

sense that you know the AI not just like

wait for you to go and use it it

actually uh personaliz and actually know

who you are and kind of understand the

context and things like that yeah so

we've always been working on agents that

are more aware of the user's context

what a user is doing uh are they in a

supermarket uh shopping for whatever

toilet paper are they talking to someone

what are they talking about right Etc

are they reading an email message and

the agent can proactively make relevant

information available to the person for

whatever the problem is the issue is

that they're dealing with right now so

fast forward after that Vision now we

have this sort of you know generative AI

agent or a system like that that are

being deploy and develop people might

argue that if you just make the AI more

powerful or smarter this question will

solve itself but in your case you still

think that we need to focus on the

interaction and the interface why is

that the case yeah so I think AI is not

just an engineering problem even though

Engineers think so yeah um AI is a as

much a human design problem and human

design is not something that you solve

after the fact when you have AGI oh

let's now like decide on colors and

fonts and bells and whistles no human

design is a big problem right we need

people to have the right level of trust

in the AI systems that help them we need

people to understand how these AI

systems function and why they come up

with a particular decision there are so

many human design issues that are

incredibly important and we are finding

that with studies that we're doing for

example people overly rely on AI

especially generative AI large language

chat like interfaces they are all very

persuasive very believable they always

come up with a great and answer very

believable answer even though it may be

totally wrong and uh

hallucinating and we find in studies

that we do that people start relying on

these um AI systems too much and they

stop thinking for themselves they stop

engaging with um the problem at hand

because they say like oh my AI seems to

know the answer sure let's accept that

so we have um several studies that we've

done that show that people's performance

can actually

degrade by using AI because they overly

trust uh the AI so there's a lot of

important issues that we have to solve

if we want AI deployments ultimately to

be successful and I would um argue that

most AI deployments will be in in the

context of people working with the AI

whether uh they'll only have some

minimal supervision and auditing of

whatever the AI does versus true

collaboration uh or even mostly the

human dealing with the problem so there

will be primarily um AI applications for

people and AI to collaborate together

not AI operating autonomously which is

why it is so critical that research how

people work with AI and how we can best

design these interfaces so that we

ultimately benefit from the intelligence

of the AI and the intelligence of uh

people well one thing that I think is is

kind of interesting and and a little um

unintuitive is that people often think

that if you have you know human

intelligence and and artificial

intelligence put them together you have

a combined super intelligence but from

study that you mentioned that we have

done in our group we show that you know

people tend to just rely on you know one

intelligence and not using the other one

what are some sort of you know

techniques or some of the interaction

method that can be used to actually you

know augment human rather than making

human complicit to this system yeah we

uh for example did a very simple study

where we had people judge newspaper

headlines and decide whether they were

fake news or True News and um we gave

some people a correct AI that gives them

correct advice and explanations like yes

this headline is right or true and

because of these reasons uh we gave

another group um basically a malicious

AI that tried to convince you of the

opposite of the truth for all the

headlines and then a third group was not

using an AI at all and we learned that

um the group that got the malicious AI

their performance their accuracy was

almost half

of their accuracy of working by

themselves without any AI assistance so

they overly relied on AI I would uh

speculate also that or hypothesize that

the more people work with AI uh the more

deskilling could happen because people

start relying on AI so much that they

forget how to deal with the issue

themselves in the first place so we have

been arguing um uh for interfaces that

uh engage the user in critically

thinking together with uh the AI in a

conversation thinking about the problem

before the AI actually says um this is

true or this is false or this is cancer

or this is not cancer you have to engage

the user in thinking about um the

problem at hand the decision at hand

before the AI actually uh gives its

classification or its recommendation so

in a way sort of you know use the

intelligence of the AI to engage people

intelligently and then see how they kind

of come up with answer together so not

just the AI just give out the answer

right away exactly I think this had

implication in education and learning as

well right what what what do you think

about that yeah well we all know that

the best method of teaching people some

skill or teaching young people a skill

is not not to tell them what the answers

are and and what all the stuff is that

they should try to uh put into their

their brains the best question is to

engage uh the best method is to engage

them engage them in asking questions

engage them in the material don't just

give them all the answers get them

excited get them thinking get them to

discover the answers themselves with the

guidance of a good teacher so I think AI

systems should play that role they

should be more um assisting us in our

decision making process without taking

over um um and and basically um with the

result that then we uh tune out

basically and you know in order to do

this it seems like you need to draw a

lot of Knowledge from Auto F right not

just engineering of the AI self but like

you know human cognitive science or

psychology or UT area how does that

again I think AI is not just an

engineering problem it's a psychology

problem it's a design problem it's a

sociology problem it's a a political

problem a a legal problem and

unfortunately all of these other domains

are sort of afterthoughts and and the

questions they think about and all the

attention is going to engineering right

now

but I believe that these other areas of

expertise should be involved um right

now in the development of AI well one

wisdom that you have shared with me

earlier was that you know in the earlier

day when AI was called intelligence or

or expert system right you say that it

already surpass human in many capability

but it wasn't adopted because of this

social or you know the human issue do

you think it's going to continue to

happen yeah so back in the late 80s and

'90s there was another wave of interest

in artificial intelligence called expert

systems and there was a lot of money

being thrown at AI not as much as today

but also significant amounts and once

again AI was dealt with as an

engineering problem at that time so

people would develop AI expert systems

that would come up with a medical

diagnosis for example and these systems

performed pretty well um at that time

not as good maybe as our systems today

but they were very useful but then

Engineers would like drop this expert

system in the hands of the target user

and that Target user say a doctor making

a diagnosis would say well I can't trust

this system I wasn't involved in

designing this system I studied for 10

or 13 years to become an expert uh in my

field why should I like even ask this

computer for a second opinion about

decisions so a big reason why expert

systems were not a success was actually

these human factors right and that were

sort of again an afterthought and then

that was of course followed by an AI

winter so I would argue meaning an AI

winter less money for AI less interest

in AI I would would argue that we risk

another AI

winter uh coming up because again we

only have engineers and money people

interested in Ai and deciding what

direction to take in as opposed to

really everybody out there not just all

disciplines but Target users Target

users should be involved we should all

have a discussion together uh really

about what AI future we want and

unfortunately that discussion is not at

all happening right now it is just

assumed that we all want AGI and that's

it and then we'll solve all the problems

later after we've deployed AGI uh to

billions of people right so I have two

more question for you so one is if

people here are interested in sort of

taking that more um human Center

approach to AI what should they do

Beyond like come to your lab and talk to

you what are the the way that they can

get start on thinking about this I think

it's important to do more um U smaller

uh user studies or studies to see how

people respond to AI how they really

work with AI because this whole um

Vision that AI if you give it to people

they'll just be super super performing

whatever it's not true uh it's not

necessarily the case people also respond

very differently from person to person

to AI some people are afraid other

people uh afraid that the AI will take

their jobs other people fall in love

with AI almost literally and think it's

sentient and so on it's just a big mess

out there right and I would say that

companies especially should be careful

and should do smaller scale experiments

to integrate AI into work processes to

see what really happens how the uh

employees really use these systems and

engage with them right because other

issue like hallucination or you know

when it make up things right still going

to continue to exist and we human I

think is is the one that going to need

to be able to navigate them so I think

there a lot of uh ideas that can be

deployed in that so the last question

for you is then what are you most

excited about or what you know what are

you most concerned about this day um

yeah yeah well one thing I'm actually

very concerned about with respect to the

state of AI and you already heard it is

that all the money is uh being spent or

most of the money in Industry uh right

now uh this goal of AGI is sucking up

all the funding out there for a goal

that I would argue is not at all a good

goal to to strive for um AI uh research

used to happen primarily l in a

university context until recently where

we have standards of discussing problems

being open about approaches um uh trying

to get a lot of feedback peer review Etc

and unfortunately AI development right

now is mostly happening in industry and

is driven purely by money interests uh

not at all by Humanity's interests I

would argue so that is a big concern

that I have about what is happening uh

today at the same time though I am still

an AI researcher and I am excited about

the potential for AI not AGI but other

forms of AI to uh really benefit uh

education and learning benefit health

and so on but that goal of AGI is not

necessarily uh what is going to bring us

there that's really that's really

important so I think that's the

conversation today um if you're excited

about augmenting human intelligence or

augmenting human capability with AI you

can talk to Patty or I guess come visit

our group on the fifth floor at the

media lab we upstairs so thank you so

much thank you thank

[Applause]

you e