Rise of AI Conference 2019 · Berlin

Overcoming Fundamental Problems of Machine Intelligence Research

Dagmar Monett · Professor of Computer Science at Berlin School of Economics and Law

At Rise of AI 2019 in Berlin, Dagmar Monett, professor of computer science at the Berlin School of Economics and Law, presents research into why artificial intelligence still has no agreed definition and what her group is doing about it. Her team ran a survey, opened in July 2017 and still running at the time of the talk, that collected more than 550 responses from people in more than 57 countries and over 180 institutions, about 80% of them from academia; 82% of respondents said they work in AI. The survey asked whether existing definitions of AI are adequate, whether human and machine intelligence need separate definitions (80% said yes), and how respondents would define intelligence themselves, producing more than 4,000 free-text comments. Monett argues the field lacks consensus because it is more than 60 years old, interdisciplinary, contextual and a moving target, since capabilities once called intelligent later stop being seen that way. She presents quality criteria her team derived for good definitions, cognitive biases that distort discussion, and a catalog of properties meant to give researchers, regulators and companies a shared vocabulary for defining intelligent systems, a precondition, she argues, for workable AI regulation and ethics frameworks.

Key takeaways

  • Monett's team surveyed 556 people across more than 57 countries and 180 institutions on how they define AI and intelligence.
  • 80% of respondents said human and machine intelligence need separate definitions rather than one shared definition.
  • The team found cognitive biases, including an anchoring effect and a tendency to comment mainly when disagreeing, that distort apparent consensus.
  • The literature definition rated highest by participants emphasizes adapting to the environment, limited resources, real-time processing and learning from experience.
  • The team built a quality-criteria catalog and vocabulary, drawing on Aristotle and software-requirements engineering, to help others define intelligent systems well.

Detailed notes

A global survey on how people define intelligence

Monett opens by asking the audience how they would define AI or machine intelligence to a child, framing the talk's central question. Her research group ran a survey, opened in July 2017 and still open at the time of the talk though the main data collection closed a year later, that asked people whether they agreed with existing literature definitions of AI, why or why not, and how they would define intelligence themselves. Of about 10,000 people invited, 556 took part, coming from more than 57 countries and over 180 institutions, roughly 80% from academia; 82% said they work in or do research or teaching in AI. The survey deliberately restricted itself to human and machine intelligence, leaving out animal or other kinds of intelligence, and asked whether the two need separate definitions: 80% of respondents said yes, arguing that what counts as intelligent behavior in machines has changed over time as capabilities that were once called intelligent stopped being seen that way.

Why AI still lacks a consensus definition

Monett lists reasons her team identified for the absence of a shared definition: AI is more than 60 years old but has been through repeated cycles of hype and reinvention; researchers from different disciplines, such as neuroscience, psychology and deep learning, bring incompatible starting assumptions; and the field is a moving target, since growing computing power and data constantly expand what systems can do. She argues that consensus still matters because regulation, governance, transparency and "responsible AI" all require agreement on what is being regulated, and because media hype creates expectations about AI that affect funding, favoring fast-moving subfields like machine learning over other AI research. Many survey respondents, however, told her team it does not pay off to seek a definition at all, comparing it to "pornography: I know it when I see it."

Quality criteria for well-formed definitions

Before proposing a definition of intelligence, the team asked what makes any definition well-formed. Monett drew on the literature on definitions going back to Aristotle and on her own work in software engineering on defining good software requirements, combining both to derive quality criteria for definitions of intelligence. Applying these criteria to more than 4,000 comments from survey respondents, the team found that the literature definition participants rated highest exemplifies intelligence through concrete cognitive capabilities: adapting to the environment, operating with insufficient knowledge and resources, processing information in real time, and learning from experience. Monett notes it surprised her that the definition participants preferred most overall came not from AI but from psychology, even though more than 80% of respondents work in AI themselves.

Cognitive biases that undermine consensus

Analyzing the free-text comments, the team found cognitive biases that shape how people argue about definitions. An anchoring or focalism effect meant definitions presented earlier in a list were agreed with more often than those presented later. Respondents were also far more likely to leave a comment when they disagreed with a definition, trying to refute it, than when they agreed, so silence in the data skews toward agreement rather than reflecting genuine consensus. Monett argues that anyone arguing about how to define intelligence, in research, standards or policy, should be aware of these biases before treating apparent agreement or disagreement as meaningful.

A catalog and vocabulary for defining intelligent systems

The team used natural language processing to analyze the semantics of the comments and compare how AI experts and non-experts talk about intelligence, then built a catalog of properties and a vocabulary intended to help people define intelligent systems well, plus an iterative process for building a definition from scratch. Monett frames this as a resource for practitioners outside AI research: policymakers, people writing standards, companies that want to add AI to their processes, and ethics work, all of which she says need a common vocabulary before they can regulate, standardize or evaluate AI meaningfully. She closes by saying the AI community is responsible for supplying that vocabulary to people making decisions about AI from outside the field.

Q&A highlights

Asked whether it is even worth debating a single definition of intelligence rather than naming specific problems like reasoning or metacognition, Monett compares the situation to definitions of fairness or transparency, which have also proliferated without consensus, and says the field's name is unlikely to change; she points to proposed subfields such as "machine behavior" as an example of terminology shifting over time, much as "configuring algorithms" in metaheuristics became "hyperparameter optimization." Asked what share of AI researchers versus government respondents wanted a definition, she says most AI researchers who answered did want one, but too few government respondents took part to draw conclusions, which is one reason the survey stays open. On how the research could inform ethical frameworks, she says the same quality criteria and bias awareness could help define what makes a system "ethical," and that the vocabulary is meant partly for software engineers with no ethics background. Asked whether definitions differ by AI subfield, she says the survey did not ask about subfield specifically, only institution, role and country, but notes that philosophers in the sample tended to want a single definition of intelligence, and mentions that the European Commission's expert group guidelines on AI, released shortly before the talk, proposed at least 200 new PhD positions in philosophy to support understanding of AI.

Quotes

How would you define AI or machine intelligence or intelligence in general if you have to tell a child what AI is?

Dagmar Monett

We think it is very important to know what the I in AI is.

Dagmar Monett

I think we have to be very careful with the terminology we use.

Dagmar Monett

Questions answered in this talk

Why is there still no consensus definition of artificial intelligence?

According to Dagmar Monett at Rise of AI 2019, AI lacks a consensus definition because the field is more than 60 years old, draws on many incompatible disciplines, and is a moving target whose boundaries shift as computing power and data grow, so capabilities once called intelligent stop being seen that way.

Should human intelligence and machine intelligence be defined the same way?

In the survey presented by Dagmar Monett at Rise of AI 2019, 80% of respondents said human and machine intelligence need separate definitions, because what counts as intelligent in a machine depends on context and changes over time.

What cognitive biases affect how people argue about defining AI?

Dagmar Monett's research, presented at Rise of AI 2019, found an anchoring or focalism effect where definitions listed earlier got more agreement, and a bias where people commented mainly to disagree, so silence does not indicate true consensus.

Mentioned

definition of artificial intelligence, machine intelligence, cognitive bias, anchoring effect, quality criteria for definitions, machine behavior, hyperparameter optimization, AI ethics guidelines, consensus definition

Keep exploring

More from Rise of AI Conference 2019

More on AGI & Future of Intelligence

Transcript

Auto-generated transcript, not edited. · 5,017 words

Show the full transcript

thank you so much for coming to my talk and I want to start it with equation and I am still in the question I heard before in the introduction so I want to know more or less how many of you have or work or have work it in AI or do AI research teaching this is great because there is there are people in the audience which maybe know about AI have suppose but I don't know the details about AI and this is what I want to talk about today so how important it is to know who our intelligence is about what machine intelligence is about for all stakeholders stakeholders so this is the second question I want you to not to answer now but I want you to think about that how would you define AI or machine intelligence or intelligence in general if you have to to to tell a child what AI is if you have to to to work with other people and you have to to death to define what an

intelligent system is how would you define that who you use definitions from others do you have all in mind or who you try to define something with your own words think about that because all the talk is about defining AI and this is what we have done in a serving in order to study how people define AI what do they think about a consensus definition because AI isn't media in the last years is a living a reborn it is not a new a new field it is more than 60 years old fields but a lot of people don't know where exactly a is an x there is a confusion in the media and the public in general cetera about what AI is at least not and we did a survey so we interviewed a lot of people or we tried to interview a lot of people and I'm very happy with the results start in July two years ago and we closed the survey one year later but it is still open so people can

still participate and till now we have more than 57 countries or people coming from 57 countries not all experts on AEI coming from more than 180 into institutions around the world mainly from academia so when almost 80% come from academia researchers in their majority but also people so educators developers etc and not everybody working in AI and this was also our our purpose or not not everybody working in AI and we collected until in July last year more than 5500 50 responses and this is what we wanted to analyze some of the questions had to do with define in AI so we had definitions from the literature and we wanted to know how people think about that if they agree with that definitions or not and very important why not and we also asked it at the end and how would you define a lie because we did a pre research study and we noticed that not everybody is happy with the definitions that are already available or in other words some people try to

define the new things the new advantages then the new the new advances in AI differently so we wanted to know maybe you have a new definition how would you define that it was my my second question to you we also had other kinds of questions for example if they have work it or not in artificial intelligence this was the first question to you and the big majority so 82% know what they are talking about and this is good but we also wanted to have the answers from other people for example philosophers or mathematicians and this was over can other question - so what do you think we've restricted the survey to only differences between machine in trillions and human insulin so we left out animal intelligences and other kinds of intelligences but we wanted to ask about do you think we need only one definition of intelligence or do you think we have to have separate definitions of intelligence for example if we define in trillions in humans they should be differently done else when

you when we define Intel is in machines and all the media of the hype sometimes we have now is artificial intelligence machine intelligence so we have to to know what these systems are and what is not an intelligent system so this is very important so the distinction should we use the same definition as for humans is valid and we wanted to know and this was not so evident so 80% of the people think we have we should have separate definitions and this is valid because it depends on the context you have and you cannot so now we say machines are intelligent but intelligent but 10 years ago we also said machines are in trillions if they do this and that and that and now that they can do that we say they are not intelligent anymore notice is not so sign of intelligence maybe if they are creative if they are something else all the properties so it changes with time it changes with the development of AI to give you an idea about some

other overview questions if these two to their orders but a lot of people think that it does not pay off to have a definition we don't have to worry about that one so I don't want to mention names now but some even think it is like pornography when I see it I know what it is and I am not so happy with that and a lot of people think that it is not important to define area I will give you some reasons why it is important to define it and a lot of people think so the ones that strongly agree or agree a with how many consensus definition to reach an agreement upon AI is possible it's very comforting so we are really happy that a lot of people think yes we can at least know the boundaries of the discourse about AI and what is not yet so this is what we took and we want to join to analyze and to drive our research we we thought about different possible reasons why there

is no definition of failure this was the first thing we wanted to study and this is a very polarizing concept since the foundation of the field more than 60 years ago so a lot of different researchers academicians as it's real try to define AI differently and then when you have a definition then comes a different person and say no but we have to add this and that and that also because a lot of different people try to give two perspective with the upstair with the with the latest state they have so they are inherently contextual and the interdisciplinarity also is a ground for having different ideas about what intelligence is for example neuroscientists psycho is etcetera I think that intelligence is something but then you have a guy in artificial intelligence doing deep learning and think no but this is not artificial intelligence this is something else so it is ok this something that happens and of course for several reasons but now we have with the emergence of the computer power more data etc

are reborn of very different old area of AI machine learning is not a new area it is a null area AI that we have different applications new applications and it is constantly evolving what we think is intelligent or not so like Lipton say this is a moving target so it changes with time with the development of AI but consumption is needed because and we see it now in D&D in the media and the news in the discourse in politics with the society etc that we need for example regulations and if we need regulations governance of AI then we cannot have regulations without knowing what we want to regulate also because of transparency sustainability responsible AI so we won systems that improve human wellbeing for that we have to define what intelligence is or what this intelligence systems are but we also want to document to understand what AI is historically and it is very difficult to track the definitions but also the media and hype which is most of the times not very favorable and

create expectations of what AI East is that it is not and this has also disadvantages for example funding going to areas ok where they are having a rapid development now but not to other areas where it's also important to do research that we long also to a eye area is not only machine learning they are a lot of other different subfields that are part of AI for that we think it is very important to know what the I in this AI is three questions that guided our research are the following ones so we wanted to distinguish and to know how the people think what do people think about different definitions of AI from the lecturers sure machining trillions but also what do people think about definitions of human intelligence in order to compare we wanted also to know about the differences between human and machine intelligence so we have really two different kinds of definitions yes or no why and also which capabilities should be considered when we when we define intelligence what is not

intelligence anymore in machines for example and why other research sub-goals it was not at the beginning of our study were to define quality criteria for definitions okay we are talking about definitions but do you know how to define a definition so we have to do to a you step back what is a well definition of anything of something not a I but also privacy for example fairness trust how do you define a concept and how do you define that well because supposedly defining something is trying to to reach consensus between a lot of parties to try to to to to to understand the concept you are defining so how do you define this well and this was some part of the research we didn't thought at the beginning it was so important and we I will talk about that too in the next in the next few minutes we also analyze it responses of the and comments from the respondents they were more than 4,000 comments we collected so you remember I told you yes

agreeing or not with the Phoenicians but we also asked it why not why do you agree and we collected more than 4,000 different opinions and we wanted to analyze that so how do people comment and we have very interesting resource so they are cognitive biases that people unconsciously for they define or discuss or argue when defining intelligence and also our interests our goal was also to - to have a catalog a vocabulary to inform the people also not coming from AI what is important when you define in Chilean systems to inform for example people that develop standards to inform people that regulate existence if you want to regulate something that this intelligence first tell me what intelligence is so we provide you with a catalogue with a vocabulary you can use you could use to define what intelligent systems are first equality criteria for definitions so we we analyze it and examine it the literature on definitions coming from Aristotle how what is definition which are the rules for defining the definition of any concept

we also analyze it and because it is also not only AI but software engineering part of my research and teaching interests different criteria for software requirements what is good software requirement is and I thought maybe a combination what it were a good criteria for a criterion for defining intelligences and we also analyzed it thoroughly all the comments provided by the respondents more than four thousand opinions and we came up with a catalogue with different properties that definitions should have in order to be considered well definitions of Internet we think if you consider this criteria your definition could be good understood and and this is good to reach consensus maybe we also proposed some iterative process if you start from zero how can we help you define in something in particular intelligence and it is some extensions from other offers what other authors have proposed so far because maybe some of you have interests yeah you know what a definition of intelligence so this was two most accepted definitions of intelligence from the literature for artificial

intelligence from one and one of the properties is you have to exemplify so you have to to give an idea of what intelligence is also by exemplifying what cognitive capabilities are of what can be done or what not and for example here adapting to the environment insufficient knowledge and resources finite processing in a timely so real-time open to new changes learning from experiences this was very important so this was one of the definitions to the participants thought it was it is one of the best it was not the first definition the first definition the most receptive one was for nap psychologists and you remember more than eighty percent of the people is working in AI so they think a definition from outside area from human intelligence is definition we should consider this was very interesting it surprised me so this is one first conclusion as the development of AI is moving then we also maybe have to change the definitions it is inherently contextual and I am NOT saying we should have only one definitions

but it would be interesting to know the boundaries what belongs to what not the cognitive biases TG second one so we analyzed it and we found a lot of different cognitive biases in G commenters in the comments from the participants one of ten for example it was strongly dependent on the position of the definitions of a on the list how participants agreed or not with the definitions so the first one the focus the anchors were more agreed upon and the last definitions disagreement this is a bias anchoring effect or focal ism for example other biases the people comment when they have disagreement when people agree they sometimes don't come and why but when they disagree and this is also a bias they tried to refute they tried to convince the others who why something is wrong and not why something is good so it is important that researchers people do in AI but also people not to in AI are aware of the cognitive biases we have when argument and when convincing people when talking

about something when defining something this is very important so we have some work also discussing which cognitive biases we found and what is important to take into account - this is one example for example when people comment these are the the agreements so they are mostly a now I have to word in German they do not agree with the definitions and when they do not comment the agreement is very very high so this is only one example so the cognitive bias biases undermine consensus if you want to reach consensus put an eye on the cognitive biases that may be present the Italians catalogue so surprised we use it AI to understand AI and we came up with a catalogue analyzed not least data from the opinions on definitions but also to suggest the definition so we use it natural language processing techniques to analyze all these comments and the semantics behind so the relationships we tried also to analyze the experience in AI and experts define and use a vocabulary vocabulary which is different from

non experts or from no B's in the field and it was very very important to try to select we shall may be the most representative world's most important capabilities that should be usage when defining the Italians and we came up with with some examples that could be used in by defining intelligence so we say that it is very important to consider experts opinions we cannot only rely on academic definitions from the literature and also definitions from standards etc we should also consider experts opinions because jdd the field is moving very fast and getting clarity of course means that that we should consider that opinions it should not be done without asking experts what do they think now and how changing the the natural of artificial intelligence is so if you really really really need a new definition then we say ok doesn't problem but you find it well and we are providing some research some results in order to help practitioners people in AI but also people not doing AI but I want to deal

with AI I want to understand AI so we have a responsibility from inside out to fail to tell others what is that and what is not and we think we can help everybody so politicians people doing standards people that like I have been my University comment to me I want to do AI and I don't know what that is but I have I want to have AI in my processes in the company so first let's reach a common vocabulary what do you understand what do you think it cetera and this is very important so thank you very much if you have some questions or you want more details about the research you can a drop a line and I will be very grateful to to help thank you as we started early we really have some time for Q&A um who would like to start with the Q&A I'll have to run because we only have one of the sent mics yeah yeah thank you for your talk my name is Hector was climbing rock

currently at the Reston button I wanted to ask about this defining the intelligence so sorry this voice quality is really bothering me but like I'll try to live with it is there even a point in I think we can all agree that there are many kinds of intelligences and for for everyday years we roughly know what we mean by when we talk about intelligence because there are some features in pattern and patterns of behavior that we are observing and those in if enough of those features fire like something looks like it had an idea behind doing that action or whatnot then we qualify it as intelligent or not so for everyday use this ambiguous definition seems to be pretty enough and then if we are talking about area in technical terms wouldn't it make more sense to just not use a big word like intelligence that has lots of meanings and lots of possible interpretations and just talk about the actual problems we are dealing with for example reasoning about the future introspection metacognition learning

and whatnot instead of using words like intelligence and whatnot so is there even a point in talking about the intelligence in this sense because I guess it might make sense in a low text but then if we take the ambiguity in the law text then the judges who don't know anything anything about this will have to ask some expert for an opinion and in the European might still be ok but for in places like the United States we know what kind of case my tendency okay so do you want to answer that yes this this is some some opinion a lot of people have and this is since the D not since the definition of the field for since the foundation of the field a lot of people think yeah why artificial intelligence why not superpower statistics processing which is who had a lot of different machine learning algorithms to today it is not nothing else than something statistically computer power powered computations and starting Chileans so maybe if we shift from the warrant alien

saying we have all no worries anymore but we have it there and we have it in the literature and defining intelligences maybe could be similar as defining transparency fairness it is an explosion of fairness definitions in the last year's and it is difficult to to to reach to to to reach agreement on what do you consider to be fair or not in an algorithm now and it is true white and concentrating our intelligence our world thought was okay the field is artificial intelligence I don't think the name of the field will change so we have to live with that maybe we have different subfields no ones now for example machine behavior is what is recently proposed we should also understand how to these systems evolve interact etc why not analyze engine like a analyzing social communities etc the behavior of human interactions so why not in machine behavior great so I think I think we have to be very careful with the terminology we use I did my PhD 15 years ago it was configuring

algorithms configuration of parameters this is very very strong in the metaheuristics community where neural networks for example or an example and now it is called hyper parameter optimization high parameter this is a configuration of the parameters of the algorithms so it changes and we have to to know about that that it changes the terms can change so we want to to stay a little outside how do you define any term well so I just a quick comment and a question so with I think to echo a bit of the sentiment of the previous speaker by the way Trenton Carnegie ocean protocol 25 years of AI work so every search so I can see there's going to be different demographics that care about definition versus not you know as you hinted yourself AI researchers are probably going to be very skeptical because definitions keep changing I'd rather just see differential equation to describe things frankly and with the definition given that would cover any feedback control system so we'll be defining our thermostats that regulate the

temperature of the room so my question to you is what percentage of AI researchers said they wanted a definition and what percentage of government people said they wanted a definition we will be happy if we had more respondents from the government because sometimes most of the times they are the ones that make decisions at the end and the majority of the AI researchers were for different definition of AI said that we had we have to distinguish yeah and and yeah unfortunately not so many people outside AI it was ten thousand people we want we send the invitations also in the media social media etc and 556 we could convince okay take part in the study but we will be very very happy if we could have more people this is why TDD survey is not closed of course it is very very specific now to artificial intelligence and not all not everybody so some people is afraid mm artificial intelligence as a field they notice is very complicated all this is not done but at

the end they want to make decisions so it is important to inform these people too and I think this should come from the AI community this is this could be you could use that in that way so and this is not so this is a reason why they are now people coming from AI for example a i8 ethic ethics people that are in these Commission's and also AI people that are in these Commission's in the European Commission's or national Commission's coming from the field that are important now a on also part of the decision and the what decisions makers stood not about that so I think this is a good sign it is not all done till now all the work that should be done but I think this this is part of something bigger and it won't stop we see in the introduction just the biggest challenge so everything trying to be better up to optimize so to do a lot of things we cannot do today if we apply mechanisms that are already

very well known or maybe new ones that will come for sure from this area from AI so I'm very optimistic a lot of people are pessimistic about the future when we deal with AI are very optimistic of course we should be very careful and also think about what could go wrong and a lot could go wrong yeah thank you some more questions from the audience I should have put on flat shoes how do you think your research can contribute to the discussion about ethical frameworks I said so I should have put on sporting shoes how this how can this contribute to the discussion which is a very very top needed discussion today ethical frameworks well the first my first instant Tripoli thought is you want to define how ethical system is parenthesis disregarding whether it is in Italian or not parenthesis close it hey this is great this is very very important it's needed now because we are seeing this so how do people or to some companies etc use the data without consent from

the users so there are a lot of different implications that should be regulated and that that we should know how fickle how ethical considering the systems I think we can contribute in a way for example trying to to to to specify better what ethics is for example the first thing if you define ethics if you define a system to be ethical or not what should you consider what children you forget what could be better and this is why the quality criteria could help in this sense but also when you will find that when you have humans talking about sounds something be aware of the cognitive biases in this way and this way and this way and also if you have some catalog some vocabulary that helps to to to to introduce new people to the field to to to deal with a lot of software engineers program which has no background in ethics maybe this is why it is very important and we include this in the courses and modules at the universities and also

outside university to teach people what ethics means for example now a lot of different concepts we win we need to be aware of we think that some vocabulary some catalogue or some guidelines how to define and what should be defining how could help so this is our our optimism to that this can contribute in some way of course if father wants to use that now but I suppose everybody's interested in in doing something a little better and we think that we have contributed to that there are some public aid publications now being reviewed and my expectation is that a lot of people can can drape them and be aware of what we have done I think you got time for one more question if somebody has one okay and keep on running to you thank you for your presentation Jesse Lurky a few Berlin I'm not sure how fine-grained your demographic data was but I was curious if you did go deep enough to note if there were differences in the definition of artificial intelligence

by researchers in different areas of artificial intelligence and whether they're working with different types of learning different types of algorithms altogether because that they might have more or less inference its interest in the data side of the definition for instance or in the experience side of the definition thank you yes so wewe didn't ask which area or a su field of AI but if you are a researcher and chemic or professor or from which institution from which country so where do you come from but also where do you work at this moment and and we have this this data we haven't been so specific in this case so what do researchers think and what do they think about the definitions we have a very general overview and their differences for example philosophers think we only need one definition and we should not be talking about differences only one definition of intelligence and that's enough and and this is very important too because now in the multidisciplinary teams we are needing we are including antithesis philosophers

and cetera and in the dearest truths were the guidelines for truth Ward's the AI from the a from the European community that were a common European Commission from the special group on AI that came out few weeks ago one of t things that they proposed is minimum 200 different 200 PhD positions in philosophy styrene and understanding a so this this is important also to now come in people coming from all the fields and this is something we have of course we don't have big day down that we have only 550 an hour now almost 600 participants but this is also very interesting now how do what the people think about the two definitions of the biases etc most of the times the discussions are coming from inside the community what people say this is not intelligence anymore this is only calculations powered by super power machines we have now CPU etc etc

Rise of AI 2027 · Berlin · May 2027

Be in the room for the next one.

Ten years of talks started as conversations in one curated room in Berlin. The next edition brings 300 selected leaders together in May 2027.