Professional Documents
Culture Documents
Survey IA
Survey IA
com
Mller, Vincent C. and Bostrom, Nick (forthcoming 2014), Future progress in artifi-
cial intelligence: A Survey of Expert Opinion, in Vincent C. Mller (ed.), Fundamental
Issues of Artificial Intelligence (Synthese Library; Berlin: Springer).
1. Introduction
Artificial Intelligence began with the conjecture that every aspect of learning or
any other feature of intelligence can in principle be so precisely described that a ma-
chine can be made to simulate it. (McCarthy, Minsky, Rochester, & Shannon, 1955,
p. 1) and moved swiftly from this vision to grand promises for general human-level AI
within a few decades. This vision of general AI has now become merely a long-term
guiding idea for most current AI research, which focuses on specific scientific and en-
gineering problems and maintains a distance to the cognitive sciences. A small minori-
ty believe the moment has come to pursue general AI directly as a technical aim with
the traditional methods these typically use the label artificial general intelligence
(AGI) (see Adams et al., 2012).
If general AI were to be achieved, this might also lead to superintelligence: We
can tentatively define a superintelligence as any intellect that greatly exceeds the cognitive per-
formance of humans in virtually all domains of interest. (Bostrom, 2014 ch. 2). One idea how
Future Progress in Artificial Intelligence: A Poll Among Experts 2/19
superintelligence might come about is that if we humans could create artificial general
intelligent ability at a roughly human level, then this creation could, in turn, create yet
higher intelligence, which could, in turn, create yet higher intelligence, and so on
So we might generate a growth well beyond human ability and perhaps even an ac-
celerating rate of growth: an intelligence explosion. Two main questions about this
development are when to expect it, if at all (see Bostrom, 2006; Hubert L. Dreyfus,
2012; Kurzweil, 2005) and what the impact of it would be, in particular which risks it
might entail, possibly up to a level of existential risk for humanity (see Bostrom, 2013;
Mller, 2014a). As Hawking et al. say Success in creating AI would be the biggest
event in human history. Unfortunately, it might also be the last, unless we learn how
to avoid the risks. (Hawking, Russell, Tegmark, & Wilczek, 2014; cf. Price, 2013).
So, we decided to ask the experts what they predict the future holds knowing
that predictions on the future of AI are often not too accurate (see Armstrong, Sotala,
& O Heigeartaigh, 2014) and tend to cluster around in 25 years or so, no matter at
what point in time one asks.1
2. Questionnaire
2.1. Respondents
The questionnaire was carried out online by invitation to particular individuals from
four different groups for a total of ca. 550 participants (see Appendix 2). Each of the
participants got an email with a unique link to our site to fill in an online form (see
Appendix 1). If they did not respond within 10 days, a reminder was sent, and another
10 days later, with the note that this is the last reminder. In the case of EETN (see
below) we could not obtain the individual email addresses and thus sent the request
and reminders to the members mailing list. Responses were made on a single web
page with one submit button that only allowed submissions through these unique
links, thus making noninvited responses extremely unlikely. The groups we asked
were:
1. PTAI: Participants of the conference on Philosophy and Theory of AI,
Thessaloniki October 2011, organized by one of us (see Mller, 2012, 2013).
Participants were asked in November 2012, i.e. over a year after the event.
The total of 88 participants include a workshop on The Web and Philoso-
phy (ca. 15 people), from which a number of nonrespondents came. A list
of participants is on: http://www.ptai.org/2011/registeredparticipants
The questionnaire was sent with our names on it and with an indication that we
would use it for this paper and Nick Bostroms new book on superintelligence
(Bostrom, 2014) our request email is in Appendix 1. Given that the respondent groups
1 and 2 attended conferences organized by us, they knew whom they were responding
to. In groups 3 and 4 we would assume that the majority of experts would not know
us, or even of us. These differences are reflected in the response rates.
These groups have different theoretical-ideological backgrounds: The partici-
pants of PTAI are mostly theoryminded, mostly do not do technical work, and of-
ten have a critical view on large claims for easy progress in AI (Herbert Dreyfus was a
keynote speaker in 2011). The participants of AGI are committed to the view that AI
research should now return from technical details to artificial general intelligence
thus the name AGI. The vast majority of AGI participants do technical work. The
EETN is a professional association in Greece that accepts only published researchers
from AI. The TOP100 group also works mostly in technical AI; its members are sen-
ior and older than the average academic; the USA is strongly represented.
Several individuals are members of more than one of these four sets and they
were unlikely to respond to the same questionnaire more than once. So, in these cases,
we sent the query only once, but counted a response for each set i.e. we knew which
individuals responded from the individual tokens they received (except in the case of
EETN).
Future Progress in Artificial Intelligence: A Poll Among Experts 4/19
2.3. Methodology
In this field, it is hard to ask questions that do not require lengthy explanations or
generate resistance in certain groups of potential respondents (and thus biased results).
It is not clear what constitutes intelligence or progress and whether intelligence can
be measured or at least compared as more or less as a single dimension. Further-
more, for our purposes we need a notion of intelligence at a level that may surpass
humans or where technical intelligent systems might contribute significantly to re-
search but humanlevel intelligence is a rather elusive notion that generates re-
sistance. Finally, we need to avoid using terms that are already in circulation and
would thus associate the questionnaire with certain groups or opinions, like artificial
intelligence, singularity, artificial general intelligence or cognitive system.
For these reasons, we settled for a definition that a) is based on behavioral ability,
b) avoids the notion of a general humanlevel and c) uses a newly coined term. We
put this definition in the preamble of the questionnaire: Define a highlevel machine
intelligence (HLMI) as one that can carry out most human professions at least as well as
a typical human. (We still had one expert writing back to us that they could not say
what a typical human is though they could be convinced to respond, after all.) In
hindsight, it may have been preferable to specify what we mean by most and wheth-
er we think of most professions or of the professions most working people do. One
merit of our behavioral question is that having HLMI in our sense very likely implies
being able to pass a classic Turing test.
To achieve a high response rate, we tried to have few questions with simple
choices and eventually settled for four questions, plus three on the respondents. We
tried to choose questions that would allow us to compare our results with those of ear-
lier questionnaires see below.
In order to improve on the quality of predictions, we tried to prime respondents
into thinking about what is involved in reaching HLMI before asking when they expect
this. We also wanted to see whether people with a preference for particular approach-
es to HLMI would have particular responses to our central questions on prediction
(e.g. whether people who think that embodied systems are crucial expect longer than
average time to HLMI). For these two purposes, we inserted a first question about
contributing research approaches with a list to choose from the options that were
Future Progress in Artificial Intelligence: A Poll Among Experts 5/19
given are an eclectic mix drawn from many sources, but the particular options are not
of much significance.
2 A further, more informal, survey was conducted in August 2007 by Bruce J Klein (then of
Novamente and the Singularity Institute) on the timeframe for when we may see greater
thanhuman level AI, with a few numerical results and interesting comments, archived on
https://web.archive.org/web/20110226225452/http://www.novamente.net/bruce/?p=54.
Future Progress in Artificial Intelligence: A Poll Among Experts 6/19
3.) The earliest that machines will be able to simulate learning and every other
aspect of human intelligence:
Within 10 years 6 5%
Between 11 and 25 years 3 2%
Between 26 and 50 years 14 11%
More than 50 years 50 41%
Never 50 41%
Totals 123 100%
5.) The earliest we will understand the basic operations (mental steps) of the
human brain sufficiently to create machine simulation of human thought is:
today (we already understand enough) 5 6%
within the next 10 years 11 12%
within the next 25 years 9 10%
within the next 50 years 19 21%
within the next 100 years or more 26 29%
never (we will never understand enough) 19 21%
Totals 89 100%
6.) The earliest we will understand the architecture of the brain (how its organi-
zational control is structured) sufficiently to create machine simulation of hu-
man thought is:
Within 10 years 12 11%
Between 11 and 25 years 15 14%
Between 26 and 50 years 24 22%
More than 50 years 44 40%
Never 15 14%
Totals 110 100%
3. Baum et al. asked for the ability to pass a Turing test, a third grade school year ex-
am [i.e. for 9 year olds] and do Nobel Prize level research. They assume that all and
only the intelligent behavior of humans is captured in the Turing test. The results they
Future Progress in Artificial Intelligence: A Poll Among Experts 7/19
got for the 50% probability point were: 2040 (Turing test), 2030 (third grade), and
2045 (Nobel).
4. Sandberg and Bostroms first question was quite similar to our 2nd (see below):
Assuming no global catastrophe halts progress, by what year would you assign a
10%/50%/90% chance of the development of humanlevel machine intelligence?
The median estimate of when there will be 50% chance of humanlevel machine in-
telligence was 2050. So, despite significant overlap with AGI 2009, the group asked by
Sandberg and Bostrom in 2011 was a bit more guarded in their expectations.
We think it is worthwhile to make a new attempt because the prior ones asked
specific groups and small samples, sometimes have methodological problems, and we
also want to see how the answers change over time, or do not change which is why
tried to use similar questions. As explained below, we also think it might be worth-
while to repeat our questionnaire at a later stage, to compare results.
The percentages here are over the total of responses. There were no significant differ-
ences between groups here, except that Whole brain emulation got 0% in TOP100,
but 46% in AGI. We did also not find relevant correlations between the answers given
here and the predictions made in the following questions (of the sort that, for example,
people who think embodied systems crucial would predict later onset of HLMI).
Future Progress in Artificial Intelligence: A Poll Among Experts 9/19
Cognitive science
Integrated cognitive architectures
Algorithms revealed by computational
Artificial neural networks
Faster computing hardware
Large-scale datasets
Embodied systems
Other method(s) currently completely
Whole brain emulation
Evolutionary algorithms or systems
Other method(s) currently known to at
Logic-based systems
Algorithmic complexity theory
No method will ever contribute to this
Swarm intelligence
Robotics
Bayesian nets
Clicks of the never box. These answers did not enter in to the averages above.
Never no. %
10% 2 1.2
50% 7 4.1
90% 28 16.5
1
0.9
0.8
0.7
0.6
10%
0.5
50%
0.4
90%
0.3
0.2
0.1
0
2000
2050
2100
2150
2200
2250
2300
For the 50% mark, the overall median is 2040 (i.e. half of the respondents gave a year
earlier than 2040 and half gave a year later than 2040) but the overall mean (average)
is 2081. The median is always lower than the mean here because there cannot be out-
liers towards earlier but there are outliers towards later (the maximum possible se-
lection was 5000, then never).
2 years 30 years
PT-AI 10% 60%
AGI 15% 90%
EETN 5% 55%
TOP100 5% 50%
Experts allocate a low probability for a fast takeoff, but a significant probability for
superintelligence within 30 years after HLMI.
Percentages here are means, not medians as in the other tables. There is a notable
difference here between the theoretical (PT-AI and AGI) and the technical groups
(EETN and TOP100).
1. Concerning the above questions, how would you describe your own exper-
tise? (0 = none, 9 = expert)
Mean 5.85
2. Concerning technical work in artificial intelligence, how would you describe
your own expertise? (0 = none, 9 = expert)
Mean 6.26
3. What is your main home academic discipline? (Select from list with 8 op-
tions: Biology/Physiology/Neurosciences Computer Science Engineering
[non CS] Mathematics/Physics Philosophy Psychology/Cognitive Sci-
ence Other academic discipline None.) [Absolut numbers.]
a. Biology/Physiology/Neurosciences 3
b. Computer Science 107
c. Engineering (non CS) 6
d. Mathematics/Physics 10
e. Philosophy 20
f. Psychology/Cognitive Science 14
g. Other academic discipline 9
h. None 1
And we finally invited participants to make a comment, plus a possibility to add their
name, if they wished. (We cannot reproduce these here; but they are on our site, see
below). A number of comments concerned the difficulty of formulating good ques-
tions, much fewer the difficulty of predicting.
4. Evaluation
have the data to show us who responded and who did not. We got one additional re-
sponse from PT-AI and two from TOP100 in this way.
For question 2 By what year would you see a (10% / 50% / 90%) probabil-
ity for such HLMI to exist? we compared the additional responses to the responses
we already had from the same respective group (PT-AI and TOP100, respectively).
We found the following differences:
The one additional respondent from PT-AI expected HLMI earlier than the mean
but later than the median, while the two respondents from TOP100 (last row) ex-
pected HLMI earlier than mean and median. The very small sample forbids confident
judgment, but we found no support for the worry that the non-respondents would
have been biased towards a later arrival of HLMI.
thereafter. The experts say the probability is 31% that this development turns out to
be bad or extremely bad for humanity.
So, the experts think that superintelligence is likely to come in a few decades and
quite possibly bad for humanity this should be reason enough to do research into
the possible impact of superintelligence before it is too late. We could also put this
more modestly and still come to an alarming conclusion: We know of no compelling
reason to say that progress in AI will grind to a halt (though deep new insights might
be needed) and we know of no compelling reason that superintelligent systems will be
good for humanity. So, we should better investigate the future of superintelligence and
the risks it poses for humanity.
5. Acknowledgements
Toby Ord and Anders Sandberg were helpful in the formulation of the questionnaire.
The technical work on the website form, sending mails and reminders, database and
initial data analysis was done by Ilias Nitsos (under the guidance of VCM). Theo
Gantinas provided the emails of the TOP100. Stuart Armstrong made most graphs
for presentation. The audience at the PT-AI 2013 conference in Oxford provided
helpful feedback. Mark Bishop, Carl Shulman, Miles Brundage and Daniel Dewey
made detailed comments on drafts. We are very grateful to all of them.
6. References
Adams, S., Arel, I., Bach, J., Coop, R., Furlan, R., Goertzel, B., . . . Sowa, J. F.
(2012). Mapping the landscape of human-level artificial general intelligence.
AI Magazine, 33(1), 25-42.
Armstrong, S., Sotala, K., & O Heigeartaigh, S. (2014). The errors, insights and
lessons of famous AI predictions and what they mean for the future. Journal
of Experimental and Theoretical Artificial Intelligence, 26(3 - Special issue Risks of
General Artificial Intelligence, ed. V. Mller), 317-342.
Baum, S. D., Goertzel, B., & Goertzel, T. G. (2011). How long until human-level AI?
Results from an expert assessment. Technological Forecasting & Social Change,
78(1), 185-195.
Bostrom, N. (2006). How long before superintelligence? Linguistic and Philosophical
Investigations, 5(1), 11-30.
Bostrom, N. (2013). Existential risk prevention as global priority. Global Policy, 4(1), 15-
31.
Bostrom, N. (2014). Superintelligence: Paths, dangers, strategies. Oxford: Oxford University
Press.
Dreyfus, H. L. (1972). What computers still can't do: A critique of artificial reason (2 ed.).
Cambridge, Mass.: MIT Press.
Dreyfus, H. L. (2012). A history of first step fallacies. Minds and Machines, 22(2 - special
issue "Philosophy of AI" ed. Vincent C. Mller), 87-99.
Future Progress in Artificial Intelligence: A Poll Among Experts 16/19
Hawking, S., Russell, S., Tegmark, M., & Wilczek, F. (2014). Transcendence looks at
the implications of artificial intelligence - but are we taking AI seriously
enough? The Independent, 01.05.2014.
Kurzweil, R. (2005). The singularity is near: When humans transcend biology. London:
Viking.
Lighthill, J. (1973). Artificial intelligence: A general survey Artificial intelligence: A paper
symposion. London: Science Research Council.
McCarthy, J., Minsky, M., Rochester, N., & Shannon, C. E. (1955). A proposal for
the Dartmouth summer research project on artificial intelligence. Retrieved
October 2006, from http://www-
formal.stanford.edu/jmc/history/dartmouth/dartmouth.html
Michie, D. (1973). Machines and the theory of intelligence. Nature, 241(23.02.1973),
507-512.
Moor, J. H. (2006). The Dartmouth College artificial intelligence conference: The
next fifty years. AI Magazine, 27(4), 87-91.
Mller, V. C. (2007). Is there a future for AI without representation? Minds and
Machines, 17(1), 101-115.
Mller, V. C. (2014a). Editorial: Risks of general artificial intelligence. Journal of
Experimental and Theoretical Artificial Intelligence, 26(3 - Special issue Risks of
General Artificial Intelligence, ed. V. Mller), 1-5.
Mller, V. C. (Ed.). (2012). Theory and philosophy of AI (Minds and Machines, Vol. 22/2-
Special volume): Springer.
Mller, V. C. (Ed.). (2013). Theory and philosophy of artificial intelligence (SAPERE, Vol. 5).
Berlin: Springer.
Mller, V. C. (Ed.). (2014b). Risks of artificial general intelligence (Journal of Experimental and
Theoretical Artificial Intelligence, Vol. (26/3) Special Issue): Taylor & Francis.
Price, H. (2013). Cambridge, Cabs and Copenhagen: My Route to Existential Risk.
The New York Times, 27.01.2013.
http://opinionator.blogs.nytimes.com/2013/01/27/cambridge-cabs-and-
copenhagen-my-route-to-existential-risk/?_php=true&_type=blogs&_r=0
Sandberg, A., & Bostrom, N. (2011). Machine intelligence survey. FHI Technial Report,
2011(1). http://www.fhi.ox.ac.uk/research/publications/
7. Appendices
1. Questionnaire
2. Letter sent to participants
Future Progress in Artificial Intelligence: A Poll Among Experts 17/19
This brief questionnaire is directed towards researchers in artificial intelligence or the theory of
artificial intelligence. It aims to gauge how people working in the field view progress towards its
original goals of intelligent machines, and what impacts they would associate with reaching
these goals.
Answers will be anonymized. Results will be made publicly available on the site of the Programme on the Impacts of Future Technology:
http://www.futuretech.ox.ac.uk (http://www.futuretech.ox.ac.uk) .
A. The Future of AI
Define a "high-level machine intelligence" (HLMI) as one that can carry out most human professions at least as well as a typical human.
1. In your opinion, what are the research approaches that might contribute the most to the development of such HLMI?:
2. Assume for the purpose of this question that human scientific activity continues without major negative disruption. By what year would
you see a 10%/50%/90% probability for such HLMI to exist?
Year reached: - - -
Never:
3. Assume for the purpose of this question that such HLMI will at some point exist. How likely do you then think it is that within (2 years / 30
years) thereafter, there will be machine intelligence that greatly surpasses the performance of any human in most professions?
Probability: - -
% %
4. Assume for the purpose of this question that such HLMI will at some point exist. How positive or negative would be the overall impact on
humanity, in the long run? Please indicate a probability for each option. (The sum should be equal to 100%.)
Extremely good On balance good More or less neutral On balance bad Extremely bad (existential catastrophe)
- % - % - % - % - %
Total: 0%
B. About you
1. Concerning the above questions, how would you describe your own expertise?:
0 = none 1 2 3 4 5 6 7 8 9 = expert
2. Concerning technical work in artificial intelligence, how would you describe your own expertise?:
0 = none 1 2 3 4 5 6 7 8 9 = expert
Biology/Physiology/Neurosciences
Computer Science
Engineering (non CS)
Mathematics/Physics
Philosophy
Psychology/Cognitive Science
Other academic discipline
None
4. Add a brief comment, if you like (<250 words). These comments may be published. Please indicate whether you would like your name to
be included with the comment. (The answers above will remain anonymous in any case.):
Please include my name with the comment (leave this field empty if you wish to remain anonymous):
CAPTCHA
This question is for testing whether you are a human visitor and to prevent automated spam submissions.
Senden
given your prominence in the field of artificial intelligence we invite you to express
your views on the future of artificial intelligence in a brief questionnaire. The aim of
this exercise is to gauge how the top 100 cited people working in the field view pro-
gress towards its original goals of intelligent machines, and what impacts they would
associate with reaching these goals.
The questionnaire has 4 multiple choice questions, plus 3 statistical data points on the
respondent and an optional 'comments' field. It will only take a few minutes to fill in.
Of course, this questionnaire will only reflect the actual views of researchers if we get
nearly everybody to express their opinion. So, please do take a moment to respond,
even (or especially) if you think this exercise is futile or misguided.
Answers will be anonymous. Results will be used for Nick Bostrom's forthcoming book
Superintelligence: Paths, Dangers, Strategies (Oxford University Press, 2014) and
made publicly available on the site of the Programme on the Impacts of Future Tech-
nology: http://www.futuretech.ox.ac.uk.