Hanson AI Expert Survey
Published 29 December, 2014; last updated 10 December, 2020
In a small informal survey running since 2012, AI researchers generally estimated that their subfields have moved less than ten percent of the way to human-level intelligence. Only one (in the slowest moving subfield) observed acceleration.
This suggests on a simple extrapolation that reaching human-level capability across subfields will take over a century (in contrast with many other predictions).
Details
Robin Hanson has asked experts in various social contexts to estimate how far we’ve come in their own subfield of AI research in the last twenty years, compared to how far we have to go to reach human level abilities. His results are listed in Table 1. He points out that on an outside view calculation, this suggests at least a century until human-level AI.
| Year added to list | Person | Subfield | Distance in 20y | Acceleration |
| 2012 | A few UAI attendees | 5-10% | ~0 | |
| 2012 | Melanie Mitchell | Analogical reasoning | 5% | ~0 |
| 2012 | Murray Shanahan | Knowledge representation | 10% | ~0 |
| 2013 | Wendy Hall | Computer-assisted training | 1% | |
| 2013 | Claire Cardie (and Peter Norvig agrees in ’14) | Natural language processing | 20% | |
| 2013 | Boi Faltings (and Peter Norvig agrees in ’14) | Constraint satisfaction | Past human-level 20 years ago | |
| 2014 | Aaron Dollar | robotic grasping manipulation | <1% | positive |
| 2014 | Peter Norvig | * | ||
| 2014 | Timothy Meese | early human vision processing | 5% | negative |
| 2015 | Francesca Rossi | constraint reasoning | 10% | negative |
| 2015 | Margret Boden | no particular subfield | 5% | |
| 2015 | David Kelley | big data analysis | 5% | positive |
| 2016 | Henry Kautz | constraint satisfaction | >100% | |
| 2016 | Henry Kautz | language | 10% | positive |
| 2016 | Jeff Legault | robotics | 5% | positive |
| 2017 | Thore Husfeldt | human-understandable explanation | <0.5% |
Table 1 : Results from Robin Hanson’s informal survey
*Hanson’s summary of Peter Norvig’s response seems hard to fit into this framework:
After coming to a talk of mine, Peter Norvig told me that he agrees with both Claire Cardie and Boi Faltings, that on speech recognition and machine translation we’ve gone from not usable to usable in 20 years, though we still have far to go on deeper question answering, and for retrieving a fact or page that is relevant to a search query we’ve far surpassed human ability in recall and do pretty well on precision.
