=Paper=
{{Paper
|id=Vol-2301/paper15
|storemode=property
|title=Linking Artificial Intelligence Principles
|pdfUrl=https://ceur-ws.org/Vol-2301/paper_15.pdf
|volume=Vol-2301
|authors=Yi Zeng,Enmeng Lu,Cunqing Huangfu
|dblpUrl=https://dblp.org/rec/conf/aaai/ZengLH19
}}
==Linking Artificial Intelligence Principles==
https://ceur-ws.org/Vol-2301/paper_15.pdf
Linking Artificial Intelligence Principles
Yi Zeng1,2,3, Enmeng Lu1,*, Cunqing Huangfu1,*
1
Institute of Automation, Chinese Academy of Sciences
2
School of Artificial Intelligence, University of Chinese Academy of Sciences
3
Berggruen Institute China Center, Peking University
{yi.zeng, enmeng.lu, cunqing.huangfu}@ia.ac.cn
Abstract to make the extracted text self-explanatory. Principle pro-
Artificial Intelligence principles define social and ethical posals are grouped by their backgrounds:
considerations to develop future AI. They come from re- • Principles from Academia, Non-profits and Non-Gov-
search institutes, government organizations and industries. ernmental Organizations: (1) Asilomar AI Principles (FLI
All versions of AI principles are with different considerations
2017). (2) General Principles in Ethically Aligned Design,
covering different perspectives and making different empha-
sis. None of them can be considered as complete and can version 2, by IEEE (IEEE 2017). (3) Principles for Algorith-
cover the rest AI principle proposals. Here we introduce mic Transparency and Accountability by ACM (USACM
LAIP, an effort and platform for linking and analyzing differ- 2017). (4) The Japanese Society for Artificial Intelligence
ent Artificial Intelligence Principles. We want to explicitly Ethical Guidelines (JSAI 2017). (5) The Montreal Declara-
establish the common topics and links among AI Principles
tion for a Responsible Development of Artificial Intelli-
proposed by different organizations and investigate on their
uniqueness. Based on these efforts, for the long-term future gence (Montreal 2017). (6) Three ideas from the Stanford
of AI, instead of directly adopting any of the AI principles, Human-Centered AI Initiative (HAI) (Stanford 2018). (7)
we argue for the necessity of incorporating various AI Prin- Three Rules for Artificial Intelligence Systems by the CEO
ciples into a comprehensive framework and focusing on how of Allen Institute for Artificial Intelligence (Etzioni 2017).
they can interact and complete each other.
(8) Harmonious Artificial Intelligence Principles (HAIP
2018). (9) Universal Guidelines for Artificial Intelligence
Artificial Intelligence Principles: Different (The Public Voice 2018). (10) Principles for the Governance
of AI (The Future Society 2017). (11) Tenets of Partnership
School of Thoughts on AI (PAI 2016). (12) Top 10 Principles For Ethical Arti-
AI ethics and social impacts have drawn serious attentions ficial Intelligence (UNI Global Union 2017). (13) AI Policy
and lots of policy frameworks have been brought up by var- Principles (ITI 2017).
ious organizations. We confine our study to different AI • Principles from Governments: (14) AI R&D Principles
principles (including guidelines, codes, and initiatives) per- (MIC 2017). (15) Draft AI Utilization Principles (MIC
taining to the general governance of AI. Typically, such 2018). (16) AI Code (House of Lords 2018). (17) Ethical
principles are literally and explicitly documented in an item- principles and democratic prerequisites, European Group on
by-item style, announced as an efforts to express the propos- Ethics in Science and New Technologies (EGE 2018).
ers’ values and attitudes towards the understanding, devel- • Principles from Industry: (18) DeepMind Ethics & So-
opment, and utilization of AI. Technically detailed discus- ciety Principles (DeepMind 2017). (19) OpenAI1 Charter
sions, including techniques oriented standards, are not in- (OpenAI 2018). (20) AI at Google: Our Principles (Google
cluded in this study. Traditional principles on robotics are 2018). (21) Microsoft AI Principles (Microsoft 2018). (22)
also not included in this study. Microsoft CEO’s 10 AI rules (Nadella 2016). (23) Princi-
Based on these considerations, we collected 27 proposals ples for the Cognitive Era (IBM 2017). (24) Principles for
of AI principles to date. For each of the collected principles, Trust and Transparency (IBM 2018). (25) Developing AI
we extract the texts of direct relevance to the author’s points for Business with Five Core Principles (Sage 2017). (26)
(in most cases this means the title words of the principles). SAP’s Guiding Principles for Artificial Intelligence. (SAP
We also include the necessary comments from the raw text 2018). (27) Sony Group AI Ethics Guidelines (Sony 2018).
1
* These authors contributed equally to this study. OpenAI identifies itself as “a non-profit AI research company”.
Semantically Linking Various AI Principles has covered the related topic. Table 1 presents 10 general
topics and related terms for AI Principles. Term expansion
We aim to link various AI principles from the perspectives efforts based on semantic similarities are introduced to ex-
that they considered in common. Common perspectives may tend the list for more comprehensive coverage.
not use exactly the same word term, and semantically equiv-
alent and similar terms should be considered.
We first identified a set of manually chosen keywords as
the core terms, which belong to 10 general topics. We use
word2vec representation of the word to find keywords with
similar meanings. Google word vector trained from news1 is
used. The similarity between the original keyword and the
other words is calculated by the cosine similarity between
the word vector of the original keyword and the other words.
A list of candidate extended keywords ranked by similarity
is generated. The first word on the list with obviously devi-
ated semantic meaning from the original keyword is selected
as the threshold point, and all words with lower similarity
Figure 1. Topic coverage of principle titles and explanatory texts
are abandoned. Some phrases with similar meanings are
based on manually chosen keywords (A) and extended keyword
added to the expanded keyword list. For example, for the
groups by semantic similarity (B).
term “collaboration”, the expanded list also includes collab-
orations, collaborative, collaboratively, collaborate, collab- Figure 1 shows the coverage of different principles on the
orates and collaborating. While for the term “fairness”, the 10 topics. The colors are related to how many times the term
expanded list also includes fair, fairer, unfair and unfairness. appeared in the proposal. As can be observed, expanding the
Table 1.Topics and Manually Chosen Keywords for AI Principles
keywords using semantic similarity significantly increased
topics found in principles, making the semantic analysis
Topics Keywords more accurate and robust against different use of similar
Humanity humanity, beneficial, well-being, human word terms and expressions. The linkages among different
value, human right, dignity, freedom, edu- AI principles are represented using Semantic Web standards
cation, common good, human-centered, (RDF/OWL) on the LAIP platform.
human-friendly
Collabora- collaboration, partnership, cooperation,
tion dialogue Complementary Considerations from Differ-
Share share, equal, equity, inequity, inequality ent Organizations and Different AI Principles
Fairness fairness, justice, bias, discrimination, prej-
udice Different principle proposals are compared by calculating
Transpar- transparency, explainable, predictable, in- their coverage on topics and keywords, as shown in Figure
ency telligible, audit, trace, opaque 2. We can observe that one of the principle proposals cov-
Privacy privacy, personal information, data pro- ered all the major topics. Among the top 10 proposals ranked
tection, informed, explicit confirmation, by keywords, 8 of them ranked top 10 on topic coverage
control the data, notice and consent ranking as well. However, SAP 2018 ranked higher on key-
Security security, cybersecurity, cyberattack, words coverage ranking (the 10th), but ranked comparatively
hacks, confidential lower on topic coverage ranking (the 14th in parallel), since
Safety safety, validation, verification, test, con- it discussed extensively about collaboration, fairness, pri-
trollability, under control, control the vacy, and safety, while may have missed the topics of share,
risks, human control accountability and AGI/ASI. HAIP 2018 covered 8 of the
10 major topics (the 7th in parallel) without going through
Accounta- accountability, responsibility
much of the details, hence ranked lower in keywords rank-
bility
ing (the 16th). We should emphasize that coverage of a pro-
AGI/ASI AGI, superintelligence, super intelligence
posal may not reflect lacking of considerations on certain
Here we define the topic coverage of a principle proposal topics, but just reflects that they may choose to have differ-
as the percentage of topics that have been mentioned in the ent emphasis. On the other hand, different considerations
proposal. If any term or expanded keyword term has ever may interact to complement with each other.
appeared in a proposal, we would mark that this proposal
1
https://drive.google.com/file/d/0B7XkCwpI5KDYNlNUT-
TlSS21pQmM/edit
the analysis of the text still remain. The ambiguities may
come from the polysemy of words and the context. For ex-
ample, “race” is used in the context of “arms race” and “race
avoiding” (FLI 2017) to represent the competition across re-
searchers and nations (thus referring to the topic of “collab-
oration”), it is also used in the context of “gender, race, sex-
ual orientation” (UNI Global Union 2017) to talk about pos-
sible biases of AI system (thus referring to the topic of “fair-
ness”). Meanwhile, the “self-improvement” of an advanced
AI system is a trait we should be very cautious about (FLI
(A) Topic Coverage Ranking 2017), yet such “self-improvement” of AI researchers is
what we ask for (JSAI 2017).
Such ambiguities also appear within a topic. For instance,
we may ask for “transparency” from the decision-making
process of the system out of our fairness concerns. We may
also ask for “transparency” from the system to make it more
safe, traceable, and controllable. The Asilomar AI principles
have made such distinctions explicitly in their discussions
(see “Judicial Transparency” and “Failure Transparency” in
(B) Keywords Coverage Ranking (FLI 2017)) while others usually seem to take one side of
the concept or mixed them up. The ambiguities in these
Figure 2. Coverage Ranking for General Topics (A) and Key-
cases can be derived from the high-level abstraction of the
words (B) for Different AI Principle Proposals
concept itself and is also a reflection of the inner linkage
According to the division of different school of thoughts between various topics.
from the types of publisher point of view, Figure 3 shows Besides the general topics those AI Principle proposals
the comparative frequency of topics mentioned in three dif- share in common, many principles also reflect the unique
ferent types of AI principle proposals. perspectives of different organizations. For example, the
Montreal Declaration has suggested promoting the well-be-
ing of “all sentient creatures”, which according to their def-
inition, includes “any being able to feel pleasure, pain, emo-
tions; basically, to feel” (Montreal 2017). The JSAI Ethical
Guidelines include that AI must abide these guidelines “in
the same manner as the members of the JSAI in order to be-
come a member or a quasi-member of society” (JSAI 2017).
The General Principles from IEEE’s report recommend that
“For the foreseeable future, A/IS should not be granted
rights and privileges equal to human rights: A/IS should al-
Figure 3. Average topic frequency in different types of publishers, ways be subordinate to human judgment and control” (IEEE
with standard error of the data. The asterisks indicate that the T- 2017). IBM takes the view that “Cognitive systems will not
test p-value of the data is less than 0.05. realistically attain consciousness or independent agency”
We can observe from Figure 3 that corporations would and thus lay their stress on promoting AI and cognitive sys-
like to mention more about collaboration, but not that much tems to “augment human intelligence” (IBM 2017). Those
for security and privacy. While governments mentioned different perspectives from different proposals reflect the di-
more about security, but would not like to mention account- versity of the whole AI community and it turns to be neces-
ability. Corporations can benefit from collaboration, but the sary to identify and incorporate such various considerations
atmosphere of collaboration may not be as good as academia, for a more comprehensive framework.
which may be the reason why they would like to mention it. Based on the analysis, we have the following suggestions
Privacy and security are sensitive issues for corporations, for future research and proposals for AI Principles:
maybe that is why corporations would not like to mention • Strengthening safety-related considerations in academia
them. And the government mentioned the topic of account- and industry. Safety issues are the core for AI governance
ability significantly less than academia. and have been realized in different government organiza-
Although in most cases, principles from different organi- tions, but many of the AI companies have not taken this
seriously. While their AI products will directly bring po-
zations usually share a common vocabulary, ambiguities in
tential risks for society.
• Long-term strategic design for AGI and ASI. Most AI IBM. 2017. Principles for the Cognitive Era. https://www.ibm
principles investigated here do not cover considerations .com/blogs/think/2017/01/ibm-cognitive-principles/.
for AGI and ASI. While most of them should have been IBM. 2018. Principles for Trust and Transparency. https://www
regarded as relatively long-term design for AI. Long-term .ibm.com/blogs/policy/trust-principles/.
planning on AGI and ASI will have clearer observations Information Technology Industry Council (ITI). 2017. AI Policy
for strategic future and could have arrangements for po- Principles. https://www.itic.org/public-policy/ITIAIPolicy
tential risks in advance. PrinciplesFINAL.pdf.
• From Human-centered to Harmonious Principle Design. Microsoft. 2018. Microsoft AI Principles. https://www.microsoft
Current AI principle proposals mainly focus on beneficial, .com/en-us/ai/our-approach-to-ai.
human-centered design, while lack of considerations that Ministry of Internal Affairs and Communications (MIC), the Gov-
the human society is on the way for transformation. More ernment of Japan. 2017. AI R&D Principles. http://www
harmonious design considering both human and future AI .soumu.go.jp/main_content/000507517.pdf.
as cognitive living systems should be considered. Ministry of Internal Affairs and Communications (MIC), the Gov-
ernment of Japan. 2018. Draft AI Utilization Principles.
http://www.soumu.go.jp/main_content/000581310.pdf.
Conclusions Nadella, S. 2016. The Partnership of the Future: Microsoft’s CEO
explores how humans and A.I. can work together to solve society’s
Different AI Principles have their own perspectives and cov- greatest challenges. https://slate.com/technology/2016/06
erage for the current and future strategies of AI. Instead of /microsoft-ceo-satya-nadella-humans-and-a-i-can-work-together
directly adopting any of the AI principles, we argue the ne- -to-solve-societys-challenges.html.
cessity of linking and incorporating various AI Principles OpenAI. 2018. OpenAI Charter. https://blog.openai.com/openai
into a comprehensive framework and focusing on how they -charter/.
can interact and complement each other. The Linking Arti- Partnership on AI (PAI). 2016. Tenets. https://www.partnership
ficial Intelligence Principles (LAIP) platform is available as onai.org/tenets/.
an online service under the address http://www.linking-ai- Sage. 2017. The Ethics of Code: Developing AI for Business with
Five Core Principles. https://www.sage.com/ca/our-news/press-re-
principles.org. It supports semantic search by keyword
leases/2017/06/designing-AI-for-business.
terms and paragraph search where semantically similar prin-
SAP. 2018. SAP's Guiding Principles for Artificial Intelligence.
ciples could be listed for exploration. https://news.sap.com/2018/09/sap-guiding-principles-for-artificial
-intelligence/.
Sony. 2018. Sony Group AI Ethics Guidelines. https://www.sony
Acknowledgement .net/SonyInfo/csr_report/humanrights/hkrfmg0000007rtj-att/AI
This study is supported by New Generation of Artificial In- _Engagement_within_Sony_Group.pdf.
telligence Development Research Center, Ministry of Sci- Stanford University. 2018. The Stanford Human-Centered AI Ini-
tiative (HAI). http://hai.stanford.edu/news/introducing_stanfords
ence and Technology of China under the project “Key issues
_human_centered_ai_initiative/.
of Social Ethics for Artificial Intelligence” from ISTIC.
The Future Society. 2017. Principles for the Governance of AI.
http://www.thefuturesociety.org/science-law-society-sls-initiative
/#1516790384127-3ea0ef44-2aae.
References
The IEEE Global Initiative on Ethics of Autonomous and Intelli-
DeepMind. 2017. DeepMind Ethics & Society Principles. gent Systems. 2017. Ethically Aligned Design, Version 2.
https://deepmind.com/applied/deepmind-ethics-society/principles/. http://standards.ieee.org/develop/indconn/ec/autonomous_
Etzioni, O. 2017. How to Regulate Artificial Intelligence. https: systems.html.
//www.nytimes.com/2017/09/01/opinion/artificial-intelligence- The Japanese Society for Artificial Intelligence (JSAI). 2017. The
regulations-rules.html. Japanese Society for Artificial Intelligence Ethical Guidelines.
European Group on Ethics in Science and New Technologies http://ai-elsi.org/wp-content/uploads/2017/05/JSAI-Ethical-
(EGE). 2018. Statement on Artificial Intelligence, Robotics and Guidelines-1.pdf.
'Autonomous' Systems. http://ec.europa.eu/research/ege/pdf/ege The Public Voice. 2018. Universal Guidelines for Artificial Intel-
_ai_statement_2018.pdf. ligence. https://thepublicvoice.org/ai-universal-guidelines/.
Future of Life Institute (FLI). 2017. Asilomar AI Principles. https: UNI Global Union. 2017. Top 10 Principles For Ethical Artificial
//futureoflife.org/ai-principles/. Intelligence. http://www.thefutureworldofwork.org/media/35420
Google. 2018. AI at Google: Our Principles. https://ai.google /uni_ethical_ai.pdf.
/principles. University of Montreal. 2017. The Montreal Declaration for a Re-
HAIP Initiative. 2018. Harmonious Artificial Intelligence Princi- sponsible Development of Artificial Intelligence. https://www
ples (HAIP). http://bii.ia.ac.cn/hai/index.php. .montrealdeclaration-responsibleai.com/the-declaration.
House of Lords, UK. 2018. AI in the UK: ready, willing and able? US Public Policy Council, Association for Computing Machinery
https://publications.parliaent.uk/pa/ld201719/ldselect/ldai/100 (USACM). 2017. Principles for Algorithmic Transparency and Ac-
/100.pdf. countability. https://www.acm.org/binaries/content/assets
/public-policy/2017_usacm_statement_algorithms.pdf.