Artificial Intelligence principles define social and ethical considerations to develop future AI. They come from research institutes, government organizations and industries. All versions of AI principles are with different considerations covering different perspectives and making different emphasis. None of them can be considered as complete and can cover the rest AI principle proposals. Here we introduce LAIP, an effort and platform for linking and analyzing different Artificial Intelligence Principles. We want to explicitly establish the common topics and links among AI Principles proposed by different organizations and investigate on their uniqueness. Based on these efforts, for the long-term future of AI, instead of directly adopting any of the AI principles, we argue for the necessity of incorporating various AI Principles into a comprehensive framework and focusing on how they can interact and complete each other.
AI ethics and social impacts have drawn serious attentions and lots of policy frameworks have been brought up by various organizations. We confine our study to different AI principles (including guidelines, codes, and initiatives) pertaining to the general governance of AI. Typically, such principles are literally and explicitly documented in an itemby-item style, announced as an efforts to express the proposers’ values and attitudes towards the understanding, development, and utilization of AI. Technically detailed discussions, including techniques oriented standards, are not included in this study. Traditional principles on robotics are also not included in this study.
Based on these considerations, we collected 27 proposals of AI principles to date. For each of the collected principles, we extract the texts of direct relevance to the author’s points (in most cases this means the title words of the principles). We also include the necessary comments from the raw text * These authors contributed equally to this study. 1 OpenAI identifies itself as “a non-profit AI research company”.
Semantically Linking Various AI Principles We aim to link various AI principles from the perspectives that they considered in common. Common perspectives may not use exactly the same word term, and semantically equivalent 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 deviated semantic meaning from the original keyword is selected as the threshold point, and all words with lower similarity are abandoned. Some phrases with similar meanings are added to the expanded keyword list. For example, for the term “collaboration”, the expanded list also includes collaborations, collaborative, collaboratively, collaborate, collaborates and collaborating. While for the term “fairness”, the expanded list also includes fair, fairer, unfair and unfairness.
Keywords humanity, beneficial, well-being, human value, human right, dignity, freedom, education, common good, human-centered, human-friendly collaboration, partnership, cooperation, dialogue share, equal, equity, inequity, inequality fairness, justice, bias, discrimination, prejudice transparency, explainable, predictable, intelligible, audit, trace, opaque privacy, personal information, data protection, informed, explicit confirmation, control the data, notice and consent security, cybersecurity, cyberattack, hacks, confidential safety, validation, verification, test, controllability, under control, control the risks, human control accountability, responsibility Accountability AGI/ASI AGI, superintelligence, super intelligence Here we define the topic coverage of a principle proposal as the percentage of topics that have been mentioned in the proposal. If any term or expanded keyword term has ever appeared in a proposal, we would mark that this proposal 1 https://drive.google.com/file/d/0B7XkCwpI5KDYNlNUTTlSS21pQmM/edit has covered the related topic. Table 1 presents 10 general topics and related terms for AI Principles. Term expansion efforts based on semantic similarities are introduced to extend the list for more comprehensive coverage.
Figure 1 shows the coverage of different principles on the 10 topics. The colors are related to how many times the term appeared in the proposal. As can be observed, expanding the keywords using semantic similarity significantly increased topics found in principles, making the semantic analysis more accurate and robust against different use of similar word terms and expressions. The linkages among different AI principles are represented using Semantic Web standards (RDF/OWL) on the LAIP platform.
Complementary Considerations from
Different Organizations and Different AI Principles
Different principle proposals are compared by calculating
their coverage on topics and keywords, as shown in Figure
2. We can observe that one of the principle proposals
covered all the major topics. Among the top 10 proposals ranked
by keywords, 8 of them ranked top 10 on topic coverage
ranking as well. However,
(A) Topic Coverage Ranking (B) Keywords Coverage Ranking
According to the division of different school of thoughts from the types of publisher point of view, Figure 3 shows the comparative frequency of topics mentioned in three different types of AI principle proposals.
We can observe from Figure 3 that corporations would like to mention more about collaboration, but not that much for security and privacy. While governments mentioned more about security, but would not like to mention accountability. Corporations can benefit from collaboration, but the atmosphere of collaboration may not be as good as academia, which may be the reason why they would like to mention it. Privacy and security are sensitive issues for corporations, maybe that is why corporations would not like to mention them. And the government mentioned the topic of accountability significantly less than academia.
Although in most cases, principles from different
organizations usually share a common vocabulary, ambiguities in
the analysis of the text still remain. The ambiguities may
come from the polysemy of words and the context. For
example, “race” is used in the context of “arms race” and “race
avoiding”
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
Besides the general topics those AI Principle proposals
share in common, many principles also reflect the unique
perspectives of different organizations. For example, the
Montreal Declaration has suggested promoting the
well-being of “all sentient creatures”, which according to their
definition, includes “any being able to feel pleasure, pain,
emotions; basically, to feel”
Based on the analysis, we have the following suggestions for future research and proposals for AI Principles: • Strengthening safety-related considerations in academia and industry. Safety issues are the core for AI governance and have been realized in different government organizations, but many of the AI companies have not taken this seriously. While their AI products will directly bring potential risks for society. • Long-term strategic design for AGI and ASI. Most AI principles investigated here do not cover considerations for AGI and ASI. While most of them should have been regarded as relatively long-term design for AI. Long-term planning on AGI and ASI will have clearer observations for strategic future and could have arrangements for potential risks in advance. • From Human-centered to Harmonious Principle Design.
Current AI principle proposals mainly focus on beneficial, human-centered design, while lack of considerations that the human society is on the way for transformation. More harmonious design considering both human and future AI as cognitive living systems should be considered.
Different AI Principles have their own perspectives and coverage for the current and future strategies of AI. Instead of directly adopting any of the AI principles, we argue the necessity of linking and incorporating various AI Principles into a comprehensive framework and focusing on how they can interact and complement each other. The Linking Artificial Intelligence Principles (LAIP) platform is available as an online service under the address http://www.linking-aiprinciples.org. It supports semantic search by keyword terms and paragraph search where semantically similar principles could be listed for exploration.
This study is supported by New Generation of Artificial Intelligence Development Research Center, Ministry of Science and Technology of China under the project “Key issues of Social Ethics for Artificial Intelligence” from ISTIC.