A Unified Framework of Five Principles for AI in Society

A Unified Framework of Five Principles for AI in Society Luciano Floridi Josh Cowls

Tag layer of the DHSS Meta-Tagging Project. The article text is reproduced from the staged copy named in sourceDesc and is not altered; the tags are held separately in standOff so that overlapping and nested spans are preserved.

A Unified Framework of Five Principles for AI in Society 2019 Harvard Data Science Review 10.1162/99608f92.8cd550d1 floridi2019-a-unified-framework-of-five-princi floridi2019-a-unified-framework-of-five-princi.txt, the project's staged copy bf26287fbfdd62712934b589fc59370e3e1c7a97 cc-by social-sciences

the paper's own words

the tag layer; a sentence written by the tagger about the paper, always anchored to a quotation but not itself a quotation

0 spans located by searching the stored text for the quote's exact words, allowing any whitespace between them SIX PUBLISHED SETS OF AI ETHICS PRINCIPLES, forty-seven principles in all, read against each other to find out whether they converge - the object is a vocabulary, not a phenomenon Quotation withheld because this record is not cleared for redistribution. how to solve 'principle proliferation': whether the competing declarations say the same thing under different names, or genuinely disagree Quotation withheld because this record is not cleared for redistribution. five principles - beneficence, non-maleficence, autonomy and justice imported from bioethics, plus EXPLICABILITY, argued to be new and necessary - claimed to capture all 47 Quotation withheld because this record is not cleared for redistribution. method Quotation withheld because this record is not cleared for redistribution. a field generating ethical declarations faster than it can reconcile them, with 'similar terms often used to mean different things' and a risk the authors name as a 'market for principles' Quotation withheld because this record is not cleared for redistribution. conceptual-analysis conceptual-analysis Quotation withheld because this record is not cleared for redistribution. the convergence is explained by a structural resemblance between the fields, and the explanation is offered rather than tested: `Of all areas of applied ethics, bioethics is the one that most closely resembles digital ethics in dealing ecologically with new forms of agents, patients, and environments` Quotation withheld because this record is not cleared for redistribution. and the fifth principle is argued for FUNCTIONALLY - explicability is needed because the other four cannot be checked without it: you cannot tell whether AI is beneficent without understanding what it does, nor hold anyone to account without knowing why an outcome arose Quotation withheld because this record is not cleared for redistribution. THE SAMPLE IS SIX DOCUMENTS AND THE PAPER SAYS WHERE THEY COME FROM: each `emerged either from initiatives with global scope, or from within western liberal democracies`, and the framework `would undoubtedly benefit from the perspectives of regions and cultures presently un- or under-represented` Quotation withheld because this record is not cleared for redistribution. the coverage claim and the distribution claim are DIFFERENT CLAIMS AND ONLY THE FIRST IS ABSOLUTE: the five principles `capture every one of the 47`, but each principle appears in `almost every statement` - and the paper's own Table 1 shows IEEE missing autonomy and justice, and the Partnership missing autonomy Quotation withheld because this record is not cleared for redistribution. THE ADOPTION EVIDENCE INCLUDES BODIES THE AUTHORS SIT ON. AI4People and the EC High-Level Expert Group are counted as uptake, and the disclosure states that FLORIDI CHAIRED AI4PEOPLE, COWLS WAS ITS RAPPORTEUR, AND FLORIDI IS A MEMBER OF THE HLEG. Declared in full, and it is still the authors' framework being adopted by the authors' committees Quotation withheld because this record is not cleared for redistribution. the funding sits on the subject: the work was supported by an EPSRC hub AND BY FACEBOOK AND GOOGLE - two of the companies whose AI the framework is written to govern. Declared Quotation withheld because this record is not cleared for redistribution. the paper leaves its central ambiguity open and says so twice: `it is not entirely clear whether it is the people developing AI, or the technology itself, which should be encouraged not to do harm - in other words, whether it is Frankenstein or his monster against whose maleficence we should be guarding` Quotation withheld because this record is not cleared for redistribution. and `justice` is left carrying more than one meaning inside the framework that is supposed to fix it: `Elsewhere ‘justice’ has still other meanings (especially in the sense of _fairness)`. The unification does not reach inside the principles Quotation withheld because this record is not cleared for redistribution. the six documents were selected by three declared criteria - recent, relevant to AI and society as a whole rather than to one sector, and published by authoritative multi-stakeholder bodies of at least national scope eliminates: ad-hoc selection as the explanation of the convergence; the criteria are stated before the result and exclude domain-specific documents · outcome: held Quotation withheld because this record is not cleared for redistribution. whether the convergence is an artefact of where the documents come from eliminates: nothing, and the authors say so: all six came from global-scope initiatives or 'from within western liberal democracies', and the framework would benefit from perspectives 'presently un- or under-represented in our sample' · outcome: failed Quotation withheld because this record is not cleared for redistribution. whether the framework's later adoption is independent evidence for it eliminates: nothing - AI4People and the EC High-Level Expert Group are counted as uptake, and the disclosure states the authors chaired, rapporteured and sat on them. THE DISCLOSURE IS COMPLETE AND THE CIRCULARITY IS STILL THERE · outcome: failed Quotation withheld because this record is not cleared for redistribution. whether 'beneficence' and 'non-maleficence' are one principle wearing two names eliminates: the collapse, explicitly: they 'may seem logically equivalent, they are not, and represent distinct principles', and the six documents treat them separately · outcome: held Quotation withheld because this record is not cleared for redistribution. principle proliferation principle proliferation Quotation withheld because this record is not cleared for redistribution. explicability explicability Quotation withheld because this record is not cleared for redistribution. intelligibility intelligibility Quotation withheld because this record is not cleared for redistribution. accountability accountability Quotation withheld because this record is not cleared for redistribution. beneficence beneficence Quotation withheld because this record is not cleared for redistribution. non-maleficence non-maleficence Quotation withheld because this record is not cleared for redistribution. bioethics bioethics Quotation withheld because this record is not cleared for redistribution. digital ethics digital ethics Quotation withheld because this record is not cleared for redistribution. artificial intelligence artificial intelligence Quotation withheld because this record is not cleared for redistribution. thinking thinking Quotation withheld because this record is not cleared for redistribution. meta-autonomy meta-autonomy Quotation withheld because this record is not cleared for redistribution. `principle proliferation` - THE CONDITION THIS PAPER NAMES AND TREATS. Either the sets are similar, `leading to unnecessary repetition and redundancy`, or they differ, and `confusion and ambiguity will result instead`. The worst case is named too: a `market for principles` in which stakeholders `SHOP` for the most appealing Quotation withheld because this record is not cleared for redistribution. THE DIAGNOSIS, WHICH IS THIS PROJECT'S OWN THESIS IN ONE CLAUSE: `the underlying meaning of each of the principles is contested, with similar terms often used to mean different things`. A shared word list is not a shared vocabulary Quotation withheld because this record is not cleared for redistribution. explicability - FIVE NAMES FOR ONE PRINCIPLE, LISTED WITH THEIR SOURCES: `TRANSPARENCY` in Asilomar and EGE, `transparency` AND `ACCOUNTABILITY` in IEEE, `INTELLIGIBILITY` in AIUK, `understandable and interpretable` in the Partnership. The new term is coined to cover a thing four documents already had Quotation withheld because this record is not cleared for redistribution. explicability, as defined - TWO SENSES DELIBERATELY BOUND TOGETHER: the EPISTEMOLOGICAL sense of intelligibility (`how does it work?`) and the ETHICAL sense of accountability (`who is responsible for the way it works?`). One term carrying two questions on purpose Quotation withheld because this record is not cleared for redistribution. AI - `A Research Area in Search of a Definition` in the paper's own heading. It keeps the 1955 Dartmouth formulation, reads it as A COUNTERFACTUAL (`WERE a human to behave in that way, that behaviour WOULD be called intelligent`), and then substitutes its own working phrase - `a reservoir of smart agency on tap` - justified by fitness for the corpus it is about to analyse Quotation withheld because this record is not cleared for redistribution. beneficence and non-maleficence - `THOUGH DO ONLY GOOD AND DO NO HARM MAY SEEM LOGICALLY EQUIVALENT, THEY ARE NOT, AND REPRESENT DISTINCT PRINCIPLES`. A near-synonymy refused explicitly, which is exactly the move most papers leave implicit Quotation withheld because this record is not cleared for redistribution. meta-autonomy - a coined term for `a DECIDE-TO-DELEGATE model`: humans keep the power to decide WHICH decisions to take, and any delegation must remain `overridable in principle (i.e., deciding to decide again)` Quotation withheld because this record is not cleared for redistribution. extends → Beauchamp & Childress's four principles of bioethics, carried into AI ethics with a fifth added and the imperfection of the translation stated Quotation withheld because this record is not cleared for redistribution. supports → the 1955 Dartmouth proposal's counterfactual formulation of the AI problem, kept as the working definition sixty-four years later Quotation withheld because this record is not cleared for redistribution. contradicts → the claim that ethics is 'Europe's silver bullet' in a 'global AI battle' - answered with two words, 'We disagree', and then with the reason Quotation withheld because this record is not cleared for redistribution. beneficence —distinct-from→ non-maleficence beneficence non-maleficence Quotation withheld because this record is not cleared for redistribution. digital ethics —distinct-from→ bioethics digital ethics bioethics Audit 2026-09-28: DOUBTFUL - read against its span and kept, but the span compares AI principles with bioethics principles; distinct-from is loose. Quotation withheld because this record is not cleared for redistribution. THE CONDITION HAS A NAME: `principle proliferation` - either the sets are similar and produce `unnecessary repetition and redundancy`, or they differ and produce `confusion and ambiguity`, with the worst case a `market for principles` where stakeholders `SHOP` for the most appealing THE CONDITION HAS A NAME: `principle proliferation` - either the sets are similar and produce `unnecessary repetition and redundancy`, or they differ and produce `confusion and ambiguity`, with the worst case a `market for principles` where stakeholders `SHOP` for the most appealing Quotation withheld because this record is not cleared for redistribution. the meaning of each principle is contested, with similar terms often used to mean different things the meaning of each principle is contested, with similar terms often used to mean different things Quotation withheld because this record is not cleared for redistribution. FOUR NAMES FOR ONE PRINCIPLE, EACH ATTRIBUTED TO ITS DOCUMENT: `transparency` in Asilomar and EGE, `transparency` and `accountability` in IEEE, `intelligibility` in AIUK, `understandable and interpretable` in the Partnership FOUR NAMES FOR ONE PRINCIPLE, EACH ATTRIBUTED TO ITS DOCUMENT: `transparency` in Asilomar and EGE, `transparency` and `accountability` in IEEE, `intelligibility` in AIUK, `understandable and interpretable` in the Partnership Quotation withheld because this record is not cleared for redistribution. the principles converge on the four core bioethics principles, which adapt well to AI but do not offer a perfect translation the principles converge on the four core bioethics principles, which adapt well to AI but do not offer a perfect translation Quotation withheld because this record is not cleared for redistribution. `do only good` and `do no harm` are argued NOT to be logically equivalent and to `represent DISTINCT principles` `do only good` and `do no harm` are argued NOT to be logically equivalent and to `represent DISTINCT principles` Quotation withheld because this record is not cleared for redistribution. six documents, 47 principles, and the five are claimed to `capture every one` - while the weaker distribution claim is hedged to `almost every statement` six documents, 47 principles, and the five are claimed to `capture every one` - while the weaker distribution claim is hedged to `almost every statement` Quotation withheld because this record is not cleared for redistribution. AI is defined by engineered outcomes and actions and treated here as a reservoir of smart agency on tap AI is defined by engineered outcomes and actions and treated here as a reservoir of smart agency on tap Quotation withheld because this record is not cleared for redistribution. the sample is declared: all six sets came from global-scope initiatives or from WITHIN WESTERN LIBERAL DEMOCRACIES the sample is declared: all six sets came from global-scope initiatives or from WITHIN WESTERN LIBERAL DEMOCRACIES Quotation withheld because this record is not cleared for redistribution. Floridi chaired AI4People, Cowls was its rapporteur, and Floridi sits on the EC High-Level Expert Group on AI Floridi chaired AI4People, Cowls was its rapporteur, and Floridi sits on the EC High-Level Expert Group on AI Quotation withheld because this record is not cleared for redistribution. Floridi's work was funded in part by Facebook and Google, as declared Floridi's work was funded in part by Facebook and Google, as declared Quotation withheld because this record is not cleared for redistribution. the central ambiguity is left open in the authors' own words: whether it is `Frankenstein or his monster against whose maleficence we should be guarding` the central ambiguity is left open in the authors' own words: whether it is `Frankenstein or his monster against whose maleficence we should be guarding` Quotation withheld because this record is not cleared for redistribution. incorporating both the epistemological sense of INTELLIGIBILITY (`how does it work?`) and the ethical sense of ACCOUNTABILITY (`who is responsible for the way it works?`) explicability Quotation withheld because this record is not cleared for redistribution. the 1955 Dartmouth formulation - making a machine behave in ways that WOULD BE CALLED intelligent IF a human were so behaving - read as a counterfactual about behaviour, not a claim about thinking artificial intelligence Quotation withheld because this record is not cleared for redistribution. a decide-to-delegate model: humans retain the power to decide which decisions to take, and delegation remains overridable in principle meta-autonomy Quotation withheld because this record is not cleared for redistribution. principles across the six documents, claimed to be captured entirely by the five 47 Quotation withheld because this record is not cleared for redistribution. high-profile initiatives analysed - Asilomar, Montreal, IEEE, EGE, AIUK, the Partnership on AI 6 Quotation withheld because this record is not cleared for redistribution. core principles in the resulting framework - four borrowed, one coined 5 Quotation withheld because this record is not cleared for redistribution.

The full text of this article is not redistributable, so it is not included here. The evidence sentences are withheld with it; the tag layer and its grounding status remain. The article itself is at https://doi.org/10.1162/99608f92.8cd550d1.