The Book of AI Fallacies


Computation Blueprints (Medieval Abacus, Houhgton Library)

Intelligence can be broadly defined as the ability of agents, live or mechanical, to deal with issues set in real, formal, and/or philosophical terms. When issues are self-contained they can be tackled through problem-solving, when they are open-ended and involve different parties and stakes they must be dealt with through decision-making processes. That simple framework provides a sound basis for the assessment of artificial intelligence.

A Sample of AI Fallacies

#1: Facts are not Information

Real issues are rooted in facts, but facts are not given as manna from heaven: they can only be observed through data, which usually provide a partial, biased, and/or temporal reflection of what happens in actual environments.

#2: Neural connectors represent facts, not meanings

Whatever the nature of nodes in neural networks their connectors bear quantitative values. That’s unquestionable for brain neurons or telecommunication nodes, likewise for the weights on connectors between tokens in documents.

#3: Information is not knowledge

Formal issues are one-sided (no parties directly involved) and bounded (unambiguous semantics) and can thus be expounded and dealt with independently of real or philosophical ones; to that end they can be expressed through symbolic languages used to define categories and rules representing hypothetical instances.

#4: Language is not knowledge

Compared to formal issues, philosophical ones may involve different parties with stakes possibly set across domains, which entail natural language for communication and representation of the relevant values, intents, and outcomes. Ignoring the distinction between expressing the issues (language) and dealing with them (knowledge) is reducing intelligence to algorithms.

#5: There is no truth in languages

In language there are only differences. Even more important: a difference generally implies positive terms between which the difference is set up; but in language there are only differences without positive terms. Whether we take the signified or the signifier, language has neither ideas nor sounds that existed before the linguistic system, but only conceptual and phonic differences that have issued from the system. Ferdinand de Saussure Course in General Linguistics.

#6: There is no reason in data

Data is made of labelled observations from physical or symbolic environments; set at the inception of cognitive processes data is meant to be free of any assumption regarding meanings and/or causations; otherwise it would introduce built-in and circular reasoning.

#7: reason is not judgement

The distinction between reason (information) and judgment (knowledge) is meant to be congruent with the one between mechanical and live agents; as such that two-pronged distinction provides a fulcrum for the artificial intelligence project, and then its ethics.

#8: judgements begin where Algorithms end

The distinction between judgement and algorithms should serve as a litmus test setting apart human and artificial intelligence. Judgement is needed for issues involving parties with intents (decision-making), as opposed to self-contained ones (problem-solving).

#9: Knowledge graphs are not Knowledge

Knowledge graphs are symbolic representations of real (e.g. communication networks), conceptual (e.g. entities and relationships), or virtual (e.g. intents and plans) realms. As languages, e.g. the Web Ontology Language (OWL), they should not be confused with what they represent.

#10: intelligence cannot emerge from Canned words

Massive increases in available computing power and data hoards have triggered a surge of interest for generative technologies. But wild and irrational expectations have come with leaps of faith and sleights of hands:

  • Leaps of faith, for the belief that knowledge can emerge from canned words
  • Sleights of hands, for getting rid of the distinction between language and knowledge, notwithstanding it being at the core of the artificial intelligence venture

These flaws induce inbreeding between language and knowledge, pushing generative AI into degenerative quagmires.

Kaleidoscope Series

Other Caminao references