Claiming that AI hallucinates – rather than, for example,
referring to the problem as a bug or glitch – shows that we are
anthropomorphising AI (viewing it as human, at least metaphorically). This is
what Dr. Henry Shevlin, an AI ethicist and philosopher of science based at the
University of Cambridge emphasizes in this video: “What Are ‘Hallucinations’
and What More Can We Expect from AI?”. The issue of anthropomorphising
computers has sparked much debate lately. IBM researchers Schneiderman &
Muller have defined anthropomorphism as “ the act of projecting human-like
qualities or behavior onto non-human entities, such as when people give
animals, objects, or natural phenomena human-like characteristics or emotions”
(“On
AI Anthropomorphism”). The researchers assert that such debates over
computers began in the 1990s. However, the controversy has reached new heights
with AI, especially after the spread of systems such as ChatGPT. Three of the
concerns over anthropomorphising AI revolve around whether a human-like
character should appear (e.g. on a screen); whether computers should imitate
humans using voice or text, as in social settings; and whether computer prompts
or responses should use the pronoun “I”.
Ben Garside, Learning Manager at the Raspberry Pi
Foundation, has warned on “How
Anthropomorphism Hinders AI in Education”. He urges that young people
studying technology must not be misled into believing these systems possess
sentience or intention. Rather, learners should take a more active role in
designing better applications for the future: “Rather than telling young people
that a smart speaker ‘listens’ and ‘understands’, it’s more accurate to say
that the speaker receives input, processes the data, and produces an output.
This language helps to distinguish how the device actually works from the
illusion of a persona the speaker’s voice might conjure for learners.”
Whether we refer to the AI-generated errors as
hallucinations or not, the errors are getting out of hand as large volumes of
information are available online and being processed, for example in news
summaries. The New York Times recently published a piece by technology
reporter Cade Metz entitled “Chatbots
May ‘Hallucinate’ More Often Than Many Realize”, warning that when
summarizing news, ChatGPT fabricates 3% of the content, according to research
by a new start-up, and that a Google system’s fabrication rate is currently
27%. Metz rightly points out that ironically AI is being used to assess the
error rate, which itself is not highly reliable! A chicken and egg situation;
user beware!