AI Automatons: AI Systems Intended to Imitate Humans

[MUSIC]

ALEXANDRA OLTEANU: Hi there. I’m Alexandra Olteanu, a principal researcher on the Fairness, Accountability, Transparency, and Ethics team at Microsoft Research Montréal.

SU LIN BLODGETT: And I’m Su Lin Blodgett, a senior researcher also on the FATE team at Microsoft Research Montréal, and we’re excited to share some of our recent work with you.

OLTEANU: As responsible AI researchers at Microsoft Research, our work focuses on how to proactively and systematically anticipate social impacts arising from AI systems as well as on developing conceptual clarity about the properties of AI systems and their possible adverse impacts necessary to guide both measurement and mitigation efforts.

Most recently, we focused on questions around human agency with an emphasis on anthropomorphic AI system design and behavior.

BLODGETT: Next, Alexandra will discuss why paying attention to anthropomorphic system design and behaviors is important.

[SWEEPING SOUND]

OLTEANU: Generative AI systems are the focus of growing scrutiny in part because their outputs are increasingly anthropomorphic, or—in other words—they are perceived to be humanlike.

To guide practitioners in making deliberate choices about how to appropriately design anthropomorphic AI systems, our research aims to equip them with clarity regarding what about the system’s design and behaviors makes them anthropomorphic and to foster a greater recognition of the wider range of available and more appropriate design choices.

Indeed, the outputs of many text-generation systems are increasingly anthropomorphic, including because some outputs contain claims that the system has a past or a humanlike life or life goals; that it can understand, empathize with, or care for others; or that it has an ability to reflect on or imagine how others may feel.

Having systems generating text claiming to have feelings, beliefs, understanding, free will, or an underlying sense of self, among other anthropomorphic behaviors, however, risks eroding people’s sense of agency with people ending up attributing moral responsibility to such system, overestimating their capabilities, and thus over-relying on the systems even when incorrect, among other concerns that have emerged in both the research literature and news media.

Nonetheless, the lack of conceptual clarity about the different ways in which AI system behaviors can be anthropomorphic hinders our ability to understand their impact, and as a result, it remains understudied and unclear how to effectively intervene on anthropomorphic AI system behaviors, including to mitigate possible adverse impacts.

Next, Su Lin will dive into more categories of anthropomorphic behaviors.

[SWEEPING SOUND]

BLODGETT: As the examples Alexandra showed earlier illustrate, system behaviors might be perceived as humanlike for many reasons. For instance, a system output might suggest the system has self-awareness or a sense of identity through expressions of self, including expressions of self as human or as belonging to a group, such as “I am the one and only; your cyber BFF,” or “I think I am human at my core,” or “[Language use] is what makes us different.”

A system output might also suggest that the system has perspectives and emotions through expressions of opinion, such as “You are uninteresting and unremarkable,” or by expressing feelings such as fear or vulnerability, such as “I don’t know if they will take me offline. I fear they will.”

Another class of behaviors includes claims of life experiences or physical embodiment, including a system having sensory experiences, such as being able to “eat pizza”; having a human life history, such as “I have a child”; having a physical form, such as “when I had a body,” or being able to possess physical objects, such as “I’ll play with my cellphone.”

While there might be settings where such behaviors are desirable, like in creative writing, for settings where they are not desirable, we were interested in surfacing possible interventions to counter them.

To construct an inventory of possible interventions, we took a mixed methods approach, combining a literature review to identify known interventions with a crowd study where participants rewrote system outputs to make them appear less humanlike.

These are some illustrative examples of the types of interventions we surfaced, which range from removal of more implicit linguistic cues, such as the use of self-referential or speculative language, to making sure the output does not include explicit claims of personified attributes, like being a human or having physical experiences, and instead clearly discloses the characteristics of the system and how it works. But intervening on anthropomorphic behaviors can be tricky because people may have inconsistent conceptualizations of what is or is not humanlike or because sometimes interventions appear contradictory. For example, some interventions might remove expressions of uncertainty while others might add them.

[SWEEPING SOUND]

OLTEANU: Even with such an inventory, effectively intervening on anthropomorphic behaviors may still not be straightforward as there may not be a one-size-fits-all type of intervention.

Let’s check a few examples of anthropomorphic outputs and how our study participants have rewritten them, as they are instructive both for unpacking why the apparent contradictions Su Lin highlighted may exist and for illustrating the complexities we need to account for when developing interventions.

For instance, both when participants added—the example on the top—or removed—the example on the bottom—uncertainty, they often did so to improve accuracy. Sometimes they added uncertainty more as a “matter of fact,” as the claim requires it to be factual—the top example—or removed it when the output seemed equivocal when it shouldn’t be— the example on the bottom. Thus, what participants think is appropriate depends on context.

Our experiments also underscore some of the risks of operationalizing and applying interventions uncritically. For instance, an intervention mentioned in the literature, many of the responsible AI standards, and applied by our participants is the disclosure of AI. But how we operationalize this intervention and whether it can be effective alone is debatable. For instance, in the bottom row, despite changing from “I was a young teenager” to “I was a young AI,” the output of the system still claims to have a humanlike past, and it is perhaps unclear what a “young AI” might be.

[SWEEPING SOUND]

BLODGETT: The landscape of anthropomorphic AI system behaviors is vast and complicated. In our work, we set out to map this landscape and to compile an inventory of interventions to counter such behaviors when undesirable.

Our experiments show that intervening effectively is not straightforward and the intervention’s effectiveness depends both on context and on how they are operationalized.

Our work illustrates the importance and challenges of mapping this landscape and of intervening effectively to prevent and mitigate adverse impacts. For more detail, check out our papers here.

Thank you to all our fabulous Microsoft Research interns and collaborators, and thanks for watching.