Bots as Virtual Confederates: Design and Ethics

Peter M. Krafft, Michael Macy, Alex Pentland

TL;DR

This paper introduces the concept of "confedabots"—automated agents used as virtual confederates to scale social science experiments in online spaces. To navigate the ethical minefield of digital deception, the authors propose a design framework and a set of "minimal risk" guidelines centered on mimicking ordinary user behavior.

What Problem Does This Paper Address?

In traditional social psychology, researchers use "confederates" (actors following a script) to manipulate social variables—think of the famous Asch conformity experiments. However, human confederates don't scale. You can't hire enough actors to interact with 50,000 people on Twitter for six months.

While online field experiments offer a solution, they face two major hurdles:

  1. Methodological: How do we precisely control the attributes and behaviors of "peers" in a digital environment?
  2. Ethical: High-profile failures (like Facebook’s "emotional contagion" study or Microsoft’s "Tay" bot) have shown that experimenting on users without informed consent can lead to massive public backlash and unintended harm.

The authors aim to provide a roadmap for using bots as a legitimate, ethical tool for causal inference in computational social science.

Methods

The paper is a conceptual and ethical framework rather than an empirical study. It defines a Design Space for confedabots across four dimensions:

  • Intervening on Actions: Randomizing specific moves, such as which posts a bot upvotes on Reddit.
  • Intervening on Attributes: Keeping behavior constant but randomizing profile traits (e.g., changing the bot's perceived gender or race to see how users respond).
  • Intervening on Algorithms: Testing entirely different behavioral profiles (e.g., a "high-activity" bot vs. a "low-activity" bot).
  • Artificial Contexts: Using groups of bots to create a synthetic social environment, such as a pre-populated forum thread.

To address the ethics, the authors map these designs against the Belmont Report principles (Respect for Persons, Beneficence, and Justice) and the Common Rule for human subject research.

Datasets

Not discussed in the paper. The authors reference various existing studies (conducted on Twitter, Facebook, and Reddit) to illustrate their points, but they do not introduce a new dataset or benchmark.

Key Insights and Ideas

  • The Scalability of Social Influence: Confedabots allow for longitudinal, large-scale experiments that were previously impossible. They provide perfect replicability because their "scripts" are literally source code.
  • The "Ordinary User" Heuristic: The authors argue that the most ethical way to deploy a bot is to ensure it acts like an "ordinary user." If a bot’s behavior falls within the range of what a user would encounter in daily life on that platform, it meets the "minimal risk" threshold required to potentially waive informed consent.
  • Contextual Integrity: Ethics are platform-dependent. Deploying a bot on Twitter or Wikipedia (where bots are pervasive and expected) is viewed as more ethical than deploying one on a private support group or a bot-free platform like Yik Yak.
  • Deception by Omission: The authors suggest that while bots are inherently deceptive (they aren't the humans they might appear to be), "deception by omission" (not explicitly stating it's a bot) is often more acceptable than "explicit deception" (lying about facts or personal feelings).

Key Results

The primary "result" is a set of Proposed Guidelines for researchers:

  1. Ordinary Behavior: Bots should only be used where they are already present, and their actions/attributes should not be "unusual" for that community.
  2. Harm Reduction: Researchers must implement technical safeguards (like keyword filters) to prevent bots from generating toxic content or harassing users.
  3. Careful Evaluation: Every experiment requires a deliberative cost-benefit analysis that considers community norms and potential "unanticipated distress."

Limitations and Future Work

  • The Consent Gap: The authors acknowledge that their "minimal risk" approach is a compromise. It doesn't fully satisfy the strictest interpretations of informed consent, which remains nearly impossible in massive online field studies.
  • Public Perception: As public awareness of bots grows, the "ordinary behavior" benchmark may shift. What is considered "normal" today might be seen as intrusive tomorrow.
  • Algorithmic Accountability: The paper suggests future work should focus on creating systems for "open public dialogue" about social science experiments to mitigate the backlash seen in previous corporate-led studies.
  • Semi-Automation: They suggest "human-in-the-loop" systems as a middle ground to prevent bots from going off the rails in sensitive social contexts.

Summary updated

BibTeX
@inproceedings{krafft2017bots,
  author = {Peter M. Krafft and Michael Macy and Alex Pentland},
  title = {{Bots as Virtual Confederates: Design and Ethics}},
  booktitle = {CSCW},
  year = {2017},
  doi = {10.1145/2998181.2998354},
}
Figure