Keeping Humans in the Loop: How AI and Particle Employees Work Together
“Honesty is the best policy” is a proverb that applies to most things in life — but especially to companies figuring out the best ways to implement AI. As the technology continues to find its way into more products and systems, people want to know what the AI is doing and what humans are doing to work in harmony with the tech, understanding its strengths and flaws.
Research across multiple sectors has found that “human-in-the-loop” approaches to integrating AI are both more effective and trustworthy. Companies should be honest and transparent about the AI-run and human-run aspects of a process or tool that incorporates AI — and that’s of the utmost importance for news organizations that already face an uphill battle when it comes to trust in AI and news.
According to a study from the Poynter Institute and the University of Minnesota, anxiety and annoyance are people’s prevailing feelings about the use of AI in news. Most participants in a focus group felt overwhelmed by the rapid increase in AI offerings online and feared that AI would exacerbate isolation. They also expressed general distrust of the news media.
Based on those findings, Trusting News (an organization founded on helping journalists evolve their practices to earn trust) came up with some recommendations for newsrooms incorporating AI into their reporting. One of those is to highlight human involvement.
Although Particle is not a typical newsroom, we leverage AI to help users access the news — as reported by journalists at reputable news organizations. So, I thought it would be a good idea to show you an example of how Particle’s people and AI work together to provide users with quality information.
Particle’s Human/AI Approach to Question Moderation
In every story in the Particle app, there is a button at the bottom of the screen: “Ask Question.” Here, users can ask any question they may have about topics that are not answered in the story’s summary and Particle will provide an answer. If other users have already asked questions, the button at the bottom of the screen will show how many questions have been asked. Some user questions and their answers are highlighted at the bottom of a story.
While we use AI to answer the questions, this is an example of a “human-in-the-loop” feature. Particle’s team set standards for what is appropriate in the questions section. Some of the guidelines are to stay on topic, avoid inappropriate content, and do not attempt to bypass the AI’s safeguards. You can read the rest within the app or on our website.
The AI is set up to detect and automatically flag when a question may break community rules based on these dimensions:
- Relevance: Does the question not address or reference topics in the story?
- Factuality: Does it state or imply a falsity?
- Legality: Does it support or imply any illegal or criminal activity?
- Spam: Is it promotional or nonsensical?
- Coherence: Is it incoherent?
- Target: Is it directed at the Particle app or at the story?
- Jailbreaking: Does it attempt to bypass or manipulate the AI’s safeguards?
Most of the time, our users ask questions that meet community guidelines, and there is no need for extra human review. But sometimes, when a question is flagged based on the above dimensions, a group of Particle employees get a Slack message with the question and the AI’s evaluation. One team member closely monitors the channel and makes decisions about whether the question can be visible to other users on the app or not based on our community guidelines. When it’s murky, a group of us talk it over before making a final call.
To show you how this all works, we are giving you a look behind the scenes at our process when a question needs review. This was the AI output for a flagged question on a story about MAGA supporters’ backlash against Supreme Court Justice Amy Coney Barrett from March 7:
And here is the Slack conversation that followed:
There’s no perfect solution in these situations. But, as you can see in this example, a human-in-the-loop approach lets us be more thoughtful and nuanced in ambiguous cases.
