Design your bot so that it respects relevant cultural norms and guards against misuse
Since bots are most often designed to appear like humans, it’s important that they are respectful of human needs, and that they have built-in safeguards to defend against and handle misuse and abuse. This is doubly important because if human history in the time of the internet has taught us anything, it is that Murphy’s Law holds true for information systems — especially systems populated by myriad, malicious, anonymous actors. Anything that can go wrong will go wrong, and any system that can be exploited will be exploited.
Design your bot so that it respects relevant cultural norms and practices
Limiting the actions of a bot to the actions it’s designed to do will reduce, but not eliminate, the chances that it will violate cultural norms. For example, if your bot is designed to order pizza, you can design the interface so conversations are limited to that topic. More open conversations can be considered high-risk, as they will require much up-front investment in content, as well as substantial monitoring.
Create a code of conduct, make it available to your users, and enforce it
Making a code of conduct that's expected to be followed is especially important if the bot allows your users to interact with each other in any way. Common and important provisions of codes of conduct include prohibitions on things like hate speech and bullying, and positive instructions on how to interact with the bot and other users on the platform.
Codes of conduct are not laws, but they can and should be enforced. While there are grey areas on the borders of any set of guidelines, beware of users who seek to weaponize these. For example, while free speech is an incredibly important moral and ethical consideration, malicious actors who seek to hide behind a facade of free speech to make hateful attacks or to peddle a racist agenda need not be allowed access to a bot. This is where the flexibility of guidelines cuts both ways, allowing for human operators to make edge-case decisions. This is one reason why moderation should never be left entirely to bots.
Consider the locales and cultures your bot will operate in
Deploying a bot for an audience of people that does not look like its developers is one area where assumption can be problematic. Large, diverse audiences for your product are one reason to have a diverse team who can raise any potential issues prior to production.
Additionally, investing in a translator will often be necessary if your bot is to be deployed in a culture or locale markedly different to the one it's developed in. A good translator or translation service will translate not only language but will identify and help solve issues around cultural practice.
Design a feedback mechanism so users can report misuse
As part of its feedback processes, your users should be able to identify and report instances of abuse. Your feedback processes should allow your bot to learn from reports that are flagged as positives. Like all feedback, the understanding that it will be taken seriously and followed through will help ensure users are engaged and acting in good faith.
Apply language and content filters, erring on the side of caution
Filters and content should be applied more broadly and then altered to suit the working environment as the bot and developers learn from users. The implementation should be two-way, including responses made by the bot as well as users. In other words: a user shouldn’t be able to swear at a bot without seeing it censored, and the bot shouldn’t be able to swear back!
The Microsoft Bot Framework includes offensive text classifiers, and can help protect your bot and its users from abuse.
It’s recommended to avoid controversial topics as entirely as possible. Truly controversial topics, or those that can cause offense or distress along the lines of race, gender, sexuality, politics, and the like, all fall under the category of consequential use cases, and should be treated with the same ethical gravity and care. Other common sensitive topics include adult content, extremism, drugs, alcohol, tobacco, vulgarity, harassment, bullying, violence and gore, and hate speech. Any conversational AI bot implementation in these areas or dealing with these topics needs careful implementation - or, often, no implementation at all.
Public facing bot APIs need to be carefully monitored
If you have or intend to have a publicly accessible bot API, this is a consequential use case. A conversational AI bot, maliciously deployed or used, can do emotional or cultural damage analogous to a computer virus. This can be extremely fraught for organizations and their customers. Appropriate safeguards need to be put in place to prevent malicious use. APIs should be reviewed to make sure they cannot be used by people outside your organization in a way that would reflect poorly on you.
If a bot’s use case doesn’t call for it to engage with culturally sensitive or potentially offensive content, don’t give it the ability to do so. A bot should only do its job; nothing more. However, this doesn’t mean these concerns aren’t of importance even in the most mundane use cases. Make sure they’re always top of mind for any bot deployment.