Will ChatGPT make content moderation a piece of cake?
Two weeks ago, my team shared with me some news from OpenAI that left me both impressed and intrigued. ChatGPT, the AI-powered language model, is now attempting to take on the vital role of content moderation.??
I expect this development to set off a cascade of questions in the minds of many. For a moment, I was worried – is AI trying to replace our superheroes, again? It brought me back to the early days of my career, when the debate revolved around whether humans or machine learning alone were the ultimate solution for content moderation. Luckily, times have changed, and so have the conversations surrounding this topic.??
Today, it is generally understood that it’s only through the combination of HI+AI (human intelligence + artificial intelligence) that we can promote successful online interactions and engage in proactive, real-time intervention.??
I read the announcement and wanted to share my thoughts, and open the conversation to others too.?
Exploring the current state of AI moderation?
I am a big fan of responsible and transparent AI. I believe that AI can — and should — help online communities thrive, make superheroes’ everyday moderation tasks timelier and more impactful, and protect them from unnecessary exposure to harmful content.??
I’m such a fan of AI for moderation that alongside Keywords Studios , I have built partnerships with the best tech companies on the market, including Modulate , Image Analyzer , Spectrum Labs (An ActiveFence Company) , Community Sift , ActiveFence , Checkstep , GoBubble , and more.??
I’m a fan of ChatGPT, too. I’ve seen firsthand how it can speed up, simplify, and improve our everyday life and work.??
But when it comes to ChatGPT content moderation, these are two topics that I would like to discuss further as an industry. (Please share your thoughts!)?
Data bias and questionable decision-making??
Obtaining data to train machine learning models on is a huge challenge. Not only is data difficult and ethically problematic to acquire (I could write pages on this!), but also because selecting, balancing, validating, and training data is incredibly time-consuming, expensive and...a "fine art"?
Simply put — gathering “the right data” is impossible without implicit bias. We need to pay close attention to this!?
Here’s a real-world example of bias: If I ask five people to choose the perfect cake for my upcoming birthday...??
In each example, the result is a cake on the table – but the process of selecting a cake type, deciding where to buy it, which ingredients to put in it, and decorations to top it, is different every time. None of the cakes are wrong, but people with nut allergies, lactose intolerance, an aversion to chocolate (who are you?!) or non-gamers won’t be able to enjoy a slice. And never mind what flavour and design the birthday girl actually wants!??
The point is, when it comes to gathering data and utilizing it for decision making, we must be extremely cautious on how we allow technology to make decisions and impact human interactions.?
Even more troubling — the lack of diversity in data models means that they are often biased toward a single gender, race, and age group, putting people at risk both online and offline. Facial detection AI that struggles to accurately identify Black faces is just one appalling example of many.??
If you are considering “just letting AI do its thing” and allowing it to make unsupervised decisions, please think about how many people are sitting at the table, staring at a cake they can’t enjoy.?
AI cannot exist without HI ?
The second thing that caught my eye in the article was the process of customizing content policies. This process has been followed in technology integrations for as long as I can remember, and it's a relatively simple one: human moderators label content as approved or rejected based on community guidelines, and then compare their decisions with the technology results.?
Spoiler alert: there are always gaps to be closed.??
领英推荐
The first gap is... you guessed it... bias between moderators. Every moderator sees cake in a different way. Even though Community Guidelines and Terms of Use are always top of mind, a moderator’s upbringing, culture, and even daily life can impact their decision-making. By that I mean, if it’s a hot day I’d like ice cream cake please, and if not… maybe apple pie for a change, instead of cake???
(Can we all just agree that the only acceptable cake is chocolate?)??
The truly valuable thing that AI can do is surface these kinds of gaps – in this case, the discrepancy between moderator’s decisions – for humans to fix. And in our example – it's probably time for a refresher on the Terms of Use, and perhaps additional training on how to avoid personal bias! ?
The point I want to drive home is that it doesn’t matter how accurate, fast, or efficient artificial intelligence becomes – when it comes to content moderation, it cannot exist independent of human intelligence. ?
AI will be at its most powerful when it can surface and flag the most egregious offensives that can only be acted upon by humans. Things like: ?
AI is a powerful tool that we should absolutely be using in our work as Trust & Safety professionals. ?
But make no mistake —?human intelligence, understanding, care, and soul will always need to be taken into consideration as we build and train AI models. ?
AI + HI will continue to make a difference between one or many cakes at the table.?
What do you think? Are you concerned about data bias in AI? How do you envision the future of AI in content moderation? And, how hungry are you for cake now? ?
Share your thoughts in the comments! ?
Not sponsored article, just amazing partners working together to better the online world:
#aiforgood #ContentModeration #EthicalAI #OnlineSafety #DataBias #HumanAI #ResponsibleTechnology #TechEthics #DigitalCommunity #OnlineInteractions #onlinecommunities #moderation #gamingindustry #responsibleai #biasawareness #trustandsafety #onlineworls=d
Co-Founder and CEO at Checkstep
12 个月I fully agree, Sharon, and thanks for sharing your thoughts! I agree AI models must be checked for biases. And an important consequence of the transparency requirements (both moral and legal) is that the AI+HI combination should also be checked for biases. Even if the AI itself is unbiased after significant effort, the overall AI+HI system could still make biased decisions if human moderators have significant implicit biases. It would be worthwhile to measure whether your team of moderators ("superheroes") tends to apply stricter policies against certain demographics, even unintentionally. Most humans have some level of implicit bias they are unaware of. We've seen this play out in practice recently, and it's a challenging issue to address fully. You can "re-train" an AI model, but you can't necessarily do the same for human biases. What I mean is that these new requirements and scrutiny on AI systems force us to also re-examine what an ideal moderation system should look like holistically.
Marketing | Gaming | Trust & Safety
1 年“AI cannot exist without HI” - fundamentally important. My understanding is that when benchmarking ChatGPT against some of the existing solutions, which are more often using HI, and a linguistic approach, that it doesn’t appear close to solving our challenges in terms of detection. Though, I do believe it has the potential to aid us with other challenges.
Amazing & fun article! ?? Always inspired by the great work you & the team at Keywords Studios are doing ??
Support, comms, production Unicorn. Hi, let's connect!
1 年Great article, thank you :) Biases are indeed a huge concern, I also wonder about measuring/preserving the integrity of the data. AI has a lot to bring to the table in terms of helpful tools but human intelligence and empathy will always be needed. Pound cake was left out ?? ??
Director of Account Management at Modulate
1 年Great insights Sharon!